Summary: The bug, introduced in https://github.com/pytorch/pytorch/pull/152765, was caused by passing the `group` parameter to the `get_rank()` function, which caused the function to return the rank of the entire group instead of the rank of the current process. The fix involves removing the `group` parameter from the `get_rank()` function call.
Test Plan: contbuild & OSS CI
Differential Revision: D74964213
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153798
Approved by: https://github.com/Skylion007
Fixes#146411, following #148654
After test, seems this could be enabled for all ipynb file.
```bash
lintrunner --take RUFF --all-files
Warning: Could not find a lintrunner config at: '.lintrunner.private.toml'. Continuing without using configuration file.
ok No lint issues.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153820
Approved by: https://github.com/Skylion007
so two things other than cleanups and refactoring
1) do not use propagate_real_tensors to resolve eval under guard_or_true/guard_or_false .
2) do not guard for dimensions of type DimDynamic.OBLIVIOUS_SIZE under guard_or_true/guard_or_false .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152657
Approved by: https://github.com/pianpwk
Fixes longstanding issue where direct references to aten operations are seen as untyped by type checkers. This is accomplished by setting attributes on several classes more consistently, so that `__getattr__` can return a single type in all other cases.
Decisions made along the way:
1. `torch.ops.higher_order` is now implemented by a single-purpose class. This was effectively true before, but the class implementing it attempted to be generalized unnecessarily. Fixing this simplified typing for the `_Ops` class.
2. `__getattr__` is only called when all other lookup methods have failed, so several constant special-cases in the function could be implemented as class variables.
The remainder of this PR is fixing up all the bugs exposed by the updated typing, as well as all the nitpicky typing issues.
Test plan: CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153558
Approved by: https://github.com/rec, https://github.com/Skylion007, https://github.com/cyyever
When loading statically launchable triton kernels from FxGraphCache, since we don't instantiate a CachingAutotuner like we do normally, we need to recheck the autotune cache based on the existing compile results. If we get a hit, we take the compile result whose config matches the best config.
Sometimes, the best config will have been from coordinate descent tuning. In this case, FxGraphCache today does not cache the resulting triton kernel, neither with static or without static cuda launcher. This is because coordinate descent tuning happens at runtime, and if the best config happens to not be one of the precompiled configs.
Test Plan:
New unit test that failed before
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153565
Approved by: https://github.com/aorenste
Summary:
## Context
See https://github.com/pytorch/pytorch/pull/149164 for more context.
Originally, this fix worked but more recently including `cmath` by itself no longer provides access to math constants on Windows platforms. I found that including `math.h` resolves this.
I'm not sure exactly what changed, but this PR updates the header to just use both includes fix the symbols not being found. It might be a bug with a recent Windows update perhaps?
Test Plan:
CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153742
Approved by: https://github.com/swolchok, https://github.com/Skylion007
In the FakeTensor cache when we get a bypass exception while computing the cache key (call this exc_1) we need to dispatch to the original operation.
It's possible for the dispatch to the original operation to get its own exception which we want to bubble up to the caller (call this exc_2).
If we directly dispatch from within the handler for exc_1 then exc_2 will have a `__context__` of exc_1 - which can cause deviations between cached and non-cached behavior - so we need to be a bit careful when we call the dispatch.
Testing:
test_aotdispatch.py::TestAOTExport::test_aot_export_predispatch_outdtype fails before this change and passes after.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153780
Approved by: https://github.com/oulgen
for messages like
```/workspace/pytorch/aten/src/ATen/native/transformers/cuda/flash_attn/flash_api.cpp:1396:38: warning: narrowing conversion of ‘(char)(& q)->at::Tensor::<anonymous>.at::TensorBase::get_device()’ from ‘char’ to ‘c10::DeviceIndex’ {aka ‘signed ```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153643
Approved by: https://github.com/Skylion007
Summary:
This change introduces a fallback path from `bmm` to `mm` when the batch dimension is `1`.
The motivation is to unlock specialized `mm` kernel paths (e.g., `decomposeK`, `persistent+TMA`, etc.) which often don't have `bmm` equivalents.
### Rationale
- **No regression:** On shapes where the fallback triggers, we see no performance loss.
- **Performance wins:** On select shapes (especially with large `K`), we observe measurable speedups by triggering `mm`-specific optimizations.
For example, on `bmm` shapes of the form `(1, H, K, H)` where `H ∈ {16, 32, 48, 64}` and `K ∈ {4096 ... 32768}`, we see an **average speedup of 10%**.
- **Prevalence in prod:** Internal workloads frequently emit `bmm` ops with `batch=1`, making this fallback broadly useful in practice.
Test Plan:
contbuild & OSS CI
Tests in test/inductor/test_torchinductor.py
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153572
Approved by: https://github.com/PaulZhang12, https://github.com/eellison
Some tests may not set the preferred backend, which leads to unexpected behavior when multiple tests are run vs. standalone
Tests that should exercise both backends should explicitly parametrize this setting
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153655
Approved by: https://github.com/ngimel
Enables a variety of misc ruff rules and fixes some incorrect indentation in the file. Now that we updated ruff recently we can enable this rule lints. Most of these lints I've already applied, but now they are out of preview can apply them as stable lints.
Including:
* Do not bother why typing union with Never as this gets cancelled otu
* Simplify nested Literal into a single Literal
* Properly use packaging to parse version instead of `map(int(`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153624
Approved by: https://github.com/atalman, https://github.com/malfet
so two things other than cleanups and refactoring
1) do not use propagate_real_tensors to resolve eval under guard_or_true/guard_or_false .
2) do not guard for dimensions of type DimDynamic.OBLIVIOUS_SIZE under guard_or_true/guard_or_false .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152657
Approved by: https://github.com/pianpwk
Prior to this PR, `_inductor/codegen/cpp_prefix.h` was copied into a new temporary directory on every inductor run utilizing the CPP backend (i.e. CPU-only), then included in the output source code. Instead, this PR puts it in an appropriate place in the torch includes, and includes it from there. This allows us to precompile it in cpp_wrapper and AOT inductor mode, saving significant compilation time.
Due to difficulties getting this to work in FBCode, the precompilation itself is only enabled in OSS PyTorch.
Differential Revision: [D69420620](https://our.internmc.facebook.com/intern/diff/D69420620)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144293
Approved by: https://github.com/desertfire
Tests:
* test_set.py
This PR adds test_set.py from the CPython 3.13 branch and ~400 files to test/dynamo_expected_failures. Most of these are expected to be fixed in upcoming PRs. Only minimal changes were made to test_set.py to enable compilation with Dynamo using the PYTORCH_TEST_WITH_DYNAMO=1 environment variable.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150792
Approved by: https://github.com/anijain2305
Summary:
Split out second pass of LayerNorm so it's more likely to show up in
profiler output. In my testing with perf, the samples from the lambda in the
current implementation are attributed somewhat haphazardly.
Differential Revision: D74181627
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153578
Approved by: https://github.com/hl475
Basically adds native _IntWrapper support to dynamo. Here's my process of trying to make symint input support work on dynamo, and how I ended up with this approach [(doc)](https://docs.google.com/document/d/1GvNRQd8BnxlMay_hrEVgEta6VUeUW_hcFeRuB7q1nDY/edit?tab=t.0).
What I did was, before passing inputs to dynamo.export, I first wrap them with a class, `_IntWrapper`. When processing dynamic shapes, I will then add the corresponding dynamic shape specification to the `dynamism` field stored on the `_IntWrapper`. If there is no dynamism specified, then this will get unwrapped back to an integer. When dynamo tracing, when we encounter an `_IntWrapper`, we will convert this to a symint if the dynamism was specified as `Dim.DYNAMIC/AUTO`. Dynamo will then trace a graph that contains symint inputs, which will get passed to AOTAutograd and so on.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152677
Approved by: https://github.com/pianpwk
DSD currently will pop tensors if these tensors are on Meta device. This forbid the use cases that users would like to let DCP to directly initialize the tensors when loading.
This PR also removes test/distributed/checkpoint/e2e/test_pipeline.py which is based on the above feature that is not realistic and is not used anywhere.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153185
Approved by: https://github.com/mori360
**Summary**
This PR adds a new op, `onednn.qbatch_norm2d`, which accepts uint8 inputs on CPU device (instead of QuantizedCPU).
The new ops are implemented with AVX512 instructions and it provides similar performance as its counterpart for QuantizedCPU device `quantized.batch_norm2d`.
The new op supports output dtypes other than uint8 (fp32, fp16 and bf16 are supported).
**Test plan**
```
pytest test/quantization/core/test_quantized_op.py -k test_int8_batch_norm_onednn
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152811
Approved by: https://github.com/leslie-fang-intel, https://github.com/jerryzh168, https://github.com/jgong5
ghstack dependencies: #152411
Fixes#147336
## Context
NCU analysis of the fp8 flex attention perf issue in #147336 showed an unexpected increase in shared memory access bank conflicts when loading the V tensor from HBM to SRAM.
Bringing this to the attention of triton developer @davidberard98 he identified the memory layout of the tensor in HBM to be causing non-pipelined loads into SRAM, causing the slowdown.
To summarize:
In flex attention when performing the FP8 GEMM `softmax_scores @ V` the right operand V must be in column-major memory layout. However, the `tl.load` of V blocks from HBM to SRAM cannot be pipelined if the V tensor isn't column-major in HBM already, leading to substantial performance degradation.
This is because triton does not perform async copies with the `cp.async` PTX instruction if the number of contiguous bytes is less than 4 (see [here](81f93f2c8e/lib/Dialect/TritonGPU/Transforms/Pipeliner/PipeliningUtility.cpp (L403))).
i.e., when loading 4 bytes of contiguous data from a tensor stored in row-major in HBM, we have to perform 4 separate non-contiguous writes to SRAM to place those bytes in their new location in the col-major layout in SRAM. Thus the load is not a candidate for pipelining w/ cp.async and just moves data to registers then performs a series of single byte stores.
## Fix summary
- To fix this, we should enforce memory layouts for Q, K, V in FlexAttention when fp8 is being used, to ensure they each exist in HBM in the necessary memory layout to facilitate pipelined loads into SRAM ahead of the FP8 GEMMs
## Benchmarks
Rerunning the repro we see fp8 runtime is reduced from 120% of bf16 to 76% of bf16 runtime.
Before fix:
```
(flex) [danvm@devgpu007.eag6 ~/ml-perf-tools/flex_attention (main)]$ rm -rf /tmp/torchinductor_${USER}; python profile_flex.py --bf16 --fp8
2025-05-11 19:07:33,402 - flex_bench - INFO - Running benchmark: bf16
2025-05-11 19:07:35,885 - flex_bench - INFO - bf16: 424.87228804347734 us
2025-05-11 19:07:35,893 - flex_bench - INFO - Running benchmark: fp8e4m3
2025-05-11 19:07:37,319 - flex_bench - INFO - fp8e4m3: 515.714000000001 us
```
After fix:
```
(flex) [danvm@devgpu007.eag6 ~/ml-perf-tools/flex_attention (main)]$ rm -rf /tmp/torchinductor_${USER}; python profile_flex.py --bf16 --fp8
2025-05-11 17:34:38,223 - flex_bench - INFO - Running benchmark: bf16
2025-05-11 17:34:41,157 - flex_bench - INFO - bf16: 423.4662032967036 us
2025-05-11 17:34:41,167 - flex_bench - INFO - Running benchmark: fp8e4m3
2025-05-11 17:34:42,917 - flex_bench - INFO - fp8e4m3: 326.3694803493453 us
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153357
Approved by: https://github.com/ngimel, https://github.com/davidberard98
Summary:
Some new errors have been showing up on the PT2 dashboard with
```
Invalid type for lengths: Expected BlobReference or torch.Tensor, got: Tensor(shape: torch.Size([10]), stride: (1,), storage_offset: 0)
```
This is caused by [this piece of code](https://fburl.com/code/5nbi9on7) which maps over a set of nodes (in this case type `IDListFeatureListField`) and turns the results into strings to be displayed later. However during pytree.tree_map we call pytree.tree_unflatten which will call the class's init function, which calls `assert_blob` (https://fburl.com/code/h3ainrn9). Because we've mapped over the values and converted them to strings, the assert_blob fails.
I initially thought to disable the assert_blob while tracing (D74684309) but then I think we should actually flatten the list first. Because tlparse will expect just a string out outputs instead of the actual structure.
Test Plan: `buck2 run mode/opt sigmoid/inference/ts_migration:pt2i_readiness_main -- --test_suite ads_all --mode test_full_model --model_id 542947220` fails with something else 😅
Differential Revision: D74744326
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153627
Approved by: https://github.com/yiming0416
C++ Reducer is silently incorrect under CA, its implementation is no-oping the collective. I'm guessing that it was no-op'd because in DDP + python reducer, the C++ reducer is still being initialized.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152735
Approved by: https://github.com/fegin
ghstack dependencies: #153300, #152689
I slap disable on the recomputation hook, otherwise the partitioner may save less/more activations and mismatch with the expected eager count in checkpoint. See code comment `Note: [compiled autograd and checkpoint unpack hook]`.
This fixes all non-nested checkpointing tests. I also wrap nested checkpointing tests, and a few of them still fail.
This also seems to fix all PYTORCH_TEST_WITH_DYNAMO checkpointing tests except for `TestAutograd.test_checkpointing_without_reentrant_custom_function_works`. For those tests, it looks like we fail to HOPify the checkpointed region and when the backward executes the unpack hooks, dynamo tried to trace them. This messed up the internal state tracking of checkpointing, some raising the _StopRecomputationError and others raising the same count mismatch error as CA.
FIXES https://github.com/pytorch/pytorch/issues/127115
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153300
Approved by: https://github.com/jansel
This enables Gloo CUDA when used with a backend that supports GPUDirect which currently is only the IBVERBS backend.
This requires some changes to Gloo which are in https://github.com/pytorch/gloo/pull/441
Since we're now depending on gloo_cuda we need to split ProcessGroupGloo into two pieces, one with the CPU bits (libtorch_cpu) and one with CUDA kernels in libtorch_cuda. This unfortunately requires some major refactoring as some CPU code is shared across both.
The gloo submodule is updated to depend on the new Gloo changes
Test plan:
```py
import os
import time
transport = "TCP"
#transport = "IBVERBS"
os.environ["GLOO_DEVICE_TRANSPORT"] = transport
rank = int(os.environ["RANK"])
os.environ["CUDA_VISIBLE_DEVICES"] = str(rank)
ibv = "mlx5_0:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_9:1,mlx5_10:1,mlx5_11:1".split(",")[rank]
ibv_name, ibv_port = ibv.split(":")
os.environ["TORCH_GLOO_IBV_NAME"] = ibv_name
os.environ["TORCH_GLOO_IBV_PORT"] = ibv_port
os.environ["TORCH_GLOO_IBV_INDEX"] = "3"
import torch
import torch.distributed as dist
dist.init_process_group("gloo")
rank = dist.get_rank()
# initial sanity check
#device = "cpu"
#t = torch.zeros(10, device=device)
#dist.all_reduce(t)
#print("sanity complete")
device = "cpu"
iters = 10
warmup_iters = 2
for nelem in [10, 100, 1000, 10000, 100000, 1000000, 10000000, 100000000]:
t = torch.zeros(nelem, device=device)
torch.cuda.current_stream().synchronize()
for i in range(warmup_iters):
dist.all_reduce(t)
torch.cuda.current_stream().synchronize()
start = time.perf_counter()
for i in range(iters):
dist.all_reduce(t)
torch.cuda.current_stream().synchronize()
dur = (time.perf_counter() - start)
qps = iters/dur
bandwidth_gb = t.nbytes * iters / dur / 1e9
gb = t.nbytes / 1e9
if rank == 0:
print(f"{transport=} {device=} {iters=} {nelem=} {qps=} {gb=} {bandwidth_gb=}\n", end="")
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153406
Approved by: https://github.com/fduwjj
Summary:
Currently things are hardcoded to only work with nccl backend. Extend it
to allow NCCL + custom plugin backend.
The split-specific methods/attributes have not been added to the base
Backend and Options as some of them are specific to backend implementations.
Instead, explicit checks have been added to the split_group method for the
expected methods and attributes.
I am open to making them part of base Backend based if folks prefer.
Test Plan:
CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152175
Approved by: https://github.com/shuqiangzhang, https://github.com/kwen2501
We handle fake tensor caching in two ways:
1. If the inputs have no symbols (SymInt, etc) then we cache on the FakeTensorMode.
2. If the inputs have symbols then we cache on the ShapeEnv.
This way the symbols in the inputs and outputs are associated with the guards in place at the time of the call.
However - it's possible to have an op where there are no symbols in the inputs but there is an unbacked symbol in the output. In this case we shouldn't cache at all because what would that really mean?
So this PR changes the caching behavior so that if there's a symbol in the output which doesn't come in some way from the input then we refuse to cache that op.
Added a test which checks for this case.
While in there I also did a couple other related changes:
1. Added negative caching - if we see that an (op, args) failed to cache previously we don't even bother trying to cache it again.
2. Reworked the inner behavior of _cached_dispatch_impl a little to make it more clear which bits we expect to be able to throw _BypassDispatchCache and add some comments.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153034
Approved by: https://github.com/masnesral, https://github.com/tugsbayasgalan
## Summary
- The unit test `pytest test/distributed/test_symmetric_memory.py -k test_fused_scaled_matmul_reduce_scatter_scatter` was not running for some reason when #149247 was merged, giving false green CI signals. When it was ran manually recently, the test failed, highlighting a bug causing incorrect numerics when `scatter_dim=1`.
- This PR fixes the bug, which was related to how we swap dims 0<=>scatter_dim at the beginning of the custom op (for more efficient cross-device data movement I believe), then swap it back prior to reduction.
## Test plan
- I confirmed the unit test `pytest test/distributed/test_symmetric_memory.py -k test_fused_scaled_matmul_reduce_scatter_scatter` is now passing.
- I confirmed e2e training w/ torchtitan looks good ([logs](https://www.internalfb.com/phabricator/paste/view/P1812054188))
- I analyzed the tlparse to verify the fused_all_gather_matmul and fused_scaled_matmul_reduce_scatter both appear at least once in the post grad graphs ([tlparse](https://manifold.edge.x2p.facebook.net/v0/read/tree/logs/.tmpVbUsdG/dedicated_log_torch_trace_65oh3qj_.log/index.html?bucketName=tlparse_reports&apiKey=tlparse_reports-key&withPayload=1&timeoutMsec=10000))
## Next steps
1. I think for async TP `fused_scaled_matmul_reduce_scatter` we may only need `scatter_dim_after_maybe_reshape` and not `orig_scatter_dim` after all. I can confirm this and refactor if it is the case.
2. This op is specifically designed for async TP, and many of the arguments don't make sense for a user trying to use this as a standalone op. IMO we should have separate standalone custom op without all the extra function args and internal logic that doesn't apply to non-async TP cases.
3. In a follow up PR I want to add shape annotations to each line (e.g. `# (B, T, H)` etc) to make this easier to debug in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153595
Approved by: https://github.com/fegin
The immediate motivation is to make map support match
ExecuTorch so we can delete ExecuTorch-specific mapping functions, but
this should also straightforwardly improve performance.
Testing: there is existing coverage for this in
vec_test_all_types.cpp. Verified that it really does cover the newly
enabled "don't convert through float" paths by temporarily adding a
TORCH_INTERNAL_ASSERT(false).
Differential Revision: [D73802126](https://our.internmc.facebook.com/intern/diff/D73802126/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152366
Approved by: https://github.com/malfet
ghstack dependencies: #152365
Change trigger for auto format to be pull_request b/c the reusable action used gets the pr number from the pull_request event context, but only run it if ciflow/autoformat is attached to the PR. Tested this on a different PR, and it seems to be working
Changed tag name because ciflow prefixed labels have special handling
Also change to run on all files so it will mimic the normal CI lintrunner call, and because lintrunner, either by itself or using -m mergebase can miss some things. Idk if it would miss for format, but it does for checking lint. Format seems to take shorter than normal lint. I don't know if the comment about making suggestions on non edited file changes is a concern. I didn't really test this part
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153289
Approved by: https://github.com/atalman, https://github.com/malfet
Summary: To support running a subset of these tests with the remote autotuning utilities, I've split out some of the tests into separate classes so that I can derive from the "main" TestMaxAutotune class when creating new tests for remote. I'm not 100% sure what some of these tests do, so please suggest if another grouping / naming might make more sense. The remaining tests in TestMaxAutotune all smelled relevant to me.
Test Plan: existing unit tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153486
Approved by: https://github.com/eellison
This makes it more obvious what's going on when TCPStore shuts down while waiting on a remote key and also shows the remote address.
Test plan:
```
[W514 18:33:36.536327028 TCPStore.cpp:138] [c10d] recvValueWithTimeout failed on SocketImpl(fd=3, addr=[localhost]:34658, remote=[localhost]:1234): Failed to recv, got 0 bytes. Connection was likely closed. Did the remote server shutdown or crash?
```
```py
import os
rank = int(os.environ["RANK"])
import time
from torch import distributed as dist
store = dist.TCPStore(
host_name="localhost",
port=1234,
is_master=(rank == 0),
wait_for_workers=False,
)
time.sleep(1)
print("starting")
if rank != 0:
store.get("foo")
else:
time.sleep(1)
print("done")
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153586
Approved by: https://github.com/XilunWu
Change logging.error to logging.exception to log additional information when relevant. A few places have slipped in logging.errors in try except since I last did a clean up here and the rule is stabilized so I am enabling it codebase wide. I have NOQA'd much of our custom exception stack trace handling for RPC calls and distributed and tried to a fix a few errors based on whether we immediately reraised it or if we didn't print any exception handling where it could be useful.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153473
Approved by: https://github.com/albanD, https://github.com/cyyever
### Summary
Fixes#148494
Explicitly prefetch the cache lines of the next `B` block to accelerate int8 WoQ (BF16 activation, int8 statically quantized weights) GEMM for small `M` dimension.
Some of this code (outer loops of the GEMM) is being ported over from Intel Extension for PyTorch. The macro-kernel* and the micro-kernel* are essentially the same, but optionally prefetch a block of B. Templatization is being used to prevent branching causing a slowdown due to unnecessary prefetching.
\* - in [BLIS](https://dl.acm.org/doi/10.1145/2764454) parlance
### Performance data with BS 1
Machine: 32 cores of one socket of a Intel Xeon SP Gen 5 machine
| Model | input tokens | output tokens | next-token latency before this PR | Next-token latency after this change | Speedup |
|-----------|-------------|-----------------|--------------------------------------|------------------------------------------|-----------|
|GPT-J | 128 | 128 | 42 ms | 38 ms | 9.52 % |
| GPT-J | 1024 | 1024 | 48 ms | 45 ms | 6.25 % |
|LLaMA 3.1 8B Instruct | 128 | 128 | 52 ms | 47 ms| 9.61% |
|LLaMA 3.1 8B Instruct | 1024 | 1024 | 57 ms | 53 ms| 7.01% |
While the input shapes of GEMMs corresponding to linear for next-token computation remain the same in case of different number of input & output tokens, the difference in next-token latency is due to attention for those cases
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149373
Approved by: https://github.com/leslie-fang-intel, https://github.com/Xia-Weiwen
Co-authored-by: Xia Weiwen <xia.weiwen@hotmail.com>
https://github.com/pytorch/pytorch/pull/129001#discussion_r1645126801 is the motivation for the whole stack of PRs. In `torch/__init__.py`, `torch._C.Type` shadows `from typing import Type`, and there is no type stub for `torch._C.Type` in `torch/_C/__init__.pyi`. So we need to use `from typing import Type as _Type`. After enabling [Generic TypeAlias (PEP 585)](https://peps.python.org/pep-0585) in the `.pyi` type stub files, we can use `type` instead of `typing.Type` or `from typing import Type as _Type`.
------
- [Generic TypeAlias (PEP 585)](https://peps.python.org/pep-0585): e.g. `typing.List[T] -> list[T]`, `typing.Dict[KT, VT] -> dict[KT, VT]`, `typing.Type[T] -> type[T]`.
- [Union Type (PEP 604)](https://peps.python.org/pep-0604): e.g. `Union[X, Y] -> X | Y`, `Optional[X] -> X | None`, `Optional[Union[X, Y]] -> X | Y | None`.
Note that in `.pyi` stub files, we do not need `from __future__ import annotations`. So this PR does not violate issue #117449:
- #117449
------
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150727
Approved by: https://github.com/aorenste
ghstack dependencies: #150726
I slap disable on the recomputation hook, otherwise the partitioner may save less/more activations and mismatch with the expected eager count in checkpoint. See code comment `Note: [compiled autograd and checkpoint unpack hook]`.
This fixes all non-nested checkpointing tests. I also wrap nested checkpointing tests, and a few of them still fail.
This also seems to fix all PYTORCH_TEST_WITH_DYNAMO checkpointing tests except for `TestAutograd.test_checkpointing_without_reentrant_custom_function_works`. For those tests, it looks like we fail to HOPify the checkpointed region and when the backward executes the unpack hooks, dynamo tried to trace them. This messed up the internal state tracking of checkpointing, some raising the _StopRecomputationError and others raising the same count mismatch error as CA.
FIXES https://github.com/pytorch/pytorch/issues/127115
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153300
Approved by: https://github.com/jansel
When using Bazel, it’s common to encounter issues like [this](https://github.com/bazelbuild/bazel/issues/14640) and [this](https://github.com/bazel-contrib/rules_python/issues/792) where the `PYTHONPATH` environment variable becomes too long and results in an error such as: `OSError: [Errno 7] Argument list too long` . To work around this, users often resort to custom logic to manipulate PYTHONPATH.
Currently, PyTorch Inductor constructs the PYTHONPATH for a subprocess using sys.path, which can lead to this issue in certain environments.
This PR introduces support for a new environment variable, `TORCH_CUSTOM_PYTHONPATH`, allowing users to override the default `PYTHONPATH` passed to the subprocess. This provides a clean way to avoid an exception when using PyTorch in Bazel.
Please let me know if I need to add some documentation to support this PR. I haven't found an open issue specific to this change but I'm confident that this change (or a similar one) would be appreciated by few.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152832
Approved by: https://github.com/masnesral
**Summary**
This PR adds two new ops, `onednn.qadd.tensor` and `onednn.qadd_relu.tensor`, for int8 elementwise add, which accepts inputs on CPU device (instead of QuantizedCPU).
The new ops are implemented with AVX512 instructions and it provides similar or better performance, depending on shape, than its counterpart for QuantizedCPU device `quantized.add` and `quantized.add_relu`.
The new op supports output dtypes other than uint8 (fp32, fp16 and bf16 are supported).
**Test plan**
```
pytest test/quantization/core/test_quantized_op.py -k test_int8_add_onednn
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152411
Approved by: https://github.com/leslie-fang-intel, https://github.com/jerryzh168
Summary:
CUDA Post: https://fb.workplace.com/groups/ai.efficiency.tools.users/permalink/2020094788475989/
# Context
In this diff, we want to enable the on-demand mode of memory snapshot to allow user to trace any remote process via dyno command line.
# Design decision
**How do we send on-demand signal to remote process**
We leverage the dyno-Kineto approach.
Since dyno is running on all machine in Meta, it can send a request to the remote machine to start the Kineto.
Kineto will start another thread for memoryProfiler (https://fburl.com/code/dxsmmrok)
**why we use different approach as CUDA**
On CUDA side, we are using pybind to load torch Module and invoke the python api to start/stop the profiling. However, this requires us to compile the whole torch binary in the predictor which is not recommended by runtime(andruwang)
Thus, we decide to use the CPP api directly to avoid un-necessary dependency
**why the snapshot is saved as json string directly instead of pickle**
Pickle is primarily designed for use with Python and doesn't have well support in cpp. Also, it is hard for user to download the snapshot file and open locally.
Due to the dependency issue, it is hard to import the gzip/pickle library to decode the data. Thus, let's use JSON for now. I will work on the visualizer to fasten the render and support other format later.
**Plan**:
* Now, we will encoded file into gz for MTIA ondemand only and update the visualizer to support both type.
* Update auto-trace and CUDA side to encode in gzip as well
* Fully remove pickle dependency.
Test Plan:
# Remote cogwheel test
Servicelab: https://fburl.com/servicelab/pckux7a3
snapshot file manifold: https://fburl.com/manifold/fnotk18c
snapshot file in pastry: P1805522232
Visualization on D74399684
{F1977786422}
# Local Predictor Test
url: https://fburl.com/pytorch_memory_visualizer/y06kskkm
{F1977787329}
Differential Revision: D74179606
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153171
Approved by: https://github.com/sraikund16
Current implementation causes silent correction problem with torch.compile when someone tries to `torch.compile` function where one of the arguments is say `np.exp(.3)`, which will be represented as torch.float64 scalar tensor
Add regssion test for this behavior
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153582
Approved by: https://github.com/dcci
Fixes#147336
## Context
NCU analysis of the fp8 flex attention perf issue in #147336 showed an unexpected increase in shared memory access bank conflicts when loading the V tensor from HBM to SRAM.
Bringing this to the attention of triton developer @davidberard98 he identified the memory layout of the tensor in HBM to be causing non-pipelined loads into SRAM, causing the slowdown.
To summarize:
In flex attention when performing the FP8 GEMM `softmax_scores @ V` the right operand V must be in column-major memory layout. However, the `tl.load` of V blocks from HBM to SRAM cannot be pipelined if the V tensor isn't column-major in HBM already, leading to substantial performance degradation.
This is because triton does not perform async copies with the `cp.async` PTX instruction if the number of contiguous bytes is less than 4 (see [here](81f93f2c8e/lib/Dialect/TritonGPU/Transforms/Pipeliner/PipeliningUtility.cpp (L403))).
i.e., when loading 4 bytes of contiguous data from a tensor stored in row-major in HBM, we have to perform 4 separate non-contiguous writes to SRAM to place those bytes in their new location in the col-major layout in SRAM. Thus the load is not a candidate for pipelining w/ cp.async and just moves data to registers then performs a series of single byte stores.
## Fix summary
- To fix this, we should enforce memory layouts for Q, K, V in FlexAttention when fp8 is being used, to ensure they each exist in HBM in the necessary memory layout to facilitate pipelined loads into SRAM ahead of the FP8 GEMMs
## Benchmarks
Rerunning the repro we see fp8 runtime is reduced from 120% of bf16 to 76% of bf16 runtime.
Before fix:
```
(flex) [danvm@devgpu007.eag6 ~/ml-perf-tools/flex_attention (main)]$ rm -rf /tmp/torchinductor_${USER}; python profile_flex.py --bf16 --fp8
2025-05-11 19:07:33,402 - flex_bench - INFO - Running benchmark: bf16
2025-05-11 19:07:35,885 - flex_bench - INFO - bf16: 424.87228804347734 us
2025-05-11 19:07:35,893 - flex_bench - INFO - Running benchmark: fp8e4m3
2025-05-11 19:07:37,319 - flex_bench - INFO - fp8e4m3: 515.714000000001 us
```
After fix:
```
(flex) [danvm@devgpu007.eag6 ~/ml-perf-tools/flex_attention (main)]$ rm -rf /tmp/torchinductor_${USER}; python profile_flex.py --bf16 --fp8
2025-05-11 17:34:38,223 - flex_bench - INFO - Running benchmark: bf16
2025-05-11 17:34:41,157 - flex_bench - INFO - bf16: 423.4662032967036 us
2025-05-11 17:34:41,167 - flex_bench - INFO - Running benchmark: fp8e4m3
2025-05-11 17:34:42,917 - flex_bench - INFO - fp8e4m3: 326.3694803493453 us
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153357
Approved by: https://github.com/ngimel, https://github.com/davidberard98
Shameful admission: I have encountered this error 1-2 times, but don't have a repro.
torch/_inductor/select_algorithm.py", line 2022, in wait_on_futures
elapsed_times[future],
~~~~~~~~~~~~~^^^^^^^^
torch._inductor.exc.InductorError: KeyError: <Future at 0x7fc4e394fb90 state=finished returned tuple>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153417
Approved by: https://github.com/Skylion007, https://github.com/ColinPeppler
Address #151097. Including below changes,
- Add XPU support package 2025.1 build and test in CI for both Linux and Windows
- Keep XPU support package 2025.0 build in CI to ensure no break issue until PyTorch 2.8 release
- Upgrade XPU support package from 2025.0 to 2025.1 in CD for both Linux and Windows
- Enable XCCL in Linux CD wheel and oneMKL integration in both both Linux and Windows
- Update XPU runtime pypi packages of CD wheels
- Remove deprecated support package version docker image build
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151899
Approved by: https://github.com/EikanWang, https://github.com/atalman
Async compile workers don't respect inductor configs generally that get changed in the middle of execution because they warm up early. StaticCudaLauncher is especially susceptible to this because it affects triton compilation without being part of the inductor meta. So we'll pass it in via extra configs on each worker run.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153382
Approved by: https://github.com/masnesral, https://github.com/jansel
MOTIVATION
This PR includes a minor change to check for TEST_HPU flag as well before falling back to CPU. Without this flag, some tests were falling back to CPU causing them to fail.
Please refer to this RFC as well: https://github.com/pytorch/rfcs/pull/66
CHANGES
add TEST_HPU flag to some of the conditions checking the environment
use DEVICE_COUNT variable instead of torch.accelerator.device_count() API since the later is not supported on out-of-tree devices like Intel Gaudi.
@ankurneog , @EikanWang , @cyyever , @guangyey
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153461
Approved by: https://github.com/EikanWang, https://github.com/cyyever, https://github.com/albanD
Fixes https://github.com/pyro-ppl/pyro/issues/3419 which is actually a `torch` bug that can be replicated by the below code:
```
from torch import rand
from torch.distributions import MixtureSameFamily, Categorical, Binomial
max_count = 20
probs = rand(10, 5)
binom_probs = rand(10, 5)
d = MixtureSameFamily(Categorical(probs=probs), Binomial(max_count, binom_probs))
d.log_prob(d.sample())
```
which results in:
```
Traceback (most recent call last):
File "test.py", line 11, in <module>
d.log_prob(d.sample())
File "pytorch\torch\distributions\mixture_same_family.py", line 168, in log_prob
self._validate_sample(x)
File "pytorch\torch\distributions\distribution.py", line 315, in _validate_sample
valid = support.check(value)
^^^^^^^^^^^^^^^^^^^^
File "pytorch\torch\distributions\constraints.py", line 307, in check
(value % 1 == 0) & (self.lower_bound <= value) & (value <= self.upper_bound)
^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: The size of tensor a (10) must match the size of tensor b (5) at non-singleton dimension 1
```
### Fix explanation (only for cases when the component distribution contains parameters with batch dimenisons)
- The failure is due to sample validation taking place before padding in `MixtureSameFamily.log_prob`, and hence the fix is to pad before doing sample validation.
- The fix itself does not alter the calculations at all. It only affects the sample validation process.
- The failure does not occur with the component distribution set to the `Normal` distribution, as its validation is not defined elementwise (the validation itself is elementwise).
- I've split the `test_mixture_same_family_log_prob` test into two tests based on the `Normal` and `Binomial` distributions.
- Initially, the `Binomial` version of the test did not fail, but this was due to the component distribution having equal batch dimensions of (5, 5) so I changed it to (10, 5).
### Updated fix explanation (for all cases)
- The previous fix caused a bug in sample shape validation (which is done correctly) due to the padding taking place before the sample validation.
- The updated fix corrects the support to reflect the fact that the support of `MixtureSameFamily` is equal to the support of its components distribution with the first event dimension removed.
- This issue was already anticipated in the [code](331423e5c2/torch/distributions/mixture_same_family.py (L127)).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151317
Approved by: https://github.com/albanD, https://github.com/fritzo
By computing matmuls of only one random non-zero batch on CPU
This reduces test runtime from 11 minutes to 14 sec
```
% python3 test/test_mps.py -v -k test_large_bmm_
test_large_bmm_bfloat16 (__main__.TestMPS.test_large_bmm_bfloat16) ... ok
test_large_bmm_float16 (__main__.TestMPS.test_large_bmm_float16) ... ok
----------------------------------------------------------------------
Ran 2 tests in 27.495s
```
TODO: Compute it over two slices when https://github.com/pytorch/pytorch/issues/153560 is fixed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153562
Approved by: https://github.com/Skylion007, https://github.com/clee2000
When we do `torch.compile(module)`, we eventually end up returning a new
`OptimizedModule` instance, whose `forward` method is the result of
`torch.compile(mod.__call__)`, meaning it already captures all the extra
logic (e.g., hook firing) for the compiled module.
`OptimizedModule` also inherits `nn.module.__call__`, and thus
has its own hook logic. This is useful for torchao, which injects module
forward hooks to run in eager for quantization purposes.
However, this might create unexpected behavior for global module hooks,
because `torch.compile(module)` causes the hook to fire one extra time
for `OptimizedModule`, when compared to eager.
To preserve BC, we simply emit a warning for this behavior, and let
users decide what to do. This is reasonable because the global module
hooks are documented to be used for debugging/profiling purposes only.
Fixes#149502
Differential Revision: [D74611716](https://our.internmc.facebook.com/intern/diff/D74611716)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152740
Approved by: https://github.com/anijain2305, https://github.com/zou3519
Summary:
### Diff Context
- PR introduces Pipes for multiprocess comms with checkpointer process.
- Pipes allow easier comms contract management due to close() API and catch-all feature when background process is dead (e.g. seg faults).
Test Plan: CI
Differential Revision: D74668559
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153488
Approved by: https://github.com/saumishr
Since we have pyproject.toml metadata for [project] and [build-requires], let's turn on the linter rules which validates this optional metadata to make sure it's properly formatted and follows the correct schema for standard Python build tools.
Right now, incorrect metadata could silently error with how our CI is invoked or only provide warnings for invalid metadata. This check will help surface those errors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153543
Approved by: https://github.com/albanD
Helion relies on torch/fx/experimental 's fake_tensor tracing but does its own dtype checking, which conflicts with some meta kernel's existing dtype checking. This PR adds a config so that we skip those dtype checking in meta kernels and rely on the calling system to do the dtype checking.
Currently it only applies to `baddbmm`, but I expect that similar changes will need to be done to other meta kernels in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153513
Approved by: https://github.com/jansel
as titled, this PR improves the device selection logic when user did not
set the device before calling the DeviceMesh constructor, as a device
manager, DeviceMesh should try to set the device for users in a good
way.
The behavior of set_device before:
* If user call init_process_group to init a world process group, we assume user already called set_device and we don't set the device for the user
* If user does not init a world process group by themselves, we init a world process group for the user and follow a heuristic to set the device.
This is ok but sometimes the set_device heuristic wouldn't work well (i.e. if user use TORCH_CUDA_VISBILE_DEVICES
So this PR improves the device selection logic to:
* If the default cuda context is initialized by the time we init DeviceMesh, then we assume user must called some cuda operation before therefore must have selected the device by themselves
* If not the above, then we check if envvars have "LOCAL_RANK" and "WORLD_SIZE" from the launcher (i.e. torchrun), if so, we use "LOCAL_RANK" to set the device for the current process, which is a very standard practice. (This solves the TORCH_CUDA_VISBILE_DEVICES issue)
* If not above, then we throw warning to users about situation, and fallback to the old heuristic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150897
Approved by: https://github.com/tianyu-l
ghstack dependencies: #150898
Previously, we launch the a2av kernel with at most 8 blocks for intra-node cases, which turns out to saturate only 57 GB/s bandwidth.
This PR adds more blocks for intra-node, up to 8 per peer, pumping up data parallelism. The kernel now achieves 350 GB/s SOL for Hopper. See figure.
It also uses a simple tuning based on input size to avoid jumping to 8 CTAs directly (i.e. 1, 2, 4, then 8)
For inter-node, we cap at 8 blocks, since 57 GB/s seems bigger than regular NIC bandwidths (400 Gb/s).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153509
Approved by: https://github.com/ngimel
ghstack dependencies: #153483
This PR adds a tensor LR variant for the CPU Adagrad(fused=True).
I copied the behavior from the tensor LR variant of CPU Adam(fused=True), where the `lr.item()` is cast to a double and passed in the default function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153078
Approved by: https://github.com/janeyx99
Updates heuristic for bmm/baddbmm and consolidates all heuristic logic in a single location
- The goal of the consolidation is to improve maintainability and readability of the heuristic logic. Instead of different parts scattered across two files, this patch centralizes everything inside `Matmul.cpp`, where there already exists heuristic-based selection for mkldnn.
- The logic of the check itself doesn't change (existing code is reused where possible) but a separate heuristic threshold for bmm/baddbmm is introduced based on newer, benchmarking data. Use the script below to see the performance improvement for bmm from the new heuristic:
```
import torch
import time
# Set below to True to use cases selected by only one of the hueristics.
USE_ONLY_DIVERGENT_TEST_CASES = True
BATCH_SIZES = [ 1, 8, 32, 64, 128, 256 ]
M_DIMS = [ 4, 8, 16, 32, 64, 256, 512 ]
N_DIMS = [ 4, 8, 16, 32, 64, 256, 512 ]
K_DIMS = [ 4, 8, 16, 32, 64, 256, 512 ]
ITERS = 50
def old_heuristic(m, n, k):
is_above_min_dims = m > 8 and n > 8 and k > 8
is_above_min_size = m*n*k > 8_192
return is_above_min_dims and is_above_min_size
def new_heuristic(b, m, n, k):
return b*b*m*n*k >= 4_194_304
def generate_test_cases():
test_cases = []
for b in BATCH_SIZES:
for m in M_DIMS:
for n in N_DIMS:
for k in K_DIMS:
if USE_ONLY_DIVERGENT_TEST_CASES:
if old_heuristic(m, n, k) != new_heuristic(b, m, n, k):
test_cases.append([b, m, n, k])
else:
test_cases.append([b, m, n, k])
return test_cases
def test(x, y):
for _ in range(5):
torch.bmm(x, y)
perf = 0.0
for _ in range(ITERS):
start = time.time()
torch.bmm(x, y)
end = time.time()
perf += (end - start) / ITERS
return perf
def main():
print(f"{'b':<10}{'m':<10}{'n':<10}{'k':<10}{'time (s)':10}")
cumulative_mean_time = 0.0
for b, m, n, k in generate_test_cases():
mean_time = test(torch.rand(b, m, n), torch.rand(b, n, k))
cumulative_mean_time += mean_time
print(f"{b:<10}{m:<10}{n:<10}{k:<10}{mean_time:10.3e}")
print(f"Cumulative mean time = {cumulative_mean_time:.4f} s")
if __name__ == "__main__":
main()
```
From the script we see that cumulative mean time from all test cases (at 16 threads) is:
- 1.6195 s for the old heuristic
- 0.7012 s for the new heuristic
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149122
Approved by: https://github.com/fadara01, https://github.com/aditew01, https://github.com/malfet
In #117066, shutdown of the rendezvous was added if a worker shuts down. This is incorrect, because the rendezvous is actually shutdown in [this file](fa6f9eb2be/torch/distributed/launcher/api.py (L290)) but should not be shutdown if a signal is received. See also [this pull request](https://github.com/pytorch/pytorch/pull/67749).
#124819 then tried to remediate the situation by fixing the faulty shutdown for the restart case. But this is only triggered if the agent restarts the training, but not if the shutdown of the rendezvous happened before.
Removing both these changes restores the original behavior. The rendezvous should only be shutdown if a run completes or fails, not for a single worker leaving.
Fixes#150916Fixes#147064
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152525
Approved by: https://github.com/kiukchung
Summary: Use pybind11::gil_scoped_acquire instead of old impl as it will automatically take care of error handling. In the original implementation we missed releasing the GIL on each possible error which could put the program in a deadlock
Test Plan: Induced error manually and saw that GIL was released
Differential Revision: D74593564
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153415
Approved by: https://github.com/Skylion007, https://github.com/cyyever
Summary:
This diff adds a justknobs check for static cuda launcher. In particular, it supports a fractional rollout where each mast job/version can be consistently enrolled in the config on or off.
It also adds a set_feature_use so we can track whether static cuda launcher is enabled on a given dynamo compile.
Test Plan: Existing unit tests. The justknobs in question are set to be disabled right now, so this diff does not launch the feature yet.
Differential Revision: D74599203
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153400
Approved by: https://github.com/oulgen
This appears to be slow in production (potentially a quadratic explosion), and
logging this explicitly in pt2_compile_events and wait_counters makes it a lot easier to see how
bad of an issue this is.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153270
Approved by: https://github.com/masnesral
Summary: I forgot to remove this unused field in D73809989.
Test Plan: `buck test 'fbcode//mode/opt' fbcode//caffe2/test:fbonly -- --exact 'caffe2/test:fbonly - test_compilation_metrics_logger_in_sync (caffe2.test.fb.test_fb.TestFBOnly)'`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153413
Approved by: https://github.com/c00w
Adds create_graph support if you don't compile or compile only with torch.compile(backend="eager").
Using a backend that uses AOTDispatch produces a post-dispatch AOT backward, where its double backward will be silently incorrect if the forward trace involved any ops that are not composite implicit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153222
Approved by: https://github.com/jansel
ghstack dependencies: #153193
Toggling on `torch._dynamo.config.compiled_autograd = True` was erroring export (optimize_assert didn't have `rebuild_ctx` defined). Separately add a way to `rebuild_ctx` for `optimize_assert` since it is a public API.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153193
Approved by: https://github.com/jansel
Summary:
For MTIA on-demand mode, since we are not using torch Module. The data upload happens in cpp and doesn't support pickle.
Thus, we store as JSON at the end and need the update visualizer to support it
Test Plan: Check Test plan in D74179606
Differential Revision: D74406209
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153166
Approved by: https://github.com/sraikund16
`Error in dlopen: /lib/x86_64-linux-gnu/libstdc++.so.6: version GLIBCXX_3.4.30 not found` error was caused by cmake migration (as conda one probably have some extra link rules), while `C++ exception with description "CUDA error: no kernel image is available for execution on the device` were caused by the fact that test were build for Maxwell, but run on SM_86
Remaining test was failing before, but was probably disabled
TODOs:
- Move build to the build step
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153423
Approved by: https://github.com/huydhn, https://github.com/cyyever
Summary: This change is to support remote autotuning. I want to use all the same benchmarking utilities in select_algorithm.py. For remote autotuning, I'll reuse the TritonBenchmarkRequest class used for subprocess autotuning because it's already serializable. That class is also used in standard, in-process autotuning, but via TritonTemplateCaller.benchmark() which sets the output_tensor param when calling the underlying TritonBenchmarkRequest. For remote, I'll be using the TritonBenchmarkRequest request directly so I want the parameter to be named 'out' to avoid "got an unexpected keyword argument 'out'".
Test Plan: Existing unit tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153169
Approved by: https://github.com/aorenste, https://github.com/eellison
Support xpu tf32 matmul using torch.bachend.mkldnn.allow_tf32, we will discuss in future if we need a new api to control matmul only
~~Support xpu tf32 matmul using torch.set_float32_matmul_precision. For conv, check https://github.com/pytorch/pytorch/pull/137570
We decide not following torch.backends.cuda.matmul.allow_tf32 because this API actually calls setAllowTF32CuBLAS to set matmul_precison to high. We also avoid other related tf32 changes (i.e. in inductor) by not introducing new API.~~
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144240
Approved by: https://github.com/EikanWang
By invalidating all variable created during the loop except for the context of iterator_cache, as storage can be done inside reduction loop and clear `IteratorRangeEntry` codegen cache.
Which results in the following kernel for `x / x.sum()` if x size is 2048 and max thread group size is 1024
```metal
[[max_total_threads_per_threadgroup(1024)]]
kernel void generated_kernel(
device half* out_ptr1,
constant half* in_ptr0,
uint2 thread_pos [[thread_position_in_grid]],
uint2 group_pos [[thread_position_in_threadgroup]]
) {
auto xindex = thread_pos.x;
auto r0_index = thread_pos.y;
threadgroup float tmp_acc_0[32];
float tmp_acc_1 = 0;
for(auto r0_0_cnt = 0; r0_0_cnt < 2; ++r0_0_cnt) {
int r0_0 = 2 * r0_index + r0_0_cnt;
auto tmp0 = static_cast<float>(in_ptr0[r0_0]);
tmp_acc_1 += tmp0;
}
auto tmp1 = c10:🤘:threadgroup_sum(tmp_acc_0, tmp_acc_1, r0_index * 1, 1024);
for(auto r0_0_cnt = 0; r0_0_cnt < 2; ++r0_0_cnt) {
int r0_0 = 2 * r0_index + r0_0_cnt;
auto tmp2 = static_cast<float>(in_ptr0[r0_0]);
auto tmp3 = tmp2 / tmp1;
out_ptr1[r0_0] = static_cast<half>(tmp3);
}
}
```
Fixes compilation report reported while running `GPUTests.test_pattern_matcher_multi_user_mps` and `GPUTests.test_weight_norm_bwd_mps`
Fixes https://github.com/pytorch/pytorch/issues/152155
Though inductor tests are still failing, need to keep refining the variable invalidation
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153362
Approved by: https://github.com/manuelcandales, https://github.com/dcci, https://github.com/jansel
During the dump of FR, due to some unknown reasons, we see cuda errors when querying events and this leads to the failures of whole FR dumps (when trying to get entries). So we do a try-catch instead of let it fails the whole process.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153414
Approved by: https://github.com/d4l3k
Improves typing so that all the optimizer subclasses (which all of them that subtype step) do not erase their type signature when this decorator is used. Now *kwarg values and returns will propogate
This complements @tsunghsienlee PR #153367 as the type signature of step() was being erased on all the optimizer subclasses by this untyped decorator
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153374
Approved by: https://github.com/janeyx99, https://github.com/tsunghsienlee
Summary:
To add PT2 information to memory snapshot we piggyback off of the Kineto implementation using record_function similar to adding the user annotations. To do this we add the following:
1. Stack implementation that we instantiate to keep track of which compile context stack we are currently in (top element of the stack). The stack will be per device and thread-local since different threads of a process can be in different compile contexts at a given time. For this reason, we do not need to add mutexes to our stack impl since no two threads will touch a given stack
2. RecordFunction hooks to properly pipe the correct events to the compile context stack. These hooks are similar to the annotation ones in the fact that we just register them lazily and DO NOT unregister them. This is done out of convenience. In the future, we should save the handles and unregister them to minimize overhead after profiling is finished. As of now, we are registering this at the FUNCTION scope which is wide; however, we treat any function that does not start with "Torch-Compiled Region" as a no-op so we anticipate the difference in performance to be negligible during and after profiling. We also hide this feature behind a flag set to off on default so existing jobs will be unaffected
3. Piping for compile context to pickle output
Test Plan:
In D74039793, we add CompileContext to the visualizer and we see the following {F1977654658}
Differential Revision: D74028214
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152707
Approved by: https://github.com/eqy
- Move community and language binding links to the horizontal bar
- Add an intro to the community page.
- Fix the link in the ogp_image
- Fix the link in the version switcher
- Clean up unneeded links
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153090
Approved by: https://github.com/albanD
Summary:
Previously, we only support non-scaling quantization, which may lead to overflow, here we support scaling quantization, and set it as the default version.
Here, we quantize activation nodes based on the size_in_mb, the default value is 100, i.e., as long as the node has at least 100MB size, we will quantize it.
Test Plan:
### how to enable
```
torch._inductor.config.post_grad_fusion_options = {
"activation_quantization_aten_pass": {
"quant_type": "torch.float8_e5m2", -> default is this type to quantize, you can change the type
"use_scaling": False, -> default is False, if you want to use scaling verison, set it to True
"size_in_mb": 0.0, -> default is 100, you can tune the value.
"exclude_primals": False, -> whether want to exclude quantize parameters, default is False
"allowed_dtypes": "torch.float16;torch.bfloat16;torch.float32", -> dtype you consider to quant, use ";" to separate, default is torch.bfloat16
},
}
```
### toy model
```
buck2 run mode/opt //scripts/qyz/autoac:quantization
```
```
Epoch [80/200], Loss: 19227.2109
Epoch [100/200], Loss: 1353.5272
Epoch [120/200], Loss: 38630.6758
Epoch [140/200], Loss: 6239.9155
Epoch [160/200], Loss: 6039.1567
Epoch [180/200], Loss: 3994.3569
Epoch [200/200], Loss: 146.3966
```
Differential Revision: D73015996
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151770
Approved by: https://github.com/Mingming-Ding
https://github.com/pytorch/pytorch/pull/146942 introduced an 8.3% regression on the `benchmark_torchbench_run_bert_pytorch_training:defaults-speedup-x1000` perf metric. This was flagged by internal CI testing (task T223596372).
The root cause seems to be that `FloorDiv` is now used to calculate the launch grid in certain scenarios, which is slower than the previously-used `//`. Since launch grid calculations happen at runtime, they can have a significant performance impact on some models.
The reason for switching to `FloorDiv` in https://github.com/pytorch/pytorch/pull/146942 was to allow the FX backend to generate runnable Python code. `FloorDiv(x, y)` maps to `x // y` in Python, whereas `sympy.floor(sympy.Rational(x,y))` maps to `floor(x/y)`, which crashes as FX doesn't know what `floor` is.
To get the best of both worlds, this PR reverts to using `//` to calculate launch grids, but then post-processes the resulting sympy expressions in the FX converter, converting `floor(x / y)` to `FloorDiv(x, y)`. Since this sympy manipulation happens at compile time, the perf impact should minimal, and should only affect the FX backend. This is similar to the approach previously explored in https://github.com/pytorch/pytorch/pull/151144, but the implementation is more minimal and self-contained.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153230
Approved by: https://github.com/jansel
The function signature of fused_scaled_matmul_reduce_scatter was changed. This PR fixes the function signature. However when scatter_dim is 1, the two outputs are not close. We need a followup on this.
Another followup is to change fused_scaled_matmul_reduce_scatter to make those newly added arguments optional. Users shouldn't need to these arguments if they don't flatten the inputs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153286
Approved by: https://github.com/kwen2501
When codegen'ed, it looks like:
```py
post_grad_custom_pass = <object at 0x12345678>
```
Which is not runnable at all. Some logic is also trying to deepcopy the
object, and not all of these objects are deepcopy-able.
This PR skips codegenning of these passes.
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153243
Approved by: https://github.com/houseroad
## Context
Suppose we have this graph like this :
```
a: "[s1 + u2, 200]"
b: "[u0, 32]"
cat: "[s1 + u2, 232]" = torch.cat([a, b], dim=1)
```
NOTE: torch.cat assumes "all tensors must either have the same shape (except in the concatenating dimension) or be a 1-D empty tensor with size (0,)."
So, we would expect u0 = s1 + u2 which is guarded on today except it's a deferred runtime assertion since unbacked symints aren't replaced today as Pian.
Notice how a has a different symbolic shape than both b and cat. Today, this will create an unexpected shape mismatch when AOTI autotunes. Here's a rough illustration where 8192 is the unbacked symint fallback value.
```
# s1 is an arbitrary integer
a = generate_example_value(size=(s1 + 8192, 200))
b = generate_example_value(size=(8192, 32))
out = generate_example_value(size=(s1 + 8192, 232))
triton_cat.run(a, b, out ...)
```
## Error
```
wrapper.py:1484: <module>: block: [443,0,0], thread: [53,0,0] Assertion `index out of bounds: 0 <= tl.broadcast_to(tmp13, [XBLOCK]) < ks0` failed.
...
wrapper.py:1484: <module>: block: [443,0,0], thread: [55,0,0] Assertion `index out of bounds: 0 <= tl.broadcast_to(tmp13, [XBLOCK]) < ks0` failed.
RuntimeError: CUDA error: device-side assert triggered
```
Differential Revision: [D74485962](https://our.internmc.facebook.com/intern/diff/D74485962)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153220
Approved by: https://github.com/desertfire
DSD currently will pop tensors if these tensors are on Meta device. This forbid the use cases that users would like to let DCP to directly initialize the tensors when loading.
This PR also removes test/distributed/checkpoint/e2e/test_pipeline.py which is based on the above feature that is not realistic and is not used anywhere.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153185
Approved by: https://github.com/mori360
Follow up to @ezyang's PR #153020 , but better uses PEP621 to reduce redundant fields and pass through metadata better to uv, setuptools, poetry and other tooling.
* Enables modern tooling like uv sync and better support for tools like poetry.
* Also allows us to set project wide settings that are respected by linters and IDE (in this example we are able centralize the minimum supported python version).
* Currently most of the values are dynamically fetched from setuptools, eventually we can migrate all the statically defined values to pyproject.toml and they will be autopopulated in the setuptool arguments.
* This controls what additional metadata shows up on PyPi . Special URL Names are listed here for rendering on pypi: https://packaging.python.org/en/latest/specifications/well-known-project-urls/#well-known-labels
These also clearly shows us what fields will need to be migrated to pyproject.toml over time from setup.py per #152276. Static fields be fairly easy to migrate, the dynamically built ones like requirements are a bit more challenging.
Without this, `uv sync` complains:
```
error: No `project` table found in: `pytorch/pyproject.toml`
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153055
Approved by: https://github.com/ezyang
Unblocks #153020 which accidentally improves the CircularImportLinter to check all Python files. It doesn't set a logname so it errors, there is another FSDP script which already defaults LOGNAME to '' if not specified, this does the same.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153324
Approved by: https://github.com/awgu
When trying to plot a trace graph, Inductor checks if "dot" is installed. Currently, the code runs a "which dot" command.
By default, Windows doesn't have the "which" command. This patch replaces it with the more portable alternative.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153259
Approved by: https://github.com/Skylion007
Fixes#152918
Before:
```
File "/data/users/bobren/a/pytorch/torch/fx/experimental/symbolic_shapes.py", line 5588, in produce_guards_verbose
raise ConstraintViolationError(
torch.fx.experimental.symbolic_shapes.ConstraintViolationError: Constraints violated (L['x'].size()[0])! For more information, run with TORCH_LOGS="+dynamic".
- You marked L['x'].size()[0] as dynamic but your code specialized it to be a constant (5). Either remove the mark_dynamic or use a less strict API such as maybe_mark_dynamic or Dim.AUTO.
```
After:
```
File "/data/users/bobren/a/pytorch/torch/fx/experimental/symbolic_shapes.py", line 5588, in produce_guards_verbose
raise ConstraintViolationError(
torch.fx.experimental.symbolic_shapes.ConstraintViolationError: Constraints violated (L['x'].size()[0])! For more information, run with TORCH_LOGS="+dynamic".
- You marked L['x'].size()[0] as dynamic but your code specialized it to be a constant (5). Either remove the mark_dynamic or use a less strict API such as maybe_mark_dynamic or Dim.AUTO.
User stack:
File "/home/bobren/local/a/pytorch/error.py", line 5, in foo
return torch.randn(5) * x
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152924
Approved by: https://github.com/pianpwk
Based on the [conversation](https://github.com/pytorch/pytorch/issues/121791), we plan to drop the "highest, high, medium" to represent fp32 internal computation data types . Instead, we will directly use the algorithm to represent it.
### Design Choice: Directly use algorithms name like "TF32", "BF16".
#### Pros
- The names are more informative. 'tf32' is more informative than a simple "high".
- Easier to extend new algorithm like `tf32x3`
#### Cons
- "HIGHEST, HIGH, MEDIUM" indicated the relative precision between different algorithms. However, we can have more documents to discuss them.
### We provide a layered structure for backends/operators.
('f32' is short for 'fp32_precision')
### We provide 3 fp32 compute precision can be set:
- **"ieee"**: Not allowed to use any other internal computation data types .
- **"tf32"**: Allowed to use tf32 as internal computation data types.
- **"bf16"**: Allowed to use bf16 as internal computation data types.
- **"none"**: Precision's are not set. Can be override by its father node.
### Overriding Precision Settings
Child node can be override by its father node if it is set to default.
For current default settings:
```
backend = generic, op = all, precision setting = none
backend = cuda, op = all, precision setting = none
backend = cuda, op = conv, precision setting = tf32
backend = cuda, op = rnn, precision setting = tf32
backend = cuda, op = matmul, precision setting = none
backend = matmul, op = all, precision setting = none
backend = matmul, op = conv, precision setting = none
backend = matmul, op = rnn, precision setting = none
backend = matmul, op = matmul, precision setting = none
```
- If the user set `torch.backends.mkldnn.fp32_precision="bf16"`, his child nodes `torch.backends.mkldnn.matmul.fp32_precision` / `torch.backends.mkldnn.conv.fp32_precision` / `torch.backends.mkldnn.rnn.fp32_precision` will also be override to "bf16".
- If the user set `torch.backends.fp32_precision="bf16"`, `torch.backends.mkldnn.fp32_precision` and his child nodes will also we override to "bf16".
### Backward Compatible
Since new API allow user to have more fine-grained control. There will be some conflict. For example, previous `torch.backends.cudnn.allow_tf32` are not enough to represent the status for `torch.backends.cudnn.rnn.fp32_precision="ieee"` and `torch.backends.cudnn.conv.fp32_precision="tf32"`. Therefore, our goal for backward compatible is
- If the user only uses previous APIs, it will work as previous expectations.
- If the user use **new** API to change the status to an **un-representable** status for old API, and try to access the status by **old** API. We will raise Runtime Error and point the document for user.
### Test Plan
```
python test/test_cuda.py -k test_fp32_precision_with_tf32
python test/test_cuda.py -k test_fp32_precision_with_float32_matmul_precision
python test/test_cuda.py -k test_invalid_status_for_legacy_api
python test/test_mkldnn.py -k test_mlkdnn_get_set
python test/test_mkldnn.py -k test_generic_precision
python test/test_mkldnn.py -k test_invalid
python test/test_mkldnn.py -k test_default_use_parent
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125888
Approved by: https://github.com/jgong5, https://github.com/albanD
Co-authored-by: Jiang, Yanbing <yanbing.jiang@intel.com>
Fixes#132644
`_batch_p2p` incorrectly assumes that `dist.batch_isend_irecv` returns a single-element list of `dist.Work`, likely due to NCCL's coalescing behaviour.
For none NCCL backends like Gloo, multiple `dist.Work` objects are returned, causing the code to discard some operations via `.pop()`. This leads to deadlocks during pipeline parallelism.
## Changes:
* Modified `_batch_p2p` to return `list[dist.Work]` instead of popping a single element.
* Added `_wait_batch_p2p` to call `wait()` on multiple `dist.Work` objects, consuming the result of `_batch_p2p`.
* Updated references from `dist.Work` to `list[dist.Work]`.
## Testing:
* `pippy_bert.py` from #132644 now works with gloo.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152938
Approved by: https://github.com/kwen2501, https://github.com/H-Huang
Summary: _fold_conv_bn_qat has logic to remove the tracking stats. Currently, this includes a check that includes only torch.nn.modules.batchnorm.BatchNorm2d. As a result, the tracking stats are not properly removed when 1D is used. This diff updates to fix this.
Test Plan:
Run N7113483 without this fix.
{F1977726982}
```
bento kernel build sensorml
```
Re-run with local version of kernel, containing this diff:
{F1977727151}
Notice that now, num_batches is removed.
Differential Revision: D74269649
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152982
Approved by: https://github.com/andrewor14, https://github.com/yushangdi
Error out if K=0 in one of the grouped gemms to avoid hangs in #152668
Also, adds meta function for _scaled_grouped_mm (TODO: do the same for _grouped_mm, unless it's done already)
One weird thing I'm seeing, when running all grouped_gemm tests, I'm erroring out with
```
File "/data/users/ngimel/pytorch/torch/_inductor/graph.py", line 1246, in call_function
out = lowerings[target](*args, **kwargs) # type: ignore[index]
File "/data/users/ngimel/pytorch/torch/_inductor/lowering.py", line 445, in wrapped
out = decomp_fn(*args, **kwargs)
File "/data/users/ngimel/pytorch/torch/_inductor/kernel/mm_scaled_grouped.py", line 444, in tuned_scaled_grouped_mm
if is_nonzero and can_use_triton_kernel(mat_a, mat_b, offs, bias):
File "/data/users/ngimel/pytorch/torch/_inductor/kernel/mm_scaled_grouped.py", line 375, in can_use_triton_kernel
offs is not None
File "/home/ngimel/.conda/envs/pytorch_monarch/lib/python3.10/site-packages/sympy/core/relational.py", line 516, in __bool__
raise TypeError("cannot determine truth value of Relational")
torch._inductor.exc.InductorError: LoweringException: TypeError: cannot determine truth value of Relational
```
which is weird, there's no relational that sympy has to evaluate in `offs is not None`, and when running this test separately (`test_scaled_grouped_gemm_2d_3d_fast_accum_True_strided_False_use_torch_compile_True_cuda`) it passes. I suspect some autotuning cache has to be reset between runs, but don't know what to look for.
Edit: that error is "fixed" by setting `dynamic=False`, now with correct meat function something's wrong with dynamic shapes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153226
Approved by: https://github.com/kwen2501
Summary: - Expose `c10_retrieve_device_side_assertion_info()` for use by external code. The motivating use case is FBGEMM kernel launcher utilities, which add FBGEMM-specific context to the errors coming out of Torch DSA
Test Plan: OSS CI
Differential Revision: D74432771
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153211
Approved by: https://github.com/Skylion007
This PR moves `mps_linear` to use MPSNDArrays and call into the MPS kernel directly instead of going through MPSGraph. It also adds a caching mechanism for reusing MPS kernels as there is also a small overhead attached to creating the kernel object.
The impact of the improvement is relatively more significant for small input kernels where the MPSGraph overhead represents a larger portion of the overall execution time of the operation but the speedup shows for both small and large input sizes as expected.
`mps_linear` before the changes:
```
input shapes: f32:[1,1,20], f32:[1,20]
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x109d67110>
func(*args, **kwargs)
Median: 199.29 us
IQR: 9.56 us (196.71 to 206.27)
979 measurements, 1 runs per measurement, 1 thread
input shapes: f32:[1,1,5120], f32:[13284,5120]
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x1063b4510>
func(*args, **kwargs)
Median: 979.29 us
IQR: 25.29 us (964.83 to 990.13)
205 measurements, 1 runs per measurement, 1 thread
```
`mps_linear` after the changes:
```
input shapes: f32:[1,1,20], f32:[1,20]
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x10693a190>
func(*args, **kwargs)
Median: 176.08 us
IQR: 15.02 us (172.42 to 187.44)
1103 measurements, 1 runs per measurement, 1 thread
input shapes: f32:[1,1,5120], f32:[13284,5120]
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x10d524dd0>
func(*args, **kwargs)
Median: 952.56 us
IQR: 15.63 us (945.47 to 961.10)
210 measurements, 1 runs per measurement, 1 thread
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152210
Approved by: https://github.com/kulinseth, https://github.com/malfet
Co-authored-by: Nikita Shulga <nshulga@meta.com>
Summary:
Solves https://github.com/pytorch/pytorch/issues/151925
A reland of https://github.com/pytorch/pytorch/pull/152125.
added a try-except around the justknob internally. Also added more documentation
Currently, AOTI only generate runtime asserts for unbacked symints. We should generate asserts for all `_assert_scalar` calls in the input graph.
Also factored out the run time assertion logic to a separate function.
We need to generate runtime asserts directly in Inductor instead of just re-using the asserts from input graphs becase we reuse the same ShapeEnv as before. In particular, on subsequent graph passes, we would immediately turn all of these assertions into noops,
because when we evaluated their expressions, we would see that because we had a deferred runtime assert in the ShapeEnv, we know "oh, of course this expression is True" already.
One example is below:
```
class Model(torch.nn.Module):
def forward(self, a, b, c):
nz = torch.nonzero(a)
ones = a.new_ones([nz.size(0), b.size(0)])
torch._check(ones.size(0) >= 1)
equals = torch.add(ones, c)
return equals
torch._dynamo.mark_dynamic(c, 0)
```
When we re-use the ShapeEnv in Inductor lowering, the check that checks a and nonzero have the same shape would be evaluted to True after we resolve unbacked bindings using the ShapeEnv.
See `test_unbacked_equals_input_size_runtime_assertion` in test_aot_inductor.
In addition to the Inductor generated runtime asserts, we also need the runtime asserts from the input graph, because some derived runtime asserts are not generated in Inductor. One example is below:
```
class Model(torch.nn.Module):
def forward(self, x):
y = x.reshape(100, -1).clone()
y = y + 1
return y
dynamic_shapes = {
"x": {0: torch.export.Dim.DYNAMIC},
}
x.shape[0] needs to be a multiple of 100.
```
See `test_aoti_runtime_asserts_backed_symint` in test_aot_inductor.
Example:
```
def forward(self):
arg0_1: "f32[s35]";
arg0_1, = fx_pytree.tree_flatten_spec([], self._in_spec)
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:11 in forward, code: y = x.reshape(100, -1).clone()
sym_size_int: "Sym(s35)" = torch.ops.aten.sym_size.int(arg0_1, 0)
#
mod: "Sym(Mod(s35, 100))" = sym_size_int % 100; sym_size_int = None
eq_2: "Sym(Eq(Mod(s35, 100), 0))" = mod == 0; mod = None
_assert_scalar = torch.ops.aten._assert_scalar.default(eq_2, "Runtime assertion failed for expression Eq(Mod(s35, 100), 0) on node 'eq'"); eq_2 = _assert_scalar = None
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:11 in forward, code: y = x.reshape(100, -1).clone()
view: "f32[100, (s35//100)]" = torch.ops.aten.reshape.default(arg0_1, [100, -1]); arg0_1 = None
clone: "f32[100, (s35//100)]" = torch.ops.aten.clone.default(view); view = None
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:12 in forward, code: y = y + 1
add_6: "f32[100, 1]" = torch.ops.aten.add.Tensor(clone, 1); clone = None
return (add_6,)
```
Generated cpp code:
```
auto inputs = steal_from_raw_handles_to_raii_handles(input_handles, 1);
auto arg0_1 = std::move(inputs[0]);
auto arg0_1_size = arg0_1.sizes();
int64_t s35 = arg0_1_size[0];
inputs.clear();
auto& kernels = static_cast<AOTInductorModelKernels&>(*this->kernels_.get());
if (!((s35 % 100L) == 0L)) { throw std::runtime_error("Expected Eq(Mod(s35, 100), 0) to be True but received " + std::to_string(s35)); }
```
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r aoti_runtime_asserts_backed_symint
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:torchinductor_dynamic_shapes -- -r test_unbacked_floordiv_simplify
TORCHINDUCTOR_SCALAR_ASSERTS_FULL=1 buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r test_sym_i64_input_codegen_cuda
TORCHINDUCTOR_SCALAR_ASSERTS_FULL=1 buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r test_unbacked_equals_input_size
```
Differential Revision: D74361799
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153182
Approved by: https://github.com/henrylhtsang
Motivation is that `AlgorithmSelectorCache.__call__` is getting very long and hard to work with. There are nested layers of local functions in it. For example, we pass `precompile_fn`, a local variable, to `do_autotuning`, a local function, which already has a pointer to choices, a local variable, and then have `do_autotuning` calls `choices` in `self.lookup`.
When I was trying to make changes to do_autotuning, I would get `UnboundLocalError: cannot access local variable 'choices' where it is not associated with a value`. But no idea why it was even working in the first place.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153212
Approved by: https://github.com/eellison
* Cleans up code in `caffe2/CMakeLists.txt` to remove individual ROCm library include paths and use `ROCM_INCLUDE_DIRS` CMake var instead
* `ROCM_INCLUDE_DIRS` CMake var is set in `cmake/public/LoadHIP.cmake` by adding all the ROCm packages that PyTorch depends on
* `rocm_version.h` is provided by the `rocm-core` package, so use the include directory for that component to be compliant with Spack
* Move `find_package_and_print_version(hip REQUIRED CONFIG)` earlier so that `hip_version.h` can be located in the hip package include dir for Spack
* `list(REMOVE_DUPLICATES ROCM_INCLUDE_DIRS)` to remove duplicate `/opt/rocm/include` entries in the non-Spack case
* Remove user-provided env var `ROCM_INCLUDE_DIRS` since `ROCM_PATH` already exists as a user-provided env var, which should be sufficient to locate the include directories for ROCm.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152569
Approved by: https://github.com/renjithravindrankannath, https://github.com/jeffdaily
Co-authored-by: Renjith Ravindran <Renjith.RavindranKannath@amd.com>
We handle fake tensor caching in two ways:
1. If the inputs have no symbols (SymInt, etc) then we cache on the FakeTensorMode.
2. If the inputs have symbols then we cache on the ShapeEnv.
This way the symbols in the inputs and outputs are associated with the guards in place at the time of the call.
However - it's possible to have an op where there are no symbols in the inputs but there is an unbacked symbol in the output. In this case we shouldn't cache at all because what would that really mean?
So this PR changes the caching behavior so that if there's a symbol in the output which doesn't come in some way from the input then we refuse to cache that op.
Added a test which checks for this case.
While in there I also did a couple other related changes:
1. Added negative caching - if we see that an (op, args) failed to cache previously we don't even bother trying to cache it again.
2. Reworked the inner behavior of _cached_dispatch_impl a little to make it more clear which bits we expect to be able to throw _BypassDispatchCache and add some comments.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153034
Approved by: https://github.com/masnesral, https://github.com/tugsbayasgalan
This patch fixes a corner case of `torch.isin` decompisition when both
inputs are scalars. This pattern showed up from #141196.
Fixes#141196.
Error stack befor this patch:
```
File "/home/ryanguo99/repos/pytorch/test/dynamo/test_misc.py", line 12503, in test_scalar_isin_decomposition
res = opt_f()
^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/eval_frame.py", line 691, in _fn
raise e.remove_dynamo_frames() from None # see TORCHDYNAMO_VERBOSE=1
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/output_graph.py", line 1618, in _call_user_compiler
raise BackendCompilerFailed(
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/output_graph.py", line 1593, in _call_user_compiler
compiled_fn = compiler_fn(gm, self.example_inputs())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/repro/after_dynamo.py", line 150, in __call__
compiled_gm = compiler_fn(gm, example_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/__init__.py", line 2365, in __call__
return compile_fx(model_, inputs_, config_patches=self.config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_inductor/compile_fx.py", line 2317, in compile_fx
return aot_autograd(
^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/backends/common.py", line 106, in __call__
cg = aot_module_simplified(gm, example_inputs, **self.kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/aot_autograd.py", line 1179, in aot_module_simplified
compiled_fn = AOTAutogradCache.load(
^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/autograd_cache.py", line 923, in load
compiled_fn = dispatch_and_compile()
^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/aot_autograd.py", line 1164, in dispatch_and_compile
compiled_fn, _ = create_aot_dispatcher_function(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/aot_autograd.py", line 576, in create_aot_dispatcher_function
return _create_aot_dispatcher_function(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/aot_autograd.py", line 826, in _create_aot_dispatcher_function
compiled_fn, fw_metadata = compiler_fn(
^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 180, in aot_dispatch_base
fw_module, updated_flat_args, maybe_subclass_meta = aot_dispatch_base_graph( # type: ignore[misc]
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 2199, in _trace_inner
t = dispatch_trace(
^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_compile.py", line 51, in inner
return disable_fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/eval_frame.py", line 872, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 1223, in dispatch_trace
graph = tracer.trace(root, concrete_args) # type: ignore[arg-type]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_dynamo/eval_frame.py", line 872, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/_symbolic_trace.py", line 850, in trace
(self.create_arg(fn(*args)),),
^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 1278, in wrapped
out = f(*tensors) # type:ignore[call-arg]
^^^^^^^^^^^
File "<string>", line 1, in <lambda>
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/traced_function_transforms.py", line 720, in inner_fn
outs = fn(*args)
^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/traced_function_transforms.py", line 419, in _functionalized_f_helper
f_outs = fn(*f_args)
^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/traced_function_transforms.py", line 81, in inner_fn
outs = fn(*args)
^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_functorch/_aot_autograd/traced_function_transforms.py", line 902, in functional_call
out = PropagateUnbackedSymInts(mod).run(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/interpreter.py", line 171, in run
self.env[node] = self.run_node(node)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/symbolic_shapes.py", line 7387, in run_node
result = super().run_node(n)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/interpreter.py", line 240, in run_node
return getattr(self, n.op)(n.target, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/interpreter.py", line 320, in call_function
return target(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 1326, in __torch_function__
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_subclasses/functional_tensor.py", line 511, in __torch_dispatch__
outs_unwrapped = func._op_dk(
^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/utils/_stats.py", line 27, in wrapper
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 1428, in __torch_dispatch__
return proxy_call(self, func, self.pre_dispatch, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 797, in proxy_call
r = maybe_handle_decomp(proxy_mode, func, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/fx/experimental/proxy_tensor.py", line 2358, in maybe_handle_decomp
out = CURRENT_DECOMPOSITION_TABLE[op](*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_prims_common/wrappers.py", line 309, in _fn
result = fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_decomp/decompositions.py", line 5108, in isin
return isin_default(elements, test_elements, invert=invert)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/ryanguo99/repos/pytorch/torch/_decomp/decompositions.py", line 5137, in isin_default
x = elements.view(*elements.shape, *((1,) * test_elements.ndim))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
TypeError: view() received an invalid combination of arguments - got (), but expected one of:
* (torch.dtype dtype)
* (tuple of ints size)
While executing %isin : [num_users=1] = call_function[target=torch.isin](args = (%x, %x), kwargs = {})
GraphModule: class GraphModule(torch.nn.Module):
def forward(self):
# File: /home/ryanguo99/repos/pytorch/test/dynamo/test_misc.py:12498 in f, code: x = torch.tensor(0)
x: "i64[][]" = torch.tensor(0)
# File: /home/ryanguo99/repos/pytorch/test/dynamo/test_misc.py:12499 in f, code: return torch.isin(x, x)
isin: "b8[][]" = torch.isin(x, x); x = None
return (isin,)
Original traceback:
File "/home/ryanguo99/repos/pytorch/test/dynamo/test_misc.py", line 12499, in f
return torch.isin(x, x)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153216
Approved by: https://github.com/williamwen42, https://github.com/peterbell10
PR #151968 adds `reorder_for_minimizing_partition` for the minimal number of partitions. If reordering two nodes cannot reduce the number of partitions, `reorder_for_minimizing_partition` should maintain the relative order of these two nodes and rely on other reorder passes for some nice features, such as shorter liveness duration or less peak memory. In an extreme case, when all nodes are on gpu and can be cudagraphed, `reorder_for_minimizing_partition` should not reorder any nodes.
This PR improves `reorder_for_minimizing_partition` for the invariant: relative order of nodes within the same graph partition are maintained. To do so, we record the index of each node in the input `nodes: list[BaseSchedulerNode]` and use a heap to pop the node with the smallest index. So we always scheduler a node with smaller index in the same graph partition and respects the invariant. Previous implementation tried to use a queue to achieve that but failed. Because node_N at the end may rely on node_1 at the start, such that node_N is added to queue once node_1 is scheduled.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153111
Approved by: https://github.com/eellison
Summary:
- Replace `C10_CUDA_KERNEL_LAUNCH_CHECK()` in the `KernelLauncher`, as the
latter does not print __FILE__ and __LINE__
The existing `C10_CUDA_KERNEL_LAUNCH_CHECK()` implementation does not print the source file and line number when a CUDA kernel launch throws an error, leaving users confused with a context-less message like `CUDA error: invalid arguments`. This new check is a slimmed re-implementation of the macro with extra context information added to the error (beyond just file and line number) so that we can at least locate the FBGEMM source file or template where the error first surfaces.
Test Plan:
```
buck2 run 'fbcode//mode/opt' fbcode//deeplearning/fbgemm/fbgemm_gpu/test/utils:kernel_launcher
buck2 run 'fbcode//mode/opt-amd-gpu' fbcode//deeplearning/fbgemm/fbgemm_gpu/test/utils:kernel_launcher
```
Reviewed By: sryap
Differential Revision: D74364031
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153178
Approved by: https://github.com/atalman, https://github.com/huydhn
Graph partition analyzes read_writes to get partition input names. However, weak dep is fake dependency and is not actually read or written. So we should not include weak dep in graph partition input names.
The following test failure is fixed by removing weak dependency from partition_input_names:
`PYTORCH_TEST_WITH_INDUCTOR=1 python test/test_torch.py TestTorchDeviceTypeCUDA.test_params_invalidated_with_grads_invalidated_between_unscale_and_step_Adam_cuda_float32`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152863
Approved by: https://github.com/eellison
Summary:
Bug fix for constant folding states. We are not setting the correct state for each updates.
One race condition would be:
(1) All threads obtain the model_exec_lock from main run.
(2) In second round of updated constant buffer, we should have set secondary as INITIALIZED but primary is mistakenly set instead.
(3) run_const_fold get called and an model_exec_lock is obtained, waiting for available at this time.
(4) main run enters INITIALIZED, waiting for unique_lock (which a shared_lock is being held by (3) at this moment)
Test Plan:
TBD
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153152
Approved by: https://github.com/jingsh, https://github.com/chenyang78
Summary: I'll need some of the benchmark-related functions surfaced so I can use them for remote autotuning. This PR just lifts the main in-process benchmarking helpers to classmethods. It wasn't strictly necessary to also move the sub-process benchmarking helper, but I think it improves readability. Also added some missing types.
Test Plan: Existing unit tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153084
Approved by: https://github.com/aorenste, https://github.com/eellison
Summary: HF loading when there is no metadata is an edge case for some users. We were previously calling safe_open(filename) to get the keys in the safetensors file, but this doesn't work with fsspec, when models have a different backend than local fs (ie. hf, s3 etc). This diff updates to open the file with fsspec.open() and then safetensors.deserialize() to get the keys
Test Plan: unit test and e2e test reading from hf
Differential Revision: D74181513
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152856
Approved by: https://github.com/joecummings
# Motivation
fixes https://github.com/pytorch/pytorch/issues/153022
The root cause is that the XPU backend registers the convolution op using `m.impl`, which bypasses the device guard logic typically added by the code generation system. This can lead to unexpected behavior if the current device isn't explicitly set.
# Additional Context
run the following script
```python
import torch
import torchvision.models as models
torch.manual_seed(0)
model = models.resnet50(weights="ResNet50_Weights.DEFAULT")
model.eval()
data = torch.rand(1, 3, 224, 224)
device = torch.device('xpu:1') # 'xpu:0'
model = model.to(device=device, dtype=torch.float16)
data = data.to(device, dtype=torch.float16)
with torch.no_grad():
ret = model(data)
print(ret)
print("Execution finished")
```
The output is
```bash
-9.2102e-02, -7.7588e-01, -1.4111e+00, -9.2383e-01, 6.4551e-01,
-6.0730e-03, -7.8271e-01, -1.1904e+00, -4.1602e-01, 3.2715e-02,
-4.9854e-01, -6.3623e-01, -8.5107e-01, -6.8555e-01, -9.4434e-01,
-8.8672e-01, -6.7969e-01, -6.9824e-01, -2.8882e-01, 2.0312e+00]],
device='xpu:1', dtype=torch.float16)
Execution finished
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153067
Approved by: https://github.com/albanD, https://github.com/EikanWang
Fixes: https://github.com/intel/torch-xpu-ops/issues/1503
`sycl/ext/oneapi/bfloat16.hpp` header file is a DPC++ compiler internal header. It's not documented for usage (see extension specification linked below) and is not guaranteed to exist. Instead, documented usage of extension suggests to rely on including `sycl/sycl.hpp` which in its turn includes `bfloat16.hpp` header (which is implementation detail).
We stepped into issues by explicitly including `bloat16.hpp` sycl header whithin user facing production environment when `intel-sycl-rt` wheel is installed (which is the dependency of `torch` wheel package built and publicly available for xpu). Compiler includes this file from `intel-sycl-rt` and due to `#pragma once` usage its content is included as well giving redefinitions of symbols in this file (previous inclusion is coming from `sycl/sycl.hpp`):
```
In file included from /workspace/lib/python3.12/site-packages/torch/include/c10/util/BFloat16.h:23:
/opt/intel/oneapi/compiler/2025.0/bin/compiler/../../include/sycl/ext/oneapi/bfloat16.hpp:60:23: error: redefinition of 'BF16VecToFloatVec'
60 | template <int N> void BF16VecToFloatVec(const bfloat16 src[N], float dst[N]) {
| ^
/workspace/include/sycl/ext/oneapi/bfloat16.hpp:60:23: note: previous definition is here
60 | template <int N> void BF16VecToFloatVec(const bfloat16 src[N], float dst[N]) {
|
```
While SYCL header files themselves can be improved (`#pragma once` dropped), we still must correct usage of sycl `bfloat16.hpp` header in pytorch, i.e. drop it. This fortunately helps to address the reported issue of redefinitions though follow up on compiler side is still required.
Also, `SYCL_EXT_ONEAPI_BFLOAT16_MATH_FUNCTIONS` used to cover inclusion of `sycl/sycl.hpp` does not make sense since it's defined in this very header. Thus, we should use `SYCL_LANGUAGE_VERSION` instead which is defined on compiler level.
See: f958dce280/sycl/doc/extensions/experimental/sycl_ext_oneapi_bfloat16_math_functions.asciidoc
CC: @EikanWang, @guangyey, @gujinghui
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152562
Approved by: https://github.com/guangyey, https://github.com/EikanWang, https://github.com/albanD
The edited comment should have the info. The code change looks large, but its copied from the install_cache script that our docker images use 6a8006472e/.ci/docker/common/install_cache.sh (L42)
Sccache stopped working on xla at some point near dec 17 2023. I am not sure what commit caused it. I think it was having trouble writing to the cache.
Either way, there is an sccache already installed on the docker image, so we should use that instead of a binary from s3 which we're probably no longer sure where it came from/what commit it was built from
The one in the docker image is installed here 69d438ee65/.github/upstream/Dockerfile (L61) and is also very old, so I have https://github.com/pytorch/xla/pull/9102 to update it
sccache still not writing properly, i will investigate, but xla build currently broken after the above xla pr, and this should fix it
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153002
Approved by: https://github.com/malfet
A typical `bmm` kernel in Helion needs to pass in symint shapes to `torch.baddbmm`. Currently `self.expand((dim1, dim2, dim3))` in baddbmm runs unconditionally and it doesn't work with symint shapes (it raises the following error):
```
Traceback (most recent call last):
File "/home/willfeng/local/helion_yf225/helion/_compiler/type_propagation.py", line 699, in propagate_call
CheckForIndexCalls.retry_call(self.value, proxy_args, proxy_kwargs),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/helion_yf225/helion/_compiler/tile_index_proxy.py", line 104, in retry_call
return fn(*proxy_args, **proxy_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/utils/_stats.py", line 27, in wrapper
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_subclasses/fake_tensor.py", line 1338, in __torch_dispatch__
return self.dispatch(func, types, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_subclasses/fake_tensor.py", line 1986, in dispatch
return self._cached_dispatch_impl(func, types, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_subclasses/fake_tensor.py", line 1450, in _cached_dispatch_impl
output = self._dispatch_impl(func, types, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_subclasses/fake_tensor.py", line 2645, in _dispatch_impl
r = func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_ops.py", line 806, in __call__
return self._op(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_prims_common/wrappers.py", line 309, in _fn
result = fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/home/willfeng/local/pytorch/torch/_meta_registrations.py", line 2172, in meta_baddbmm
self = self.expand((dim1, dim2, dim3))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: /home/willfeng/local/pytorch/build/aten/src/ATen/RegisterCompositeExplicitAutograd_0.cpp:5025: SymIntArrayRef expected to contain only concrete integers
```
This PR changes it so that we don't run `expand()` when not necessary, which makes the Helion use case (i.e. no broadcasting) work.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153112
Approved by: https://github.com/jansel
Follow up to @ezyang's PR #153020 , but better uses PEP621 to reduce redundant fields and pass through metadata better to uv, setuptools, poetry and other tooling.
* Enables modern tooling like uv sync and better support for tools like poetry.
* Also allows us to set project wide settings that are respected by linters and IDE (in this example we are able centralize the minimum supported python version).
* Currently most of the values are dynamically fetched from setuptools, eventually we can migrate all the statically defined values to pyproject.toml and they will be autopopulated in the setuptool arguments.
* This controls what additional metadata shows up on PyPi . Special URL Names are listed here for rendering on pypi: https://packaging.python.org/en/latest/specifications/well-known-project-urls/#well-known-labels
These also clearly shows us what fields will need to be migrated to pyproject.toml over time from setup.py per #152276. Static fields be fairly easy to migrate, the dynamically built ones like requirements are a bit more challenging.
Without this, `uv sync` complains:
```
error: No `project` table found in: `pytorch/pyproject.toml`
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153055
Approved by: https://github.com/ezyang
Summary:
As discussed with @malfet , we're porting nativert code to torch/nativert/.
Following up some concerns over the new directory, I'm trying to setup the tooling on OSS so various things (like linters) can run on torch/nativert/ properly.
Test Plan: CI
Differential Revision: D74407808
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153164
Approved by: https://github.com/dolpm, https://github.com/Skylion007
Summary: The autotuner is using zero-filled tensors to autotune
_scaled_grouped_mm and that's not appropriate for the offsets tensor, since it
essentially corresponds to "no input" and thus yields invalid perf results.
We can't really use the actual input tensors, since we might be compiling this
op in the context of an entire graph.
So instead, I decided to create a synthetic offsets tensor assuming that each
group is (roughly) the same size. I don't have data but I'd guess this
approach is OK for MoE since we're generally hoping to load-balance the
experts; I'm not sure how well it applies to other scenarios that might be more
heavy-tailed.
Test Plan:
```
pytest test_matmul_cuda.py -k test_scaled_grouped_gemm_
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152968
Approved by: https://github.com/ngimel
Flex Attention may have symints in subgraph inputs and outputs. Existing code implicitly captures these symints but does not explicitly store it in TritonTemplateBuffer. This leads to error when analyzing symints used in Flex Attention as a TritonTemplateBuffer. This PR fixes the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152878
Approved by: https://github.com/drisspg
Summary:
Added some logging and captured the indexing. See below image.
{F1977773416}
This is why the saved module path is called `/tmp/jimwan/minimizer_a_acc.pt`
Now the updated module paths are `/tmp/jimwan/minimizer_addmm_default_103_acc.pt`.
Test Plan:
```
MTIAC_USE_DIST_REF_KERNELS=all buck2 run @//mode/opt mtia/accuracy/minimizer:mtia_minimizer_runner -- --mode sequential --compare_fn allclose --pt_save_dir /tmp/debug3 --atol 1e-4 --rtol 1e-4 --all_outputs --start_idx native_layer_norm_default_80 --end_idx getitem_272 2>&1 | tee ~/test.log
```
{F1977773610}
Reviewed By: qcyuan
Differential Revision: D74369107
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153130
Approved by: https://github.com/Skylion007
As in title
idk how the install_cmake script is used because I see it being called with 3.18 but when I look at the build jobs some say 3.18 and others 3.31
Just make everything install cmake via the requirements-ci.txt. I don't know if the comment at 5d36485b4a/.ci/docker/common/install_conda.sh (L78) still holds, but pretty much every build has CONDA_CMAKE set to true, so I'm just defaulting to installing through pip
Also defaulting to 4.0.0 everywhere except the executorch docker build because executorch reinstalls 3.31.something
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152537
Approved by: https://github.com/cyyever, https://github.com/atalman, https://github.com/malfet
Summary:
We should generate the kernel for const graph and main graph separately.
The reason is that when we run autotuning, we would create separate
kernel calls and we should make sure that main graph also contains the
runner.
Test Plan:
python test/inductor/test_aot_inductor.py -k test_autotune_with_constant_folding
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D74347765](https://our.internmc.facebook.com/intern/diff/D74347765)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153040
Approved by: https://github.com/angelayi
Summary:
X-link: https://github.com/pytorch/gloo/pull/437
This provides a new "UnboundBuffer" implementation for Gloo ibverbs backend so it can be used with PyTorch.
This currently is passing basic tests such as `reduce_test` and `send_recv_test` but there are a number of failures. Putting this up for review so the follow up fixes are less of a mega PR and also so we can start doing some initial testing with this E2E with PyTorch.
Known issues:
* support recv from any is not supported
* AllreduceBcubeBase2 is failing
Test Plan:
```
buck2 run mode/dbgo //gloo/test:send_recv_test_ibverbs
buck2 test //gloo/test:
GLOO_DEVICE_TRANSPORT=IBVERBS buck2 run @//mode/opt //caffe2/test/distributed:c10d -- -r '.*gloo.*' -f
```
We can't run any of the gloo tests in CI since none of our CI machines have ibverbs so they're disabled by default and need to be manually run.
Differential Revision: D73291471
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153015
Approved by: https://github.com/fduwjj
ShapeEnv.evaluate_expr() behaves differently based on the (tls) global "suppress_guards" - so its cache key needs to include that value.
This came up because #152662 triggered it in the test `test/dynamo/test_exc.py::ExcTests::test_trigger_bisect_on_error` - fixing this caused that test to work again.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152661
Approved by: https://github.com/laithsakka
## Summary
This PR updates the docstring for `CosineAnnealingLR` to accurately reflect its recursive learning rate schedule. The previous docstring displayed only the SGDR closed-form expression, which doesn't match the actual recursive implementation in code.
Changes:
- Added the recursive update formula used in `get_lr()`
- Retained the original closed-form SGDR expression for reference
- Clarified that warm restarts are not implemented in this scheduler
This addresses confusion raised in issue #152081.
## Related issue
[#152081](https://github.com/pytorch/pytorch/issues/152081)
## Testing
Doc-only change. Ran pre-commit to verify formatting.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152936
Approved by: https://github.com/janeyx99
This diff hardens StaticCudaLauncher in the event a cubin file gets deleted under us. We store the raw cubin on the static cuda launcher, and reload it as needed. On cold start, this can happen if the cubin file is created by triton, and gets deleted before we can load the kernel on the parent process.
We don't want to store the entire cubin both in file format and in memory for caching purposes, so we delete it before caching the data. In the unfortunate/unlikely event where we can't load/find the necessary file on warm start, skip the stored triton launcher, falling back to regular triton.
This comes at a cost to worker memory, but it's not more memory than regular triton workers already take, so it should be okay.
Tests:
- Make test_static_cuda_launcher always delete the cubin path and reload it
Fixes#153030
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153064
Approved by: https://github.com/oulgen, https://github.com/jansel
The reason why we did this before is because that's how our older
autograd.Function x Dynamo interaction work, but we've since adopted
newer designs that don't actually need the autograd.Function.ctx proxied
into the graph.
We still need a fx.Proxy for the autograd.Function.ctx object, so
whenever we do I create one via discard_graph_changes.
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152621
Approved by: https://github.com/oulgen
As in title
idk how the install_cmake script is used because I see it being called with 3.18 but when I look at the build jobs some say 3.18 and others 3.31
Just make everything install cmake via the requirements-ci.txt. I don't know if the comment at 5d36485b4a/.ci/docker/common/install_conda.sh (L78) still holds, but pretty much every build has CONDA_CMAKE set to true, so I'm just defaulting to installing through pip
Also defaulting to 4.0.0 everywhere except the executorch docker build because executorch reinstalls 3.31.something
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152537
Approved by: https://github.com/cyyever, https://github.com/atalman, https://github.com/malfet
Fixes#122757
## Test Result
```python
import torch
model_output = torch.randn(10, 5).cuda()
labels = torch.randint(0, 5, (10,)).cuda()
weights = torch.randn(5)
loss_fn = torch.nn.CrossEntropyLoss(weight=weights)
loss = loss_fn(input=model_output, target=labels)
print(loss)
Traceback (most recent call last):
File "/home/zong/code/pytorch/../loss2.py", line 17, in <module>
loss = loss_fn(input=model_output, target=labels)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/loss.py", line 1297, in forward
return F.cross_entropy(
^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/functional.py", line 3494, in cross_entropy
return torch._C._nn.cross_entropy_loss(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Expected all tensors to be on the same device, but got weight is on cpu, different from other tensors on cuda:0 (when checking argument in method wrapper_CUDA_nll_loss_forward)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150750
Approved by: https://github.com/malfet
Summary: We enable the activation quantization in the forward pass, and users can customize the dtype they want to quantize.
Test Plan:
# unit test
```
buck2 test 'fbcode//mode/dev-nosan' fbcode//caffe2/test/inductor:quantization -- test_activation_quantization_aten
```
Buck UI: https://www.internalfb.com/buck2/776d3911-bb86-4ac8-a527-540cf1510b9d
Test UI: https://www.internalfb.com/intern/testinfra/testrun/4785074873051017
Network: Up: 4.3MiB Down: 42MiB (reSessionID-fef7e727-68b1-4645-a519-5652854df38d)
Executing actions. Remaining 0/4 6.7s exec time total
Command: test. Finished 2 local
Time elapsed: 3:11.5s
Tests finished: Pass 2. Fail 0. Fatal 0. Skip 0. Build failure 0
# E2E
### how to enable (you can overrite the dtype, if nothing given, the default is fp8)
```
post_grad_fusion_options={
"activation_quantization_aten_pass": {"quant_type": "torch.float8_e5m2"}
},
```
Differential Revision: D70522237
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148380
Approved by: https://github.com/Mingming-Ding, https://github.com/Hahu803
Summary:
Solves https://github.com/pytorch/pytorch/issues/151925
Currently, AOTI only generate runtime asserts for unbacked symints. We should generate asserts for all `_assert_scalar` calls in the input graph.
Also factored out the run time assertion logic to a separate function.
We need to generate runtime asserts directly in Inductor instead
of just re-using the asserts from input graphs becase we reuse the
same ShapeEnv as before. In particular, on subsequent graph passes,
we would immediately turn all of these assertions into noops,
because when we evaluated their expressions, we would see that
because we had a deferred runtime assert in the ShapeEnv, we
know "oh, of course this expression is True" already.
One example is below:
```
class Model(torch.nn.Module):
def forward(self, a, b, c):
nz = torch.nonzero(a)
ones = a.new_ones([nz.size(0), b.size(0)])
torch._check(ones.size(0) >= 1)
equals = torch.add(ones, c)
return equals
torch._dynamo.mark_dynamic(c, 0)
```
When we re-use the ShapeEnv in Inductor lowering, the check that checks
a and nonzero have the same shape would be evaluted to True after we resolve
unbacked bindings using the ShapeEnv.
See test_unbacked_equals_input_size_runtime_assertion in test_aot_inductor.
In addition to the Inductor generated runtime asserts, we also
need the runtime asserts from the input graph, because some derived
runtime asserts are not generated in Inductor. One example is
below:
```
class Model(torch.nn.Module):
def forward(self, x):
y = x.reshape(100, -1).clone()
y = y + 1
return y
dynamic_shapes = {
"x": {0: torch.export.Dim.DYNAMIC},
}
x.shape[0] needs to be a multiple of 100.
```
See test_aoti_runtime_asserts_backed_symint in test_aot_inductor.
Example:
```
def forward(self):
arg0_1: "f32[s35]";
arg0_1, = fx_pytree.tree_flatten_spec([], self._in_spec)
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:11 in forward, code: y = x.reshape(100, -1).clone()
sym_size_int: "Sym(s35)" = torch.ops.aten.sym_size.int(arg0_1, 0)
#
mod: "Sym(Mod(s35, 100))" = sym_size_int % 100; sym_size_int = None
eq_2: "Sym(Eq(Mod(s35, 100), 0))" = mod == 0; mod = None
_assert_scalar = torch.ops.aten._assert_scalar.default(eq_2, "Runtime assertion failed for expression Eq(Mod(s35, 100), 0) on node 'eq'"); eq_2 = _assert_scalar = None
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:11 in forward, code: y = x.reshape(100, -1).clone()
view: "f32[100, (s35//100)]" = torch.ops.aten.reshape.default(arg0_1, [100, -1]); arg0_1 = None
clone: "f32[100, (s35//100)]" = torch.ops.aten.clone.default(view); view = None
# File: /data/users/shangdiy/fbsource/buck-out/v2/gen/fbcode/73a672eb896e7996/scripts/shangdiy/__pt__/pt#link-tree/scripts/shangdiy/pt.py:12 in forward, code: y = y + 1
add_6: "f32[100, 1]" = torch.ops.aten.add.Tensor(clone, 1); clone = None
return (add_6,)
```
Generated cpp code:
```
auto inputs = steal_from_raw_handles_to_raii_handles(input_handles, 1);
auto arg0_1 = std::move(inputs[0]);
auto arg0_1_size = arg0_1.sizes();
int64_t s35 = arg0_1_size[0];
inputs.clear();
auto& kernels = static_cast<AOTInductorModelKernels&>(*this->kernels_.get());
if (!((s35 % 100L) == 0L)) { throw std::runtime_error("Expected Eq(Mod(s35, 100), 0) to be True but received " + std::to_string(s35)); }
```
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r aoti_runtime_asserts_backed_symint
```
Differential Revision: D73596786
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152125
Approved by: https://github.com/henrylhtsang, https://github.com/jingsh
### Changes
- Detect NVSHMEM install location via `sysconfig.get_path("purelib")`, which typically resolves to `<conda_env>/lib/python/site-packages`, and NVSHMEM include and lib live under `nvidia/nvshmem`
- Added link dir via `target_link_directories`
- Removed direct dependency on mlx5
- Added preload rule (following other other NVIDIA libs)
### Plan of Record
1. End user experience: link against NVSHMEM dynamically (NVSHMEM lib size is 100M, similar to NCCL, thus we'd like users to `pip install nvshmem` than torch carrying the bits)
2. Developer experience: at compile time, prefers wheel dependency than using Git submodule
General rule: submodule for small lib that torch can statically link with
If user pip install a lib, our CI build process should do the same, rather than building from Git submodule (just for its header, for example)
3. Keep `USE_NVSHMEM` to gate non-Linux platforms, like Windows, Mac
4. At configuration time, we should be able to detect whether nvshmem is available, if not, we don't build `NVSHMEMSymmetricMemory` at all.
For now, we have symbol dependency on two particular libs from NVSHMEM:
- libnvshmem_host.so: contains host side APIs;
- libnvshmem_device.a: contains device-side global variables AND device function impls.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153010
Approved by: https://github.com/ngimel, https://github.com/fduwjj, https://github.com/Skylion007
Fixes an error message in test/test_optim.py
Current behavior: If running the test with Adagrad, the error message reads: "SGD does not currently support capturable".
Fix: The error message now says correctly: "Adagrad does not currently support capturable".
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153076
Approved by: https://github.com/janeyx99
This fixes an issue in the TORCH_CHECK error message in the FusedSgdKernel.
Current behavior: If the LR tensor is not on the same device as the parameters, the error message reads: "found_inf must be on the same GPU device as the params".
Fix: The error message now correctly points out "lr must be on the same GPU device as the params".
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153074
Approved by: https://github.com/Skylion007, https://github.com/janeyx99
Although Dim.AUTO covers the cases that a user sets more axes to be dynamic than the model actually needs, it silently falls back to STATIC when DYNAMIC fails. This increases the difficulty of debugging.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153065
Approved by: https://github.com/justinchuby
Differential Revision: D74218923
Running on A100 seems to result in precision loss from decompose_k. This was root caused to the fp16/bf16 reduction setting, which establishes a less precise baseline than decompose_k, as decompose_k uses the bmm.dtype overload for fp32 output.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152897
Approved by: https://github.com/eellison
Current FR code is built with `USE_C10D_NCCL` we should remove it to make it generic. And we keep existing API used by NCCL so that we can have some bc compatibility because lots of use cases are around FR with NCCL. The generic version with C10::Event can then be used for other backend like Gloo, etc.
The current Unit test should cover the change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152563
Approved by: https://github.com/kwen2501, https://github.com/d4l3k
ghstack dependencies: #152585
Tests serialization for RANGE_ITERATOR_MATCH; includes no non-test changes.
This PR handles iterator exhaustion issues by utilizing the janky solution from #152865; it passes a function to generate kwargs and `frame_state.f_locals` is updated with fresh iterators through a second kwarg generation pass after initial tracing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152872
Approved by: https://github.com/jansel
ghstack dependencies: #152725, #152727, #152728, #152730, #152865
Tests serialization for TUPLE_ITERATOR_LEN; includes no non-test changes.
Passing a tuple iterator as input results in the iterator being exhausted during testing. I threw together a super janky workaround via accepting a func for kwarg generation and replacing `frame_state.f_locals` with newly-generated kwargs to get fresh iterators, but insights into a better approach are welcome!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152865
Approved by: https://github.com/jansel
ghstack dependencies: #152725, #152727, #152728, #152730
Fix
```
CMake Warning (dev) in cmake/Codegen.cmake:
A logical block opening on the line
/var/lib/jenkins/workspace/cmake/Codegen.cmake:393 (if)
closes on the line
/var/lib/jenkins/workspace/cmake/Codegen.cmake:401 (endif)
with mis-matching arguments.
```
by removing the condition in `endif`.
We could instead fix it, however, that is not best practice. For example, cmake_lint warns that, and CMake says
```
The optional <condition> argument is supported for backward compatibility only.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153023
Approved by: https://github.com/aditew01, https://github.com/Skylion007
Prep change for getting rid of `_mac-test-mps.yml`
A complete no-op for now, but will be used by PR above the stack, but they should be landed few days apart to avoid forcing lots of people to rebase their PRs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153012
Approved by: https://github.com/wdvr
Issue was that
- symbol-ids appeared out-of-order w.r.t to the order of the forward inputs
```
def forward(arg0 # [(s3 - 1) + s4, 32], arg1 #[(s3 - 1)] ..)
```
- this causes codegen to fail because it expects all the base symbols `s4,s3` to have been codegen-ed already.
- well, we can skip codegen-ing sympy expr with many symbols e.g. `(s3 - 1) + s4` because `s3` and `s4` will be codegen-ed by other inputs.
```
# for example
s3 = arg1.size(0) + 1
s4 = argN.size(0)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152579
Approved by: https://github.com/jingsh, https://github.com/desertfire
Noticed some of these ops were contributing to a big chunk of the runtime for OpenLLama as well as a few other benchmarks
At the op level, moving to a TensorIterator-based Metal kernel gives a 20x speedup. Will migrate the inverse trigonometric functions & log ops in a follow-up PR, as this one is already a bit large
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152876
Approved by: https://github.com/malfet
When we do `torch.compile(mod)`, we eventually end up returning a new
module instance, whose `forward` method is the result of
`torch.compile(mod.__call__)`, meaning it already captures all the extra
logic (e.g., hook firing) from the default `torch.nn.Module.__call__`.
As a result we can't reuse the inherited default `__call__` as is,
because we'd end up running the logic twice.
This patch makes the returned `OptimizedModule` override the default
`__call__`, and directly calls into its compiled `forward` method.
Fixes#149502
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152740
Approved by: https://github.com/anijain2305
## Summary
- Added --no-install-recommends flag to all apt-get install commands to reduce unnecessary dependencies
- Added apt-get clean after package installations to remove package cache and reduce image size
- Combined multiple apt commands into single instructions to reduce Docker image layers
## Test plan
Test by building the devcontainer and verifying functionality while ensuring reduced image size
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152882
Approved by: https://github.com/cyyever, https://github.com/atalman, https://github.com/Skylion007
## Summary
- Replaced miniconda base image with base Ubuntu 22.04 image
- Installed Python and required dependencies using apt
- Replaced conda-based CUDA installation with apt-based version
- Updated paths in install-dev-tools.sh to reflect the new non-conda environment
- Removed conda-specific files and added requirements.txt for Python dependencies
## Test plan
Test by building and running the devcontainer in VS Code with both CPU and CUDA configurations
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152881
Approved by: https://github.com/atalman
## Summary
- Changed the devcontainer context path from '../..' to './' for both CPU and CUDA configurations
- Added workspace mount configuration to properly mount the repository in the container
- Added containerEnv to disable implicit --user pip flag
## Test plan
Test by building and running the devcontainer in VS Code
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152880
Approved by: https://github.com/atalman
Teach the graph printer how to allow overriding printing SymTypes (`SymInt`, `SymFloat`, `SymBool`) and then use that to reuse the fast SymNode printing from `torch._inductor.utils.sympy_str()` to make computing the cache key faster.
On my computer the repro from #151823 goes from 480s -> 80s (still terrible... but better).
Fixes#151823
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151928
Approved by: https://github.com/laithsakka
A lot of last minute bugfixes for CUTLASS blackwell that we should upstream. It's a header only library and a minor release so this should strictly improve compiler support and fix some bugs. Needed to update some instruction numbers in torch compile baselines for the new kernels
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152779
Approved by: https://github.com/henrylhtsang
This PR refactors CompiledFxGraph by adding a new post_compile step that only runs on cache hit. This refactors a bunch of code in _lookup_graph to its own function so that we can use it in BundledAOTAutogradCacheEntry. No difference in behavior here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152839
Approved by: https://github.com/oulgen
ghstack dependencies: #152836
The purpose of this stack is to create a new BundledAOTAutogradCacheEntry, which is an AOTAutogradCacheEntry that is self contained, i.e. it contains all of the CompiledFxGraph directly in the entry, instead of relying on FxGraphCache._lookup_graph.
Because this woudl balloon the size of the actual cache entry to do this, our goal is not to use BundledAOTAutogradCacheEntry in cache scenarios: only for precompile use cases. Thus, it's important we make this whole setup generic, to be able to support these two workflows clearly.
This PR genericizes AOTAutogradCacheEntry considerably, so that it can take in different types of Forwards and Backwards.
Each GenericAOTAutogradCacheEntry is composed of two parts, a TForward and a TBackward. The forward and backward can be loaded in multiple ways, either via FxGraphCache._lookup_graph, or by saving the entire CompiledFxGraph.
For simplicify, this PR only implements the generic code refactors needed, but does not fully implement BundledAOTAutogradCacheEntry, which is an AOTAutogradCacheEntry that takes a full CompiledForward. We'll handle and implement BundledAOTAutogradCacheEntry in the PR above this, for easier review.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152836
Approved by: https://github.com/oulgen
This PR updates `GuardsStatePickler.reducer_override()` in `torch/_dynamo/guards.py` to handle reconstruction of traceable wrapper subclasses. It's intended to work recursively and handle any level of subclass instance nesting (e.g. subclass instances that contain subclass instances, etc.)
This PR tests the guard on several traceable wrapper tensor subclasses:
* `LocalSubclass`: used to ensure the correct error message is thrown when the subclass is not defined globally
* `torch.testing._internal.two_tensor.TwoTensor`: defines None for its extra metadata
* `SubclassWithMeta`: stores non-trivial extra metadata
* `SubclassWithCustomMetadataGuard`: stores non-trivial extra metadata and defines a custom `__metadata_guard__` classmethod
* `SubclassWithSubclassInnerTensors`: used to test recursiveness; this subclass contains subclass inner tensor components
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152626
Approved by: https://github.com/jansel
# Sub-PRs
These PRs contain refactors from the main one. They should be reviewed and merged first.
- https://github.com/pytorch/pytorch/pull/150458
- https://github.com/pytorch/pytorch/pull/152391
- https://github.com/pytorch/pytorch/pull/152587
# Feature
The goals of this PR are twofold.
## Goal 1: Introduce Wrapper IR as an intermediate step in wrapper codegen.
In addition to Triton/C++/Halide kernels, Inductor also generates "wrapper" code which allocates memory and calls the kernels. Originally, this wrapper code was fairly standard Python which resembled a user-written PyTorch program. Over time, various wrapper code generators have been added to accommodate things like AOTInductor, which prefers C++ code for static compilation. This complexity has bled into other parts of the codebase, as we now need if/else statements to choose between Python and C++ macros. (See an example [here](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/ir.py#L5515-L5522).) Since most of these code generation steps are conceptually identical across target languages, it seems reasonable to refactor them into some kind of intermediate representation which can be shared between the various backends. This might also make it easier to develop out-of-tree backends which cannot put their own macros in core Inductor components.
This PR takes some initial steps to formalize Inductor's wrapper codegen by generalizing the existing Memory Planning IR into a fully fledged Wrapper IR. This is pretty much identical to the existing Memory Planning IR, but it supports a richer set of ops for things like kernel definitions and calls. This refactor could help encapsulate wrapper codegen. Ideally, we don't need to worry about direct Python/C++ codegen in the main compiler files such as `ir.py`, and can instead defer these to classes like `PythonWrapperCodegen` and `CppWrapperCpu`, which operate on the Wrapper IR.
## Goal 2: Convert Wrapper IR into FX IR.
One of the main benefits of Wrapper IR is to enable more diverse Inductor backends. This PR introduces a converter from Wrapper IR into [FX IR](https://pytorch.org/docs/stable/fx.html), which is the intermediate representation most commonly used in PyTorch graph compilers. The purpose of this is to enable out-of-tree backends to consume Inductor's output in FX IR, which would hopefully make Inductor easier to leverage in novel compilers, hardware accelerators, etc.
It's not trivial to generate Python or C++ code which Inductor can compile and run, and doing so may require changes to other core Inductor files, for the reasons outlined in the previous section. The goal of supporting FX output is to enable something like `torch.compile`'s [custom backend](https://pytorch.org/docs/stable/torch.compiler_custom_backends.html) system, in which an out-of-tree backend can receive an optimized FX graph from Inductor, and compile and run it however it likes.
The typical users of this feature would likely not be part of PyTorch, and may or may not support running a kernel in eager mode. However, they can understand what `torch.empty_strided` means, compile and run Triton kernels, etc. So we just need to present them with an FX graph saying what code Inductor wants to run, which should be easier to analyze and transform in a third party system than Python or C++ source.
Since FX IR is fairly stable, this mechanism should hopefully isolate third-party backends, hardware accelerators, etc. from the implementation details of Inductor, and vice versa.
# Current status
Things that seem to work:
- Converted a lot of the most common Python codegen lines to Wrapper IR lines.
- Handled the following cases, in addition to what was already in the Memory Planning IR:
- Comments
- Triton kernels
- Extern/fallback kernels
- Freeing tensors (`del buf0`)
- MultiOutput
- Graph outputs
- ReinterpretView / StorageBox, for both call args and outputs.
- FX conversion asserts that the program only contains Wrapper IR lines, and not strings of Python/C++ code.
- Prototype FX converter which can handle some of the most common use cases.
- Defining Triton kernels, and putting them in a side table using TorchDynamo's existing [utilities](https://dev-discuss.pytorch.org/t/higher-order-operators-2023-10/1565).
- Calling wrapped Triton kernels.
- Calling extern kernels and certain types of fallback kernels.
- Support both `extern_kernels.*` and `aten.*`.
- Support multi-output kernels like `torch.topk`.
- Graphs with multiple inputs/outputs.
- Training i.e. calling `Tensor.backward()` in a compiled function.
- Graph breaks (training).
- Run the `torch.fx.GraphModule` on GPU using the standard `__call__` method. This makes it easy to test the correctness of FX codegen.
Things that don't work:
- Both Wrapper IR and Wrapper -> FX coverage are currently best effort. There are still features which aren't captured as Wrapper IR lines, and fall back to plain strings. This representation is functionally correct but probably not rich enough to achieve the goals outlined in the previous sections.
- Fallback kernels seem like the most difficult thing to fully cover, since they each define their own Python/C++ macros that would need to be converted to FX.
- Size/alignment asserts are currently disabled via the config file. It's possible to generate FX IR for these, but it seems reasonable to defer these sanity checks to a later PR.
- CommBuffer's and distributed communication are not yet supported. An earlier version of this PR attempted to implement this by calling `empty_strided_p2p`. However, building and testing distributed support seems non-trivial, so it's probably better to defer this.
# Out-of-tree compilers
With this PR, out of tree backends will be able to do further compilation on the FX graphs by subclassing `WrapperFxCodegen` and overriding the `compile_graph` function. This follows the same API as torch.compile's [custom backends](https://pytorch.org/docs/stable/torch.compiler_custom_backends.html), where the user simply returns a callable running the graph. The callable need not be a method of `GraphModule` or any other PyTorch class. See an example below.
```
from torch._inductor.codegen.wrapper_fxir import WrapperFxCodegen
class MyCustomBackend(WrapperFxCodegen):
def compile_graph(self, gm):
# Add 1 to the graph's outputs
def compiled_fn(*args):
return [x + 1 for x in gm.graph.forward(*args)]
return compiled_fn
```
# Example FX graphs
This section contains some example FX graphs generated by Inductor. The correctness of these graphs was verified against eager mode by calling the corresponding `GraphModule`.
Here's an FX graph calling a basic Triton kernel. Notice how outputs are allocated with `torch.empty_strided`, and the Triton kernel is called by reference to Dynamo's triton side table.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((8,), (1,)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(8,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg1_1, in_ptr1: %arg0_1, out_ptr0: %buf0, xnumel: 8, XBLOCK: 8}})
return (buf0,)
```
Here's a more complicated graph that calls a `torch.addmm` extern kernel.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=2] = placeholder[target=arg1_1]
%buf0 : [num_users=3] = call_function[target=torch.empty_strided](args = ((), ()), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(1,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg1_1, out_ptr0: %buf0, xnumel: 1, r0_numel: 129, XBLOCK: 1}})
%buf2 : [num_users=2] = call_function[target=torch.empty_strided](args = ((129, 1), (1, 1)), kwargs = {dtype: torch.float32, device: cuda:0})
%addmm : [num_users=0] = call_function[target=torch.addmm](args = (%buf0, %arg0_1, %arg1_1), kwargs = {alpha: 1, beta: 1, out: %buf2})
%delete : [num_users=0] = call_function[target=torch._inductor.codegen.wrapper_fxir.delete](args = (%buf0,), kwargs = {})
return (buf2,)
```
Here's a graph which indexes into a tuple using `operator.getitem`. This is necessary to use the output of the `torch.topk` operation.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%buf0 : [num_users=3] = call_function[target=torch.ops.aten.topk.default](args = (%arg0_1, 2), kwargs = {})
%buf1 : [num_users=2] = call_function[target=operator.getitem](args = (%buf0, 0), kwargs = {})
%buf2 : [num_users=2] = call_function[target=operator.getitem](args = (%buf0, 1), kwargs = {})
%delete : [num_users=0] = call_function[target=torch._inductor.codegen.wrapper_fxir.delete](args = (%buf0,), kwargs = {})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(2,)], tma_descriptor_metadata: {}, kwargs: {in_out_ptr0: %buf1, xnumel: 2, XBLOCK: 2}})
%triton_kernel_wrapper_mutation_1 : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 1, constant_args_idx: 1, grid: [(2,)], tma_descriptor_metadata: {}, kwargs: {in_out_ptr0: %buf2, xnumel: 2, XBLOCK: 2}})
return (buf1, buf2)
```
Here's a graph that reinterprets an output tensor using `torch.as_strided`. This is one way to handle Inductor's `ReinterpretView` op.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((2, 4), (4, 1)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(8,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg0_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: 8, XBLOCK: 8}})
%buf0_view_buf0_0 : [num_users=1] = call_function[target=torch.as_strided](args = (%buf0, (8,), (1,), 0), kwargs = {})
return (buf0_view_buf0_0,)
```
Here's a graph with dynamic shapes. This one is a little bit funky. Inductor provides a graph input for each shape symbol, which we map to a placeholder, in this example `s6`. Then, shape expressions in the generated code can refer to the symbol `s6`. The size hint for `s6` is stored in `node.meta["val"]` where `node` is the placeholder defining it. This works out in the generated python code because the placeholder defines a Python variable with the name `s6`.
```
graph():
%s6 : [num_users=0] = placeholder[target=s6]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%arg2_1 : [num_users=1] = placeholder[target=arg2_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((s6,), (1,)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [[-(((-s6)//8)), 1, 1]], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg2_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: s6, XBLOCK: 8}})
return buf0
```
Here's another graph, this time with dynamic shapes and strides. The grid expression is more complex since the numel is a product of dimensions.
```
graph():
%s10 : [num_users=0] = placeholder[target=s10]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%arg2_1 : [num_users=1] = placeholder[target=arg2_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ([s10, s10], [s10, 1]), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [[-(((s10**2)//(-64))), 1, 1]], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg2_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: s10**2, XBLOCK: 64}})
return buf0
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146942
Approved by: https://github.com/jansel
# Motivation
Fix https://github.com/pytorch/pytorch/issues/152301
When XPU is not available, calling `torch.xpu.is_bf16_supported()` still returns `True`, which is inconsistent with the expected behavior (should be False).
# Solution
Align to other backend, adding `including_emulation` to `torch.xpu.is_bf16_supported` and,
- return `False` if XPU is not available
- return `True` if `including_emulation` is True
- return `torch.xpu.get_device_properties().has_bfloat16_conversions` if `including_emulation` is False, it means if the device could generate SPIRV code for bf16.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152317
Approved by: https://github.com/EikanWang
so two things other than cleanups and refactoring
1) do not use propagate_real_tensors to resolve eval under guard_or_true/guard_or_false .
2) do not guard for dimensions of type DimDynamic.OBLIVIOUS_SIZE under guard_or_true/guard_or_false .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152657
Approved by: https://github.com/pianpwk
Fixes#122757
## Test Result
```python
import torch
model_output = torch.randn(10, 5).cuda()
labels = torch.randint(0, 5, (10,)).cuda()
weights = torch.randn(5)
loss_fn = torch.nn.CrossEntropyLoss(weight=weights)
loss = loss_fn(input=model_output, target=labels)
print(loss)
Traceback (most recent call last):
File "/home/zong/code/pytorch/../loss2.py", line 17, in <module>
loss = loss_fn(input=model_output, target=labels)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/module.py", line 1762, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/modules/loss.py", line 1297, in forward
return F.cross_entropy(
^^^^^^^^^^^^^^^^
File "/home/zong/code/pytorch/torch/nn/functional.py", line 3494, in cross_entropy
return torch._C._nn.cross_entropy_loss(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: Expected all tensors to be on the same device, but got weight is on cpu, different from other tensors on cuda:0 (when checking argument in method wrapper_CUDA_nll_loss_forward)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150750
Approved by: https://github.com/mikaylagawarecki
Summary:
Torch Native Runtime RFC: https://github.com/pytorch/rfcs/pull/72
This diff moves `TensorMeta.cpp` and `TensorMeta.h` to PyTorch core under `torch/nativert/graph/`
Existing `torch::_export::TensorMeta` in `torch/csrc/utils/generated_serialization_types.h` is auto-generated from the export serde schema and therefore only containing the most basic serializable types. We need the newly added `TensorMeta.cpp` to deserialize the metadata into a in-memory class with c10 types so that it can be consumed by the runtime later.
Test Plan:
Added test under `test/cpp/nativert/test_tensor_meta.cpp`
Differential Revision: D73820548
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152475
Approved by: https://github.com/albanD
**Context**:
bucketize is relatively expensive, computationally. So it's not always profitable to fuse it if it means doing extra computation. For example, this repro:
https://gist.github.com/davidberard98/7fd6af7e6291787c246c705945a25554
shows a slowdown from 56us (eager) to ~100us (torch.compile-d): instead of computing 2\*\*15 binary searches, the fused version does 2\*\*15 * 384 - one for each of the broadcasted outputs.
**Solution**:
Realize the output of bucketize (and searchsorted, which also uses inductor's ops.bucketize). If there's an opportunity to do non-broadcasted fusions, the scheduler can still apply such fusions later on.
After this PR, instead of a slowdown, we see an improvement from 56us (eager) to 33us (compiled).
**Retry**
Original PR (https://github.com/pytorch/pytorch/pull/152644) was reverted due to internal bisect blaming this change, but the bisect was a false positive (and is marked as such)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152858
Approved by: https://github.com/aakhundov
Summary: The default action for ls for the local filesystem is with details=False, but this isn't the case for all filesystems (eg. huggingface), so setting details=False explicitly so that the return type of ls is a list of strings, and not a list of dictionaries, which is what it would be with details=True.
Test Plan: tested in notebook
Differential Revision: D74080572
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152693
Approved by: https://github.com/joecummings
After the CI change from 12.4 -> 12.6 around mid-March, the foreach tests have been flaky and hard to repro due to nondeterminism. Per @davidberard98's suggestion, let's try to add a synchronize before checking profiler results to see whether this fixes the flake! The hope is that the 48 currently open foreach flaky issues will close from this change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152857
Approved by: https://github.com/davidberard98
Summary:
Allow TMA workspace creation allow specification for `num_programs`, which defaults to `num_sms` when not specified.
We need a total `num_programs * num_tma_descriptors` no. of descriptors for a kernel.
Test Plan: CI.
Differential Revision: D74189599
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152844
Approved by: https://github.com/drisspg
Summary:
1. Adding `input` field to `_adapt_flat_args` function
2. In `process_forward_inputs`, `reorder_kwargs` will now do nothing if no kwargs are provided (previously would error)
3. Pass `args` as input to `_adapt_flat_args`
These changes are made to update the InputAdapter
see more context in D73811508
Test Plan: see D73811508
Differential Revision: D73945419
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152575
Approved by: https://github.com/angelayi
Summary: Replace the key (path) from `<hash>.best_config` to `<parent_dir>/<hash>.best_config` to ensure that Autotune artifacts in MegaCache are loaded to the correct location locally.
Test Plan: NA
Differential Revision: D74052400
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152678
Approved by: https://github.com/oulgen
# Sub-PRs
These PRs contain refactors from the main one. They should be reviewed and merged first.
- https://github.com/pytorch/pytorch/pull/150458
- https://github.com/pytorch/pytorch/pull/152391
- https://github.com/pytorch/pytorch/pull/152587
# Feature
The goals of this PR are twofold.
## Goal 1: Introduce Wrapper IR as an intermediate step in wrapper codegen.
In addition to Triton/C++/Halide kernels, Inductor also generates "wrapper" code which allocates memory and calls the kernels. Originally, this wrapper code was fairly standard Python which resembled a user-written PyTorch program. Over time, various wrapper code generators have been added to accommodate things like AOTInductor, which prefers C++ code for static compilation. This complexity has bled into other parts of the codebase, as we now need if/else statements to choose between Python and C++ macros. (See an example [here](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/ir.py#L5515-L5522).) Since most of these code generation steps are conceptually identical across target languages, it seems reasonable to refactor them into some kind of intermediate representation which can be shared between the various backends. This might also make it easier to develop out-of-tree backends which cannot put their own macros in core Inductor components.
This PR takes some initial steps to formalize Inductor's wrapper codegen by generalizing the existing Memory Planning IR into a fully fledged Wrapper IR. This is pretty much identical to the existing Memory Planning IR, but it supports a richer set of ops for things like kernel definitions and calls. This refactor could help encapsulate wrapper codegen. Ideally, we don't need to worry about direct Python/C++ codegen in the main compiler files such as `ir.py`, and can instead defer these to classes like `PythonWrapperCodegen` and `CppWrapperCpu`, which operate on the Wrapper IR.
## Goal 2: Convert Wrapper IR into FX IR.
One of the main benefits of Wrapper IR is to enable more diverse Inductor backends. This PR introduces a converter from Wrapper IR into [FX IR](https://pytorch.org/docs/stable/fx.html), which is the intermediate representation most commonly used in PyTorch graph compilers. The purpose of this is to enable out-of-tree backends to consume Inductor's output in FX IR, which would hopefully make Inductor easier to leverage in novel compilers, hardware accelerators, etc.
It's not trivial to generate Python or C++ code which Inductor can compile and run, and doing so may require changes to other core Inductor files, for the reasons outlined in the previous section. The goal of supporting FX output is to enable something like `torch.compile`'s [custom backend](https://pytorch.org/docs/stable/torch.compiler_custom_backends.html) system, in which an out-of-tree backend can receive an optimized FX graph from Inductor, and compile and run it however it likes.
The typical users of this feature would likely not be part of PyTorch, and may or may not support running a kernel in eager mode. However, they can understand what `torch.empty_strided` means, compile and run Triton kernels, etc. So we just need to present them with an FX graph saying what code Inductor wants to run, which should be easier to analyze and transform in a third party system than Python or C++ source.
Since FX IR is fairly stable, this mechanism should hopefully isolate third-party backends, hardware accelerators, etc. from the implementation details of Inductor, and vice versa.
# Current status
Things that seem to work:
- Converted a lot of the most common Python codegen lines to Wrapper IR lines.
- Handled the following cases, in addition to what was already in the Memory Planning IR:
- Comments
- Triton kernels
- Extern/fallback kernels
- Freeing tensors (`del buf0`)
- MultiOutput
- Graph outputs
- ReinterpretView / StorageBox, for both call args and outputs.
- FX conversion asserts that the program only contains Wrapper IR lines, and not strings of Python/C++ code.
- Prototype FX converter which can handle some of the most common use cases.
- Defining Triton kernels, and putting them in a side table using TorchDynamo's existing [utilities](https://dev-discuss.pytorch.org/t/higher-order-operators-2023-10/1565).
- Calling wrapped Triton kernels.
- Calling extern kernels and certain types of fallback kernels.
- Support both `extern_kernels.*` and `aten.*`.
- Support multi-output kernels like `torch.topk`.
- Graphs with multiple inputs/outputs.
- Training i.e. calling `Tensor.backward()` in a compiled function.
- Graph breaks (training).
- Run the `torch.fx.GraphModule` on GPU using the standard `__call__` method. This makes it easy to test the correctness of FX codegen.
Things that don't work:
- Both Wrapper IR and Wrapper -> FX coverage are currently best effort. There are still features which aren't captured as Wrapper IR lines, and fall back to plain strings. This representation is functionally correct but probably not rich enough to achieve the goals outlined in the previous sections.
- Fallback kernels seem like the most difficult thing to fully cover, since they each define their own Python/C++ macros that would need to be converted to FX.
- Size/alignment asserts are currently disabled via the config file. It's possible to generate FX IR for these, but it seems reasonable to defer these sanity checks to a later PR.
- CommBuffer's and distributed communication are not yet supported. An earlier version of this PR attempted to implement this by calling `empty_strided_p2p`. However, building and testing distributed support seems non-trivial, so it's probably better to defer this.
# Out-of-tree compilers
With this PR, out of tree backends will be able to do further compilation on the FX graphs by subclassing `WrapperFxCodegen` and overriding the `compile_graph` function. This follows the same API as torch.compile's [custom backends](https://pytorch.org/docs/stable/torch.compiler_custom_backends.html), where the user simply returns a callable running the graph. The callable need not be a method of `GraphModule` or any other PyTorch class. See an example below.
```
from torch._inductor.codegen.wrapper_fxir import WrapperFxCodegen
class MyCustomBackend(WrapperFxCodegen):
def compile_graph(self, gm):
# Add 1 to the graph's outputs
def compiled_fn(*args):
return [x + 1 for x in gm.graph.forward(*args)]
return compiled_fn
```
# Example FX graphs
This section contains some example FX graphs generated by Inductor. The correctness of these graphs was verified against eager mode by calling the corresponding `GraphModule`.
Here's an FX graph calling a basic Triton kernel. Notice how outputs are allocated with `torch.empty_strided`, and the Triton kernel is called by reference to Dynamo's triton side table.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((8,), (1,)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(8,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg1_1, in_ptr1: %arg0_1, out_ptr0: %buf0, xnumel: 8, XBLOCK: 8}})
return (buf0,)
```
Here's a more complicated graph that calls a `torch.addmm` extern kernel.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=2] = placeholder[target=arg1_1]
%buf0 : [num_users=3] = call_function[target=torch.empty_strided](args = ((), ()), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(1,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg1_1, out_ptr0: %buf0, xnumel: 1, r0_numel: 129, XBLOCK: 1}})
%buf2 : [num_users=2] = call_function[target=torch.empty_strided](args = ((129, 1), (1, 1)), kwargs = {dtype: torch.float32, device: cuda:0})
%addmm : [num_users=0] = call_function[target=torch.addmm](args = (%buf0, %arg0_1, %arg1_1), kwargs = {alpha: 1, beta: 1, out: %buf2})
%delete : [num_users=0] = call_function[target=torch._inductor.codegen.wrapper_fxir.delete](args = (%buf0,), kwargs = {})
return (buf2,)
```
Here's a graph which indexes into a tuple using `operator.getitem`. This is necessary to use the output of the `torch.topk` operation.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%buf0 : [num_users=3] = call_function[target=torch.ops.aten.topk.default](args = (%arg0_1, 2), kwargs = {})
%buf1 : [num_users=2] = call_function[target=operator.getitem](args = (%buf0, 0), kwargs = {})
%buf2 : [num_users=2] = call_function[target=operator.getitem](args = (%buf0, 1), kwargs = {})
%delete : [num_users=0] = call_function[target=torch._inductor.codegen.wrapper_fxir.delete](args = (%buf0,), kwargs = {})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(2,)], tma_descriptor_metadata: {}, kwargs: {in_out_ptr0: %buf1, xnumel: 2, XBLOCK: 2}})
%triton_kernel_wrapper_mutation_1 : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 1, constant_args_idx: 1, grid: [(2,)], tma_descriptor_metadata: {}, kwargs: {in_out_ptr0: %buf2, xnumel: 2, XBLOCK: 2}})
return (buf1, buf2)
```
Here's a graph that reinterprets an output tensor using `torch.as_strided`. This is one way to handle Inductor's `ReinterpretView` op.
```
graph():
%arg0_1 : [num_users=1] = placeholder[target=arg0_1]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((2, 4), (4, 1)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [(8,)], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg0_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: 8, XBLOCK: 8}})
%buf0_view_buf0_0 : [num_users=1] = call_function[target=torch.as_strided](args = (%buf0, (8,), (1,), 0), kwargs = {})
return (buf0_view_buf0_0,)
```
Here's a graph with dynamic shapes. This one is a little bit funky. Inductor provides a graph input for each shape symbol, which we map to a placeholder, in this example `s6`. Then, shape expressions in the generated code can refer to the symbol `s6`. The size hint for `s6` is stored in `node.meta["val"]` where `node` is the placeholder defining it. This works out in the generated python code because the placeholder defines a Python variable with the name `s6`.
```
graph():
%s6 : [num_users=0] = placeholder[target=s6]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%arg2_1 : [num_users=1] = placeholder[target=arg2_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ((s6,), (1,)), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [[-(((-s6)//8)), 1, 1]], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg2_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: s6, XBLOCK: 8}})
return buf0
```
Here's another graph, this time with dynamic shapes and strides. The grid expression is more complex since the numel is a product of dimensions.
```
graph():
%s10 : [num_users=0] = placeholder[target=s10]
%arg1_1 : [num_users=1] = placeholder[target=arg1_1]
%arg2_1 : [num_users=1] = placeholder[target=arg2_1]
%buf0 : [num_users=2] = call_function[target=torch.empty_strided](args = ([s10, s10], [s10, 1]), kwargs = {dtype: torch.float32, device: cuda:0})
%triton_kernel_wrapper_mutation : [num_users=0] = call_function[target=torch.ops.higher_order.triton_kernel_wrapper_mutation](args = (), kwargs = {kernel_idx: 0, constant_args_idx: 0, grid: [[-(((s10**2)//(-64))), 1, 1]], tma_descriptor_metadata: {}, kwargs: {in_ptr0: %arg2_1, in_ptr1: %arg1_1, out_ptr0: %buf0, xnumel: s10**2, XBLOCK: 64}})
return buf0
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146942
Approved by: https://github.com/jansel
Make Functorch interpreters serializable most of the time, so that we can save the guards on functorch states.
## Test Cases:
0. torch.compile() without functorch layers present. Guard should fail with any layer being pushed.
1. torch.compile() nested in vmap.
2. torch.compile() nested in grad.
3. torch.compile() nested in jvp + vmap
4. torch.compile() nested functionalize
5. torch.compile() nested in vmap + grad
Differential Revision: [D74008787](https://our.internmc.facebook.com/intern/diff/D74008787/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152616
Approved by: https://github.com/zou3519
ghstack dependencies: #152615
Change CI docker images to be `ci-image:<image name>-<folder sha>` instead of `<image name>:<folder sha>` so we never have to make a new ecr repo ever again
Pros:
never have to make a new ecr repo ever again
Cons:
if it aint broken, dont fix it?
Don't need to change linux-test images since they use the "full name" of the image with the docker registry and the tag
In order to prevent others needing to rebase past this PR, also push the image to the "old name". This can be removed after this PR has been in main for a while
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152209
Approved by: https://github.com/seemethere, https://github.com/atalman
There's an "are we compiling" check in SAC, which we rely on to know when to propagate recompute tags during tracing.
This check was a bit brittle, and missed cases where input ops accept list of tensors - I updated it to check if a `FunctionalTensorMode` is active, which should be a 100% reliable way to know if AOTDispatcher is in the middle of running.
There is a long-standing followup here around unifying `torch.compiler.is_compiling()` to work in all cases. We should probably just update it to always check if FakeMode/FunctionalMode are active and use it there. This has a bit of BC risk though so I opted for the more local fix to SAC.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152195
Approved by: https://github.com/soulitzer
Summary: This diff implements an AsyncManifoldCache class that performs cache write and update ttl operations in an async manner. Essentially we are ok with the fire and forget approach where we dont guarantee that we can observe our writes, this gives us better runtime latency.
Test Plan: added new unit test
Reviewed By: jamesjwu
Differential Revision: D73867797
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152452
Approved by: https://github.com/jamesjwu
By implementing `div_floor` and `div_trunc` . Do not mark `div_trunc` as OPMATH, to align following output with CPU(if division is performed in fp32, than result will be truncated to 25
```
import torch
print(torch.tensor([[-7.4688, -3.1289]], dtype=torch.float16,device="cpu").div(torch.tensor([-0.2988, -0.8789], dtype=torch.bfloat16,device="cpu"), rounding_mode="trunc"))
tensor([[24., 3.]])
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152758
Approved by: https://github.com/dcci
ghstack dependencies: #152663, #152515, #152737, #152743
### Summary
Recreating #151990 to mitigate easyCLA failure
compute_global_tensor_shape util function takes in local tensor shape, device mesh
and placements. We all gather the shapes from the shards and according to the placement
type we construct the global shape.
Note: currenty only implemented for placement type Shard and Replicate, TODO for StridedShared
### Test
`pytest test/distributed/tensor/test_utils.py`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152751
Approved by: https://github.com/XilunWu
Adds the `is_triton_capable` and `raise_if_triton_unavailable` class methods to the device interface, to allow device types to run their own checks for Triton _capability_ (which means a device can actually support Triton in the first place) and _availability_ (if the correct backend of Triton is installed and is functional for the device).
Using the device interface allows us to do these checks in a device-agnostic way, allow external backends to attest their Triton support by simply implementing those methods. The intention is for this to back things like the `has_triton` utility method.
This has been split from #139171.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152529
Approved by: https://github.com/jansel
Two error messages in the codebase instruct the user to use `Tendor.dense()`. This method doesn't exist, but `Tensor.to_dense()` does, and this is what the user should be using instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152631
Approved by: https://github.com/jansel
This is my suggestion for resolving #152087
This PR extends the constructor of `AOTIModelPackageLoader` with an (optional) device index. The device type is still determined by `metadata_["AOTI_DEVICE_KEY"]`, but the `device_index` argument can be used to move an AOTI model package to different devices like `cuda:0`, `cuda:1`, ... in a convenient way. AFAIK, this is not possible so far using `AOTIModelPackageLoader` alone. The default case (no device index specified) with `metadata_["AOTI_DEVICE_KEY"] == "cuda"` would lead to the current behavior, i.e., the model is loaded to device `cuda`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152093
Approved by: https://github.com/desertfire
`torch/csrc/utils.h` should be device-independent. Currently, it contains CUDA-related implementations, which indirectly causes the [failure of ROCm testing](https://github.com/pytorch/pytorch/pull/151914#issuecomment-2839691038) (The reason is that the ROCm test environment shouldn`t expose HIP-related header files, which causes the JIT compilation to fail during testing)
Therefore, move CUDA-related implementations to `torch/csrc/cuda/utils.h`.
**Question:**
This change may introduce BC-breack.
I searched for this function globally on github and I think the impact is very small.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152521
Approved by: https://github.com/Skylion007, https://github.com/albanD
ghstack dependencies: #152512, #152513
Summary:
change set
1. a ShardedTensor could have 0 size initially, the current check won't pass if the size is 0, added here
2. when we call ShardedTensor._init_from_local_shards, it will assume all the metadata is correct, all_gather to double check. In the new case, the metadata could be all 0 size, and the tensor has actual size, we need to provide such capability to recalculate the local/global metadata from the local tensor by all_gathering the information
Test Plan: i don't see a UT is associated, I have tested this with diff stack, D73274786.
Differential Revision: D73903933
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152583
Approved by: https://github.com/q10, https://github.com/fduwjj
What initially supposed to be a very straightforward change resulted in small refactor of binary op tensor generators when invoked for mixed dtype, which surfaced via `test_output_grad_match_sinc_mps_float16` test failure.
If operands are of different dtype (in particular float16 tensor and float32 scalar), one must perform an operation with `opmath_t` (or `TensorIterator::common_dtype()`) precision, rather than casting both operands to output dtype and performing it then, which can be demonstrated via the following example:
```
>>> torch.tensor([-1.8633, 6.2031, -2.2500, -3.3926, 8.5938, 5.9766], dtype=torch.half).mul(torch.pi)
tensor([ -5.8555, 19.4844, -7.0703, -10.6562, 27.0000, 18.7812],
dtype=torch.float16)
>>> torch.tensor([-1.8633, 6.2031, -2.2500, -3.3926, 8.5938, 5.9766], dtype=torch.half).mul(torch.tensor(torch.pi, dtype=torch.float16))
tensor([ -5.8516, 19.4844, -7.0664, -10.6562, 26.9844, 18.7656],
dtype=torch.float16)
```
Solve this problem for now, but introducing `REGISTER_OPMATH_BINARY_OP` that indicates that operands must be cast to opmath_t, before performing the computation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152515
Approved by: https://github.com/Skylion007, https://github.com/kulinseth, https://github.com/dcci
ghstack dependencies: #152663
As a result of adding subgraph as a choice to inductor https://github.com/pytorch/pytorch/pull/149761 and enabling FP32 output from PyTorch GEMMs from FP16/BF16 inputs: https://github.com/pytorch/pytorch/pull/150812, this PR enables decompose_k as an autotuning choice for Inductor in generating the fastest matmuls with Triton. DecomposeK is currently only enabled for `torch.compile`.
Followups:
* decompose_k does not currently support epilogue fusion, which will take some work to enable
* Enable autotuning the bmm with Triton Templates as well without requiring tons of more compile time, async compilation. Anecdotal evidence shows that Triton BMM performs better usually than aten BMM
* Add for addmm
* Enable for Inference and AOTI
Below are the results of running TritonBench for Split-K shapes, comparing the aten performance versus pt2_triton, which now autotunes on decompose_k, seeing >10% speedup compared to aten on average, and for some shapes over 3x the performance of the best Triton mm previously:
<img width="929" alt="Screenshot 2025-04-28 at 9 15 39 PM" src="https://github.com/user-attachments/assets/27d85bbc-4f3a-43a6-a8fa-d4a5bbb8c999" />
TorchInductor Benchmark Dashboard:
<img width="1727" alt="Screenshot 2025-04-30 at 2 02 53 PM" src="https://github.com/user-attachments/assets/4acd7ffc-407f-4cfd-98bb-2e3d8b1f00b3" />
We see speedups across all runs for training. Compile time increased as expected, with more `mm` options to tune over.
Differential Revision: [D73820115](https://our.internmc.facebook.com/intern/diff/D73820115)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150654
Approved by: https://github.com/eellison
This PR:
- cleans up some existing comments that don't make sense anymore
- hooks up the "custom_op_default_layout_constraint" back (that seems to
have broken)
- cleans up the "lazy registration path" which seems to never get hit
anymore
- adds dislike_padding to nodes that require exact strides
Test Plan:
- tests + CI
disable padding
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148104
Approved by: https://github.com/shunting314, https://github.com/eellison
I saw this warning when compiling a 3rd party lib and did not agree with it. I'm not sure the original reason why we would want to force people to pass in TORCH_CUDA_ARCH_LIST to cmake vs set it as an env var. As a developer, it's much easier to set it as an env var or have it be autodetected. I also realized this warning was from before 2018!!! 7 years ago! And there are no plans to actually enforce this (nor should there be), so let's remove this misleading warning.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152715
Approved by: https://github.com/malfet, https://github.com/zou3519
Summary:
This PR fixes a bug in the Triton kernel invocation path where the `workspace_tensor` was inserted before the unpacked `extra_args` list in the final kernel argument list. This broke the expected ordering of arguments when dynamic shape size hints are emitted.
When dynamic shapes are used, `extra_args` contains both size hint arguments and grid arguments. The kernel expects the argument list to follow the order: **size hints → workspace tensor → grid args**. But previously, the `workspace_tensor` was inserted before unpacking `extra_args`, resulting in: **workspace tensor → size hints → grid args**, which is incorrect.
This fix constructs the workspace tensor earlier, allowing it to be slotted in after the size hints and before the grid arguments, restoring the expected argument layout.
Test Plan:
contbuild and OSS CI
Reviewers: paulzhan
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152660
Approved by: https://github.com/PaulZhang12, https://github.com/drisspg
Preparatory refactor for https://github.com/pytorch/pytorch/pull/146942.
This PR introduces a new wrapper IR line to represent symbolic call args. This deletes a little bit of duplicated code between the Python and C++ backends. In the main PR, having a Wrapper IR line for this also tells the FX backend what this part of the wrapper code is doing. Before this PR, symbolic call args generated raw Python lines, which confuse the FX converter.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152587
Approved by: https://github.com/jansel
**Context**:
bucketize is relatively expensive, computationally. So it's not always profitable to fuse it if it means doing extra computation. For example, this repro:
https://gist.github.com/davidberard98/7fd6af7e6291787c246c705945a25554
shows a slowdown from 56us (eager) to ~100us (torch.compile-d): instead of computing 2\*\*15 binary searches, the fused version does 2\*\*15 * 384 - one for each of the broadcasted outputs.
**Solution**:
Realize the output of bucketize (and searchsorted, which also uses inductor's ops.bucketize). If there's an opportunity to do non-broadcasted fusions, the scheduler can still apply such fusions later on.
After this PR, instead of a slowdown, we see an improvement from 56us (eager) to 33us (compiled).
Differential Revision: [D74036850](https://our.internmc.facebook.com/intern/diff/D74036850)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152644
Approved by: https://github.com/benjaminglass1, https://github.com/eellison
Summary:
### Diff Context
- Sometime background process can be stuck processing async checkpoint request, and trainer shutdown can occur before the background process completes.
- Fix, timeout the thread while reading the IPC queue for a response from background process.
Differential Revision: D74017700
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152629
Approved by: https://github.com/saumishr
This PR enabled fp8 distributed tests on MI300.
For testing the added feature, ran distributed.tensor.parallel.test_micro_pipeline_tp test and all the tests passed successfully, and no tests were skipped.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151977
Approved by: https://github.com/jeffdaily
definitely_true is almost same as guard_or_false, the potential differences are not meaningful to a degree that justify the
existence of both. same for definitely_false, it can be expressed with guard_or_true and guard_or_false.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152463
Approved by: https://github.com/bobrenjc93
In some internal frameworks, on second attempts the actual code is copied to a different path than previous attempts.
but its still the same. PGO will not work on those cased due to the following, sate entries before this PR used to be identified by (filepath, function name, line number).
after this PR they are identified by (hash(filepath) , function name, line number). This way PGO will work for those jobs on future attempts and re-compilations of static versions will be avoided.
Sometimes we do not have access to the source code, (file does not exists)
This seems to happen mostly when we re-trace a compiled function but generally it can happen .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152628
Approved by: https://github.com/oulgen
Add additional conditions to `build_pytorch_libs.py` to avoid fetching NCCL when `USE_CUDA` or `USE_NCCL` are disabled. While at it, adjust the existing condition for `USE_SYSTEM_NCCL` to use the utility function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152533
Approved by: https://github.com/albanD
So when we use mark_unbacked the graph will have an unbacked inputs symInt. Right now,
deferred runtime assertions that uses those is never generated.
This PR changes that, such that in the forward graph we consider those and generate the corresponding
runtime assertions of them. We still ignore them for backward which is not ideal
The way we generate runtime assertion is by emitting them when all the defined unbacked symbols used
in them are seen.
We previously skipped placeholder, because for backward we have a wacky approach were we
ignore input defined unbacked symbols and assumes assertions that uses them are already emitted
in forward and we try to emit all other runtime assertions again. see [Note [Backwards runtime asserts]
Doing that we ends up only emitting the runtime assertions that depends on things defined solely in backward, but we could miss checks that spans inputs defined in both backward and forward, i.e one symbol defined in forward passed as input to backward., and another that is defined in backward.) .This is not ideal an ideal approach could be something like this https://github.com/pytorch/pytorch/pull/151919 but it require more work .
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152231
Approved by: https://github.com/aorenste
I tried `beginAllocateToPool` instead of `_cuda_beginAllocateCurrentStreamToPool` and the error in #151199 does not happen any more.
However, this approach is unsafe for multithreading. When multiple run_eager happens concurrently, we expect memory allocation to different mem_pool. Since beginAllocateToPool does not check stream, these memory allocation may happen on the same mem_pool.
So, I use `_cuda_beginAllocateCurrentThreadToPool` to direct all memory allocation on the same thread to a given mem_pool. In particular, `_cuda_beginAllocateCurrentThreadToPool` records the launching thread id, and during runtime checks if the current thread id matches the launching thread id.
Fixes#151199
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152472
Approved by: https://github.com/eellison, https://github.com/ngimel
As a result of adding subgraph as a choice to inductor https://github.com/pytorch/pytorch/pull/149761 and enabling FP32 output from PyTorch GEMMs from FP16/BF16 inputs: https://github.com/pytorch/pytorch/pull/150812, this PR enables decompose_k as an autotuning choice for Inductor in generating the fastest matmuls with Triton. DecomposeK is currently only enabled for `torch.compile`.
Followups:
* decompose_k does not currently support epilogue fusion, which will take some work to enable
* Enable autotuning the bmm with Triton Templates as well without requiring tons of more compile time, async compilation. Anecdotal evidence shows that Triton BMM performs better usually than aten BMM
* Add for addmm
* Enable for Inference and AOTI
Below are the results of running TritonBench for Split-K shapes, comparing the aten performance versus pt2_triton, which now autotunes on decompose_k, seeing >10% speedup compared to aten on average, and for some shapes over 3x the performance of the best Triton mm previously:
<img width="929" alt="Screenshot 2025-04-28 at 9 15 39 PM" src="https://github.com/user-attachments/assets/27d85bbc-4f3a-43a6-a8fa-d4a5bbb8c999" />
TorchInductor Benchmark Dashboard:
<img width="1727" alt="Screenshot 2025-04-30 at 2 02 53 PM" src="https://github.com/user-attachments/assets/4acd7ffc-407f-4cfd-98bb-2e3d8b1f00b3" />
We see speedups across all runs for training. Compile time increased as expected, with more `mm` options to tune over.
Differential Revision: [D73820115](https://our.internmc.facebook.com/intern/diff/D73820115)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150654
Approved by: https://github.com/eellison
Compiled Autograd retraces AOT's bw_module at backward runtime into a larger graph, and today this runs into an issue on warm cache runs because the bw_module is not restored. This PR adds it to the cache, by first stripping it bare from unserializable metadata. I also intentionally differentiate the cached and non-cached versions to avoid accidental attempts of AOT compilation with a restored bw_module (would probably crash).
Note that since the cache entry may be used by runs that use compiled autograd and runs that do not, we need to cache both the lowered backward and the bw_module.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151860
Approved by: https://github.com/jamesjwu
ghstack dependencies: #149707
Today, we mark graph outputs as maybe dynamic, this lets a compilation to communicate to future compilations whether certain graph inputs are dynamic. Similarly, we can do this to saved activations, which may be used in future compilations as well. This is especially prevalent in compiled autograd, where tensor activations will always become graph inputs.
Changes to the tests were mainly cosmetic, with the exception of tests that relied on duck shaping. By annotating tensor dims, we prevent them from reusing pre-existing symbols, so this change will make graphs use duck shapes less than before, which affects some of the caching tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149707
Approved by: https://github.com/bdhirsh
Summary: Discovered when attempting to resolve arvr builds, should resolve issues around utilizing functorch through export.
Test Plan:
```
buck2 test arvr/mode/linux/opt //arvr/libraries/xrrp/ml/python/test:convert_to_etvk_test
```
Differential Revision: D74013898
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152620
Approved by: https://github.com/zou3519
When multiple checkpoint regions are back-to-back with no operations in-between, we enforce the operation at the boundary to be force-saved, see 7ea0da2d57/torch/_functorch/partitioners.py (L772-L807)
When using the `memory_budget` formulation on a graph which already has AC inside, we should respect the boundaries of the AC decision (which is set to `MUST_SAVE`), and thus ban those nodes from possible recomputation.
Adding tests would be nice, but not sure what's the best way to test this right now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141684
Approved by: https://github.com/bdhirsh
Summary: Previously D70489427 changed tanh impl to `.tanh()`, and this is causing some meta internal workload perf regression. This diff will introduce a config so we can set it based on need.
Differential Revision: D73909371
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152542
Approved by: https://github.com/desertfire
Fixes#151582
example warning for Dim.AUTO:
```
torch/_export/non_strict_utils.py:499] dimension inputs['x'].shape[1] 0/1 specialized; Dim.AUTO was specified along with a sample input with hint = 1.
```
example error when Dim.DYNAMIC specializes:
```
- Received user-specified dim hint Dim.DYNAMIC(min=None, max=None), but export 0/1 specialized due to hint of 0 for dimension inputs['x'].shape[0].
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151827
Approved by: https://github.com/angelayi
Adding NVSHMEM as a backend for `SymmetricMemory`, implementation of which is in `NVSHMEMSymmetricMemory.cu`.
Moving some helper functions in `CUDASymmetricMemory.cu` to `CUDASymmetricMemoryUtils.cpp`, so that they can be shared by `NVSHMEMSymmetricMemory`. These functions are mostly side-band exchange helpers (`store_all_gather`, `IpcChannel`, etc).
Adding `TORCH_SYMMEM` to control which implementation to use for CUDA tensors, currently support: `CUDA` (in-house impl), `NVSHMEM`.
The NVSHMEM feature is gated by build-time flag: `USE_NVSHMEM=1`. And `NVSHMEM_HOME` setting is required (TODO).
Ported most code from #146593.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151261
Approved by: https://github.com/fegin, https://github.com/fduwjj
We need to make function schema proxyable to trace a the auto_functionalized hop that takes function schema as inputs. The implementation basically follows how we support torchbind object:
1. upon seeing an untracked function schema arg, we creates a constant get_attr node
2. we track the function schema argument in export to support lift/unlift.
3. we need to support serde for functional schema. We'll add support for this in follow-up PRs.
However, compared with torchbind object:
1. we don't need a dynamo implementation, because the function schema is added when we auto_functionalize a hop to the argument of auto_functionalized. One potential use case is users re-traces an exported program with strict mode. Since non-strict is the default now, we don't see a use case yet.
2. we don't need an inductor implementation, because the function schema will go away after auto_functionalized re-inplacing pass.
edit: we greatly simplifies (and generalizes) the implementation following @zou3519 's suggestion of using pytree.register_constant
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152073
Approved by: https://github.com/zou3519
ghstack dependencies: #152072
Test Plan:
Dumped the local net torch.package to local
Ran
```
buck2 run scripts/shengqin:test_model_export -- /tmp/mtia_local_torch_package {\"local\":null}
```
succeeded
Reviewed By: hongyang-zhao
Differential Revision: D73405271
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152039
Approved by: https://github.com/houseroad
Summary:
Same with D71358949, but removing newly added log to avoid test failures.
Port over replace_lce_with_matmul and replace_first_lce_with_fused_matmul_lce to PT2 pre_grad pass.
Original dper pass diffs: D67884534, D68123479, D68384238
Test Plan:
Test 1. Covers replace_lce_with_matmul and case 1 of replace_first_lce_with_fused_matmul_lce
```
CUDA_VISIBLE_DEVICES=6 TORCH_LOGS=+inductor,aot TORCH_COMPILE_DEBUG=1 TORCHINDUCTOR_MAX_AUTOTUNE=1 buck2 run mode/opt-split-dwarf mode/inplace -c fbcode.platform010_cuda_version=12 -c fbcode.nvcc_arch=h100 caffe2/torch/fb/model_transform/experimental/benchmark:mts_gpu_benchmark -- --model-path=manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/669809193/0/gpu_lowering/input.predictor.disagg.gpu.merge --lower-backend="AOT_INDUCTOR" --add_passes="use_matmul_fuse_lce_replace_first_LCE,use_contiguous_linear_reduction_replace_linear_reduction" --batch-size=3072 --gpu-trace --disable_acc_tracer=true 2>&1 | tee ~/logs/disable_acc_tracer/aoti_cmf_ctr_triton_669809193_0_diable_acc.log
```
Log: P1798246938
Test 2. Covers replace_lce_with_matmul and case 2 of replace_first_lce_with_fused_matmul_lce
```
CUDA_VISIBLE_DEVICES=7 TORCH_LOGS=+inductor,aot TORCH_COMPILE_DEBUG=1 TORCHINDUCTOR_MAX_AUTOTUNE=1 buck2 run mode/opt-split-dwarf mode/inplace -c fbcode.platform010_cuda_version=12 -c fbcode.nvcc_arch=h100 caffe2/torch/fb/model_transform/experimental/benchmark:mts_gpu_benchmark -- --model-path=manifold://ads_storage_fblearner/tree/user/facebook/fblearner/predictor/677734158/9/gpu_lowering/input.predictor.disagg.gpu.merge --lower-backend="AOT_INDUCTOR" --add_passes="use_matmul_fuse_lce_replace_first_LCE,use_matmul_lce_replace_normal_LCE" --batch-size=3072 --gpu-trace --disable_acc_tracer=true 2>&1 | tee ~/logs/disable_acc_tracer/aoti_cmf_ctr_triton_677734158_9_diable_acc.log
```
Log: P1798246675
Seeing logs like
`[Pre grad(predispatch IR)] Apply use_matmul_fuse_lce_replace_first_LCE pass, save before/after graph to /tmp/tmp8lyzoh79, graph before/after are the same = False`
Differential Revision: D73934142
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152536
Approved by: https://github.com/wdvr
These are the tests for torch._inductor.compile, so I renamed the file
test_compile. This is to avoid confusion with
torch._inductor.standalone_compile, which is now a lot more standalone
than torch._inductor.compile.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152103
Approved by: https://github.com/oulgen
Today `cpp_extensions` makes heavy use of printing to stderr, this makes our life harder in KernelBot where we typically rely on stderr to only surface real errors but instead today cpp_extensions leverages stderr for updates that could be qualified as INFO, WARNING, ERROR
Now instead we'll recommend users of our cpp extension system to do something like
```python
import logging
cpp_ext_logger = logging.getLogger("torch.utils.cpp_extension")
cpp_ext_logger.setLevel(logging.WARNING)
```
While this dramatically reduces log spew, it can be viewed as a BC breaking change if people were relying on certain strings being present in stdout or stderr
Considering different teams might want to silence errors differently, this PR proposes replacing all `print()` statements with `logging` statements with the same heuristics that the python logging module recommends
1. DEBUG: For things like detailed compilation steps or reading filepaths - by default gets logged on stdout
2. INFO: Build progress - by default gets logged on stdout
3. WARNING: Surfacing issues that might cause bad performance or slow compilation times - by default gets logged on stdout
4. ERROR: Problems that prevent proper functioning - by default gets logged on stdout
Note that warnings.warn is a different library and is not hooked up to the python logging module by default
So the goal of this PR is to make it possible for teams to set the logging that is most appropriate to them. One annoying thing is logger throws ruff errors if you try to use it in conjunction with f strings or .format so have to use old school %s
An unrelated improvement I'd be happy to push to a seperate PR is adding support for "native" in `TORCH_CUDA_ARCH_LIST` which would just pick the ARCH for the current device
An example of what's in stderr today
```
Using /root/.cache/torch_extensions/py311_cu124 as PyTorch extensions root...
Detected CUDA files, patching ldflags
Emitting ninja build file /root/.cache/torch_extensions/py311_cu124/grayscale/build.ninja...
/usr/local/lib/python3.11/site-packages/torch/utils/cpp_extension.py:2059: UserWarning: TORCH_CUDA_ARCH_LIST is not set, all archs for visible cards are included for compilation.
If this is not desired, please set os.environ['TORCH_CUDA_ARCH_LIST'].
warnings.warn(
Building extension module grayscale...
Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)
Loading extension module grayscale...
/usr/local/lib/python3.11/site-packages/torch/_dynamo/variables/functions.py:679: UserWarning: Graph break due to unsupported builtin grayscale.PyCapsule.grayscale. This function is either a Python builtin (e.g. _warnings.warn) or a third-party C/C++ Python extension (perhaps created with pybind). If it is a Python builtin, please file an issue on GitHub so the PyTorch team can add support for it and see the next case for a workaround. If it is a third-party C/C++ Python extension, please either wrap it into a PyTorch-understood custom operator (see https://pytorch.org/tutorials/advanced/custom_ops_landing_page.html for more details) or, if it is traceable, use torch.compiler.allow_in_graph.
torch._dynamo.utils.warn_once(msg)
```
Whereas after this PR users can do
`python benchmark_load_inline.py > >(tee stdout.txt) 2> >(tee stderr.txt >&2)`
```python
import os
import sys
from pathlib import Path
import shutil
import tempfile
import torch
from torch.utils.cpp_extension import load_inline
import logging
cpp_ext_logger = logging.getLogger("torch.utils.cpp_extension")
cpp_ext_logger.setLevel(logging.WARNING)
os.environ["TORCH_CUDA_ARCH_LIST"] = "native"
cpp_code = """
torch::Tensor to_gray(torch::Tensor input);
"""
cuda_kernel_code = """
torch::Tensor to_gray(torch::Tensor input) {
auto output = torch::epty({input.size(0), input.size(1)}, input.options());
return output ;
}
"""
# Avoid caching results
with tempfile.TemporaryDirectory() as build_dir:
cuda_module = load_inline(
name="to_gray_cuda",
cpp_sources=cpp_code,
cuda_sources=cuda_kernel_code,
functions=["to_gray"],
with_cuda=True,
verbose=True,
extra_cflags=["-std=c++17"], # "-ftime-report", "-H"],
extra_cuda_cflags=["-arch=sm_89"],
build_directory=build_dir,
)
```
## New logs
### On failure
Which gives a much more reasonable stdout
```
[1/3] /usr/local/cuda-12.8/bin/nvcc --generate-dependencies-with-compile --dependency-output cuda.cuda.o.d -DTORCH_EXTENSION_NAME=to_gray_cuda -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1016\" -isystem /home/marksaroufim/pytorch/torch/include -isystem /home/marksaroufim/pytorch/torch/include/torch/csrc/api/include -isystem /usr/local/cuda-12.8/include -isystem /usr/local/cuda/targets/x86_64-linux/include -isystem /home/marksaroufim/.conda/envs/nv/include/python3.10 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_80,code=sm_80 --compiler-options '-fPIC' -arch=sm_89 -std=c++17 -c /tmp/tmpbg_xzv0r/cuda.cu -o cuda.cuda.o
FAILED: cuda.cuda.o
/usr/local/cuda-12.8/bin/nvcc --generate-dependencies-with-compile --dependency-output cuda.cuda.o.d -DTORCH_EXTENSION_NAME=to_gray_cuda -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1016\" -isystem /home/marksaroufim/pytorch/torch/include -isystem /home/marksaroufim/pytorch/torch/include/torch/csrc/api/include -isystem /usr/local/cuda-12.8/include -isystem /usr/local/cuda/targets/x86_64-linux/include -isystem /home/marksaroufim/.conda/envs/nv/include/python3.10 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_80,code=sm_80 --compiler-options '-fPIC' -arch=sm_89 -std=c++17 -c /tmp/tmpbg_xzv0r/cuda.cu -o cuda.cuda.o
/tmp/tmpbg_xzv0r/cuda.cu(6): error: namespace "torch" has no member "epty"
auto output = torch::epty({input.size(0), input.size(1)}, input.options());
^
1 error detected in the compilation of "/tmp/tmpbg_xzv0r/cuda.cu".
[2/3] c++ -MMD -MF main.o.d -DTORCH_EXTENSION_NAME=to_gray_cuda -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1016\" -isystem /home/marksaroufim/pytorch/torch/include -isystem /home/marksaroufim/pytorch/torch/include/torch/csrc/api/include -isystem /usr/local/cuda-12.8/include -isystem /usr/local/cuda/targets/x86_64-linux/include -isystem /home/marksaroufim/.conda/envs/nv/include/python3.10 -fPIC -std=c++17 -std=c++17 -c /tmp/tmpbg_xzv0r/main.cpp -o main.o
ninja: build stopped: subcommand failed.
```
And stderr
```
Traceback (most recent call last):
File "/home/marksaroufim/pytorch/torch/utils/cpp_extension.py", line 2874, in _run_ninja_build
subprocess.run(
File "/home/marksaroufim/.conda/envs/nv/lib/python3.10/subprocess.py", line 526, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['ninja', '-v']' returned non-zero exit status 1.
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/home/marksaroufim/load_inline_slow/benchmark_load_inline.py", line 30, in <module>
cuda_module = load_inline(
File "/home/marksaroufim/pytorch/torch/utils/cpp_extension.py", line 2261, in load_inline
return _jit_compile(
File "/home/marksaroufim/pytorch/torch/utils/cpp_extension.py", line 2367, in _jit_compile
_write_ninja_file_and_build_library(
File "/home/marksaroufim/pytorch/torch/utils/cpp_extension.py", line 2528, in _write_ninja_file_and_build_library
_run_ninja_build(
File "/home/marksaroufim/pytorch/torch/utils/cpp_extension.py", line 2892, in _run_ninja_build
raise RuntimeError(message) from e
RuntimeError: Error building extension 'to_gray_cuda'
```
### On success
stdout
```
[1/3] c++ -MMD -MF main.o.d -DTORCH_EXTENSION_NAME=to_gray_cuda -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1016\" -isystem /home/marksaroufim/pytorch/torch/include -isystem /home/marksaroufim/pytorch/torch/include/torch/csrc/api/include -isystem /usr/local/cuda-12.8/include -isystem /usr/local/cuda/targets/x86_64-linux/include -isystem /home/marksaroufim/.conda/envs/nv/include/python3.10 -fPIC -std=c++17 -std=c++17 -c /tmp/tmpxv_ovlrf/main.cpp -o main.o
[2/3] /usr/local/cuda-12.8/bin/nvcc --generate-dependencies-with-compile --dependency-output cuda.cuda.o.d -DTORCH_EXTENSION_NAME=to_gray_cuda -DTORCH_API_INCLUDE_EXTENSION_H -DPYBIND11_COMPILER_TYPE=\"_gcc\" -DPYBIND11_STDLIB=\"_libstdcpp\" -DPYBIND11_BUILD_ABI=\"_cxxabi1016\" -isystem /home/marksaroufim/pytorch/torch/include -isystem /home/marksaroufim/pytorch/torch/include/torch/csrc/api/include -isystem /usr/local/cuda-12.8/include -isystem /usr/local/cuda/targets/x86_64-linux/include -isystem /home/marksaroufim/.conda/envs/nv/include/python3.10 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -gencode=arch=compute_80,code=compute_80 -gencode=arch=compute_80,code=sm_80 --compiler-options '-fPIC' -arch=sm_89 -std=c++17 -c /tmp/tmpxv_ovlrf/cuda.cu -o cuda.cuda.o
[3/3] c++ main.o cuda.cuda.o -shared -L/home/marksaroufim/pytorch/torch/lib -lc10 -lc10_cuda -ltorch_cpu -ltorch_cuda -ltorch -ltorch_python -L/usr/local/cuda-12.8/lib64 -lcudart -o to_gray_cuda.so
```
And an empty stderr as expected
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152260
Approved by: https://github.com/albanD
This PR adds support for `sm_121` of the DGX Spark. The `sm_121` is binary compatible with `sm_120` (just like `sm_89` and `sm_86`), therefore a compilation targeting `sm_121` is not required.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152314
Approved by: https://github.com/eqy
I tried `beginAllocateToPool` instead of `_cuda_beginAllocateCurrentStreamToPool` and the error in #151199 does not happen any more.
However, this approach is unsafe for multithreading. When multiple run_eager happens concurrently, we expect memory allocation to different mem_pool. Since beginAllocateToPool does not check stream, these memory allocation may happen on the same mem_pool.
So, I use `_cuda_beginAllocateCurrentThreadToPool` to direct all memory allocation on the same thread to a given mem_pool. In particular, `_cuda_beginAllocateCurrentThreadToPool` records the launching thread id, and during runtime checks if the current thread id matches the launching thread id.
Fixes#151199
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152472
Approved by: https://github.com/eellison
This relaxes the guard introduced in #100444 (which aggressively guard
on the object id, despite Dynamo is just tracing its `__call__` method.
This allows users to bypass the high compilation time issue in #150706
by compiling transformer blocks only. Without this patch, we'd get lots
of unnecessary recompilation, as the block has difference attention
processor instances.
Compiling blocks only _significantly_ speeds up compilation process
(from ~310s to ~32s), and even speeds up e2e performance for some reason
(7.83s to 7.67s).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152395
Approved by: https://github.com/anijain2305
ghstack dependencies: #152369
Seems there was a typo where `set_device` was called when the intent was to use `current_device`
As-is the test will fail on multigpu systems with
`TypeError: set_device() missing 1 required positional argument: 'device'`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152474
Approved by: https://github.com/Skylion007
These are the tests for torch._inductor.compile, so I renamed the file
test_compile. This is to avoid confusion with
torch._inductor.standalone_compile, which is now a lot more standalone
than torch._inductor.compile.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152103
Approved by: https://github.com/oulgen
The standard requires that the argument to functions like `isdigit`, `isalpha`, and similar must be either `EOF` or an `unsigned char`; otherwise, the behavior is undefined (UB).
To avoid out-of-bounds reads, modern implementations of some libraries (such as glibc) deliberately pad their internal tables to guarantee valid memory access even for negative values. However, this is implementation-specific, and other libraries may not do this.
Properly casting the argument to `unsigned char` is good practice to avoid potential issues on some platforms.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152360
Approved by: https://github.com/cyyever, https://github.com/Skylion007
For 3rd-party devices now, `` instantiate_device_type_tests()`` with explicitly passing ``str`` obj (rather than `List[str]/Tuple[str]`) to argument ``only_for`` or ``except_for`` would causes unexpected results.
For example, if calling ``instantiate_device_type_tests(TestXXX, globals(), only_for="cpu")``, then it goes into [filter_desired_device_types()](f38dae76ee/torch/testing/_internal/common_device_type.py (L729)) and results in ``only_for=['c', 'p', 'u']`` because ``only_for`` we passed is a "cpu" string.
This PR fixes the above unexpected behavior for ``str`` case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152177
Approved by: https://github.com/albanD
First of all, by extending `c10:🤘:cast_to` to work correctly with complex dtypes, by introducing two more specializations: one that casts complex to scalar, and another that casts scalar to complex (as default metal typecast will turn `float x` into `float2(x, x)`)
Add ComplexHalf and ComplexFloat enum values to `c10:🤘:ScalarTypes` and handle them in `val_at_offs(ptr, offs, type)`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152504
Approved by: https://github.com/dcci
ghstack dependencies: #152443, #152466, #152479
As well as `.add`/`.sub` with complex alpha
Before this change `python3 -c "import torch;print(torch.rand(10, device='mps', dtype=torch.complex64).add(torch.rand(10, device='mps', dtype=torch.complex64), alpha=.5j))"` used to fail with
```
RuntimeError: value cannot be converted to type double without overflow
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152479
Approved by: https://github.com/dcci
ghstack dependencies: #152443, #152466
Differential Revision: D73626732
In this PR, we add support for bmm dynamic shape, provided that the batch stride is the biggest in the stride for A, B, and D. For example, for A of size `(B, M, K)`, we support stride `(M*K, K, 1)` and `(M*K, 1, M)`. With this assumption, we can infer the batch stride from existing arguments.
The reason is we don't want to add 2-3 more runtime params. The concerns are complexity and possible perf regression, though we didn't verify the latter.
We can revisit this if there is a need for that.
We also remove `B = 1` for normal mm and addmm. We tested it and didn't see perf regression. But open to revisiting this as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152393
Approved by: https://github.com/ColinPeppler
Summary:
Fixes https://github.com/pytorch/pytorch/issues/151476
The `custom_meta` collected from `mod` has keys that follow name of nodes in `mod`, which are inconsistent with the node names after the naming pass. For example a constant `b` will become `c_b`.
Test Plan: buck2 run caffe2/test:test_export -- -r test_run_decompositions_keep_tensor_constant_metadata
Differential Revision: D73703068
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152241
Approved by: https://github.com/angelayi
Power System build is failing with below error.
After this commit it is failing:
912102b4ec
Fix the build error along with test cases that are failing for complex double and complex float data type.
Build Failure Logs:
```
vec_base.h:790:6: error: use of deleted function ‘at::vec::DEFAULT::ComplexDbl& at::vec::DEFAULT::Vectorized<c10::complex >::operator’
790 | c[i] = a[i] * b[i];
| ~^
error: use of deleted function ‘at::vec::DEFAULT::ComplexDbl& at::vec::DEFAULT::Vectorized<c10::complex >::oper
ator’
802 | c[i] = a[i] / b[i];
| ~^
error: use of deleted function ‘at::vec::DEFAULT::ComplexFlt& at::vec::DEFAULT::Vectorized<c10::complex >::opera
tor’
790 | c[i] = a[i] * b[i];
| ~^
error: use of deleted function ‘at::vec::DEFAULT::ComplexFlt& at::vec::DEFAULT::Vectorized<c10::complex >::opera
tor’
802 | c[i] = a[i] / b[i];
| ~^
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152402
Approved by: https://github.com/malfet
When PyTorch is built with OpenBLAS support and libopenblas is ldrectly linked with libgomp.so the libtorch_cpu.so ends up getting multiple omp runtimes linked against it. This may result in unexpected runtime behaviour /regression. This patch fixes this by avoiding linking against libomp.so if OpenBLAS is linked against libgomp.so
Fixes#146603
Pull Request resolved: https://github.com/pytorch/pytorch/pull/147725
Approved by: https://github.com/albanD
This PR implements the second codegen task of CUTLASS EVT: translating inductor epilogue nodes into python code that will be traced by the EVT infra.
Details:
The implementation uses a simple ops wrapper which only supports add and mul pointwise ops today (to be extended in the future). This ops wrapper generates python code from inner_fn of the epilogue nodes in the format EVT expects. The main caveat is that one of the outputs needs to be named "D" and the accumulator input needs to be named "acc". Reads/writes are named according to the inductor buffer names otherwise.
Previously merged:
* #150904
* #150903
* #150346
* #150345
* #150344
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150905
Approved by: https://github.com/eellison
ghstack dependencies: #152305, #152306
This is a new pass to replace the pre-existing passes. It has the same
basic goal, to achieve communication overlap (latency hiding), but also
constrains the solution to not increase peak memory.
The principles of operation are detailed in code comments, but
summarized here:
- never reorder collectives relative to each other (TBD if we should
relax this later)
- before performing reordering, push all comm and wait nodes as late as possible, respecting data dependencies
- estimate peak memory and current memory at each scheduler node
- move collective nodes forward one position at a time, if the move does
not increaes curr memory beyond peak memory
The pass logs a summary table for each graph to TORCH_LOGS=overlap.
e.g. (exact format may have been tweaked but this shows the idea).
```
rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] Collective node initial exposed final exposed improvement limiting factor moves
[rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] ----------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------- --------------- ------------- ------------------- -------
[rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] ExternKernelSchedulerNode(name='op2') (torch.ops._c10d_functional.all_gather_into_tensor.default) (size=[2256, 256], stride=[256, 1]) (buf2) (12142 ns) 12141.6 6514.53 5627.08 prefetch limit 75
[rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] ExternKernelSchedulerNode(name='op6') (torch.ops._c10d_functional.reduce_scatter_tensor.default) (size=[282, 256], stride=[256, 1]) (buf7) (32266 ns) 32265.8 28429.2 3836.61 data dependency 78
[rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] ExternKernelSchedulerNode(name='op9') (torch.ops._c10d_functional.all_gather_into_tensor.default) (size=[256], stride=[1]) (buf11) (10801 ns) 10800.6 10732.3 68.254 peak memory 1
[rank0]:[rank0]:I0210 17:24:28.494000 2711253 torch/_inductor/comms.py:195] [0/0] [__overlap] ExternKernelSchedulerNode(name='op14') (torch.ops._c10d_functional.reduce_scatter_tensor.default) (size=[32], stride=[1]) (buf17) (10810 ns) 10809.5 10809.5 0 data dependency 4
[rank
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146562
Approved by: https://github.com/eellison
ghstack dependencies: #152060, #146561
Summary: This PR adds a private configuration to the partitioner that ensures that the decision taken is the same across all ranks. This is a temporary workaround, as when size_hints are also taken into account in compiler collectives this workaround will not be needed anymore.
Test Plan:
This has been tested on some internal models, but I haven't added any tests in PyTorch (yet?)
T
Differential Revision: D73666017
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152264
Approved by: https://github.com/bdhirsh
Installs setuptools since I get
https://github.com/pytorch/pytorch/actions/runs/14736804186/job/41364832984#step:5:60
```
+ python3 -m tools.generate_torch_version --is_debug=false
Traceback (most recent call last):
File "<frozen runpy>", line 198, in _run_module_as_main
File "<frozen runpy>", line 88, in _run_code
File "/home/ec2-user/actions-runner/_work/pytorch/pytorch/tools/generate_torch_version.py", line 9, in <module>
from setuptools import distutils # type: ignore[import]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'setuptools'
```
It should be a no op in the normal lint workflow since setuptools is in the docker image
Switched from using python3.10 to system python, which should be python3.9
Use venv to put deps not in the base?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152433
Approved by: https://github.com/huydhn
Summary: I'm just trying to fix the test again. It's out of date because it's disabled and some dynamo_timed-related fields are gone now.
Test Plan: `python test/dynamo/test_utils.py -k dynamo_timed`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152387
Approved by: https://github.com/anijain2305
This PR adds detailed logging of each triton kernel we compile, and its autotune result, to every kernel we compile with triton. We add these results to a global variable that we then clear after each triton kernel compile.
We can't keep these objects around after compile time, so we can't record the autotune cache save or coordinate descent tuning, unfortunately, but we can log at least:
- The duration of compilation
- Whether or not autotune cache hit
- The best autotuning config, if there's only one.
Example triton kernel info: https://gist.github.com/jamesjwu/493bdd0f36b0b7e3ca327f87bd6c2c75
See internal diff for an example log for internal model.
Differential Revision: [D73674443](https://our.internmc.facebook.com/intern/diff/D73674443)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152197
Approved by: https://github.com/oulgen, https://github.com/eellison
Run into this problem while re-enabling `test_float_repr_dynamic_shapes`, where `_print_Max` were called for integer and long argument which resulted in the following compilation error
```
error: call to 'max' is ambiguous
out_ptr0[x0 + x1*metal::max(1, ks0)] = static_cast<float>(tmp26);
^~~~~~~~~~
/System/Library/PrivateFrameworks/GPUCompiler.framework/Versions/32023/Libraries/lib/clang/32023.619/include/metal/metal_integer:2477:16: note: candidate function
METAL_FUNC int max(int x, int y)
^
/System/Library/PrivateFrameworks/GPUCompiler.framework/Versions/32023/Libraries/lib/clang/32023.619/include/metal/metal_integer:3686:17: note: candidate function
METAL_FUNC long max(long x, long y)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152430
Approved by: https://github.com/dcci
ghstack dependencies: #152363
This pr adds an optimal reordering for minimizing #partitions.
## Optimal reordering for minimizing #partitions
A bfs could minimize #partitions (ignore peak memory for now):
1. For each node, compute node_to_indegree: dict[node, int].
2. Maintain 2 queues: cudagraphable_nodes, and non_cudagraphable_nodes. Iterate through all nodes and add nodes to one of these 2 queues if node_to_indegree[node] == 0.
3. While non_cudagraphable_nodes is not empty: Pop 1 node, schedule it, update the indegree of all its successors, and add its successor nodes to one of the queues if node_to_indegree[successor] == 0.
4. While cudagraphable_nodes is not empty: Pop 1 node, schedule it, update the indegree of all its successors, and add its successor nodes to one of the queues if node_to_indegree[successor] == 0.
5. Repeat step 3 & 4 until all nodes have been scheduled.
We call this strategy `reorder_for_minimizing_partition`.
**Q: Why is this optimal?**
Suppose this is not optimal, we have a counter example with 2 non_cudagraphable regions:
```
[non_cudagrable1, cudagraphable2, non_cudagraphable3]
```
where we can reorder to only 1 non_cudagraphable region:
```
[non_cudagrable1, non_cudagraphable3, cudagraphable2]
```
This reorder means non_cudagraphable3 does not depend on cudagraphable2. So after we scheduled non_cudagraphable1, both non_cudagraphable3 and cudagraphable2 have in_degree as 0. If this is true, Step 3 should have already scheduled non_cudagraphable3 before cudagraphable2 such that the counter example cannot exist.
This shows we cannot find such a counter example and the bfs is optimal on minimizing #partitions.
## Minimize peak memory
`reorder_for_peak_memory` currently uses topological_sort_dfs, topological_sort_lpmf, and topological_sort_bfs, where the later 2 are bfs. ILP brings small benefits and it can hardly scale to more than 100 nodes, according to @xuanzhang816. So ILP is not used for peak memory reorder in the inductor.
Heuristics strategy:
- Conduct reorder_for_peak_memory as the default order
- Conduct reorder_for_minimal_partitions and get results as list[tuple[partition, bool]], where partition: list[BaseSchedulerNode] and bool for cudagraphable.
- If the reorder increases peak memory too much, we use the default order.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151968
Approved by: https://github.com/eellison
I frequently come to CONTRIBUTING.md to copy paste the below snippet to rebuild pytorch which in zsh gives this error because zsh interprets # as a command. These comments add nothing so just removing
```
error: pathspec 'sync' did not match any file(s) known to git
error: pathspec 'the' did not match any file(s) known to git
error: pathspec 'submodules' did not match any file(s) known to git
Building wheel torch-2.8.0a0+git9c01c87
invalid command name '#'
```
```
git submodule update --init --recursive # very important to sync the submodules
python setup.py develop # then try running the command again
git submodule update --init --recursive
python setup.py develop
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152259
Approved by: https://github.com/janeyx99
Remove conda usage from TD llm retriever job
python3 in the base is python3.9 right now. I'm not sure what the best way to deal with a potentially different python version would be, dnf install?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152338
Approved by: https://github.com/huydhn
Fixes#152344
Leak seems to be on the MPS Graph side, even though there is an identity tensor it seems like it's no longer enough to bypass the SDPA sequence which seems to leak memory.
Even adding 0.0f seems to be optimized to be ignored and still take the sdpa sequence(that's the reason for adding 1e-20)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152371
Approved by: https://github.com/malfet
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Add precedence to the infix printing done by sympy_str.
Without this change sympy_str will print the same string for both `a+b*(c+d)` and `(a+b)*(c+d)`.
While there I also cleaned up the printing for `-a` and `a - b`.
Added some tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151920
Approved by: https://github.com/jansel
Summary:
bucketization involves comparing an input with border values. Without careful consideration of dtypes, this can cause dangerous implicit casting.
aten.bucketize resolves this via dtype promotion. We enable dtype promotion for the inductor bucketization pass so as to maintain alignment with the aten op.
Test Plan:
```
python3 test/inductor/test_torchinductor.py -k "bucketize"
```
Fixes#145929
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150634
Approved by: https://github.com/davidberard98, https://github.com/eellison
Replaces the janky way of using the IntArrayRef to create an NSArray to ask for it to provide its contents in a string format with use of stringstream.
This speeds up the call for getting the key string for caching (or reading from cache) for shaped inputs by ~5x. While the actual wall time, depending on the number of input tensors, is only some microseconds this time represents non-negligible chunk of the overall time spent in preparing to dispatch work to the GPU. And since this function gets called on every time a (cacheable) operation in MPS is used it should be a small but broadly impacting time saver.
Using mps_linear as an example. Note this is before PR https://github.com/pytorch/pytorch/pull/152199 so it only captures the CPU time spent in the op call:
Before the change:
```
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x1108f07d0>
func(*args, **kwargs)
Median: 22.75 us
IQR: 0.87 us (22.50 to 23.38)
8361 measurements, 1 runs per measurement, 1 thread
```
After the change:
```
torch.linear time: <torch.utils.benchmark.utils.common.Measurement object at 0x108875350>
func(*args, **kwargs)
Median: 18.67 us
IQR: 0.46 us (18.50 to 18.96)
10342 measurements, 1 runs per measurement, 1 thread
```
Which aligns with the observed change for getTensorStringKeys() taking ~1us instead of ~5us in mps_linear op I got from a point measurement sandwiching the function call with `std::chrono::high_resolution_clock`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152202
Approved by: https://github.com/Skylion007
When CompiledFxGraph is deallocated, its cudagraphifed fn (i.e., `current_callable`) is expected to also be deallocated.
Without graph partition, this is true since the cudagraphified fn is only refered by compiled_fx_graph.current_callable.
However, with graph partition, runner.partitions hold cudagraphified fns while compiled_fx_graph.current_callable holds the runner.call. Thus the cudagraphied fn may not be deallocated when CompiledFxGraph is deallocated. This leads to errors in several unit tests (e.g., test_unaligned_static_input_no_cudagraphs and test_unaligned_static_input_non_trees).
In this PR, we also clean up runner.partitions when CompiledFxGraph is deallocated. This fixes the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152066
Approved by: https://github.com/eellison
Summary:
We suspect that switching the NVCC host compiler from GCC to Clang, while targeting multiple architectures, is causing issues because only _CUDA_ARCH_LIST_ is being passed, without _CUDA_ARCH_.
To resolve this c10 compilation error, we should first fix the problem and then switch the NVCC host compiler from GCC to Clang. Once this is done, the errors no longer occur.
Test Plan: CI
Reviewed By: zhuhan0
Differential Revision: D73383236
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152030
Approved by: https://github.com/cyyever, https://github.com/ZainRizvi
as titled, we can just set new_local_tensor to be the local tensor and
remove the None check, as there would be cases where there's no
transformation needed (i.e. src_placements and dst_placements are the same,
and we still want to return the original local_tensor)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152303
Approved by: https://github.com/awgu
Summary: The HF storage reader/writer component can work for any back-end in theory, so we shouldn't enforce the token to be passed into fsspecreader/writer, because the specific fsspec implementation may not handle tokens. Specifically, manifold doesn't accept a token arg, but we're passing one in always, which is throwing
Test Plan: signals
Differential Revision: D73130679
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151464
Approved by: https://github.com/Skylion007
`test_reproduce_121253_issue_addmm_fusion_check` checks for "mkl._mkl_linear" being found in the generated source which cannot be there when MKL isn't available.
Add skip marker similar to other tests in this file.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152322
Approved by: https://github.com/Skylion007
Adds explicit error checking during sharding propagation for view ops
rather than relying on runtime errors during local op execution.
Before:
An error is thrown by aten.view op called by DTensor dispatch, because
the local shard size is incompatible with the (incorrectly calculated)
args to the view op.
`RuntimeError: shape '[384]' is invalid for input of size 512`
After:
We raise more specific errors for cases of incompatible view operations
during sharding propagation, before getting to runtime dispatch.
`RuntimeError: Attempted to flatten an unevenly sharded dimension, which would require resharding the input. Please explicitly redistribute the tensor instead.`
Change Summary:
add 'strict_view' kwarg to the helper methods that implement
view/reshape op shard prop rules, so it can be decided op-by-op whether
to raise these new errors
enabled errors just for the 'view' op in this PR
added two specific checks/errors that can occur during view ops.
Details:
- View ops are never allowed to flatten a dimension that is unevenly
sharded, since that would likely change the size/content of the
local_tensor and require redistribute
- View ops are also never allowed to flatten two dims if the rightmost
dim is a Shard() placment, becuase it would cause contiguity errors
without redistribution
Notes:
- Disables support for several ops in test_dtensor_ops.py test, which
decompose to an illegal view that only works by performing a
redistribution: cartesian_prod, flatten, ravel, reshape, reshape_as, view, view_as, take_along_dim, kron
Follow Ups:
- triage other view-like ops (besides aten::view) for using strict_view
- look for other gaps where view-like ops could still perform
redistribution (ban them all, and document this)
Fixes#143372
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149764
Approved by: https://github.com/wanchaol, https://github.com/XilunWu
ghstack dependencies: #152045
Using `logging.basicConfig` to set root logger's level is not a good behavior. Fix common_distributed.py to set level for current logger only, because it affects downstream's 3rd-party testing plugins.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152319
Approved by: https://github.com/Skylion007
this was discussed with @eellison and he recommended using statically_known_true here, the intuition is. We already have 0/1 specializations in place, if we reach those checks with dynamic shapes that are not already specialized
then we do not want them to specialize them, "a recompilation here is not justified".
Those are all non-semantic changing optimizations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148893
Approved by: https://github.com/eellison
Summary:
# Context:
When memory leak happens, it usually trigger the OOM in the later iterations. The snapshot of full iteration will be huge and hard to interpret.
On CUDA side, they provide OOM observer which generates snapshot when OOM happens with latest 1,500,000 entries for debugging.
In this diff, we want to implement the feature on MTIA side
Test Plan:
Run this test with last diff in the stack.
```
buck run @//mode/opt kineto/libkineto/fb/mtia/integration_tests:mtia_memory_auto_trace_test
```
As shown, the memory_snapshot is generated when oom happens
Log: P1794792326
Snapshot: https://fburl.com/pytorch_memory_visualizer/lx73y6s3 {F1977402355}
Differential Revision: D71993315
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152160
Approved by: https://github.com/sraikund16
Fixes#151522
This PR fixes the issue that Dynamo fails to trigger a graph break for sparse tensors in certain code paths. I added an additional check to handle this case, and it resolves the original problem.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151897
Approved by: https://github.com/jansel
Fix: #135099
This PR changes how we map the original inputs into the new set of
inputs that take in the tensor input's base instead of their aliases.
**Problem:** in order to create this mapping, we had a dictionary that
mapped the hashed arguments into their respective indices. However, if
there's a group of equal arguments, we will have only one mapping for
such an argument. This breaks the assumption that there will be one
mapping for each argument.
**Solution:** map the hashed arguments into a list of indices. Then, we
will be able to correctly reconstruct the parameters for the new calling
convention.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146275
Approved by: https://github.com/bdhirsh
Although torch.cuda.Event and torch.xpu.Event have cuda_event and sycl_event fields respectively, the event_id exposed from the base class torch.Event is always 0, which can confuse users.
The memory of torch.Event is not useful to torch.cuda.Event and torch.xpu.Event, but we still need to inherit from torch.Event because CPython will check it.
Repro with cuda:
```
>>> import torch
>>> event = torch.cuda.Event()
>>> event.cuda_event
0
>>> event.event_id
0
>>> event.record()
>>> event.cuda_event
127982096
>>> event.event_id
0
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151226
Approved by: https://github.com/albanD, https://github.com/guangyey
ghstack dependencies: #151404, #151221, #151411
MemPool is a separate pool of memory handled by the caching allocator. This PR adds the option let the caching allocator try to use this pool as a last resort instead of OOMing by associating a use_on_oom bool with each MemPool.
Usage:
Users can optionally specify a ``use_on_oom`` bool (which is False by default) during MemPool creation. If true, then the CUDACachingAllocator will be able to use memory in this pool as a last resort instead of OOMing.
```
pool = torch.cuda.MemPool(allocator, use_on_oom=True)
with torch.cuda.use_mem_pool(pool):
a = torch.randn(40 * 1024 * 1024, dtype=torch.uint8, device="cuda")
del a
# at the memory limit, this will succeed by using pool's memory in order to avoid the oom
b = torch.randn(40 * 1024 * 1024, dtype=torch.uint8, device="cuda")
```
Testing:
```
python test/test_cuda.py -k test_mempool_limited_memory_with_allocator
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151487
Approved by: https://github.com/eqy, https://github.com/syed-ahmed, https://github.com/ngimel
Without MKL there is only 1 epilogue, not 2 because `addmm` is used instead of `packed_linear/_mkl_linear`.
This fails first at `TestSelectAlgorithmCPU.test_linear_with_in_out_buffer_batch_size_8_in_features_3_in_features2_192_image_size_224_out_features_64_bias_True_cpu_float32`
Instead of skipping the whole test just adjust the count for the single check.
Final numbers of `test/inductor/test_cpu_select_algorithm.py` without MKL:
```
Ran 1337 tests
OK (skipped=1211)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151548
Approved by: https://github.com/jansel
Solves the following problems of caffe2 HIP tests building on Windows:
1. HIP tests now use `hip_add_executable` to be built with custom_command invoking hip compiler, due to lack of cmake support for HIP in 3.18 (currently used).
2. failing with "Command line too long" which resulted from `hip_add_executable` adding the same flags over and over on top of `HIP_HIPCC_FLAGS` with every test added.
3. Disables `HasSameArgTypes` test on Windows, as `at::native::modern::detail` is nowhere to be found in the codebase (I think it must be a legacy thing). Perhaps the whole test should be removed/rewritten?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152014
Approved by: https://github.com/jeffdaily
Extensions can still rely on it, and we should decorate it with deprecated, but it is a C++20 feature.
XPU still uses it, so exclude XPU builds until https://github.com/intel/torch-xpu-ops/pull/1615 is merged
Test plan:
- 0def9b4acc should fail MPS builds
```
/Users/ec2-user/runner/_work/pytorch/pytorch/aten/src/ATen/native/mps/OperationUtils.mm:975:44: error: no template named 'optional' in namespace 'c10'; did you mean 'std::optional'?
c10::optional<int64_t> extra) {
^~~~~~~~~~~~~
std::optional
```
- a769759dd4 should fail CUDA builds
```
/var/lib/jenkins/workspace/torch/csrc/distributed/c10d/CUDASymmetricMemoryOps.cu(530): error: namespace "c10" has no member "nullopt"
input, c10::nullopt, reduce_op, group_name, out);
^
1 error detected in the compilation of
```
Fixes https://github.com/pytorch/pytorch/issues/150313
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150464
Approved by: https://github.com/atalman
Extensions can still rely on it, and we should decorate it with deprecated, but it is a C++20 feature
Test plan:
- 0def9b4acc should fail MPS builds
```
/Users/ec2-user/runner/_work/pytorch/pytorch/aten/src/ATen/native/mps/OperationUtils.mm:975:44: error: no template named 'optional' in namespace 'c10'; did you mean 'std::optional'?
c10::optional<int64_t> extra) {
^~~~~~~~~~~~~
std::optional
```
- a769759dd4 should fail CUDA builds
```
/var/lib/jenkins/workspace/torch/csrc/distributed/c10d/CUDASymmetricMemoryOps.cu(530): error: namespace "c10" has no member "nullopt"
input, c10::nullopt, reduce_op, group_name, out);
^
1 error detected in the compilation of
```
Fixes https://github.com/pytorch/pytorch/issues/150313
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150464
Approved by: https://github.com/atalman
Summary: I'm investigating differences in total torch.compile overhead in our two main internal sources: dynamo_compile and pt2_compile_events. One source of discrepancy is due to cudagraphs overheads. Currently, we have a context manager that optionally attributes a dynamo_timed region to a cudagraph-related column logged to dynamo_compile, but _all_ dynamo_timed regions show up in pt2_compile_events (hence the discrepancy; pt2_compile_events is overcounting). We could filter out these specific events from pt2_compile_events when measuring overall overhead. But I'm going to argue that those timed regions that we DO NOT consider as a compiler-related overhead don't have much value in logging in the first place. So I'm suggesting we just remove those instances.
Here's the production job with the discrepancy:
* dynamo_compile: https://fburl.com/scuba/dynamo_compile/3604eypl
* pt2_compile_events: https://fburl.com/scuba/pt2_compile_events/c2dv8sty
Test Plan:
torchbench nanogpt:
* tlparse: https://fburl.com/h1n2ascc
* dynamo_compile: https://fburl.com/scuba/dynamo_compile/sandbox/u37yrynp
* pt2_compile_events: https://fburl.com/scuba/pt2_compile_events/s7avd0di
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152136
Approved by: https://github.com/BoyuanFeng
Summary:
Various fixes to make fbcode work w/ ArmPL's cblas header:
1) Avoid re-declaring prototypes for internal blas methods which ArmPL already declares.
2) Fix `std::complex` conversion when using these methods.
3) Drop `extern "C"` around include fo `cblas.h`.
Test Plan: CI
Differential Revision: D72808561
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151126
Approved by: https://github.com/Skylion007
By instantiating it implicitly, otherwise attempts to run something like
```
% python3 -c "import torch; print(torch.special.entr(torch.testing.make_tensor(10, dtype=torch.bool, device='mps')))"
```
will fail with
```
Failed to created pipeline state object, error: Error Domain=AGXMetalG14X Code=3 "Compiler encountered an internal error"
```
Similar in spirit to https://github.com/pytorch/pytorch/pull/149123
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152204
Approved by: https://github.com/dcci
This is a proof-of-concept of how we could serialize a guard and deserialize it back from the bytes.
The main behavioral change introduced in this diff is on CheckFunctionManager:
```
check_fn_manager = CheckFunctionManager(code, output_graph, guards_serialization_mode="save")
guards_state: bytes = check_fn_manager.guards_state
```
Once `guards_serialization_mode` is set to `save`, CheckFunctionManager will return an addtional `bytes` object called `guards_state` which should contain all the information needed for deserializing guards later.
When we load back guards state, we will set `guards_serialization_mode` is set to `load`:
```
output_graph_state = pickle.loads(guards_state)
check_fn_manager = CheckFunctionManager(code, output_graph_state, guards_serialization_mode="load")
```
# TENSOR_MATCH
Since we have many types of guards to support, we will break the work into small diffs instead of a single diff to support every guards.
We kick off the work from TENSOR_MATCH from this diff.
# Testing
For each type of guard we will test it like the following:
1. Use guard_filter_fn to select 1 type of guard each time.
2. Call InstructionTranslator directly on an example function to get OutputGraph and CheckFunctionManager (reference guard manager)
3. Serialize->deserialize the output graph state and re-build the guards with a new CheckFunctionManager (loaded guard manager)
4. Throw a set of example inputs to both reference and loaded guard manager to see if their behavior match.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151318
Approved by: https://github.com/jansel, https://github.com/anijain2305
When e.g. OpenBLAS is used instead of MKL the differences get to large:
> Greatest absolute difference: 5.91278076171875e-05 at index (7,) (up to 1e-05 allowed)
> Greatest relative difference: 3.468156592134619e-06 at index (7,) (up to 1.3e-06 allowed)
I traced some of the matmul operations and there are differences of around 8e-6 between MKL and OpenBLAS but I haven't found where exactly the backward pass is calculated which is where the actual differences arise. So I couldn't check if there is some difference in the low-level BLAS function used by the autograd.
However it seems odd that there is a difference at all: For the MKL case it seems to be zero up to the accuracy shown by Python.
So it seems the AOT compilation has some differences when MKL is not available.
Maybe this is also the reason why it fails for ARM and hence the test is skipped there. Maybe @zou3519 knows more as he introduced those skip markers in https://github.com/pytorch/pytorch/pull/85565
Is there any documentation how and where `matmul_backward(_out)` is generated and how AOT transforms it with and without MKL?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152106
Approved by: https://github.com/zou3519
**Summary**
Add a new op, `onednn.qmul.tensor`, for int8 elementwise mul, which accepts inputs on CPU device (instead of QuantizedCPU).
The new op is implemented by AVX512 instructions and it provides similar or better performance, depending on shape, than its counterpart for QuantizedCPU device `quantized.mul`.
The new op supports output dtypes other than uint8 (fp32, fp16 and bf16 are supported).
**Test plan**
```
pytest test/quantization/core/test_quantized_op.py -k test_int8_mul_onednn
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151112
Approved by: https://github.com/leslie-fang-intel, https://github.com/jerryzh168
# Motivation
We propose adding support for the Python with statement on `torch.accelerator.device_index` to enable device switching functionality. This enhancement would simplify writing device-agnostic code and provide benefits across all accelerators. Its device-specific counterparts include [`torch.cuda.device`](00199acdb8/torch/cuda/__init__.py (L482)) and [`torch.cuda._DeviceGuard`](00199acdb8/torch/cuda/__init__.py (L469)).
**Design Philosophy**
It accepts either an `Int` or `None` as input. When `None` is passed, no device switch is performed. Supporting `None` is important for compatibility, as it's possible to encounter `None` values from `torch.device.index`.
Therefore, with this PR, we can do like this
```python
src = 0
dst = 1
# Set src to current device
torch.accelerator.set_device_index(src)
with torch.accelerator.device_index(dst):
# Inside with statement, we set dst to current device
assert torch.accelerator.get_device_index() == dst
# Here the current device should be src
assert torch.accelerator.get_device_index() == src
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148864
Approved by: https://github.com/albanD
Adding `torch.ops.fbgemm` to GraphPickler's allowlist. Otherwise, the fx graph module containing `fbgemm` node will return "Unable to pickle non-standard op" error.
The validation is done on the model and the difference appears only on the graph name not the node.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152079
Approved by: https://github.com/aorenste
**Summary**
For int8 GEMM Template, the micro GEMM will calculate in u8s8s32 and we will do the scale/zp compensation in the epilogue. In general, it will be calculated as:
```
temp = micro_gemm_output * x_scale * w_scale
temp = temp - (x_scale * w_scale * x_zp) * sum(w, 0)
```
For case when `x_scale, w_scale, x_zp` are constant, we can pre-calculate the compensation to save runtime calculation.
**Performance**
Test with 4 cores of XEON-5 and shapes from VIT model
Before
```
GEMM(M=197,N=768,K=768) compile: 0.0939 ms (2.48 TOPS, 18.13 GB/s)
GEMM(M=197,N=3072,K=768) compile: 0.4275 ms (2.17 TOPS, 13.90 GB/s)
GEMM(M=197,N=768,K=3072) compile: 0.2677 ms (3.47 TOPS, 22.20 GB/s)
GEMM(M=1,N=1000,K=768) compile: 0.0148 ms (0.10 TOPS, 99.10 GB/s)
```
After
```
GEMM(M=197,N=768,K=768) compile: 0.0597 ms (3.90 TOPS, 28.53 GB/s)
GEMM(M=197,N=3072,K=768) compile: 0.2126 ms (4.37 TOPS, 27.95 GB/s)
GEMM(M=197,N=768,K=3072) compile: 0.2282 ms (4.07 TOPS, 26.04 GB/s)
GEMM(M=1,N=1000,K=768) compile: 0.0149 ms (0.10 TOPS, 98.71 GB/s)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152000
Approved by: https://github.com/Xia-Weiwen, https://github.com/CaoE, https://github.com/jansel
This is a follow-up PR of the reverted one https://github.com/pytorch/pytorch/pull/147019 :
Modified TorchInductor’s autotuning flow so that each best_config JSON file also includes the Triton “base32” (or base64) cache key.
Motivation
Debugging & Analysis: With this change, we can quickly identify which compiled binary and IRs belongs to a given best config.
The impact is minimal since it is only an extra field in .best_config. It can help advanced performance tuning or kernel-level debugging.
Also, since Triton already stores cubin/hsaco in its cache, developers/researchers can avoid to set store_cubin = True since they can get the cubin/hsaco in the Triton cache and with the code provided in this PR, they can easily match the best_config with the right Triton cache directory for the "best" kernel.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148981
Approved by: https://github.com/davidberard98
In https://github.com/pytorch/pytorch/issues/151746, users ran into an error where a custom triton op cannot be resolved into an operator from string target. We improve the error message by reminding users to register the same custom operator at de-serialization time.
Now the error looks like this:
```python
torch._export.serde.serialize.SerializeError: We failed to resolve torch.ops.triton_kernel.add.default to an operator. If it's a custom op/custom triton op, this is usally because the custom op is not registered when deserializing. Please import the custom op to register it before deserializing. Otherwise, please file an issue on github. Unsupported target type for node Node(target='torch.ops.triton_kernel.add.default', inputs=[NamedArgument(name='x', arg=Argument(as_tensor=TensorArgument(name='linear')), kind=1), NamedArgument(name='y', arg=Argument(as_tensor=TensorArgument(name='mul')), kind=1)], outputs=[Argument(as_tensor=TensorArgument(name='add'))], metadata={'stack_trace': 'File "/data/users/yidi/pytorch/test.py", line 50, in forward\n output = triton_add(dense_output, bias)', 'nn_module_stack': 'L__self__,,__main__.SimpleModel', 'torch_fn': 'add.default_1;OpOverload.add.default'}, is_hop_single_tensor_return=None): <class 'str'>.```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152029
Approved by: https://github.com/jingsh
spot checked builds for line like `Found CUSPARSELT: /usr/local/cuda/lib64/libcusparseLt.so`. I don't know if there's another way to do it
I am slowly trying to reduce the duplicated code in docker image installs
Pros:
* less dup code
Cons:
* more docker copies
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150600
Approved by: https://github.com/atalman
This is a proof-of-concept of how we could serialize a guard and deserialize it back from the bytes.
The main behavioral change introduced in this diff is on CheckFunctionManager:
```
check_fn_manager = CheckFunctionManager(code, output_graph, guards_serialization_mode="save")
guards_state: bytes = check_fn_manager.guards_state
```
Once `guards_serialization_mode` is set to `save`, CheckFunctionManager will return an addtional `bytes` object called `guards_state` which should contain all the information needed for deserializing guards later.
When we load back guards state, we will set `guards_serialization_mode` is set to `load`:
```
output_graph_state = pickle.loads(guards_state)
check_fn_manager = CheckFunctionManager(code, output_graph_state, guards_serialization_mode="load")
```
# TENSOR_MATCH
Since we have many types of guards to support, we will break the work into small diffs instead of a single diff to support every guards.
We kick off the work from TENSOR_MATCH from this diff.
# Testing
For each type of guard we will test it like the following:
1. Use guard_filter_fn to select 1 type of guard each time.
2. Call InstructionTranslator directly on an example function to get OutputGraph and CheckFunctionManager (reference guard manager)
3. Serialize->deserialize the output graph state and re-build the guards with a new CheckFunctionManager (loaded guard manager)
4. Throw a set of example inputs to both reference and loaded guard manager to see if their behavior match.
Differential Revision: [D72987485](https://our.internmc.facebook.com/intern/diff/D72987485/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151318
Approved by: https://github.com/jansel, https://github.com/anijain2305
Summary: In the "too big to optimize" error message, tell the user that they should use the torch._inductor.config.aot_inductor.compile_wrapper_opt_level = 'O0' flag
Test Plan:
This is not added to unit test cases because it runs for a little longer time before the expected failure
```
def test_runtime_checks_error_msg(self):
with torch.library._scoped_library("mylib", "FRAGMENT") as lib:
torch.library.define(
"mylib::foo",
"(Tensor a, Tensor b) -> Tensor",
tags=torch.Tag.pt2_compliant_tag,
lib=lib,
)
torch.library.impl("mylib::foo", "cpu", lib=lib)
def foo(a: torch.Tensor, b: torch.Tensor) -> torch.Tensor:
return a + b
torch.library.impl_abstract("mylib::foo", lib=lib)
def foo_fake_impl(a, b):
return a + b
class Model(torch.nn.Module):
def __init__(self) -> None:
super().__init__()
def forward(self, x):
for i in range(10000):
x = torch.ops.mylib.foo(x, x)
return x
inputs = (torch.ones(8, 8, 8), )
model = Model()
with self.assertRaisesRegex(Exception, "torch._inductor.config.aot_inductor.compile_wrapper_opt_level"):
with torch.no_grad():
AOTIRunnerUtil.compile(
model,
inputs,
)
```
Differential Revision: D72323380
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151855
Approved by: https://github.com/desertfire
Followup work on top https://github.com/pytorch/pytorch/pull/149480
Wrapper on top of nvrtc inspired by https://gist.github.com/malfet/2c9a25976dd7396430c38af603f791da from @malfet
Compiling toy kernels with this setup takes 0.01s vs 90s using `load_inline()` on my local H100. This was primarily motivated by the timeouts I was seeing in the popcorn leaderboard but would also be useful to integrate into KernelBench
This PR is in the same spirit as https://github.com/pytorch/pytorch/pull/148972 which was a similar UX for Metal
For now we are planning on landing this as a private function because we expect to iterate both on the user facing API and the internals implementation, will open up a seperate issue to discuss the path towards making this work public and give a broader overview of the state of custom cuda kernel authoring in PyTorch
Future work, as a prereq to making the work public
* divup primitive
* support multiple kernels
* Expose _get_nvrtc_version from native code
* interop with torch.compile
* AMD support
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151484
Approved by: https://github.com/malfet
As reported in https://github.com/pytorch/pytorch/issues/149292, according to manual, `vfmsq_f32` implements `c - a * b` rather than `a * b - c`, so it's call must be prefixed with `vnegq_f32`
Also, adjust the tests to use OpMath for FMA computation to avoid accuracy error accumulation due to non-fused multiply-and-add over lower precision dtypes
Note that `Vectorized::fmsub` is not currently instantiated anywhere, so it could safely remain broken
TODO:
- Enable C++ testing on MacOS and/or aarch64 platforms (right now Mac tests are build without C++ tests)
Fixes https://github.com/pytorch/pytorch/issues/149292
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152075
Approved by: https://github.com/swolchok
ghstack dependencies: #151955
Implements layernorm forward pass as a metal kernel instead of MPSGraph ops. Speed ups are indicated on the chart below:
Script for generating times, need to build torch with old/new codebase and then run this with different file name indicated at the end of the script
```python
import csv
import time
import numpy as np
import torch
import torch.nn.functional as F
matrix_sizes = [32, 64, 128, 256, 512, 1024, 2048, 4096, 8192]
batch_sizes = [1]
elementwise_affine = [False, True]
num_runs = 50
warmup_runs = 3
def create_input_tensor(n, batch_size):
torch.manual_seed(42)
return torch.randn(batch_size, n, dtype=torch.float32)
def run_layer_norm(A, normalized_shape, elementwise_affine):
torch.mps.synchronize()
start = time.perf_counter()
out = F.layer_norm(A, normalized_shape)
torch.mps.synchronize()
end = time.perf_counter()
return out, end - start
results = {"N": [], "elementwise_affine": [], "batch_size": [], "mean_time": [], "std_time": []}
for el_aff in elementwise_affine:
for n in matrix_sizes:
for batch_size in batch_sizes:
print(f"\nBenchmarking LayerNorm for input size N={n}, batch_size={batch_size}, elementwise_affine={el_aff}")
try:
A_cpu = create_input_tensor(n, batch_size)
A_mps = A_cpu.to("mps")
normalized_shape = (n,)
for _ in range(warmup_runs):
_, _ = run_layer_norm(A_mps, normalized_shape, el_aff)
times = []
for _ in range(num_runs):
_, t = run_layer_norm(A_mps, normalized_shape, el_aff)
times.append(t)
mean_time = np.mean(times)
std_time = np.std(times)
results["N"].append(n)
results["elementwise_affine"].append(el_aff)
results["batch_size"].append(batch_size)
results["mean_time"].append(mean_time)
results["std_time"].append(std_time)
print(f"Mean time: {mean_time:.4f}s ± {std_time:.4f}s")
except RuntimeError as e:
print(f"Error for N={n}, batch_size={batch_size}: {e}")
continue
with open("layernorm_benchmark_times_new.csv", "w", newline="") as f:
writer = csv.writer(f)
writer.writerow(["N", "elementwise_affine", "batch_size", "mean_time", "std_time"])
for i in range(len(results["N"])):
writer.writerow(
[
results["N"][i],
results["elementwise_affine"][i],
results["batch_size"][i],
results["mean_time"][i],
results["std_time"][i],
]
)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152010
Approved by: https://github.com/malfet
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Summary: The TORCH_LIBRARY_THREAD_UNSAFE_LAZY_INIT feature is thread unsafe for calling the initializers, but we want to allow the deferred initializer call to be safe from multiple threads. Add a mutex to ensure we have thread safe construction of the libraries post launch.
Differential Revision: D73457714
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151938
Approved by: https://github.com/swolchok, https://github.com/zou3519
Summary:
We reverted D72896450 due to a weird error happens at a seemingly unrelated test "buck2 run apf/data/tests:preproc_state_serializer_test -- --filter-text "test_load_artifact"
"
I did some investigation and found that moving import AOTConfig and create_joint inside the create_fw_bw_grap causes a delay of importing the recursively imported modules in AOTConfig create_joint from test construction time to the test running time. The path.exists mock gets called multiple times due to the inspect.getsource calls in multiple places of torch.
Specifically, we set a breakpoint at the sideeffect of mocked os.path.exists. P1787425831 shows the importing stack trace before the change. P1787431638 shows the importing stacktrace after the change.
The notable difference is that in the second pastry, we trigger an os.path.exists when somewhere in triton we called inspect.getsourcelines when we construct OnDiskPreprocStateSerializer, which gets recorded by the mock.
Looking at the test, it seems what the test actualy wants to test is the deserialize step. So we reset_mock before the step to avoid mocking things happened at import time.
Test Plan:
buck2 run apf/data/tests:preproc_state_serializer_test -- --filter-text "test_load_artifact"
and existing tests for map.
Differential Revision: D73138415
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151479
Approved by: https://github.com/angelayi, https://github.com/zou3519
This PR reapplies #151490 and #151753 together, and adds some missing checks when applying the fast path.
Previously missed checks:
1) indexing path has the stride in the indexed dimension in bytes, gather path has the stride in the indexed dimension in elements. When checking if fast path is applicable, I didn't take this difference into account, and still multiplied the indexing stride by element size. Fixed and test added
2) We want to take fast path only when we are copying contiguous equally spaced slices of inputs + all the necessary alignment requirements. The effective tensor size should be 2d (after all possible flattening is applied), the index stride in the last dimension should be 0, and, since in the kernel we are not applying non-indexing-related offsets to src tensor, the src tensor stride in the second dimension should be 0. This automatically happens for gather with dim=0, so I didn't put in an explicit condition for this. Sometimes all conditions except first dim "effective" stride equal to 0 are satisfied for scatter on non-zero dim, when index size in the indexing dimension is 1 and thus it is collapsed (dimensions of size 1 are always collapsed), e.g.
```
# test gather along 1st dim that can accidentally trigger fast path
# because due to index dimension in the gather dim being 1
# an unexpected squashing in tensorIterator happens
src = make_tensor((16, 2, 16), device=device, dtype=dtype)
ind = torch.randint(2, (16, 1), device=device).view(16, 1, 1).expand(16, 1, 16)
res = torch.gather(src, dim=1, index=ind)
if res.device.type == "cuda":
ref_cpu = torch.gather(src.cpu(), dim=1, index=ind.cpu())
self.assertEqual(res.cpu(), ref_cpu, atol=0, rtol=0)
```
Note that if index size here was (16, 2, 16) instead of (16, 1, 16) then the middle dimension could not be collapsed and we wouldn't end up incorrectly taking fast path.
We could update the kernel to take this stride into account when computing offsets into src tensor, or we could specifically disallow non-zero stride on the first dimension. I took the second path for now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151917
Approved by: https://github.com/eqy, https://github.com/malfet, https://github.com/Skylion007
Summary:
This PR introduces additional autotuning configurations for the persistent+TMA version of Triton `mm` and `addmm` operations. The new configurations are as follows:
* `(128, 128, 64, 5, 8)`
* `(256, 128, 64, 4, 8)`
* `(128, 128, 64, 5, 4)`
These configurations were selected based on exhaustive autotuning performed on commonly used shapes from an internal foundational model.
While these new configs are generally more performant across the board, we see notable gains a few specific cases:
* In scenarios where `n >> m, k`, the configurations `(128, 128, 64, 5, 8)` and `(256, 128, 64, 4, 8)` tend to produce an additional 5-10% speedup over the aten baseline compared to the original configurations.
* Similarly, the configuration `(128, 128, 64, 5, 4)` yields approximately an 8% improvement in scenarios where k >> m, n.
These enhancements are expected to provide performance benefits across diverse use cases, particularly when compared to the original set of configurations.
Test Plan:
contbuild & OSS CI
Reviewers: paulzhan
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150587
Approved by: https://github.com/PaulZhang12, https://github.com/drisspg, https://github.com/eellison
Summary: Previously, when attr is defined, "if attr" will try to evaluate the data of attr, which is not intendended and we get a ugly error stack if the attr is not evaluable (like a fake tensor) before the callable(attr) check.
Test Plan: Existing tests.
Reviewed By: yushangdi, henryoier
Differential Revision: D73460905
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151944
Approved by: https://github.com/yushangdi
Summary:
I can confirm that `torch.jit.Error.mro()` contains `Exception` in the inheritance hierarchy.
This avoids a bunch of `pyre-ignore`s in D73352417.
Test Plan: Sandcastle
Differential Revision: D73464544
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151947
Approved by: https://github.com/Skylion007
Implement traceable config patching for Dynamo: enables restricted patching of Dynamo config where user can use a context manager/decorator to change tracing behavior for parts of the code.
The new `dont_skip_tracing` decorator/context manager for ignoring most trace rules is easily implemented with this more generic traceable config patching feature.
Implementation:
- Create a new specialized context manager class representing a wrapper around torch._dynamo.config.patch
- Dynamo doesn't trace into the context manager but updates config at compile time
- Correctness is based on our correctness for handling supported context managers
- Implementation is inspired by how `GradModeVariable` is implemented.
Previous attempts: https://github.com/pytorch/pytorch/pull/148736 (decorator-only global approach) and https://github.com/pytorch/pytorch/pull/149439 (decorator-only traceback approach)
See https://docs.google.com/document/d/1vWNwKL_jpg-PLopifcaSa338wks3GqSVF4GHRguybGg/edit?tab=t.0 for more details on implementation - including previous approaches.
NOTE: this PR fixes a bug where skipped code objects were not tracked by convert_frame.py, leading to cases where code objects would be automatically skipped even after `torch._dynamo.reset()`. This exposed some latent dynamo-wrapped test failures in CI that previously passed in CI but not locally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150586
Approved by: https://github.com/jansel, https://github.com/zou3519, https://github.com/anijain2305
As requested by Shuai. I also included an additional refactor to capture
changes in the whitelist over time since previously the first time it
was set, it was impossible override when a new config was set.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151766
Approved by: https://github.com/pianpwk
**Summary**
It's part of the task to enable max-autotune with GEMM template for WoQ INT4 GEMM on CPU.
This PR adds AMX-based GEMM templates for `torch.ops.aten_weight_int4pack_mm_for_cpu`. It brings performance benefits on platforms where AMX is available.
**Validation results**
We have run GPT-J-6B and Llama-3-8B-Instruct on a 6th gen Xeon with 96 cores. Results show that the AMX-based microkernel outperforms AVX512-based one by >5x for prefill stage with 1024 input length.
**Test plan**
```
python test/inductor/test_cpu_select_algorithm.py -k test_int4_woq_mm_amx
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150603
Approved by: https://github.com/leslie-fang-intel, https://github.com/jansel
Fix https://github.com/pytorch/pytorch/issues/151589
It's interesting that the Q4_K dequantization example in the referred GH issue does not crash even if Inductor pass triton the wrong alignment information. I dig this a bit. The main reason is, there are 2 things in triton that decides the vectorization size
1. alignement
2. max number of contiguous elements a thread need to process
Here is the triton code that decides vectorization size [link](c5fed8e1ca/third_party/nvidia/lib/TritonNVIDIAGPUToLLVM/LoadStoreOpToLLVM.cpp (L147-L157)), and here is the triton code that considers contiguity for vectorization [link](c5fed8e1ca/lib/Analysis/AxisInfo.cpp (L1250-L1269))
When Inductor wrongly tell triton that a unaligned tensor is aligned, Triton may not do vectorization (or not do full vectorization) because of the second restriction.
Check this test:
```
@parametrize(
"size",
(
128,
1024,
1024 * 1024,
),
)
def test_slice_view_dtype(self, size):
offset = 1
def f(x):
return x[2:].view(dtype=torch.float32) + 1
x = torch.randn((size + offset) * 2, dtype=torch.bfloat16, device=self.device)
self.common(f, (x,), reference_in_float=False)
```
Before the fix, Inductor would tell Triton that the output of aten.view.dtype tensor is aligned even though it's not. That tensor will be passed to the triton kernel for the aten.add. Triton may do different vectorization decision depending on the tensor size
1. when size = 128, triton pick ld.global.b32 to load data from global memory
2. when size = 1024, triton uses ld.global.v2.b32
4. when size = 1024 * 1024, triton uses ld.global.v4.b32
So whether wrong alignment metadata causes issue depends on if triton picks the vectorized instructions. The latter depends on the triton config (block size) decided by inductor and triton internal logic (how they assign elements to each thread). We'd better to make sure Inductor always generate correct metadata to make sure such hidden issues does not turn into crash later.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151859
Approved by: https://github.com/jansel, https://github.com/eellison
ghstack dependencies: #151841
**Summary**
Fixes [#151290](https://github.com/pytorch/pytorch/issues/151290) and [#151523](https://github.com/pytorch/pytorch/issues/151523), which are regressions introduced by [#144020](https://github.com/pytorch/pytorch/pull/144020). That PR enabled parallelization at the inner loop level.
However, a currently unsupported case arises when parallel reduction occurs under the vectorization loop level, specifically in patterns like:
```
for vec_loop_level:
do_parallel_reduction
```
In such cases, a temporary buffer `tmp_acc_array` is allocated for tail scalar kernels, and another temporary buffer `tmp_acc_array` is also defined for parallel reduction. This results in a conflict due to overlapping temporary buffers. This PR disables the problematic case to avoid the conflict until proper support is implemented.
**Test Plan**
```
python test/inductor/test_flex_attention.py -k test_make_block_mask_cpu
python test/inductor/test_cpu_repro.py -k test_parallel_reduction_vectorization
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151887
Approved by: https://github.com/jansel
Summary:
I had a minor annoyance when debugging graphs using EXIR dialect ops,
that all the function normalization went away. For functions with > 5 arguments,
some of which are just simple bools and ints, it's very helpful to have
the kwarg names attached.
Enhance `normalize_target` to handle EdgeOpOverload targets. To avoid
a circular dependency on Executorch from pytorch core, I just use a `hasattr`
check for "_op". This only happens if the target is not already a recognized
torch function.
Also, I noticed that the new `fx.Node.normalized_arguments` function
didn't forward an important kwarg to `normalize_target`, so I fixed that too.
Test Plan: Tested with FxGraphDrawer and an fx Graph containing EXIR nodes.
Differential Revision: D67545909
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143689
Approved by: https://github.com/angelayi
Filtering out the stacktrace so that the stacktrace on nodes when using fx.Tracer looks nicer. I just copied the filtering we have in [proxy_tensor.py](6720d23969/torch/fx/experimental/proxy_tensor.py (L1903-L1931)).
Previously the stacktrace looked like:
```
File "/data/users/angelayi/pytorch/moo.py", line 3964, in <module>
run_tests()
File "/data/users/angelayi/pytorch/torch/testing/_internal/common_utils.py", line 1342, in run_tests
unittest.main(argv=argv)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/main.py", line 101, in __init__
self.runTests()
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/main.py", line 271, in runTests
self.result = testRunner.run(self.test)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/runner.py", line 184, in run
test(result)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/suite.py", line 84, in __call__
return self.run(*args, **kwds)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/suite.py", line 122, in run
test(result)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/suite.py", line 84, in __call__
return self.run(*args, **kwds)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/suite.py", line 122, in run
test(result)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/case.py", line 650, in __call__
return self.run(*args, **kwds)
File "/data/users/angelayi/pytorch/torch/testing/_internal/common_utils.py", line 3324, in run
self._run_custom(
File "/data/users/angelayi/pytorch/torch/testing/_internal/common_utils.py", line 3296, in _run_custom
super_run(result=result)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/case.py", line 591, in run
self._callTestMethod(testMethod)
File "/home/angelayi/.conda/envs/pytorch-3.10/lib/python3.10/unittest/case.py", line 549, in _callTestMethod
method()
File "/data/users/angelayi/pytorch/torch/testing/_internal/common_utils.py", line 3156, in wrapper
method(*args, **kwargs)
File "/data/users/angelayi/pytorch/moo.py", line 1495, in test_stack_trace
gm = torch.fx.GraphModule(m, tracer.trace(m))
File "/data/users/angelayi/pytorch/torch/fx/_symbolic_trace.py", line 837, in trace
(self.create_arg(fn(*args)),),
File "/data/users/angelayi/pytorch/moo.py", line 1485, in forward
x = x * 2
File "/data/users/angelayi/pytorch/torch/fx/proxy.py", line 716, in impl
return tracer.create_proxy("call_function", target, args, kwargs)
File "/data/users/angelayi/pytorch/torch/fx/proxy.py", line 248, in create_proxy
proxy.node.stack_trace = "".join(CapturedTraceback.extract().format())
```
Now it looks like:
```
File "/data/users/angelayi/pytorch/moo.py", line 1485, in forward
x = x * 2
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151029
Approved by: https://github.com/jfix71, https://github.com/zou3519, https://github.com/jingsh
With `AdditionalInputs`, the behavior is the same as with tensors:
```python
class M(torch.nn.Module):
def forward(self, x, y):
return x + y
additional_inputs = torch.export.AdditionalInputs()
additional_inputs.add((5, 5))
additional_inputs.add((3, 5))
additional_inputs.add((5, 4))
ep = torch.export.export(
M(), (6, 7), dynamic_shapes=additional_inputs, strict=False
)
```
With `ShapesCollection`, we now need to wrap integer inputs as `_IntWrapper` so that we can have a unique identifier for each integer input.
```python
class M(torch.nn.Module):
def forward(self, x, y):
return x + y
from torch.export.dynamic_shapes import _IntWrapper
args = (_IntWrapper(5), _IntWrapper(5))
# Or we can do `args = pytree.tree_map_only(int, lambda a: _IntWrapper(a), orig_args)`
shapes_collection = torch.export.ShapesCollection()
shapes_collection[args[0]] = Dim.DYNAMIC
shapes_collection[args[1]] = Dim.DYNAMIC
ep = torch.export.export(
M(), args, dynamic_shapes=shapes_collection, strict=False
)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151842
Approved by: https://github.com/pianpwk
By reusing `c10/metal/atomic.h`
This also fixes `GPUTests.test_index_put_fallback[12]_mps` that is unrolled by inductor, so no need for dedicated atomic_add support
TODOs:
- Get rid of indexing kernel and compute it directly when kernel is run
- Simulate atomic_add for int64 types as series of int32 atomic-add-and-fetch
- Setup tolerances correctly to pass float16/bfloat16 tests (as CPU always takes sequential strategy)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151869
Approved by: https://github.com/Skylion007, https://github.com/dcci
In this approach, we are catching any lane within a wave that is doing fastatomics to the same destination address and computing the sum on the CU. This is leading to 3x improvement in scatter_add performance and 2x improvement in index_select.
scatter_add performance on MI300x:
dtype|Baseline (before optimizations)|opportunistic fastatomics
-------|----------------------------------|----------------------------------
f32|1.389425039|0.430447996
fp16|2.195472956|0.779729486
bf16|2.194051027|0.784599513
Using the following reproducer
```
import torch
import triton
def main():
dtype = torch.float32
dim = 1305301
a = torch.rand(100, device="cuda", dtype=dtype)
index = torch.randint(0, 100, (dim,), device="cuda")
src = torch.rand(dim, device="cuda", dtype=dtype)
print("=" * 20)
print(
triton.testing.do_bench(
lambda: a.scatter_add(0, index, src),
return_mode="median",
)
)
print("=" * 20)
if __name__ == "__main__":
main()
```
co-authored by: @amd-hhashemi
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146264
Approved by: https://github.com/jeffdaily, https://github.com/mxz297
Co-authored-by: Hashem Hashemi <hashem.hashemi@amd.com>
Should be the last part of https://github.com/pytorch/pytorch/pull/150558, except for maybe s390x stuff, which I'm still not sure what's going on there
For binary builds, do the thing like we do in CI where we tag each image with a hash of the .ci/docker folder to ensure a docker image built from that commit gets used. Previously it would use imagename:arch-main, which could be a version of the image based on an older commit
After this, changing a docker image and then tagging with ciflow/binaries on the same PR should use the new docker images
Release and main builds should still pull from docker io
Cons:
* if someone rebuilds the image from main or a PR where the hash is the same (ex folder is unchanged, but retrigger docker build for some reason), the release would use that image instead of one built on the release branch
* spin wait for docker build to finish
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151706
Approved by: https://github.com/atalman
standalone_compile needs to get dynamic shape information from
somewhere. We add a new `dynamic_shapes` argument with three options:
1. from the passed-in graph (dynamic="from_graph"). This is the default.
2. from the example inputs, thereby specializing on them. (dynamic="from_example_inputs")
3. from the current tracing context (dynamic="from_tracing_context")
1 and 3 are not exactly the same. 2 can also be used for more advanced
things... (specialize on one input but not the other).
Most of this PR is tests.
Test Plan:
- a lot of new tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151788
Approved by: https://github.com/oulgen
fixes https://github.com/pytorch/pytorch/issues/151055. Thanks @desertfire for the patch that fixed this.
I was a bit careful about the test - I wanted to make sure the test accurately ensures that we don't regress and our error message is not spammy when users enter an invalid `TORCH_LOGS=....` argument. But I tried to avoid using expecttests, since people occasionally add new logging artifacts and I didn't want to add to much churn by forcing this to fail CI.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151678
Approved by: https://github.com/desertfire, https://github.com/zou3519
Fixes#150367
This PR makes decomposition table from onnx registry, which includes registered ops not only ATen and prim. This will help to keep the custom ops that are specified in the custom_translation table from decomposition during ONNX export.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151826
Approved by: https://github.com/justinchuby
Summary: tree_flatten_with_map will internally call unflatten function with user supplied function. But this function was not returning anything causing the leaves to be None. This is wrong when the constructor is sensitive to this behaviour
Test Plan: CI
Differential Revision: D73388529
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151824
Approved by: https://github.com/bdhirsh
Summary: Also check the module's named buffers and parameters when resolving name collision
Test Plan:
```
buck2 run mode/dev-nosan caffe2/test/inductor:test_aot_inductor -- -r aoti_constant_tensor_name_collision
```
Differential Revision: D73264885
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151684
Approved by: https://github.com/angelayi
Triton codegen currently [sorts vars by divisor](ae6f6b8efb/torch/_inductor/codegen/simd.py (L233-L237)). When there are two vars with the same divisor, the order is undecided.
```python
nodes.sort(
key=lambda x: V.graph.sizevars.size_hint(
x.divisor, fallback=config.unbacked_symint_fallback
)
)
```
The test case leads to the following nodes:
```
(Pdb) nodes[0]
IterationRangesEntry(x1, ((s37 + 127)//128), 2, (xindex//ps0), {x0: ((s37 + 127)//128), x1: 2, x2: ((s12 + 127)//128), x4: 2*(((s12 + 127)//128))*(((s37 + 127)//128)), x5: 0, x6: 2, x7: (((s12 + 127)//128))*(((s37 + 127)//128))})
(Pdb) nodes[1]
IterationRangesEntry(x0, 1, ((s37 + 127)//128), ModularIndexing(xindex, 1, ps0), {x0: ((s37 + 127)//128), x1: 2, x2: ((s12 + 127)//128), x4: 2*(((s12 + 127)//128))*(((s37 + 127)//128)), x5: 0, x6: 2, x7: (((s12 + 127)//128))*(((s37 + 127)//128))})
(Pdb) nodes[2]
IterationRangesEntry(x2, 2*(((s37 + 127)//128)), ((s12 + 127)//128), (xindex//(2*(((s37 + 127)//128)))), {x0: ((s37 + 127)//128), x1: 2, x2: ((s12 + 127)//128), x4: 2*(((s12 + 127)//128))*(((s37 + 127)//128)), x5: 0, x6: 2, x7: (((s12 + 127)//128))*(((s37 + 127)//128))})
(Pdb) V.graph.sizevars.statically_known_equals(nodes[0].length, 2)
True
(Pdb) V.graph.sizevars.statically_known_equals(nodes[1].length, 1)
True
(Pdb) V.graph.sizevars.statically_known_equals(nodes[2].length, 1)
True
(Pdb) V.graph.sizevars.statically_known_equals(nodes[0].divisor, 1)
True
(Pdb) V.graph.sizevars.statically_known_equals(nodes[1].divisor, 1)
True
(Pdb) V.graph.sizevars.statically_known_equals(nodes[2].divisor, 2)
True
```
Since x1 and x0 both have divisor 1, the relative order is random across runs.
In some runs, we have order [x1, x0, x2] with divisors as [1,1,2] and lengths as [2,1,1]. After x1, we have [divisor = divisor * node.length](ae6f6b8efb/torch/_inductor/codegen/simd.py (L246)) = 1 * 2 = 2. Then, when processing x0, we have node.divisor=1, divisor=2, and [FloorDiv(node.divisor, divisor)](ae6f6b8efb/torch/_inductor/codegen/simd.py (L251)) = 0, which indicates an iteration length of 0 and leads errors later.
The fix is to sort by both divisor and length_is_one. So for two nodes with the same divisor, we process the node with length=1 first.
Fixes#149789
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151634
Approved by: https://github.com/Skylion007, https://github.com/drisspg
Summary: There are a few issues I'm solving:.
1. It's too hard to measure total pt2 overhead using the dynamo_compile table because users need to know the columns representing all the top-level events (dynamo_cumulative_compile_time_us, etc.). Instead, let's populate the existing duration_us field for all top-level events. The complication is that runtime events in particular (Triton autotuning, cudagraphify) can be collapsed into a single row, with gaps in between, so we can't simply use `end_time - start_time` in all cases. Instead, we'll sum durations for all outer events when updating the compile-time or runtime metrics context. Introduce a 'depth' counter in TLS to track the nesting of CompilationMetrics events.
2. The existing implementation relies on callers of dynamo_timed to specify whether the event is a runtime or compile-time event. That doesn't work because some methods can be called in both situations, e.g., `CachingAutotuner.benchmark_all_configs`. For example `TORCHINDUCTOR_BENCHMARK_FUSION=1` enables benchmarking during compile-time. Instead, we can figure out automatically whether we're measuring a compile-time or runtime event and log accordingling.
3. If `log_compilation_events` were to throw an exception, we'd fail to clear the aggregated counters for runtime logs and they could be attributed to the wrong compile ID. I didn't actually find evidence of this in practice, but I added exception handling for extra safety.
Test Plan:
Ran internal models and compared dynamo_compile to pt2_compile_events:
`TORCHINDUCTOR_BENCHMARK_FUSION=0`
* tlparse: https://fburl.com/itciwnxc
* dynamo_compile: https://fburl.com/scuba/dynamo_compile/yvkif5vb
* pt2_compile_events: https://fburl.com/scuba/pt2_compile_events/segijet7
`TORCHINDUCTOR_BENCHMARK_FUSION=1`
* tlparse: https://fburl.com/jgurcvkw
* dynamo_compile: https://fburl.com/scuba/dynamo_compile/uum91ceb
* pt2_compile_events: https://fburl.com/scuba/pt2_compile_events/x4xnisez
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151749
Approved by: https://github.com/Skylion007
# Motivation
This PR aims to deprecate the host allocator legacy API and recommend users to use the unified API `getHostAllocator(device_type)` APIs, such as:
```cpp
at::getHostAllocator(device_type)->allocate(...);
at::getHostAllocator(device_type)->empty_cache();
at::getHostAllocator(device_type)->record_event(...);
at::getHostAllocator(device_type)->get_stats();
at::getHostAllocator(device_type)->reset_accumulated_stats();
at::getHostAllocator(device_type)->reset_peak_stats();
```
# Additional Context
TODO:
- [ ] Move is_pinned from `AcceleratorHookInterface` to `HostAllocator`
- [ ] Deprecate `getPinnedMemoryAllocator` inside `AcceleratorHookInterface` and recommend using `getHostAllocator` instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151437
Approved by: https://github.com/EikanWang, https://github.com/albanD
ghstack dependencies: #151403, #151431
This somewhat complicated PR does a few things:
- It separates out a lot of the guard checking logic into its own class, GuardedCache[T]
- It adds a new `check_guard_hit` lambda to FXGraphCache._lookup_graph, which allows callers to define their own guard checking logic
- It then uses these two combined parts to lift guard checking to AOTAutogradCache. This means that AOTAutogradCache stores its own guard expressions and evaluates them.
- FXGraphCache's guard checking logic is completely unchanged, just refactored. As part of the work, I'm able to extend a bit of the logging functionality of AOTAutogradCache into FXGraphCache, so that you can know if FXGraphCache missed due to a guard failure or a full cache miss.
# Why do this?
Lifting guards to AOTAutogradCache has a few benefits:
- First, it fixes a long standing bug in guard checking logic. Backward passes can have different symint inputs than forward passes depending on forward output, if AOTAutograd chooses to store symints for the backward. These symint inputs have the same underlying symbols as the forward, but on AOTAutogradCache hit, we don't have access to the hints backing these exact symints (we only have hints for the symints on the forward function). By lifting guard checking logic to AOTAutogradCache, we no longer need to check the backward guards, as they'll be included in the AOTAutogradCache guard expression. **I've added a unit test that failed before my diff, and now passes, as an example of this**
- Secondly, this is the first step necessary to bundle CompiledFxGraph into AOTAutogradCache. Doing so will simplify our cache logic significantly, and also make precompile logic simpler, as precompiles will only need to store AOTAutogradCacheEntrys, without needing to match them up with inductor FXGraphCache entries.
- Finally, adding guard checking logic to AOTAutogradCache my allow us in the future to handle more complicated cases like a single forward with multiple backwards, as guard checks are now storable on the cache entry itself.
# Guard checking logic of AOTAutogradCache
When AOTAutogradCache evaluates guard expressions, it no longer needs to evaluate the forward/backward guards in the FXGraphCacheEntry (since the AOTAutogradCache guard expressions will encompass them). Because of this, we still need a way for AOTAutogradCache to distinguish between multiple FXGraphCache local entries. To do so, AOTAutogradCache stores the guard string from FXGraphCache, which it uses as a second "cache key". It doesn't need to **evaluate** these guards, it just needs to find the cache entry from FXGraphCache that had the same guards as when it was stored.
After this, I will work on putting the FXGraphCache entries directly into AOTAutogradCache. If I can put CompiledFxGraphs in the cache directly, I no longer need this complicated `check_guard_hit` overriding logic.
## Test Plan
Added a new unit test. There are comprehensive guard checking unit tests in `test_aot_autograd_cache` already, and those pass.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151563
Approved by: https://github.com/oulgen
Summary:
This PR introduces additional autotuning configurations for the persistent+TMA version of Triton `mm` and `addmm` operations. The new configurations are as follows:
* `(128, 128, 64, 5, 8)`
* `(256, 128, 64, 4, 8)`
* `(128, 128, 64, 5, 4)`
These configurations were selected based on exhaustive autotuning performed on commonly used shapes from an internal foundational model.
While these new configs are generally more performant across the board, we see notable gains a few specific cases:
* In scenarios where `n >> m, k`, the configurations `(128, 128, 64, 5, 8)` and `(256, 128, 64, 4, 8)` tend to produce an additional 5-10% speedup over the aten baseline compared to the original configurations.
* Similarly, the configuration `(128, 128, 64, 5, 4)` yields approximately an 8% improvement in scenarios where k >> m, n.
These enhancements are expected to provide performance benefits across diverse use cases, particularly when compared to the original set of configurations.
Test Plan:
contbuild & OSS CI
Reviewers: paulzhan
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150587
Approved by: https://github.com/PaulZhang12, https://github.com/drisspg, https://github.com/eellison
Chatting with Bob the goal of this is to const fold the floats that where tensorified by calling
guard_scalar(val) on them and then replacing their usages by their values.
Hence we do not need to do this for nodes with no float symbols.
We do not want todo proper const folding because we need to preserve statements that deferred
runtime asserts depend on. (see the added test)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151494
Approved by: https://github.com/bobrenjc93
When bicubic interpolation was added to grid_sampler in #44780, `GridSampleFuncOptions` was not updated to allow a user to use bicubic mode in LibTorch, even though the function could handle it. This PR fixes the parity such that LibTorch's `torch::nn::functional::grid_sample` behaves the same as PyTorch's `torch.nn.functional.grid_sample`.
Existing users can directly use `torch::grid_sampler` but must know what int to pass for the interpolation (2 for bicubic) and padding mode parameters, which is not ideal.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150817
Approved by: https://github.com/Skylion007
Summary:
Similar to what we did previously in D70033166
Previously, for cpu we decompose addmm if
```
check_device(mat1, mat2, device="cpu")
and statically_known_true(mat1.shape[0] == 1)
and statically_known_true(mat2.shape[0] <= 64)
and statically_known_true(mat2.shape[1] <= 512)
```
We have a new case where `mat1.shape[0] = 80`, and benchmark shows that it will beneficial if we decompose, so update the condition to
```
check_device(mat1, mat2, device="cpu")
and statically_known_true(mat1.shape[0] == 1)
and statically_known_true(mat2.shape[0] <= 128)
and statically_known_true(mat2.shape[1] <= 512)
```
Differential Revision: D73292985
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151730
Approved by: https://github.com/kflu, https://github.com/houseroad
Fixes#135998
Adds support for fp8. These are compared bitwise, without atol and rtol. The implementation uses the same comparison functions, just with atol and rtol forced to zero. The error message is different from the default case; it only tells the user the first mismatch. This is to avoid triggering the error from #135998.
Test Plan:
New unit test covers new code paths.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150002
Approved by: https://github.com/cyyever, https://github.com/zou3519
By constructing tensor on that device, because it does not call `self.common` but rather executes test directly.
Otherwise `test_add_complex3_mps` will test CPU inductor, rather than MPS one
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151732
Approved by: https://github.com/dcci
Although torch.cuda.Event and torch.xpu.Event have cuda_event and sycl_event fields respectively, the event_id exposed from the base class torch.Event is always 0, which can confuse users.
The memory of torch.Event is not useful to torch.cuda.Event and torch.xpu.Event, but we still need to inherit from torch.Event because CPython will check it.
Repro with cuda:
```
>>> import torch
>>> event = torch.cuda.Event()
>>> event.cuda_event
0
>>> event.event_id
0
>>> event.record()
>>> event.cuda_event
127982096
>>> event.event_id
0
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151226
Approved by: https://github.com/albanD
This PR adds support for submatrices in offline tuning for:
- GEMM
- GEMM and bias
- ScaledGEMM
- Batch Strided GEMM
New UTs to cover submatrices. Submatrices for strided batch API is not part of this PR and will be done seperately.
There is also a bug fix for offline tuning for full matrix for GEMM and bias in the `NT` case. Offline and online UTs were updated to cover this corner case.
To improve code readability, swapped definition of transA and transB.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151138
Approved by: https://github.com/jeffdaily
By constructing tensor on that device, because it does not call `self.common` but rather executes test directly.
Otherwise `test_add_complex3_mps` will test CPU inductor, rather than MPS one
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151732
Approved by: https://github.com/dcci
By building wheel with USE_DISTRIBUTED=1
Otherwise attempt to run
```
python3 benchmarks/dynamo/torchbench.py --performance --only hf_T5 --backend inductor --inference --devices mps
```
wil fail with
```
File "/Users/nshulga/Library/Python/3.10/lib/python/site-packages/transformers/modeling_utils.py", line 40, in <module>
import torch.distributed.tensor
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/tensor/__init__.py", line 4, in <module>
import torch.distributed.tensor._ops # force import all built-in dtensor ops
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/tensor/_ops/__init__.py", line 2, in <module>
from ._conv_ops import * # noqa: F403
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/tensor/_ops/_conv_ops.py", line 5, in <module>
from torch.distributed.tensor._dtensor_spec import DTensorSpec, TensorMeta
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/tensor/_dtensor_spec.py", line 6, in <module>
from torch.distributed.tensor.placement_types import (
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/tensor/placement_types.py", line 8, in <module>
import torch.distributed._functional_collectives as funcol
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/_functional_collectives.py", line 9, in <module>
import torch.distributed.distributed_c10d as c10d
File "/Users/nshulga/git/pytorch/pytorch/torch/distributed/distributed_c10d.py", line 23, in <module>
from torch._C._distributed_c10d import (
ModuleNotFoundError: No module named 'torch._C._distributed_c10d'; 'torch._C' is not a package
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151721
Approved by: https://github.com/wdvr, https://github.com/dcci, https://github.com/huydhn
Summary: Further testing the script, we found that we shouldn't always assume rank 0 is the first rank, so we need to check all entries and see if it P2P op for this coalesced group.
Test Plan: Directly test with corner case.
Differential Revision: D73266257
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151683
Approved by: https://github.com/fegin
The name was updated by https://github.com/pytorch/pytorch/pull/151155. The benchmark results weren't updated on the dashboard otherwise.
For PT2 compiler perf benchmark, we are still relying on this old workflow. To get rid of this, we need to update PT2 benchmark dashboard to use the new benchmark database (cc @yangw-dev)
The results are there on the new database:
```
SELECT
*
FROM
oss_ci_benchmark_v3
WHERE
workflow_id = 14510035576
```
but not on the old database:
```
SELECT
*
FROM
inductor_torch_dynamo_perf_stats
WHERE
workflow_id = 14510035576
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151698
Approved by: https://github.com/seemethere, https://github.com/atalman
Summary: See internal Diff for more details.
In ExternKernel, the FakeTensors do not have associated real tensors, because they are just created from ir.Node's shape and stride.
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r aoti_data_dependent_ex
buck2 run mode/dev-nosan fbcode//caffe2/test/inductor:aot_inductor_arrayref_cpu -- -r data_dependent_extern_kernel_op
```
Differential Revision: D73002775
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151377
Approved by: https://github.com/angelayi
Summary: It looks symmetric memory only supports cuda12.3+. We do have the definition w/ 12.3- but we don't have implementation. So maybe a good idea to even disable the definition.
Test Plan: CI
Reviewed By: jianyuh, houseroad, ngimel, jiawenliu64
Differential Revision: D72936993
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151203
Approved by: https://github.com/ngimel, https://github.com/houseroad
This PR fixes two things.
The first problem is that in the vLLM style standalone_compile is
called from within a custom torch.compile backend. If there already is a
FakeTensorMode (which there is), we shouldn't create a new
FakeTensorMode with the same shape_env, instead we should just reuse the
same FakeTensorMode.
The second thing is that compile_fx can mutate the passed in gm, so we
deepcopy (since standalone_compile should be standalone)
Test Plan:
- new test
- updated old tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151502
Approved by: https://github.com/oulgen
ghstack dependencies: #151501, #151551
We were only returning the first one. There's an edge case on what to do
if the original function returns a single Tensor. capture(f) returns a
function that returns a tuple of one Tensor in this case and we were
originally converting this back to one single Tensor. I think it's fine
to return a tuple of one Tensor (that is what the graph passed to
standalone_compile asked for!) but we can revisit.
fine
Test Plan:
- modified one test to used multiple outputs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151551
Approved by: https://github.com/Skylion007, https://github.com/oulgen
ghstack dependencies: #151501
To follow pattern set by CPU and CUDA impls: define common_dtype and optionally casts `elements` and `test_elements` to common dtype if needed
- Add regression test, though skip it on MacOS-13, as `isin` seems to produce garbage there even for same dtypes
```
>>> import torch
>>> x=torch.arange(4.0, device='mps')
>>> y=torch.arange(1.0, 3.0, device='mps')
>>> x, y, torch.isin(x, y), torch.isin(y, x)
(tensor([0., 1., 2., 3.], device='mps:0'), tensor([1., 2.], device='mps:0'), tensor([False, True, False, False], device='mps:0'), tensor([False, False], device='mps:0'))
>>> torch.__version__
'2.6.0'
```
- Cleanup code a bit
Fixes https://github.com/pytorch/pytorch/issues/151443
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151600
Approved by: https://github.com/Skylion007, https://github.com/dcci, https://github.com/kulinseth
If pip is not installed:
### Before
```console
> python3 torch/utils/collect_env.py
Collecting environment information...
Traceback (most recent call last):
File "/Users/Adam/pytorch/torch/utils/collect_env.py", line 694, in <module>
main()
~~~~^^
File "/Users/Adam/pytorch/torch/utils/collect_env.py", line 677, in main
output = get_pretty_env_info()
File "/Users/Adam/pytorch/torch/utils/collect_env.py", line 672, in get_pretty_env_info
return pretty_str(get_env_info())
~~~~~~~~~~~~^^
File "/Users/Adam/pytorch/torch/utils/collect_env.py", line 497, in get_env_info
pip_version, pip_list_output = get_pip_packages(run_lambda)
~~~~~~~~~~~~~~~~^^^^^^^^^^^^
File "/Users/Adam/pytorch/torch/utils/collect_env.py", line 450, in get_pip_packages
for line in out.splitlines()
^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'splitlines'
```
### After
```console
> python3 torch/utils/collect_env.py
Collecting environment information...
PyTorch version: N/A
Is debug build: N/A
CUDA used to build PyTorch: N/A
ROCM used to build PyTorch: N/A
OS: macOS 15.4 (arm64)
GCC version: Could not collect
Clang version: 20.1.0
CMake version: version 3.31.6
Libc version: N/A
Python version: 3.13.2 (main, Apr 8 2025, 15:27:33) [Clang 17.0.0 (clang-1700.0.13.3)] (64-bit runtime)
Python platform: macOS-15.4-arm64-arm-64bit-Mach-O
Is CUDA available: N/A
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: Could not collect
Nvidia driver version: Could not collect
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: N/A
CPU:
Apple M2 Pro
Versions of relevant libraries:
[pip3] Could not collect
[conda] Could not collect
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151607
Approved by: https://github.com/malfet
This case creates subprocess in a subprocess. In Windows it can't load function at this scenario hence I have to skip it
```
File "C:\ProgramData\miniforge3\envs\lfq\lib\multiprocessing\spawn.py", line 116, in spawn_main
exitcode = _main(fd, parent_sentinel)
File "C:\ProgramData\miniforge3\envs\lfq\lib\multiprocessing\spawn.py", line 126, in _main
self = reduction.pickle.load(from_parent)
AttributeError: Can't get attribute 'run_model' on <module '__main__' (built-in)>
Traceback (most recent call last):
File "<string>", line 25, in <module>
File "<string>", line 16, in test_multi_process
AssertionError
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150999
Approved by: https://github.com/guangyey, https://github.com/EikanWang
Co-authored-by: Yu, Guangye <106960996+guangyey@users.noreply.github.com>
So far it's only for `gather`, but we'll move index_select and index to this implementation too. Torchtitan and fbgemm have noticed that gather/index_select perf is bad, this PR brings core implementation to be on par with those customized implementations. Added benefits: all dtypes are supported, a bit less strict on the tensor dimensions/contiguity because we pick the fast path after TensorIterator collapsed the dimensions.
Biggest part of this PR is not even the kernel (it's dumb, just vectorized loads are enough), but moving utilities for vectorized loads and stores from SymmetricMemory to be generally accessible in MemoryAccess.cuh.
Additional tests are coming to make sure this implementation doesn't break anything
`gather` is equivalent to x[indices] for 1d indices via
```
def fn_gather(x, indices):
return torch.gather(x, dim=0, index=indices.unsqueeze(1).expand(-1, x.shape[1]))
def fn_index(x, indices):
return x[indices]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151490
Approved by: https://github.com/Skylion007, https://github.com/eqy
# Motivation
This PR introduces a unified parent class `HostAllocator` with the following goals:
1. Enable backend-specific host allocator registration, including support for out-of-tree backends.
2. Provide a unified and extensible API surface for host memory management across all backends, especially accelerators.
The new interface includes:
- `at::getHostAllocator()->allocate`
- `at::getHostAllocator()->empty_cache`
- `at::getHostAllocator()->record_event`
- `at::getHostAllocator()->get_stats`
- `at::getHostAllocator()->reset_accumulated_stats`
- `at::getHostAllocator()->reset_peak_stats`
# Additional Context
We plan to deprecate legacy APIs such as `at::cuda::CachingHostAllocator_emptyCache` and recommend users migrate to the new backend-specific API, for example:
```cpp
at::getHostAllocator(at::kCUDA)->empty_cache();
```
This refactor will help standardize host memory management across devices and simplify backend integration in the future.
Another key improvement I am going to do is move the `is_pinned` functionality into the `HostAllocator` class, which enables centralized pinned memory verification through calls like `at::getHostAllocator(at::kCUDA)->is_pinned(ptr)`.
Benefits include:
- Consistent host memory handling across all device backends
- Decouple pinned memory functionality with `AcceleratorHooksInterface` in a more modular way
- Clearer separation between device memory allocation and pinned host memory management
This architecture makes the system more maintainable and extensible for future device support.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151431
Approved by: https://github.com/albanD
ghstack dependencies: #151403
This adds a non-blocking mode to queue_pop. This allows for workers to poll if work is ready without blocking the main loop. This is useful for the case where you want to have a GPU have maximum utilization when something only periodically is sent on the queue.
We also expose a `torch.distributed.QueueEmptyError` so users can catch the error and handle it accordingly.
Test plan:
```
pytest test/distributed/test_store.py -k queue -v -s -x
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151485
Approved by: https://github.com/fduwjj, https://github.com/tianfengfrank
collective APIs accept either group or global rank for src/dst rank.
We provide a helper `_canonicalize_group_rank` which converts from maybe
group or maybe global to one particular format (defined by the kwarg
return_global: bool=False).
In this PR we stop performing the mapping lookup that converts group to
global or global to group in the case that the caller wants us to return
the same value that was passed in. The PR should be functionally
equivalent, except in cases where the mapping itself would raise an
exception but the mapping was not necessary in the first place.
This has come up in cases where people create new process groups outside
of 'init_process_group' APIs and group-specific ranks may not have a
valid mapping to the 'global' rank.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151373
Approved by: https://github.com/xunnanxu, https://github.com/d4l3k
Summary: add one option to allow skipping all reduce unused parameters, this could help improve training throughput significantly when the number of unused parameters is large in the model.
Test Plan: unit tests, CI
Differential Revision: D72282069
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151503
Approved by: https://github.com/mrshenli
This is part of splitting up https://github.com/pytorch/pytorch/pull/150558 into smaller chunks, please see that for more context
Use the binary docker build action from https://github.com/pytorch/pytorch/pull/151471
Change the workflow trigger to be all of .ci/docker so it will make a new image + tag whenever it changes.
build script:
* change to be independent of the CUDA_VERSION env var, since all the info should be in the imagename:tag
* remove docker push parts since that will happen during the workflow
* clean up a bit
* make the build script more like the CI build script (use a temp image name)
I don't think this image is actually used anywhere
Also push docker image to imagename:tag, I got rid of it in the PR making the reusable workflow since I thought it was not in the original scripts but it actually is there
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151483
Approved by: https://github.com/ZainRizvi
If a custom operator does not contain a fake impl, currently draft-export will use the real-tensor propagation to get an output for the operator and continue tracing. However if we retrace the exported model using `ep.run_decompositions`, or `export`, or run the exported program with fake tensors, we'll still fail because there's no fake impl.
With this PR, after draft-export we will generate an operator profile for each operator call that we encounter, and store this on the report attached to the exported program `ep._report.op_profiles`. Users can then use `torch._library.fake_profile.register_fake_profile` to temporarily generate and register a fake impl based on these operator profiles. This way future fake tensor retracing will work.
The workflow would look something like:
```python
class M(torch.nn.Module):
def forward(self, a, b):
res = torch.ops.mylib.foo8(a, b) # no fake impl
return res
ep = export(M(), (torch.ones(3, 4), torch.ones(3, 4)) # this fails bc no fake impl
ep = draft_export(M(), (torch.ones(3, 4), torch.ones(3, 4))
ep.run_decompositions() # this fails bc no fake impl
# this registers fake impls based on the profiles
with torch._library.fake_profile.register_fake_profile(ep._report.op_profiles):
decomp = ep.run_decompositions() # this works
new_inp = (
torch.ones(2, 3, 4),
torch.ones(2, 3, 4),
)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150809
Approved by: https://github.com/zou3519
This implements epilogue visitor tree argument generation (example type [here](3fe62887d8/include/cutlass/epilogue/fusion/sm90_callbacks_tma_warpspecialized.hpp (L332))).
Details:
The codegen task here is to implement a function which can generate a tree of C++ structs and properly extract the correct properties from Inductor buffers and write them to the correct locations in the generated struct. To implement this with the minimum amount of code, I generate the cutlass DAGIR (the EVT internal represenation) which specifically has a pass, [pass_argument_type.py ](5e497243f7/python/cutlass/backend/evt/passes/pass_argument_type.py (L4)) which generates a nested tree of custom argument types for each node in the DAGIR. This nested tree of constructors is then passed kwargs to fill in the proper values, where the node's name is used to differentiate between different values in the kwarg dictionary. This however is non-customizable; the nested tree of EVT args is a nested tree of ctypes which looks for *actual values* so that this object can be passed directly to the cutlass-python C++ runner. Inductor on the other hand needs to fill this struct with string C++ expressions representing the values (or extracting the values from kernel launcher args). So `_render_argument_type` implements this: it iterates over the tree of types created by pass_argument_type.py and generates a string representing the nested structs, filling in C++ expressions representing the different fields.
Long term plan:
Long term, I will ask the nvidia to provide an overridable [visitor_factory](5e497243f7/python/cutlass/backend/evt/passes/pass_argument_type.py (L82)) which could allow us to override the behavior of pass_argument_type.py to generate the string we would like during DAGIR generation.
Previously merged:
* #150346
* #150345
* #150344
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150903
Approved by: https://github.com/henrylhtsang, https://github.com/eellison
This is similar to how we handle protobufs and it makes it more convenient for bazel users to handle their version of flatbuffers. (Flatbuffers is very picky about the generated code matching the runtime). Instead of using the checked in generated code, we generate it on the fly.
This is relevant to https://github.com/pytorch/pytorch/issues/112903, because having the version of flatbuffers tied to pytorch will make pytorch difficult to use as an external workspace.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151364
Approved by: https://github.com/malfet
I have noticed that there are some errors like
```
UnicodeDecodeError: 'utf-8' codec can't decode byte 0x95 in position 169302: invalid start byte
```
I havent been able to repro this locally yet, this change should fix the encoding issues
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151472
Approved by: https://github.com/masnesral
Summary:
We add states in the constant folding process for AOTInductor.
Basically, there's 3 states, which is
(1) None: The state when no constants are loaded and uninitialized.
(2) Initialized: The state when constants are loaded, but not yet
folded.
(3) Folded: The state where the model is fully ready with folded
constants.
Note that even if constant folding is not enabled, we still only run
when state is FOLDED, this is okay because without constant folding, the
transition from INITIALIZED to FOLDED is just a pass-throught.
Test Plan:
python test/inductor/test_aot_inductor.py -k test_constant_folding_with_update
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D73002538](https://our.internmc.facebook.com/intern/diff/D73002538)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151273
Approved by: https://github.com/jingsh, https://github.com/desertfire
Now that we have bc_lint in CI, this script is no longer needed (nor has it ever been conclusive). I've already updated the Runbook to not need this script.
Suppressing bc_lint as this script is not shipped as a part of torch--it is not user facing! For context, this script is (rarely) used by the release notes manager to ensure BC across releases. It had been broken for at least since 2.6.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151453
Approved by: https://github.com/albanD, https://github.com/jbschlosser
Which is a regression, introduced by https://github.com/pytorch/pytorch/issues/150629#issue-2970312779 which I should have reviewed more thoroughly.
- Defined `_fused_rms_norm`, added MPS-only implementation for it and dispatch from `rms_norm_symint`, which is registered as `CompositeImplicitAutograd`, i.e. it is not supposed to do any computations over Tensor, only dispatch to other ops
-
- Register `_fused_rms_norm` as a fallback in `torch/_inductor/lowering.py`
- Added unit test to avoid those regressions in the future
TODO:
- Get rid of this op, change `rms_norm_symint` definition to `CompositeExplicitAutograd` and implement backward function in `tools/autograd/derivatives.yaml`
- Benchmark compiler and re-enable decomp as follows when compiled code is faster
```python
@register_decomposition(aten._rms_norm_fused)
def rms_norm_fused(
self: torch.Tensor, ndim: int, weight: torch.Tensor, eps: float
) -> torch.Tensor:
dtr = [self.dim() - i - 1 for i in range(ndim)]
return self * weight * (self.pow(2).mean(dtr, keepdim=True).add(eps).rsqrt())
```
Fixes https://github.com/pytorch/pytorch/issues/150629
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150661
Approved by: https://github.com/manuelcandales, https://github.com/jansel
As the title stated.
**Changes Before:**
```C++
[999/1526] Building CUDA object caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/int4mm.cu.o
/root/Git.d/pytorch/pytorch/aten/src/ATen/native/cuda/int4mm.cu(142): warning #177-D: variable "at::native::kWarpSize" was declared but never referenced
constexpr int32_t kWarpSize = 32;
^
Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151427
Approved by: https://github.com/Skylion007, https://github.com/malfet
Summary: In most instances, this action would take ~33% of the total run time, which means that our benchmark would previously differ from the end results by a lot.
Test Plan:
We can compare the benchmark results for
```
CUDA_VISIBLE_DEVICES=4,5 buck run mode/opt -c python.package_style=inplace -c fbcode.enable_gpu_sections=true -c fbcode.nvcc_arch=h100a //caffe2/torch/fb/model_transform/experimental/benchmark:mts_gpu_benchmark -- --model-snapshot-id=672308665_0 --lower-backend=AOT_INDUCTOR --node-replacement-dict="{'torch.nn.Linear':{'(autotune)': 'fp8_float_model_dynamic_quantization_rowwise'}}" --trace-aot-inductor-module=True --disable-acc-tracer=False --batch-size=1024
```
before and after the diff, and notice that on average, the benchmark results decrease by ~0.1ms per iteration, which is more closely aligned with the lowered modules.
Differential Revision: D72469845
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150696
Approved by: https://github.com/frank-wei
`has_triton()` returns True if Triton is present on the system and supports _any_ backend we care about. In this case, that means we _always_ check gradients, even though the intended behavior is to skip gradients when testing on CPU.
Fixes a bug from #146911.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151435
Approved by: https://github.com/masnesral
Which is a regression, introduced by https://github.com/pytorch/pytorch/issues/150629#issue-2970312779 which I should have reviewed more thoroughly.
- Defined `_fused_rms_norm`, added MPS-only implementation for it and dispatch from `rms_norm_symint`, which is registered as `CompositeImplicitAutograd`, i.e. it is not supposed to do any computations over Tensor, only dispatch to other ops
-
- Register `_fused_rms_norm` as a fallback in `torch/_inductor/lowering.py`
- Added unit test to avoid those regressions in the future
TODO:
- Get rid of this op, change `rms_norm_symint` definition to `CompositeExplicitAutograd` and implement backward function in `tools/autograd/derivatives.yaml`
- Benchmark compiler and re-enable decomp as follows when compiled code is faster
```python
@register_decomposition(aten._rms_norm_fused)
def rms_norm_fused(
self: torch.Tensor, ndim: int, weight: torch.Tensor, eps: float
) -> torch.Tensor:
dtr = [self.dim() - i - 1 for i in range(ndim)]
return self * weight * (self.pow(2).mean(dtr, keepdim=True).add(eps).rsqrt())
```
Fixes https://github.com/pytorch/pytorch/issues/150629
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150661
Approved by: https://github.com/manuelcandales, https://github.com/jansel
Summary:
When we have an exported program that looks like this:
```
ExportedProgram:
class GraphModule(torch.nn.Module):
def forward(self, b__tensor_constant0: "f32[1]", ... c_lifted_tensor_0: "i64[925]", …. , tupleized_input_0_0: "f32[10, 2139]",
clone: "i64[925]" = torch.ops.aten.clone.default(c_lifted_tensor_0); c_lifted_tensor_0 = None
index_select: "f32[10, 925]" = torch.ops.aten.index_select.default(tupleized_input_0_0, 1, clone); clone = None
```
The graph after `aot_export_module` could have a name collision, notice that `_tensor_constant0` arg of `clone` is different from the `_tensor_constant0` in the input module .
```
def forward(self):
arg9_1: "f32[10, 2139]"
_tensor_constant0: "f32[1]" = self._tensor_constant0 # this should be int64, conflicted with the original _tensor_constant0, had a clone on this constant before lifting
index: "f32[10, 925]" = torch.ops.aten.index.Tensor(arg9_1, [None, _tensor_constant0]); _tensor_constant0 = None
```
This caused the `tensors used as indices must binary, int...` aoti error on PT2I dashboard because later we used `clone` as index.
We had this error because we created a new `_tensor_constant0` at [here](https://github.com/pytorch/pytorch/blob/main/torch/fx/_symbolic_trace.py#L403-L412), and the new `_tensor_constant0` overrides the original `_tensor_constant0` on the input Module in `_unlift_graph`. The `arg` for `clone` is created at `create_proxy` in `proxy.py`.
To fix this, we do a graph pass before we unlift the graph inputs to avoid name collision
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:torchbind -- -r aot_compile_constant_folding
buck2 run mode/dev-nosan caffe2/test/inductor:test_aot_inductor -- -r aoti_constant_tensor_name_collision
```
Differential Revision: D72761937
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151123
Approved by: https://github.com/tugsbayasgalan, https://github.com/jingsh
# Motivation
This stack of PRs aims to generalize and improve PyTorch host allocator code.
This PR introduces a `DeleterFnPtr` template parameter to `CachingHostAllocatorInterface` to resolve circular dependency issues. This change allows for better code reuse and simplifies the implementation of host allocators.
# Additional Context
TODO:
- [ ] Unify host allocator related API
- [ ] Deprecate those device-specific legacy API
- [ ] Move `is_pinned` to host allocator
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151403
Approved by: https://github.com/gujinghui, https://github.com/albanD
This PR reduces #graph partitions by reordering nodes when the `should_partition` nodes have simple dependencies. Specifically, for `should_partition` nodes:
a. If a node has no dependency or only depends on graph inputs: move to the front. Use case is when we move symints to cuda tensor for PaddedTensorSubclass
b. If the only user of a node is OutputNode: move it to the end.
#### Example
The following example shows a padded tensor subclass use case where we copy symint to a cuda tensor (aka mask) in the middle of function. Reordering still generates 1 cudagraph by moving the mask to the front.
```python
import torch
torch._inductor.config.graph_partition = True
# Two reasons for this:
# 1. We want to reuse the same mask for many masked_fill calls
# 2. Prevent inductor from fusing this op into other ops (e.g. masked_fill)
# so we can still reorder in scheduler
@torch.library.custom_op("mylib::create_mask", mutates_args=(), tags=(torch._C.Tag.cudagraph_unsafe,))
def create_mask(padded_size: int, original_size: int, device: torch.device) -> torch.Tensor:
mask = torch.zeros((padded_size,), dtype=torch.bool, device=device)
mask[original_size:] = True
return mask
@create_mask.register_fake
def _(padded_size, original_size, device):
return torch.empty((padded_size,), dtype=torch.bool, device=device)
def f(padded_tensor, original_tensor, weight):
original_size = original_tensor.size()[0]
padded_size = padded_tensor.size()[0]
# element wise op so we don't care padding value
padded_tensor = padded_tensor + 1
padded_tensor = torch.nn.functional.relu(padded_tensor)
# dot product requires padding with 0
dot_res = padded_tensor.dot(weight)
padded_tensor += dot_res
# min requires padding with inf, so we create mask now
mask = create_mask(padded_size, original_size, padded_tensor.device)
min_res = torch.min(
torch.ops.aten.masked_fill(padded_tensor, mask, float("inf"))
)
# max requires padding with inf. we can reuse previous mask
max_res = torch.max(
torch.ops.aten.masked_fill(padded_tensor, mask, -float("inf"))
)
return min_res+max_res+padded_tensor
compiled_f = torch.compile(f, mode="reduce-overhead")
def run(padded_size, original_size):
padded_tensor = torch.randn(padded_size, device="cuda")
padded_tensor[original_size:] = 0
original_tensor = torch.randn(original_size, device="meta")
weight = torch.randn(padded_size, device="cuda")
eager_out = f(padded_tensor, original_tensor, weight)
compiled_out = compiled_f(padded_tensor, original_tensor, weight)
assert torch.allclose(eager_out[0], compiled_out[0])
assert torch.allclose(eager_out[1], compiled_out[1])
# new cudagraph
run(8, 4)
# new cudagraph due to recompile
run(8, 6)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150814
Approved by: https://github.com/eellison
# Feature
This fixes a bug related to block pointer stores. Since Triton's block pointer stores don't support implicit broadcasting, in certain cases we need to generate a `reshape->broadcast->reshape` pattern to ensure that the tensor being stored has the same shape as the block pointer. This happens when the block indexing expression involves strides of 0 or dimensions of 1, both of which we eliminate from the block pointer.
The existing logic missed an important edge case. We may need a broadcast prior to the first `reshape` of this pattern, in case the tensor comes from a load with implicit broadcasting. For example, if the range trees have shape `[YBLOCK, XBLOCK]`, but the load has a shape `[1, XBLOCK]`, we need to broadcast this to `[YBLOCK, XBLOCK]` prior to storing. See the example kernel below, which comes from `expand` -> `clone` with 3D tiling. The load has an implicit broadcast, and the store has a reshape. Thus, we need to insert an explicit broadcast between them.
```
@triton.jit
def triton_poi_fused_clone_0(in_ptr0, out_ptr0, znumel, ynumel, xnumel, ZBLOCK : tl.constexpr, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
znumel = 32
ynumel = 1
xnumel = 32
zoffset = tl.program_id(2) * ZBLOCK
zindex = zoffset + tl.arange(0, ZBLOCK)[:, None, None]
zmask = zindex < znumel
yoffset = tl.program_id(1) * YBLOCK
yindex = yoffset + tl.arange(0, YBLOCK)[None, :, None]
ymask = tl.full([ZBLOCK, YBLOCK, XBLOCK], True, tl.int1)
xoffset = tl.program_id(0) * XBLOCK
xindex = xoffset + tl.arange(0, XBLOCK)[None, None, :]
xmask = xindex < xnumel
x1 = xindex
z0 = zindex
tmp0 = tl.load(tl.make_block_ptr(in_ptr0, shape=[32], strides=[1], block_shape=[XBLOCK], order=[0], offsets=[xoffset]), boundary_check=[0], eviction_policy='evict_last')[None, None, :]
tl.store(tl.make_block_ptr(out_ptr0, shape=[32, 32], strides=[32, 1], block_shape=[ZBLOCK, XBLOCK], order=[1, 0], offsets=[zoffset, xoffset]), tl.reshape(tl.broadcast_to(tmp0, [ZBLOCK, YBLOCK, XBLOCK]), [ZBLOCK, XBLOCK]).to(tl.float32), boundary_check=[0, 1])
''', device_str='cuda')
```
The tricky part is that we don't want to emit redundant broadcasts in the store. This PR reworks the logic a bit to make sure we don't emit a second broadcast unless it actually changes the shape.
# Test plan
Added a CI test for this case, which would fail on trunk. Checked that only one broadcast was emitted.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151399
Approved by: https://github.com/jansel, https://github.com/eellison
Given an exception in torch.export, I want to try/catch it to add the message "hey try out draft-export!". Currently I only add this message for errors that draft-export is known to fix, like DataDependentErrors, ConstraintViolationErrors, and no fake impl.
Originally the error message looks like:
```
File "/data/users/angelayi/pytorch/torch/_library/custom_ops.py", line 626, in fake_impl
raise RuntimeError(
RuntimeError: There was no fake impl registered for <CustomOpDef(mylib::foo2)>. This is necessary for torch.compile/export/fx tracing to work. Please use `foo2_impl.register_fake` to add an fake impl.
```
Now, the error msg now looks something like:
```
File "/data/users/angelayi/pytorch/torch/_library/custom_ops.py", line 626, in fake_impl
raise RuntimeError(
RuntimeError: There was no fake impl registered for <CustomOpDef(mylib::foo2)>. This is necessary for torch.compile/export/fx tracing to work. Please use `foo2_impl.register_fake` to add an fake impl.
The error above occurred when calling torch.export.export. If you would like to view some more information about this error, and get a list of all other errors that may occur in your export call, you can rerun your program with the `DRAFT_EXPORT=1` envvar, or replace your `export()` call with `draft_export()`.
```
In python versions >= 3.11, we can use `exception.add_note` to add to the error message. However with previous versions I did a hack to modify `e.args`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151065
Approved by: https://github.com/pianpwk
ghstack dependencies: #151051
As the title stated.
**Changes:**
- Remove unnecessary header file
- Remove unnecessary registry logic about PrivateUse1HooksRegistry，such as TORCH_DECLARE_REGISTRY, C10_DEFINE_REGISTRY, etc,.
- using static + global variable to do initialization instead of call_one
**Next Step:**
Enable test_openreg.py in CI/CD to guard the quality of PrivateUse1
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151005
Approved by: https://github.com/albanD
Implement traceable config patching for Dynamo: enables restricted patching of Dynamo config where user can use a context manager/decorator to change tracing behavior for parts of the code.
The new `dont_skip_tracing` decorator/context manager for ignoring most trace rules is easily implemented with this more generic traceable config patching feature.
Implementation:
- Create a new specialized context manager class representing a wrapper around torch._dynamo.config.patch
- Dynamo doesn't trace into the context manager but updates config at compile time
- Correctness is based on our correctness for handling supported context managers
- Implementation is inspired by how `GradModeVariable` is implemented.
Previous attempts: https://github.com/pytorch/pytorch/pull/148736 (decorator-only global approach) and https://github.com/pytorch/pytorch/pull/149439 (decorator-only traceback approach)
See https://docs.google.com/document/d/1vWNwKL_jpg-PLopifcaSa338wks3GqSVF4GHRguybGg/edit?tab=t.0 for more details on implementation - including previous approaches.
NOTE: this PR fixes a bug where skipped code objects were not tracked by convert_frame.py, leading to cases where code objects would be automatically skipped even after `torch._dynamo.reset()`. This exposed some latent dynamo-wrapped test failures in CI that previously passed in CI but not locally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150586
Approved by: https://github.com/jansel, https://github.com/zou3519, https://github.com/anijain2305
Summary:
D72906445 seemed to cause a SIGABRT when running the test in the test plan. The change I narrowed it down to was where in fake_impls the [`deregister_fake_kernel` no longer calls `lib.destroy`](https://github.com/pytorch/pytorch/pull/150806/files#diff-7fd3f4222276c63b91f3a895530bb5efe137fd23165b48f25afcf3c06a5d2a8fL65-L69).
Calling `lib.destroy` in that handle results in a maximum recursion error where someone calls library.destroy which calls the handle which calls back to library.destroy.
So I compared the implementation of this _del_library and lib.destroy and it seemed like the main thing different was deleting `self.m`. So adding that fixed my issue!
Side note, I feel like we can combine `_del_library` and `library._destroy`? But I won't do it in this diff to make sure we don't break too many things 😅
Test Plan:
`buck test 'fbcode//mode/opt' fbcode//aiplatform/gmpp/bulk_eval/reader/service/tests:reader_service_handler_tests -- --exact 'aiplatform/gmpp/bulk_eval/reader/service/tests:reader_service_handler_tests - aiplatform.gmpp.bulk_eval.reader.service.tests.reader_service_handler_tests.ReaderServiceHandlerTests: test_add_preproc_output_into_queue'`
https://www.internalfb.com/intern/testinfra/testrun/10977524170296078
Differential Revision: D73017613
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151299
Approved by: https://github.com/zou3519
Currently, `torch._chunk_cat` only supports contiguous inputs (due to `.view()` usage in `_pad_chunk()` supporting only contiguous tensor). This doesn't work for internal models where there can be non-contiguous input tensors:
- size=[8192, 16416], stride=[16448, 1] # stride[0] is larger than size[1]
- size=[1152, 384], stride=[1, 1152] # column-major tensor
In this PR, we relax the assumption on contiguous input tensor, by switching from `.view()` to `.reshape()`. Note that since `.reshape()` will try to use `.view()` under the hood whenever possible, this should not cause regression to existing use cases.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151263
Approved by: https://github.com/BoyuanFeng
After https://github.com/pytorch/pytorch/pull/150652, we still see some ranks missing dumps. Upon looking further, the case is that FR dump timed out for its first attempt:
watchdog thread: notify FR dump -> wait for 1 mins -> throw watchdog timeout -> notify elastic to kill process
FR dump thread: received FR dump signal -> timeout after 1 mins with first attempt -> started 2nd attempt -> got killed.
So we want to make the FR dump timeout shorter, in reality, the log shows that the dump finished within one sec. Even if we consider a very slow speed like 200K/s the usual size FR (1MB at most) takes around 5 secs, so 15 secs is like 3 times buffer.
Also we still let watchdog sleep for 1 min so that we can wait enough time for two dump to timeout and the following check like GIL checker to execute.
Also, if we get stuck in getting GIL or cuda hang, 15 seconds should be enough to detect the hang.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151329
Approved by: https://github.com/fegin
Finishes up the work started in #121686 + adds test
Update: this was not as straightforward as I originally imagined. Context below.
**TL;DR:** `TestFoo{CPU, CUDA}` now actually derive from `TestFoo`! Also, `{CPU, CUDA}TestBase` setup / teardown logic is now always called (it is required to set the primary device), regardless of whether `super().setUpClass()` / `super().tearDownClass()` are called or not.
**Background:** The typical way to get device-specific tests is to write a generic `TestFoo` and call `instantiate_device_type_tests(TestFoo, locals())` to get `TestFooCPU`, `TestFooCUDA`, etc. After this, generic tests (e.g. `TestFoo.test_bar()`) become `TestFooCPU.test_bar_cpu()` / `TestFooCUDA.test_bar_cuda()`.
Behind the scenes, this was historically accomplished by creating a `TestFooCUDA` that derives from both a `CUDATestBase` and an *empty class* called `TestFoo_base`. This `TestFoo_base` has the same bases as `TestFoo`, but none of the test functions (e.g. `test_bar()`). The documented reason for this is to avoid things like a derived `TestFooCUDA.test_bar()` being discovered in addition to the real device-specific test `TestFooCUDA.test_bar_cuda()`.
(1) A reason this matters is because it should be possible to call e.g. `super().setUpClass()` from a custom setup / teardown classmethod. If the generated TestFooCUDA does not derive from TestFoo, but instead derives from the empty class described above, this syntax does not work; in fact there is no way to form a proper `super()` call that works across the device-specific test variants. Here's an example that breaks in the OpInfo tests:
070f389745/test/test_ops.py (L218-L221)
(2) Further, there is some precedent within a custom `setUpClass()` impl for storing things on the `cls` object to be accessed at test time. This must be the device-specific test class (`TestFooCUDA`) and not `TestFoo` for this to work. As an example, the open device registration tests load a module during setup and use it in the test logic:
070f389745/test/test_cpp_extensions_open_device_registration.py (L63-L77)070f389745/test/test_cpp_extensions_open_device_registration.py (L79-L80)
To accomplish both (1) and (2) at the same time, I decided to revisit the idea of utilizing a proper inheritance hierarchy for `TestFoo` -> `{TestFooCPU, TestFooCUDA}`. That is: have TestFooCPU / TestFooCUDA **actually** derive from `TestFoo`. This achieves both (1) and (2). The only thing left is to make sure the generic tests (e.g. `TestFoo.test_bar()`) are not discoverable, as was the stated reason for diverging from this in the first place. It turns out we can simply `delattr()` these generic tests from `TestFoo` once `TestFooCPU` / `TestFooCUDA` have been setup with the device-specific variants, and all works well. The `instantiate_device_type_tests(...)` logic already deletes `TestFoo` from scope, so I don't see a problem with deleting generic tests from this base class as well (CI will prove me right or wrong ofc).
**Side note:** I was encountering a weird race condition where sometimes the custom `setUpClass()` / `tearDownClass()` defined & swapped in [here](4a47dd9b3f/torch/testing/_internal/common_device_type.py (L940-L955)) would be used, and sometimes it wouldn't. This non-deterministic behavior was called out previously by @ngimel here:
4a47dd9b3f/test/inductor/test_torchinductor_dynamic_shapes.py (L128-L130)
To address this, I moved this block of logic to before the first call to `instantiate_test()`, as that method queries for the primary device, and the primary device identification logic may manually invoke `setUpClass()` (see [here](4a47dd9b3f/torch/testing/_internal/common_device_type.py (L381-L384))). Goal: define the `setUpClass()` / `tearDownClass()` we want for correctness before they're ever called. This seems to work and the behavior is deterministic now AFAICT.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151129
Approved by: https://github.com/janeyx99, https://github.com/masnesral, https://github.com/malfet
Introduced by https://github.com/pytorch/pytorch/pull/149512
Before this change, following warning was generated
```
/Users/nshulga/git/pytorch/pytorch/aten/src/ATen/native/transformers/attention.cpp:452:71: warning: extra ';' outside of a function is incompatible with C++98 [-Wc++98-compat-extra-semi]
452 | REGISTER_HPU_DISPATCH(_fused_sdp_choice_stub, &_fused_sdp_choice_meta);
| ^
1 warning generated.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151367
Approved by: https://github.com/drisspg
This PR adds standalone_compile API that does precompilation via caching to support vLLM use case in the short term while we work on the longer term precompilation solution.
```
standalone_compile(gm, example_inputs, options) -> CompiledArtifact
CompiledArtifact.save(path, format: binary|unpacked = binary)
CompiledArtifact.load(path, format: binary|unpacked = binary)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150670
Approved by: https://github.com/jamesjwu, https://github.com/zou3519
By suppressing `missing-template-arg-list-after-template-kw` warning, which seems to be required to compile Google's libnop, which is in a semi-abandoned state now
```
In file included from /Users/malfet/git/pytorch/pytorch/third_party/tensorpipe/third_party/libnop/include/nop/base/variant.h:21:
/Users/malfet/git/pytorch/pytorch/third_party/tensorpipe/third_party/libnop/include/nop/types/variant.h:241:30: error: a template argument list is expected after a name prefixed by the template keyword [-Wmissing-template-arg-list-after-template-kw]
241 | index_ = value_.template Construct(std::forward<Args>(args)...);
| ^
/Users/malfet/git/pytorch/pytorch/third_party/tensorpipe/third_party/libnop/include/nop/types/variant.h:258:26: error: a template argument list is expected after a name prefixed by the template keyword [-Wmissing-template-arg-list-after-template-kw]
258 | if (!value_.template Assign(TypeTag<T>{}, index_, std::forward<U>(value))) {
| ^
/Users/malfet/git/pytorch/pytorch/third_party/tensorpipe/third_party/libnop/include/nop/types/variant.h:265:26: error: a template argument list is expected after a name prefixed by the template keyword [-Wmissing-template-arg-list-after-template-kw]
265 | if (!value_.template Assign(index_, std::forward<T>(value))) {
| ^
3 errors generated.
```
Fixes https://github.com/pytorch/pytorch/issues/151316
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151344
Approved by: https://github.com/ZainRizvi, https://github.com/seemethere
Summary:
This fixes the error in https://fb.workplace.com/groups/1075192433118967/permalink/1640802133224658/
I tried really hard but I couldn't come up with a test case to repro the issue, but I confirmed with the OP that this issue has been fixed.
```
Traceback (most recent call last):
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/compile_fx.py", line 746, in _compile_fx_inner
mb_compiled_graph = fx_codegen_and_compile(
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/compile_fx.py", line 1343, in fx_codegen_and_compile
return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/compile_fx.py", line 1232, in codegen_and_compile
compiled_module = graph.compile_to_module()
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/graph.py", line 2087, in compile_to_module
return self._compile_to_module()
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/graph.py", line 2095, in _compile_to_module
self.codegen_with_cpp_wrapper() if self.cpp_wrapper else self.codegen()
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/graph.py", line 2002, in codegen
self._update_scheduler()
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/graph.py", line 1996, in _update_scheduler
self.scheduler = Scheduler(self.operations)
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/scheduler.py", line 1954, in __init__
self._init(nodes)
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/scheduler.py", line 1974, in _init
self.update_zero_dim_cpu_tensor()
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/scheduler.py", line 4433, in update_zero_dim_cpu_tensor
and buffer.get_size() == []
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/ir.py", line 3903, in get_size
return [*self.get_layout().size]
File "/dev/shm/uid-99/d2b830f6-seed-nspid4026547915_cgpid362302-ns-4026547912/torch/_inductor/ir.py", line 3914, in get_layout
raise NotImplementedError(type(self.layout).__name__)
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
NotImplementedError: NoneLayout
```
Test Plan: OP said the issue is fixed
Differential Revision: D72575808
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151321
Approved by: https://github.com/BoyuanFeng
We revisited how coalesced collective is working in https://github.com/pytorch/pytorch/pull/151243 and we now want to enable the script to work for slow path. The change is indeed bc-breaking but this is needed to make it work and the API is an internal use API. It is not user facing. For slow path the individual has input-sizes and output sizes recorded but no state. The final one has the state ready. We check the correctness of each individual collective one by one but we don't check the state match for these collectives, we can only check the state match for the last one which is the work item with coalesced label.
Added more unit test for slow path.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151247
Approved by: https://github.com/d4l3k, https://github.com/XilunWu
Sometimes the python script didn't exit normally and the lock file remains in the path. In this case, the `file_baton.py` may sleep forever waiting for the lock file to release. This PR will add a warning to show the existing lock file path, let the user better understand which file to delete when the waiting time is too long.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149382
Approved by: https://github.com/soulitzer
Finishes up the work started in #121686 + adds test
Update: this was not as straightforward as I originally imagined. Context below.
**TL;DR:** `TestFoo{CPU, CUDA}` now actually derive from `TestFoo`! Also, `{CPU, CUDA}TestBase` setup / teardown logic is now always called (it is required to set the primary device), regardless of whether `super().setUpClass()` / `super().tearDownClass()` are called or not.
**Background:** The typical way to get device-specific tests is to write a generic `TestFoo` and call `instantiate_device_type_tests(TestFoo, locals())` to get `TestFooCPU`, `TestFooCUDA`, etc. After this, generic tests (e.g. `TestFoo.test_bar()`) become `TestFooCPU.test_bar_cpu()` / `TestFooCUDA.test_bar_cuda()`.
Behind the scenes, this was historically accomplished by creating a `TestFooCUDA` that derives from both a `CUDATestBase` and an *empty class* called `TestFoo_base`. This `TestFoo_base` has the same bases as `TestFoo`, but none of the test functions (e.g. `test_bar()`). The documented reason for this is to avoid things like a derived `TestFooCUDA.test_bar()` being discovered in addition to the real device-specific test `TestFooCUDA.test_bar_cuda()`.
(1) A reason this matters is because it should be possible to call e.g. `super().setUpClass()` from a custom setup / teardown classmethod. If the generated TestFooCUDA does not derive from TestFoo, but instead derives from the empty class described above, this syntax does not work; in fact there is no way to form a proper `super()` call that works across the device-specific test variants. Here's an example that breaks in the OpInfo tests:
070f389745/test/test_ops.py (L218-L221)
(2) Further, there is some precedent within a custom `setUpClass()` impl for storing things on the `cls` object to be accessed at test time. This must be the device-specific test class (`TestFooCUDA`) and not `TestFoo` for this to work. As an example, the open device registration tests load a module during setup and use it in the test logic:
070f389745/test/test_cpp_extensions_open_device_registration.py (L63-L77)070f389745/test/test_cpp_extensions_open_device_registration.py (L79-L80)
To accomplish both (1) and (2) at the same time, I decided to revisit the idea of utilizing a proper inheritance hierarchy for `TestFoo` -> `{TestFooCPU, TestFooCUDA}`. That is: have TestFooCPU / TestFooCUDA **actually** derive from `TestFoo`. This achieves both (1) and (2). The only thing left is to make sure the generic tests (e.g. `TestFoo.test_bar()`) are not discoverable, as was the stated reason for diverging from this in the first place. It turns out we can simply `delattr()` these generic tests from `TestFoo` once `TestFooCPU` / `TestFooCUDA` have been setup with the device-specific variants, and all works well. The `instantiate_device_type_tests(...)` logic already deletes `TestFoo` from scope, so I don't see a problem with deleting generic tests from this base class as well (CI will prove me right or wrong ofc).
**Side note:** I was encountering a weird race condition where sometimes the custom `setUpClass()` / `tearDownClass()` defined & swapped in [here](4a47dd9b3f/torch/testing/_internal/common_device_type.py (L940-L955)) would be used, and sometimes it wouldn't. This non-deterministic behavior was called out previously by @ngimel here:
4a47dd9b3f/test/inductor/test_torchinductor_dynamic_shapes.py (L128-L130)
To address this, I moved this block of logic to before the first call to `instantiate_test()`, as that method queries for the primary device, and the primary device identification logic may manually invoke `setUpClass()` (see [here](4a47dd9b3f/torch/testing/_internal/common_device_type.py (L381-L384))). Goal: define the `setUpClass()` / `tearDownClass()` we want for correctness before they're ever called. This seems to work and the behavior is deterministic now AFAICT.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151129
Approved by: https://github.com/janeyx99, https://github.com/masnesral, https://github.com/malfet
This PR intends to rework the dispatching of the autograd key.
I.e., currently the DispatchKey.Autograd of the HOPs was triggered, even if non of the operands of the HOP have `requires_grad=True`. With this rework, the autograd is bypassed if non of the operands require gradients and only invoked if any of the operands require gradients.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151107
Approved by: https://github.com/ydwu4
There are only 118 failures atm, mark them all with xfail to avoid new regressions
Add `xfail_if_mps_unimplemented` decorator to distinguish between tests that call unimplemented eager op vs ones that fail for some other reason.
Added `aten._scaled_dot_product_attention_math_for_mps` fallback to make test behavior consistent between MacOS-15 (where falback is in place) and MacOS-14
Weird MacOS-14 specific skips:
- test_torchinductor.py::GPUTests::test_cat_extern_kernel_mps
- test_torchinductor.py::GPUTests::test_sort_transpose_mps (likely an eager bug)
- test_torchinductor.py::GPUTests::test_unaligned_input_mps
Numerous MacOS-13 skips, including few eager hard crashes, for example running `test_torchinductor.py::GPUTests::test_scatter5_mps` causes
```
/AppleInternal/Library/BuildRoots/c651a45f-806e-11ed-a221-7ef33c48bc85/Library/Caches/com.apple.xbs/Sources/MetalPerformanceShaders/MPSNDArray/Kernels/MPSNDArrayScatter.mm:309: failed assertion `Rank of destination array (1) must be greater than or equal to inner-most dimension of indices array (3)'
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150821
Approved by: https://github.com/ZainRizvi, https://github.com/dcci
ghstack dependencies: #151224, #151246, #151272, #151282, #151288
Summary:
as title
`export._trace._WrapperModule` is used to wrap functions into a Module so we can export the function.
We add `export._wrapper_utils` to `dynamo`'s `MOD_INLINELIST` so dynamo traces into `_WrapperModule`
Fixes https://github.com/pytorch/pytorch/issues/146867
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test:test_export -- -r wrapper_module
```
Differential Revision: D72986826
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151264
Approved by: https://github.com/angelayi
Preparatory refactor for https://github.com/pytorch/pytorch/pull/146942.
# Feature
This PR refactors the existing wrapper codegen into `WrapperLine` subclasses, extending the existing Memory Planning IR into a fully-fledged Wrapper IR. See the diagram below.
The IR currently supports the following ops:
- All existing memory planning IR ops (`AllocateLine`, `FreeIfNotReusedLine`, etc.)
- Reinterpret views (`ReinterpretLine`)
- Kernel definitions (`KernelDefinitionLine`)
- Calls to defined kernels (`KernelCallLine`)
- Calls to extern kernels (`ExternKernelLine`, `ExternKernelAllocLine`)
- Ops with multiple outputs (`MultiOutputLine`)
- Tensor cleanup at the end of a graph (`FreeLine`)
- Leaving comments in code (`CommentLine`)
There are two main motivations for this refactor:
1. Unlike free-form C++ and and Python code, Wrapper IR lines provide structured information about what the wrapper code does. This serves as a natural extension point for other types of wrapper codegen. For example, the parent PR generates FX IR from Wrapper IR. Wrapper IR aims to give new backends enough information to generate wrapper code without needing to modify core Inductor files such as `ir.py`.
2. This design will hopefully promote stronger modularity and encapsulation.
a. Inductor's core compilation passes don't need to worry about whether they're targeting Python, C++, FX or anything else. They can simply focus on generating Wrapper IR, and target-specific code can be refactored into the various backends.
b. Backends do not need to know about all the details and internal state of `V.graph` IR. For example, they don't need to consider whether a buffer has been removed from the graph when generating code. Wrapper IR will hopefully provide a simpler interface for generating wrapper code, which abstracts away the details of device code.
# Implementation details
The implementation mainly consists of separating direct C++/Python codegen into two phases:
1. Emit Wrapper IR lines describing what the wrapper code is supposed to do.
2. Inside the `codegen()` method of each `WrapperLine`, call backend methods which generate pure Python/C++ code using the information stored in the Wrapper IR line. For example, `KernelCallLine` calls `wrapper._generate_kernel_call_helper`, which is overriden by the various Python and C++ backends to generate the final wrapper code.
The main difficulty in implementing this is that we need to be careful that code is generated in the correct order. Wrapper codegen happens in two passes: first we write code into `self.lines` which mainly contains wrapper IR, but can also contain raw Python or C++ lines in some situations. Then, we convert the wrapper IR into the final Python/C++ code in `self.wrapper_call`. Since the same macros may be used in both passes, it's difficult to ensure that code is written to the correct buffer. The easiest solution for this was to implement a context manager overriding the `writeline` method to write to `self.wrapper_call` after memory planning is finished. This way, `writeline` writes to `self.lines` in the first pass, and `self.wrapper_call` in the second. This obviated the need to pass `code` or `writeline` variables all the way through the call stack, which would have touched most of the existing macros.
# Test plan
Since this refactor touches all the existing wrapper codegen classes, the existing CI provides good coverage.
The parent PR introduces new tests for the FX IR backend. Among other things, these tests assert that `self.lines` only contains Wrapper IR lines, and no free-form code. While this would not be true of all programs today, the tests suggests that the IR implemented in this PR is sufficient to cover basic PyTorch usage.
# Future directions
These two goals are only partially realized by this PR. These are several important steps which still undergo direct Python/C++ codegen in core files:
- User-defined Triton kernels.
- Reinterpret views on outputs, from `gen_output_refs()`. (In the parent PR, the FX converter has a custom way of handling this. This can eventually be ported into Wrapper IR.)
- Fallback ops with custom `codegen()` methods, e.g. `ScatterFallback`.
- Misc. C++ lines emitted by the various cpp backends, e.g. declaring constants.
These cases will gradually be handled in subsequent PRs, as the Inductor->FX converter expands its coverage. Given that these refactors are pretty tricky to do, it seems wiser to execute them in stages, as opposed to porting everything to Wrapper IR at once.Some Python and codegen still lives in core files such as `ir.py`, as described in previous sections. Hopefully, this PR will serve as a starting point which moves the codebase towards a more modular design. Over time, we can gradually refactor the remaining codegen (mainly in `ir.py`) into backend classes.
One limitation of this PR is that codegen still happens in two phases during `PythonWrapperCodegen`. First, we generate Wrapper IR into `self.lines`, and from there we generate Python or C++ code into `self.wrapper_call`, `self.header`, etc. In the long term, it would be cleaner to split wrapper IR into its own class which doesn't deal with Python/C++ codegen at all. (See the diagram at the top.) That would strictly enforce the boundary between Wrapper IR and Python/C++ wrapper code. However, this would probably be a much larger refactor.
Another limitation of the current code is that the helper functions have a lot of call args. It's also possible to clean this up by passing Wrapper IR ops e.g. `KernelCallLine` into helper functions like `_generate_kernel_call_helper`, since they store all the arguments. However, that change would likely be prone to merge conflicts, so I would like to save it for follow-up PRs if possible.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150458
Approved by: https://github.com/eellison
PT2 benchmark scripts has a pattern like:
```
def forward_and_backward_pass(self, mod, inputs, collect_outputs=True):
cloned_inputs = clone_inputs(inputs)
self.optimizer_zero_grad(mod)
with self.autocast(**self.autocast_arg):
pred = mod(**cloned_inputs)
loss = self.compute_loss(pred)
self.grad_scaler.scale(loss).backward()
self.optimizer_step()
if collect_outputs:
return collect_results(mod, pred, loss, cloned_inputs)
return None
```
for training.
The collect_outputs argument is True only for accuracy testing and it's false for performance testing.
For HF benchmark suite, a model usually returns tuple (loss, logits). For performance testing, even though the logits is never used anywhere, dynamo has to keep it due to the control flow.
A few bad things if we keep logits here
1. the peak memory will be higher since the logits is large and we can not release its memory earlier.
2. we can not do optimization like chunking for the logits because the tensor needs to be returned from the pre-grad graph
Actually I think it's fine to not return logits at all.
- For training cases, checking loss and gradients for accuracy is good enough. It's hard to see two runs have mismatch logits but matching loss/gradients.
- Also, discarding logits as soon as possible for perf benchmarking makes it more fair for us.
On the other hand, it may be interesting to let dynamo support something like dynamo.constexpr (similar to tl.constexpr). A variable annotated as dynamo.constexpr will be specialized at compile time and we can do more optimization (DCE e.g.) at compile time. (A small [repro](https://gist.github.com/shunting314/0912a8947028a904c34f361021b8024d))
Benchmark results here [link](https://hud.pytorch.org/benchmark/compilers?dashboard=torchinductor&startTime=Fri%2C%2004%20Apr%202025%2018%3A03%3A26%20GMT&stopTime=Fri%2C%2011%20Apr%202025%2018%3A03%3A26%20GMT&granularity=hour&mode=training&dtype=amp&deviceName=cuda%20(h100)&lBranch=gh/shunting314/204/head&lCommit=fe25dab3f65e1b0e9db0af03f7664af70fcc9c66&rBranch=main&rCommit=55e62ff74ad5614faf80b060c7bfc551e3b7af5a)
- HF 15% (1.51 -> 1.66 compression ratio) peak memory improvement
- I also see 5% (2.74 -> 2.79x) perf win for HF. It could be true. We may generate more efficient kernels since we don't need keep logits and return it from the pre-grad graph. But I'll double check
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151075
Approved by: https://github.com/eellison, https://github.com/jansel
Retry of https://github.com/pytorch/pytorch/pull/150957, which was reverted due to internal meta failures
Credit to @mgmtea who wrote the initial version of this PR: https://github.com/pytorch/pytorch/pull/146604
Context: CUPTI is the NVIDIA library that Kineto uses for collecting GPU-side info during profiling. The intended usage is to register a callback while you want profiling to occur, and then unregister the callback when you want profiling to stop. But a bug would cause crashes if CUPTI callbacks were de-registered when used with cudagraphs. The workaround was to disable "CUPTI_LAZY_REINIT" and "CUPTI_TEARDOWN" in Kineto - which prevents crashes, but can result in slower execution after profiling has occurred and completed.
This bug is believed to be fixed in CUDA >= 12.6, so this PR qualifies that DISABLE_CUPTI_LAZY_REINIT=1 and CUPTI_TEARDOWN=0 should only be applied if CUDA >= 12.6. Additionally, `profiler_allow_cudagraph_cupti_lazy_reinit_cuda12()` is added as an escape hatch so that we can add a killswitch in case we see more crashes related to this.
Differential Revision: [D72842114](https://our.internmc.facebook.com/intern/diff/D72842114/)
**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D72842114/)!
Differential Revision: [D72842114](https://our.internmc.facebook.com/intern/diff/D72842114)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151124
Approved by: https://github.com/sraikund16
If optree is less than the minimum version, we should pretend it doesn't
exist.
The problem right now is:
- Install optree==0.12.1
- `import torch._dynamo`
- This raise an error "min optree version is 0.13.0"
The fix is to pretend optree doesn't exist if it is less than the min
version.
There are ways to clean up this PR more (e.g. have a single source of
truth for the version, some of the variables are redundant), but I am
trying to reduce the risk as much as possible for this to go into 2.7.
Test Plan:
I verified the above problem was fixed. Also tried some other things,
like the following, which now gives the expected behavior.
```py
>>> import torch
>>> import optree
>>> optree.__version__
'0.12.1'
>>> import torch._dynamo
>>> import torch._dynamo.polyfills.pytree
>>> import torch.utils._pytree
>>> import torch.utils._cxx_pytree
ImportError: torch.utils._cxx_pytree depends on optree, which is
an optional dependency of PyTorch. To u
se it, please upgrade your optree package to >= 0.13.0
```
I also audited all non-test callsites of optree and torch.utils._cxx_pytree.
Follow along with me:
optree imports
- torch.utils._cxx_pytree. This is fine.
- [guarded by check] f76b7ef33c/torch/_dynamo/polyfills/pytree.py (L29-L31)
_cxx_pytree imports
- [guarded by check] torch.utils._pytree (changed in this PR)
- [guarded by check] torch/_dynamo/polyfills/pytree.py (changed in this PR)
- [guarded by try-catch] f76b7ef33c/torch/distributed/_functional_collectives.py (L17)
- [guarded by try-catch] f76b7ef33c/torch/distributed/tensor/_op_schema.py (L15)
- [guarded by try-catch] f76b7ef33c/torch/distributed/tensor/_dispatch.py (L35)
- [guarded by try-catch] f76b7ef33c/torch/_dynamo/variables/user_defined.py (L94)
- [guarded by try-catch] f76b7ef33c/torch/distributed/tensor/experimental/_func_map.py (L14)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151257
Approved by: https://github.com/malfet, https://github.com/XuehaiPan
sdp on xpu will fallback to math path in some cases (i.e. training). In dynamo benchmark, we prefer to use fp16 for better performance. Although `allow_fp16_bf16_reduction_math_sdp` is under backends.cuda, its implementation is for all device.
I didn't add if device == xpu here, I suppose cuda devices will not run into math path anyway
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150996
Approved by: https://github.com/drisspg, https://github.com/EikanWang
Goal is to have a way to compare if a change make it better or worse.
```
Average edge over aten (max(-edge, 0), higher is better):
triton: 8.596507086950552 (from 6 valid values)
triton_persistent_tma: 9.517193693923307 (from 6 valid values)
cutlass_lvl_default: 3.3234737908691785 (from 6 valid values)
cutlass_lvl_1111: 7.088173348313991 (from 6 valid values)
cutlass_lvl_2222: 7.291869722320318 (from 6 valid values)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150639
Approved by: https://github.com/ColinPeppler
This PR is to enable FR for all coalesce ops for fast path. (batch p2p is enabled in the current script, so we will mainly focus on non-P2P ops). To explain what is fast path, let's revisit how coalesced collective is working today:
For non-P2P coalesced ops, there are are several ways to call it (due to legendary reasons):
- Way one: Directly call python api like all_reduce_coalesced in python, this will be deprecated soon.
- Way two: Directly call api inside PGNCCL like allreduce_coalesced. The way case 1 will eventually call into this. This is not deprecated and will not be deprecated, IIUC.
- Way three: Using _coalescing_manager in python, like:
```
with _coalescing_manager():
for i in range(num_colls):
dist.all_reduce(tensors[i])
```
This way has two path:
- Fast path: when users call all-reduce, all-gather-into-tensor or reduce-scatter, we will only launch one big collective by calling the api from case 1.
- Slow path: we call startCoalescing() in the beginning and then a bunch of collectives (each one will generate a FR entry) and then endCoalescing(). Inside startCoalescing(), groupStart() is called and inside endCoalescing(), groupEnd() is then called. So although this is going to be one collective, we call into PGNCCL for each collective coalesced in the slow path case.
- For uneven all-gather (allgather_v) and reduce-scatter, it follows the pattern mention in slow path. It directly call cpp api inside PGNCCL.
This PR addressed the fast path because this is just an easy case, we store the collectives info on the python side, and we will only call into PGNCCL once so there will only be one work and one FR entry. We can just treat them as regular coalesced collective.
We add some e2e unit test for build_db function so that the change to FR is more thoroughly tested.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151243
Approved by: https://github.com/d4l3k, https://github.com/wz337
Fixes#151216, #151215
Previously I forgot to revert the timeout after setting it for the timeout test.
To prevent this in the future I split the test into 3 different tests so timeout testing is isolated.
Test plan:
Stress tested
```
pytest test/distributed/test_store.py -k queue -v -s --minutes 10
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151252
Approved by: https://github.com/XilunWu
By using Metal `as_type` which according to documentation does exactly
that:
> Metal adds an as_type<type-id> operator to allow any scalar or vector data type (that is not
a pointer) to be reinterpreted as another scalar or vector data type of the same size. The bits in
the operand are returned directly without modification as the new type. The usual type
promotion for function arguments is not performed.
Using `reinterpret_cast` created a potential silent correctness error when dtypes of different sizes were bitcast to each other
Add expicit cast to src_type to avoid errors due to type promotion (i.e.
soemthing like (x+1).view(dtype=torch.float16) would work correctly in
eager mode for int16 dtype, but would fail in compile, as arithmetic
operations will promote int16 to int32
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151272
Approved by: https://github.com/dcci
ghstack dependencies: #151224, #151246
This PR adds standalone_compile API that does precompilation via caching to support vLLM use case in the short term while we work on the longer term precompilation solution.
```
standalone_compile(gm, example_inputs, options) -> CompiledArtifact
CompiledArtifact.save(path, format: binary|unpacked = binary)
CompiledArtifact.load(path, format: binary|unpacked = binary)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150670
Approved by: https://github.com/jamesjwu, https://github.com/zou3519
To avoid accuracy issues when small reductions are unrolled, cast half to float during the `load` op
As `op_math_t<half>` is indeed float
This fixes `test_unroll_small_reduction` for reduced precision types
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151246
Approved by: https://github.com/dcci
ghstack dependencies: #151224
This PR implements _allgather_base, reduce_scatter, and _reduce_scatter_base in the MPI backend (ProcessGroupMPI), enabling support for Fully Sharded Data Parallel (FSDP) in environments that use MPI for distributed communication.
### Context
As noted in https://github.com/pytorch/pytorch/issues/85628, FSDP currently supports only the NCCL backend. Due to this limitation, FSDP cannot run on legacy HPC environments or clusters that rely on MPI.
By implementing just these three collective operations, we can enable FSDP to work with the MPI backend. These collectives are implemented in a similar manner to existing operations such as allgather.
### Testing
We validated this PR using pytorch/build/bin/ProcessGroupMPITest with OpenMPI, and all tests passed successfully.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150162
Approved by: https://github.com/H-Huang
Added a context manager, `torch._library.fake_profile.register_fake_profile(op_profiles)`, where given an operator profile, it will generate and register a fake impl for the operator based on the operator profile.
The input to `register_fake_profile` is a dictionary mapping operator name to a set of profiles which describe the input and outputs of the operator. Here's an example of a profile for `mylib.foo.default`:
```
"mylib.foo.default": {
OpProfile(
args_profile=(
TensorMetadata(rank=2, dtype=torch.float32, device=torch.device("cpu"), layout=torch.strided,),
TensorMetadata(rank=2, dtype=torch.float32, device=torch.device("cpu"), layout=torch.strided,),
),
out_profile=TensorMetadata(rank=2, dtype=torch.float32, device=torch.device("cpu"), layout=torch.strided,),
)
}
```
`foo`'s profile contains only one profile, which says that for 2 input tensors of rank 2, dtype float32, device cpu, we will return one tensor of rank 2, dtype float32, and device cpu.
This will then generate a fake kernel where given 2 input tensors of rank 2 (and the other tensor metadata), we will output one tensor of rank 2 (and the other tensor metadata). If the operator also supports other input ranks, then we can add to the profile for the fake impl to support more input types.
This profile can either be manually written or created by draft-export, and then checked into the codebase.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150807
Approved by: https://github.com/zou3519
ghstack dependencies: #150806
- Added a test to guard bfloat16. The optimizer incorrectly turns bfloat16 initializers into uint16, but this is not relevant to export logic.
- Fix bfloat16 support in onnx_program callable
Tested with the following with cuda
```py
import torch
class BfloatModel(torch.nn.Module):
def __init__(self):
super().__init__()
self.param = torch.nn.Parameter(torch.tensor(2.0, dtype=torch.bfloat16))
def forward(self, x):
return x * torch.tensor(1.0, dtype=torch.bfloat16) * self.param
input = torch.randn(1, 10, dtype=torch.bfloat16)
model = BfloatModel()
onnx_program = torch.onnx.export(model, (input,), dynamo=True, optimize=False, verify=True)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151121
Approved by: https://github.com/titaiwangms
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Fixes#147846. Previously there is no error out under out variant of`tensordot` while `requires_grad=True`. This can cause potential issue when out tensor is part of a computation graph.
Enforces the out variant of tensordot to run without setting `requries_grad=True`. Change same to #117067
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150270
Approved by: https://github.com/soulitzer
Summary:
There are a number of places in the code checking for the existence of `_boxed_call` instead of checking for a `True` value. This is somewhat dangerous because one would assume that setting it to `None` or `False` would be the same as not setting it (output_code.py does this, for example).
Change `hasattr()` to `getattr(..., False)` for these cases.
Test Plan: unit tests pass
Differential Revision: D72806693
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151130
Approved by: https://github.com/Skylion007
This has been pretty helpful for the size-oblivious rewrite. Wanted the variadic args version to avoid `sym_or(a, sym_or(b, sym_or(c, d)))` in favor of `sym_or(a, b, c, d)`. Happy to change this to ban the 1-arg version.
This is better than plain and/or because the whole symbolic expression gets preserved, and if we guard on it or defer as a runtime assert, we preserve all branches.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150456
Approved by: https://github.com/laithsakka
By adding `pass` in front of the comment for fake set_device call
Which fixes `TestGPU.test_zero_element_mutation_mps`, which previously
failed with
```
torch._inductor.exc.InductorError: RuntimeError: Failed to import /var/folders/sc/2thx6_x95h7_h9qs8s48yh140000gn/T/tmp2emka_sx/7k/c7kmnwhb363ysalhewglr3cwtej6tiz3t4ppqa4bvhubaokmlprw.py
IndentationError: expected an indented block after 'with' statement on line 38 (c7kmnwhb363ysalhewglr3cwtej6tiz3t4ppqa4bvhubaokmlprw.py, line 40)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151224
Approved by: https://github.com/Skylion007, https://github.com/jansel, https://github.com/dcci
This PR adds standalone_compile API that does precompilation via caching to support vLLM use case in the short term while we work on the longer term precompilation solution.
```
standalone_compile(gm, example_inputs, options) -> CompiledArtifact
CompiledArtifact.save(path, format: binary|unpacked = binary)
CompiledArtifact.load(path, format: binary|unpacked = binary)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150670
Approved by: https://github.com/jamesjwu, https://github.com/zou3519
Motivation: for the following script:
```
// demo.py
import torch
import json
from transformers import BertModel, BertConfig
CONFIG = """
{
"architectures": [
"BertForMaskedLM"
],
"attention_probs_dropout_prob": 0.1,
"gradient_checkpointing": false,
"hidden_act": "gelu",
"hidden_dropout_prob": 0.1,
"hidden_size": 768,
"initializer_range": 0.02,
"intermediate_size": 3072,
"layer_norm_eps": 1e-12,
"max_position_embeddings": 512,
"model_type": "bert",
"num_attention_heads": 12,
"num_hidden_layers": 12,
"pad_token_id": 0,
"position_embedding_type": "absolute",
"transformers_version": "4.6.0.dev0",
"type_vocab_size": 2,
"use_cache": true,
"vocab_size": 30522
}
"""
config = json.loads(CONFIG)
bloom_config = BertConfig(**config)
model = BertModel(bloom_config).half().cuda()
torch.compiler.reset()
torch.cuda.empty_cache()
compiled_fn = torch.compile(model)
vocab_size = 30522
for b in range(1, 3):
for s in range(1, 10):
print(f"🚀 {b} {s}")
input_ids = torch.randint(0, vocab_size, (b, s)).cuda()
attention_mask = torch.ones(b, s).cuda()
with torch.no_grad():
out = compiled_fn(input_ids, attention_mask).last_hidden_state
```
when we run it with:
```
time TORCH_LOGS=recompiles python demo.py
```
We can see there are 7 recompilations and it takes 2 mins (fresh build) or 1 min (cached build) in my machine.
One root cause of the recompilations is, there are guards to check the alignments of the inputs (see the patch). So there are unexpected recompilations for `(1, 4)`, `(1, 8)`, `(2, 4)` and `(2, 8)` inputs.
In this patch, we always try to always pad the inputs if we don't know its shape at compilation to avoid the guards on alignment. It is fine to always pad the tensor. It won't change the semantics.
Now there are only 3 recompilations and it takes 1 min (fresh build) and 17s (cached build) in my machine.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150403
Approved by: https://github.com/drisspg
This change enables basic NestedTensor operations on HPU,
fixing the runtime error when creating a NestedTensor on HPU.
- Extended `NestedTensorImpl` to recognize `hpu` as a valid storage device.
- Added `NestedTensorHPU` to `DispatchKey` parsing in `DispatchKey.cpp`.
- Updated `torchgen/model.py` to include `NestedTensorHPU` in `dispatch_keys`.
- Modified `native_functions.yaml` to enable `NestedTensorHPU` support for various ops.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148659
Approved by: https://github.com/jeromean, https://github.com/albanD, https://github.com/sujoysaraswati
This can save ~0.2ms on non cuda devices by skip calling `amp_definitely_not_available()`. It can improve small models in torchbench like lennard_jones on xpu 10% on both eager and inductor in dynamo benchmarks.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151111
Approved by: https://github.com/soulitzer
As titled, this pr adds additional `forward_dtype` and `backward_dtype` conversion in DTensor `redistribute` API to enable SimpleFSDP's mixed precision training.
In this forward pass, the DTensor can be configured to be cast to `forward_dtype`; in the backward pass, the DTensor can be configured to be cast to `backward_dtype`.
1. **Correctness**: The end-to-end SimpleFSDP mixed precision training integration has been proved to work properly in the PR from this fork: https://github.com/tianyu-l/pytorch_intern24/pull/20. We are now migrating the code to official PyTorch DTensor.
2. **Example Usage**: There is an example in TorchTian's SimpleFSDP implementation: https://github.com/pytorch/torchtitan/pull/1060.
In the example below, a DTensor `x` is all-gather'ed along the `self.compute_placements`, with datatype cast to `self.param_dtype`. In the backward pass, additionally, the computed gradients are reduce-scatter'ed along the `self.grad_placements`, with datatype cast to `self.reduce_dtype`.
```python
output = x.redistribute(
placements=self.compute_placements,
forward_dtype=self.param_dtype,
backward_dtype=self.reduce_dtype,
).to_local(grad_placements=self.grad_placements)
```
Under the hood, in `class Redistribute(torch.autograd.Function):`, the `forward` function first takes `x`'s local tensor, convert it to `forward_dtype`, before all-gather `x`.
The `backward` function take `grad_output` and convert it to `backward_dtype`, before reduce-scatter `grad_output`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150740
Approved by: https://github.com/tianyu-l
Summary: CK doesn't support FP32 attention, but aotriton does. If we prefer CK, and the input dtype is FP32, we'll select mem efficient attention but CK doesn't support it. So we'll exclude mem eff attention and pick math.
Differential Revision: D72880985
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151132
Approved by: https://github.com/yoyoyocmu
I want to format and refactor the csrc file of pytorch_openreg. To make the code review clearer and easier to understand, I divide the code refactoring into two parts:
- Part 1: Code formatting
- Part 2: Code refactoring and optimization (Next PR)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151004
Approved by: https://github.com/albanD
ghstack dependencies: #151000
Summary:
as title
`export._trace._WrapperModule` is used to wrap functions into a Module so we can export the function.
We add `export._wrapper_utils` to `dynamo`'s `MOD_INLINELIST` so dynamo traces into `_WrapperModule`
Fixes https://github.com/pytorch/pytorch/issues/146867
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test:test_export -- -r wrapper_module
```
Differential Revision: D69434316
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146919
Approved by: https://github.com/angelayi
Prior to this PR, `rng_state` is in `V.graph.graph_inputs` but not in read_writes of any IRNode. As a result, it is not identified as a partition inputs:
```python
def partition_0(args):
primals_2, primals_1 = args
...
buf0 = torch.ops.higher_order.graphsafe_run_with_rng_state(torch.ops.aten.rand.default, [4, 4], dtype=torch.float32, device=device(type='cuda', index=1), pin_memory=False, rng_state=fwd_rng_state_0)
# <----- access fwd_rng_state_0 but it's not an input
...
def call(self, args):
primals_1, primals_2, fwd_rng_state_0 = args
...
partition0_args = [primals_2, primals_1]
(buf2, primals_2, primals_1) = self.partitions[0](partition0_args)
# <---- fwd_rng_state_0 is graph_inputs but is not passed to partitions[0]
...
```
This PR fixes this issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150958
Approved by: https://github.com/eellison
Added a flag, `allow_override`, to allow overriding existing kernel implementations in `torch.library.register_fake` `library.impl`. The default is false, where if a user tries to register a kernel to a dispatch key that already contains a kernel, it will error. This flag doesn't apply to CustomOpDefs, where overriding a fake kernel is already allowed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150806
Approved by: https://github.com/zou3519
Preparatory refactor for https://github.com/pytorch/pytorch/pull/146942.
# Feature
This PR refactors the existing wrapper codegen into `WrapperLine` subclasses, extending the existing Memory Planning IR into a fully-fledged Wrapper IR. See the diagram below.
The IR currently supports the following ops:
- All existing memory planning IR ops (`AllocateLine`, `FreeIfNotReusedLine`, etc.)
- Reinterpret views (`ReinterpretLine`)
- Kernel definitions (`KernelDefinitionLine`)
- Calls to defined kernels (`KernelCallLine`)
- Calls to extern kernels (`ExternKernelLine`, `ExternKernelAllocLine`)
- Ops with multiple outputs (`MultiOutputLine`)
- Tensor cleanup at the end of a graph (`FreeLine`)
- Leaving comments in code (`CommentLine`)
There are two main motivations for this refactor:
1. Unlike free-form C++ and and Python code, Wrapper IR lines provide structured information about what the wrapper code does. This serves as a natural extension point for other types of wrapper codegen. For example, the parent PR generates FX IR from Wrapper IR. Wrapper IR aims to give new backends enough information to generate wrapper code without needing to modify core Inductor files such as `ir.py`.
2. This design will hopefully promote stronger modularity and encapsulation.
a. Inductor's core compilation passes don't need to worry about whether they're targeting Python, C++, FX or anything else. They can simply focus on generating Wrapper IR, and target-specific code can be refactored into the various backends.
b. Backends do not need to know about all the details and internal state of `V.graph` IR. For example, they don't need to consider whether a buffer has been removed from the graph when generating code. Wrapper IR will hopefully provide a simpler interface for generating wrapper code, which abstracts away the details of device code.
# Implementation details
The implementation mainly consists of separating direct C++/Python codegen into two phases:
1. Emit Wrapper IR lines describing what the wrapper code is supposed to do.
2. Inside the `codegen()` method of each `WrapperLine`, call backend methods which generate pure Python/C++ code using the information stored in the Wrapper IR line. For example, `KernelCallLine` calls `wrapper._generate_kernel_call_helper`, which is overriden by the various Python and C++ backends to generate the final wrapper code.
The main difficulty in implementing this is that we need to be careful that code is generated in the correct order. Wrapper codegen happens in two passes: first we write code into `self.lines` which mainly contains wrapper IR, but can also contain raw Python or C++ lines in some situations. Then, we convert the wrapper IR into the final Python/C++ code in `self.wrapper_call`. Since the same macros may be used in both passes, it's difficult to ensure that code is written to the correct buffer. The easiest solution for this was to implement a context manager overriding the `writeline` method to write to `self.wrapper_call` after memory planning is finished. This way, `writeline` writes to `self.lines` in the first pass, and `self.wrapper_call` in the second. This obviated the need to pass `code` or `writeline` variables all the way through the call stack, which would have touched most of the existing macros.
# Test plan
Since this refactor touches all the existing wrapper codegen classes, the existing CI provides good coverage.
The parent PR introduces new tests for the FX IR backend. Among other things, these tests assert that `self.lines` only contains Wrapper IR lines, and no free-form code. While this would not be true of all programs today, the tests suggests that the IR implemented in this PR is sufficient to cover basic PyTorch usage.
# Future directions
These two goals are only partially realized by this PR. These are several important steps which still undergo direct Python/C++ codegen in core files:
- User-defined Triton kernels.
- Reinterpret views on outputs, from `gen_output_refs()`. (In the parent PR, the FX converter has a custom way of handling this. This can eventually be ported into Wrapper IR.)
- Fallback ops with custom `codegen()` methods, e.g. `ScatterFallback`.
- Misc. C++ lines emitted by the various cpp backends, e.g. declaring constants.
These cases will gradually be handled in subsequent PRs, as the Inductor->FX converter expands its coverage. Given that these refactors are pretty tricky to do, it seems wiser to execute them in stages, as opposed to porting everything to Wrapper IR at once.Some Python and codegen still lives in core files such as `ir.py`, as described in previous sections. Hopefully, this PR will serve as a starting point which moves the codebase towards a more modular design. Over time, we can gradually refactor the remaining codegen (mainly in `ir.py`) into backend classes.
One limitation of this PR is that codegen still happens in two phases during `PythonWrapperCodegen`. First, we generate Wrapper IR into `self.lines`, and from there we generate Python or C++ code into `self.wrapper_call`, `self.header`, etc. In the long term, it would be cleaner to split wrapper IR into its own class which doesn't deal with Python/C++ codegen at all. (See the diagram at the top.) That would strictly enforce the boundary between Wrapper IR and Python/C++ wrapper code. However, this would probably be a much larger refactor.
Another limitation of the current code is that the helper functions have a lot of call args. It's also possible to clean this up by passing Wrapper IR ops e.g. `KernelCallLine` into helper functions like `_generate_kernel_call_helper`, since they store all the arguments. However, that change would likely be prone to merge conflicts, so I would like to save it for follow-up PRs if possible.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150458
Approved by: https://github.com/eellison
`compute_local_shape_and_global_offset` util computes the local shape of
a particular shard of a DTensor, and the global offset (which describes
how the shard fits into the global tensor).
When the tensor dim does not evenly divide into the mesh dim, uneven
sharding occurs. In some cases, uneven sharding results in an empty
shard.
e.g.
tensor dim size: 4096
mesh dim size: 30
ranks 0..27 have local size 18
rank 28 has local size 8
rank 29 has local size 0 <--- empty shard
The global offset for an empty shard was previously undefined and
returned values that were computed based on logic that assumes no empty
shards. This caused DCP to fail to save a checkpoint, becuase
deduplication logic could 'throw away' real (non-empty) shards thinking
they were duplicates of zero-sized shards with the same offset.
Now, we define the global offset of an empty shard to be the dim-size,
which is out of bounds of the tensor and can't overlap with any
non-empty shards.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150862
Approved by: https://github.com/teja-rao, https://github.com/XilunWu
[dynamo] Deprecate enable_cpp_framelocals_guard_eval config variable - default: True
Reading the feature enabling param `enable_cpp_framelocals_guard_eval `at the CPP level is time consuming and slows down the operation of the dynamo as it is done every time the function using this param is called. Reading the value only once at init isn’t an option as it would disable the modification of this param at the runtime. Since this feature is enabled by default for some time and it doesn’t cause known issues, the `enable_cpp_framelocals_guard_eval `configuration param will be deprecated by this commit and its value is hardcoded to true.
Local microbenchmark dynamo_guard_eval.py:
- 931.9 us -> 538.9 us (3.10)
@williamwen42 @jansel @anijain2305
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151008
Approved by: https://github.com/williamwen42
Summary:
When creating an `OpaqueTensorImpl`, currently there's only an option to create it for a non-view tensor, but it can be useful to create one for view tensors as well.
View tensors should contain the same autograd parameters as the original tensor, whereas non-view tensors get created with whatever `inference_mode` option is currently enabled. For this reason, `TensorImpl` has a special view constructor that takes `TensorImpl::ImplType` as its first parameter, so adding a new constructor to `OpaqueTensorImpl` that does the same thing allows us to create views with it.
Test Plan: CI
Reviewed By: scottxu0730
Differential Revision: D71748460
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151028
Approved by: https://github.com/scottxu0730, https://github.com/chaos5958
This adds queue operations as described in https://github.com/pytorch/pytorch/issues/150943.
This works by adding two new operations `queue_push` and `queue_pop`. The semantics are designed to be blocking with a timeout. Pushing will always succeed as the queue is infinite size. Popping will first call `wait` until the key is ready and then pop the value from the queue.
This implements queues for only: HashStore, TCPStore w/ libuv. FileStore and the legacy backends are not supported.
`wait` and `check` work for queue operations though queue_push will only wake up the first waiter rather than all of them.
This also has a few cleanups to error types/documentation in related code.
Example trace:
```
[I409 16:51:43.963833529 TCPStoreLibUvBackend.cpp:829] [c10d - trace] validate magic:1015412686 address:[localhost]:55816
[I409 16:51:43.963845838 TCPStoreLibUvBackend.cpp:842] [c10d - trace] ping nonce:2840795 address:[localhost]:55816
[I409 16:51:43.963902914 TCPStoreLibUvBackend.cpp:911] [c10d - trace] add key:init/ val:1 address:[localhost]:55816
[I409 16:51:43.963939389 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:init/ address:[localhost]:55816
[I409 16:51:43.963974842 TCPStoreLibUvBackend.cpp:893] [c10d - trace] get key:init/ address:[localhost]:55816
[I409 16:51:43.964071909 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/test_queue_support address:[localhost]:55816
[I409 16:51:43.964080221 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964108584 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964123207 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964128194 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964156347 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964187493 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964217709 TCPStoreLibUvBackend.cpp:1133] [c10d - trace] queue_pop key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964324300 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964354495 TCPStoreLibUvBackend.cpp:1133] [c10d - trace] queue_pop key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964416299 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964458733 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/non_existant address:[localhost]:55816
[W409 16:51:43.974516585 socket.cpp:460] [c10d] waitForInput: poll for socket SocketImpl(fd=75, addr=[localhost]:55816, remote=[localhost]:46641) returned 0, likely a timeout
[W409 16:51:43.974559169 socket.cpp:485] [c10d] waitForInput: socket SocketImpl(fd=75, addr=[localhost]:55816, remote=[localhost]:46641) timed out after 10ms
[I409 16:51:43.974600451 TCPStoreLibUvBackend.cpp:1101] [c10d - trace] cancel_wait address:[localhost]:55816
```
Test plan:
```
$ pytest test/distributed/test_store.py -k queue -v -s
test/distributed/test_store.py::FileStoreTest::test_queues SKIPPED [0.4351s] (Store does not support queues)
test/distributed/test_store.py::HashStoreTest::test_queues PASSED [0.0009s]
test/distributed/test_store.py::PrefixFileStoreTest::test_queues SKIPPED [0.0006s] (Store does not support queues)
test/distributed/test_store.py::TCPStoreTest::test_queues SKIPPED [0.0012s] (Store does not support queues)
test/distributed/test_store.py::LibUvTCPStoreTest::test_queues PASSED [0.0014s]
test/distributed/test_store.py::PrefixTCPStoreTest::test_queues PASSED [0.0014s]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150969
Approved by: https://github.com/XilunWu, https://github.com/fduwjj
Add the option for providing a Subgraph as an autotuning choice in Inductor. This is crucial for implementing the split-k optimization for GEMMs by decomposing a mm -> bmm. https://github.com/pytorch/pytorch/pull/150654 uses these changes to add decomposeK as a default autotuning choice for aten.mm in Inductor.
Using https://github.com/pytorch/pytorch/pull/150654 and a simple script:
```
import torch
def f(a, b):
return torch.matmul(a, b)
def decompose_func(a_in, b_in):
M, K = a_in.shape
K, N = b_in.shape
# TODO: Ideally we want to autotune over this parameter
kPartitions = 256
assert K % kPartitions == 0, "K must be divisible by Kmini"
B = K // kPartitions
a_reshaped = a_in.reshape(M, B, kPartitions).transpose(
0, 1
) # Shape: (B, M, kPartitions)
b_reshaped = b_in.reshape(B, kPartitions, N) # Shape: (B, kPartitions, N)
result = torch.bmm(a_reshaped, b_reshaped) # Shape: (B, M, N)
return result.sum(dim=0).to(torch.float16) # Sum over B dimension, Shape: (M, N)
for k in [4096, 8192, 12288, 16384, 20480, 24576, 28672, 32768]:
a = torch.randn(32, k, dtype=torch.float16, device="cuda", requires_grad=True)
b = torch.randn(k, 32, dtype=torch.float16, device="cuda", requires_grad=True)
compiled_res = torch.compile(f, dynamic=False)(a, b)
decompose_res = decompose_func(a, b)
print(f"Compiled mm result close to aten: {torch.allclose(f(a, b), compiled_res, atol=1e-5, rtol=0.5)}")
print(f"Compiled mm result close to decompose: {torch.allclose(decompose_res, compiled_res, atol=1e-5, rtol=0.5)}")
```
we are able to autotune the decomposeK optimization to aten and the traditional Triton templates in Inductor. DecomposeK is faster than aten by about ~10% on average and > 4x speedup over the best Triton templates on an H100 machine, e.g.:
```
AUTOTUNE mm(32x28672, 28672x32)
decompose_k_mm 0.0126 ms 100.0%
mm 0.0144 ms 87.5%
triton_mm_69 0.0579 ms 21.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=4
triton_mm_75 0.0677 ms 18.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=4, num_warps=4
triton_mm_76 0.0850 ms 14.8% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=4
triton_mm_68 0.1444 ms 8.7% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=5, num_warps=4
triton_mm_72 0.1546 ms 8.1% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=64, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
triton_mm_74 0.1819 ms 6.9% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=4, num_warps=4
triton_mm_67 0.1917 ms 6.6% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=2, num_warps=4
triton_mm_73 0.2766 ms 4.5% ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=32, BLOCK_M=32, BLOCK_N=32, EVEN_K=True, GROUP_M=8, num_stages=3, num_warps=4
```
https://pastebin.com/g3FMaauT is the generated code from Inductor containing the subgraph decomposition for aten.mm.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150653
Approved by: https://github.com/eellison
This is the final PR, where everything comes together.
The problem I'm trying to solve is the following: when we register a MemPool with the NCCL ProcessGroup, it calls `ncclCommRegister` on all the allocations that are _currently_ in the pool. However, any later allocation will _not_ be registered with the NCCL communicator!
This is terribly inconvenient, because it means that every piece of code that allocates a tensor must be changed to become aware of whether it's doing so within a private pool, and it must become aware of NCCL and of all the PGs in existence, in order to re-register that pool with them.
Moreover, I believe there can be performance implications because allocating tensors is usually done in the critical path (i.e., during the forward and backward of every step of a training), whereas registering memory is a slow operation that should be done once at init time.
With this PR, once the user registers a Mempool with the NCCL PG, we install some hooks into the CachingAllocator in order to listen for all future memory allocations and, if they belong to the pool, we automatically call `ncclCommRegister` on them! (In fact, we reuse the hooks that already exist for `TORCH_NCCL_USE_TENSOR_REGISTER_ALLOCATOR_HOOK`).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150684
Approved by: https://github.com/kwen2501
ghstack dependencies: #150683
In the NCCL ProcessGroup we want to support being able to "register" with NCCL all the allocations that belong to a certain private MemPool. In order to do so on-the-fly for every new allocation, we register a hook for the CachingAllocator's TraceEvents. However, we were lacking a way to know whether a given TraceEvent belonged to the MemPool that we cared about or not. With this PR, we add a MempoolId_t field to the TraceEvents.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150683
Approved by: https://github.com/syed-ahmed, https://github.com/kwen2501
Differential Revision: [D72760205](https://our.internmc.facebook.com/intern/diff/D72760205/)
We hardcoded to only use GEMM anyway.
This also raises the problem with high instantiation level. As the instantiation level goes higher (here it is 3333), the time it takes to list the configs might be long already (here it is >3 minutes).
If we know exactly what configs we care, we should have a way to generate them without calling generators. But let's see if we need that.
using this script
```
import os
os.environ["TORCH_LOGS"] = "inductor"
import torch
import torch._inductor.config
torch._inductor.config.max_autotune = True
torch._inductor.config.force_disable_caches = True
torch._inductor.config.max_autotune_gemm_backends = "Aten,CUTLASS"
# intentionally use no cutlass ops
torch._inductor.config.cuda.cutlass_max_profiling_configs = 0
torch._inductor.config.cuda.cutlass_instantiation_level = "3333"
def main():
M = 128
dtype = torch.float16
A = torch.randn(M, M, device="cuda", dtype=dtype)
B = torch.randn(M, M, device="cuda", dtype=dtype)
compiled_model = torch.compile(torch.mm)
_ = compiled_model(A, B)
print("done")
if __name__ == "__main__":
main()
```
before, with logs:
```
CUTLASS library generated 7 operations in 235.03 seconds
Got cutlass configs: total number of ops: 4753. Filtering took 10.51 seconds
```
after:
```
CUTLASS library generated 1 operations in 207.39 seconds
Got cutlass configs: total number of ops: 4753. Filtering took 9.53 seconds
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150973
Approved by: https://github.com/ColinPeppler
These two events are really common, and also make up a huge portion of logs (~70%) we get internally in PT2 Compile Events. I don't think it's actually that useful to aggregate them, so instead of logging them to PT2 Compile Events, lets just only log them to chromium.
These two events will still be visible from tlparse: they just won't be in our internal tables. Please let me know if folks disagree.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151053
Approved by: https://github.com/oulgen, https://github.com/masnesral
### Compilation error
The issue is that u0 (an unbacked symint) can come from a smaller int dtype e.g. int16, int32.
```
error: no matching function for call to ‘min(int64_t&, short int&)’
759 | call_add_kernel_with_scaling_0(... std::min(100L, s97, u0) ...);
```
### Diff
The fix is to explicitly specify `int64_t` in the std::min template.
```
int64_t s97 = arg0_1_size[0];
int16_t u0_raw; # not a long
auto u0 = u0_raw;
# Before
std::min({100L, s97, u0})
# After
std::min<int64_t>({100L, s97, u0})
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150894
Approved by: https://github.com/desertfire
This PR is the duplicated one for https://github.com/pytorch/pytorch/pull/139975.
This PR is to add torch._scaled_mm for CPU backend.
_scaled_mm_out_cpu and _scaled_mm_cpu are new added and included in torch._scaled_mm CPU dispatch. We also add _scaled_mm_out_cpu_emulated as a fallback function if the current platform cannot run FP8 matmul using oneDNN. And this PR also updates the various UTs related to FP8 to support CPU tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150410
Approved by: https://github.com/atalman
If input is channels last than MPS will return a channels last output
This fixed `GPUTests.test_convolution_4_mps` from test_torchinductor.py
That previous failed with
```
AssertionError: expected size 3==3, stride 1==192 at dim=1; expected size 12==12, stride 48==16 at dim=2; expected size 16==16, stride 3==1 at dim=3
```
As FakeTensor implementation of conv returned `Contiguous`, rather than `ChannelLast` layout on MacOS-15 or later.
This doesn't seem to be very well documented, so will try to document the call path for `ExternKernel` invocation for `aten::convolution`:
- First inductor decomp defined here is called
c93e4b8290/torch/_inductor/kernel/conv.py (L424-L425)
- Then it goes thru FakeTensor decomposition implemented here
320914f1b6/torch/_subclasses/fake_impls.py (L739-L740)
- Finally it goes down to convolution meta registrations implemented here
320914f1b6/torch/_meta_registrations.py (L2416-L2417)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151042
Approved by: https://github.com/dcci
Summary:
- Flip default value of `strict` argument from True to False on torch.export.export_for_training API
- All callsites have been updated to provide this argument explicitly to avoid behavior change.
- If you see any breakages, that means you may have a new callsite that is missed, please set `strict=True` explicitly to the callsite to mitigage.
Test Plan: CI
Differential Revision: D72724975
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150941
Approved by: https://github.com/ydwu4
Summary:
The reference quantized modules for linear / conv / etc fail to torchscript due to two issues
(1) The type of torch.qscheme doesn't script
(2) The "_DTYPE_TO_QVALUE_BOUNDS" values were resolving to union[float, int] instead of just int. We fix that with a hard cast.
See: <internal post> + comments for more context
Test Plan: unit tests + fixing this NB N6923590
Differential Revision: D72652616
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150870
Approved by: https://github.com/jerryzh168
While fixing the memory leak in https://github.com/pytorch/pytorch/pull/145757, we accidentally close the socket for the case when nread == 0 and thought it is the case when connection is closed. This is not true. According to libuv doc: https://docs.libuv.org/en/v1.x/stream.html#c.uv_read_cb.
> nread might be 0, which does not indicate an error or EOF. This is equivalent to EAGAIN or EWOULDBLOCK under read(2).
We found this bug when debugging a broken pipe issue when users first call a set and then wait for all keys right afterwards on 128 ranks. This might also cause other broken pipe issues we have seen in the prod jobs recently.
Added a unit test to test this case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150987
Approved by: https://github.com/d4l3k, https://github.com/XilunWu
## Summary of changes
1. Change assertion to a warning, when no all gather or reduce scatter patterns are found, and remove the corresponding unit test. It seems some valid TP graphs may not have any pattern matches, from what I can see.
2. Fix wrong variable name being referenced (`A_with_scatter_dim_0` instead of just `A`)
3. Simplify reshaping to target output shape (don't need to recalculate output shape)
4. When "A" tensor is 2D, so we are doing doing a 2D x 2D scaled mm, we need to fix our handling of the case where the scatter dim is 0. When scatter dim is 0 for the 2D scaled mm output shape, this is actually dim 1 in the unreduced stacked partial scaled mm outputs, which has a (logical) shape of `(group_size, M//group_size, N)`. To summarize:
- Unreduced stacked partials are of shape `(M, N)`
- We view as `(group size, M//group_size, N)` and reduce along the scatter dim (`group_size` / dim 0).
- Reduced output (`reduced_out`) has shape (M//group_size, N)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150935
Approved by: https://github.com/lw
This modification is to support XPU kernels for depthwise_conv2d and depthwise_conv3d.
Currently, when running depthwise_conv on XPU devices, it is calculated with Mkldnn via the ConvBackend::Overrideable path.
After this modification, depthwise_conv will be calculated directly using XpuDepthwise3d when the Mkldnn backend is disabled.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149114
Approved by: https://github.com/guangyey, https://github.com/albanD
Credit to @mgmtea who wrote the initial version of this PR: https://github.com/pytorch/pytorch/pull/146604
Context: CUPTI is the NVIDIA library that Kineto uses for collecting GPU-side info during profiling. The intended usage is to register a callback while you want profiling to occur, and then unregister the callback when you want profiling to stop. But a bug would cause crashes if CUPTI callbacks were de-registered when used with cudagraphs. The workaround was to disable "CUPTI_LAZY_REINIT" and "CUPTI_TEARDOWN" in Kineto - which prevents crashes, but can result in slower execution after profiling has occurred and completed.
This bug is believed to be fixed in CUDA >= 12.6, so this PR qualifies that DISABLE_CUPTI_LAZY_REINIT=1 and CUPTI_TEARDOWN=0 should only be applied if CUDA >= 12.6. Additionally, `profiler_allow_cudagraph_cupti_lazy_reinit_cuda12()` is added as an escape hatch so that we can add a killswitch in case we see more crashes related to this.
Differential Revision: [D72745929](https://our.internmc.facebook.com/intern/diff/D72745929)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150957
Approved by: https://github.com/aaronenyeshi, https://github.com/Skylion007
This PR remove the usage of guard_size_oblivious in vector_norm by inlining it in the runtime check,
this prevent any data dependent error from ever appearing here at the locations where guard_size_oblivious
used to exist. Before this PR it used to break potentially. This is NOT BC breaking or changing of semantics from eager.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148809
Approved by: https://github.com/bobrenjc93
# Root cause
The barrier timeout set to 0.1 is too short, some threads may not have enough time to reach the barrier.
# How to reproduce
Adding some sleep will be easy to reproduce.
```python
def test_barrier_timeout_rank_tracing(self):
N = 3
store = dist.HashStore()
def run_barrier_for_rank(i: int):
if i != 0:
import time;time.sleep(1) # Let some thread sleep for a while
try:
store_util.barrier(
store,
N,
key_prefix="test/store",
barrier_timeout=0.1,
rank=i,
rank_tracing_decoder=lambda x: f"Rank {x} host",
trace_timeout=0.01,
)
except Exception as e:
return str(e)
return ""
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150768
Approved by: https://github.com/d4l3k
PR https://github.com/pytorch/pytorch/pull/149665 did a change to the optimized_add that is causing an issue internally.
In general make_optimized should be only be called with valid new_args, new_args can become None
when elements already exists also, we should break out of the loop in that case.
Note that I also only maintained the optimized summation when both lhs and rhs lengths are <=2.
This is ok because the optimization is based on the inductive property of adding one symbol at a time.
the [2]+[2] here is serving as base case ( i feel we can also remove it ) .
Note that keeping it for all sizes while correct, I am not sure if tis as efficient (we will do N log(n) insertions).
there is no current justification for that.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150955
Approved by: https://github.com/Mingming-Ding, https://github.com/atalman, https://github.com/bobrenjc93
# Motivation
Adapt `torch.accelerator.device_count` for multi-process usage. For example, `torch.cuda.device_count` avoids poisoning fork, then `torch.accelerator.device_count` should meet the same requirement.
Now that `torch.get_device_module(device).device_count` supports this, `torch.accelerator.device_count` should align with this behavior as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149924
Approved by: https://github.com/albanD
ghstack dependencies: #147507
This adds a new `clone()` method to Store which will return a new Store instance that can be used from a different thread.
This is intended to better support multiple threads with stores such as when ProcessGroupNCCL needs a store to do error propagation.
Related issue: https://github.com/pytorch/pytorch/issues/150943
Test plan:
```
pytest test/distributed/test_store.py -k PythonStore
pytest test/distributed/test_store.py -k clone
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150966
Approved by: https://github.com/fduwjj
Summary:
- We are saying the minimum version of pytree that PyTorch can use is
0.13.0
- If a user imports torch.utils._cxx_pytree, it will raise an
ImportError if optree doesn't exist or exists and is less than the
minimum version.
Fixes https://github.com/pytorch/pytorch/issues/150889. There are
actually two parts to that issue:
1. dtensor imports torch.utils._cxx_pytree, but the optree installed in
the environment might be too old. Instead, raising ImportError in
torch.utils._cxx_pytree solves the issue.
2. We emit an "optree too low version" warning. I've deleted the
warning in favor of the more explicit ImportError.
Test Plan:
- code reading
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150956
Approved by: https://github.com/albanD, https://github.com/atalman, https://github.com/XuehaiPan
If you try to use torch in c++ using modules then it will not compile due to static function not being supported in MSVC when using modules https://developercommunity.visualstudio.com/t/10323558.
It's also aligned with [C++20 standard](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/n4849.pdf) (ISO/IEC 14882:2020) 10.2.7 Export declaration [module.interface]: "Exported names have either external linkage or no linkage".
Fixes https://github.com/pytorch/pytorch/issues/71309
Tested using the following code.
```c++
export module testModule;
import <torch/torch.h>;
import <memory>;
import <string>;
import <tuple>;
import <iostream>;
export namespace testModule
{
export void test()
{
torch::Tensor tensor1 = torch::rand({ 2, 3 });
torch::Tensor tensor2 = torch::rand({ 3, 2 });
// Perform tensor multiplication
torch::Tensor result = torch::matmul(tensor1, tensor2);
// Print the tensors
std::cout << "Tensor 1: " << tensor1 << std::endl;
std::cout << "Tensor 2: " << tensor2 << std::endl;
std::cout << "Result of multiplication: " << result << std::endl;
}
}
```
```c++
import testModule;
int main()
{
testModule::test();
return 0;
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148675
Approved by: https://github.com/albanD, https://github.com/malfet
Co-authored-by: mantaionut <ionut@janeasystems.com>
Summary:
We add the functionality to allow users to directly pass in a at::Tensor
into AOTInductor, that would be used as the constant.
This user managed buffer skips the copying step in AOTInductor, and let
users to directly manage the memory usage themselve.
Test Plan:
LD_LIBRARY_PATH=/data/users/$USER/pytorch/build/lib
/data/users/$USER/pytorch/build/bin/test_aoti_inference
Reviewers:
Subscribers:
Tasks:
Tags:
Differential Revision: [D72589514](https://our.internmc.facebook.com/intern/diff/D72589514)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150276
Approved by: https://github.com/chenyang78, https://github.com/desertfire
Summary: We need real_tensor on the FakeTensor in node.meta["val"] in order to aot_compile the draft exported programs. Otherwise, we cannot propagate real tensors even when fake_mode.propagate_real_tensors = True.
This also fixes real tensor propagation in `run_decomposition()`.
Test Plan:
```
buck2 run @mode/dev-nosan caffe2/test:test_export -- -r test_dedup_data_dependent_failure
```
Differential Revision: D72732714
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150948
Approved by: https://github.com/angelayi
This enables using FSDP+TP on parameters with dimensions that aren't
evenly divisible by the DP/TP mesh sizes.
- this may not support all possible combinations of strided shardings
and shardings, but the support before this PR is not complete anyway
This contains several fixes for different aspects of DTensor behavior
relating to uneven strided sharding:
- original creation of the strided tensor requires fixes in
StridedShard._split_tensor
- full_tensor() reconstruction requries fixes in
StridedShard._to_replicate_tensor to correctly reshuffle the data into
the original pre-sharded order
- Distributed Checkpointing support requires correct computation of the
compute_local_shape_and_global_offset util so it knows how a local
shard maps to the global tensor, for reconstruction during
load/reshard.
This PR also adds a util `_explicit_order_placements` which converts a list of
placements with StridedSharding into a list of placements with only
regular sharding, with the order shuffled such that it is equivalent.
Builds on and completes the work started in https://github.com/pytorch/pytorch/pull/148894
Uneven Sharding Example
-------
(copied from _StridedShard._to_replicate_tensor docstring)
mesh = (DP=2, TP=2)
original = torch.arange(5)
**Applying Sharding**
Step 1 - Apply TP sharding
`tp = distribute_tensor(x, world_mesh['tp'], [Shard(0)])`
local_tensors:
rank0: [0,1,2] rank1: [3,4]
rank1: [0,1,2] rank3: [3,4]
Step 2 - Apply FSDP sharding
`dp_tp = ...` (the process of creating a strided-shard tensor is skipped over as it is hacky and complicated)
dp_tp has placement (_StridedShard(0, split_factor=2), Shard(0))
local_tensors:
rank0: [0,1] rank1: [3]
rank1: [2] rank3: [4]
**Reconstructing the Full Tensor**
Now, say someone wants to reconstruct dp_tp's full tensor. This will invoke 'redistribute' to replicate.
redistribute will first replicate the "Shard(0)" placement on the rightmost mesh dim, then replicate the
StridedShard placement second, which is implemented by this function.
So our starting point (`local_tensor` arg) is the result of replicating the Shard(0) placement across the
TP dim, which looks like this.
Note the discrepancy with the 'tp sharded tensor' line above! We'll fix it by locally shuffling data.
local_tensors:
rank0: [0,1,3] rank1: [0,1,3]
rank1: [2,4] rank3: [2,4]
Step 1: replicate over the DP dimension. Afterwards, each rank can locally sort the values.
note: we need padding to do this allgather, and we'll need to keep track of the padding amount for later
local_tensors:
rank0: [0,1,3,2,4] rank1: [0,1,3,2,4]
rank1: [0,1,3,2,4] rank3: [0,1,3,2,4]
Step 2: chunk and shuffle values around to account for the wrong order of operations above
and get the original tensor content back
01324# <- our allgather includes padding, if padding was applied in step 1
01324 <- Remove the padding
013, 24 <- chunk once, 'undoing' the DP allgather
01, 3, 2, 4 <- chunk each chunk, 'undoing' the initial (wrong) TP allgather performed by Shard(0)->Replicate()
012, 34 <- interleave with stride=TP mesh dim size
01234 <- concatenate
Co-authored-by: Luca Wehrstedt <lw@meta.com>
Co-authored-by: Will Constable <whc@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150490
Approved by: https://github.com/wanchaol, https://github.com/XilunWu
## Improvements to `docstring_linter`
* Add a "grandfather list" of existing undocumented classes and functions (`--grandfather`, `--grandfather-tolerance`, `--no-grandfather`, `--write-grandfather`)
* In classes, now just one of the class itself or its `__init__()` method needs to be documented (`--lint-init` turns the old behavior back on)
* Now classes and functions defined local to other functions do not need to be documented (`--lint-local` turns the old behavior back on)
* New `--report` flag produces a compact report of long, undocumented classes or function definitions: see attached example run over all pytorch: [pytorch-docs.json](https://github.com/user-attachments/files/18455981/pytorch-docs.json)
## Help text
```
$ python tools/linter/adapters/docstring_linter.py --help
usage: docstring_linter.py [-h] [-l] [-v] [--grandfather GRANDFATHER] [--grandfather-tolerance GRANDFATHER_TOLERANCE] [--lint-init]
[--lint-local] [--lint-protected] [--max-class MAX_CLASS] [--max-def MAX_DEF]
[--min-docstring MIN_DOCSTRING] [--no-grandfather] [--report] [--write-grandfather]
[files ...]
`docstring_linter` reports on long functions, methods or classes without docstrings
positional arguments:
files A list of files or directories to lint
optional arguments:
-h, --help show this help message and exit
-l, --lintrunner Run for lintrunner and print LintMessages which aren't edits
-v, --verbose Print more debug info
--grandfather GRANDFATHER, -g GRANDFATHER
Set the grandfather list
--grandfather-tolerance GRANDFATHER_TOLERANCE, -t GRANDFATHER_TOLERANCE
Tolerance for grandfather sizes, in percent
--lint-init, -i Lint __init__ and class separately
--lint-local, -o Lint definitions inside other functions
--lint-protected, -p Lint functions, methods and classes that start with _
--max-class MAX_CLASS, -c MAX_CLASS
Maximum number of lines for an undocumented class
--max-def MAX_DEF, -d MAX_DEF
Maximum number of lines for an undocumented function
--min-docstring MIN_DOCSTRING, -s MIN_DOCSTRING
Minimum number of characters for a docstring
--no-grandfather, -n Disable the grandfather list
--report, -r Print a report on all classes and defs
--write-grandfather, -w
Rewrite the grandfather list
```
---
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145834
Approved by: https://github.com/amjames, https://github.com/eellison
This adds lazy initialization support to ProcessGroupGloo via `TORCH_GLOO_LAZY_INIT` or via `create_device(..., lazy_init=True)`
This is still a draft PR as there's one race condition when doing coalesced operations that needs to be fixed upstream in Gloo first. Depends on https://github.com/facebookincubator/gloo/pull/427 landing first
This also updates the gloo submodule to include the required changes.
Test plan:
added lazy init test variants
```
pytest -v test/distributed/test_c10d_gloo.py -k Lazy
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150801
Approved by: https://github.com/fduwjj
Summary:
When we divide a FakeTensor by an integer using the fast op implementation, the type promotion should be `ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT` so we get a float when dividing an int FakeTensor by an integer.
```
FAST = get_fast_op_impls()
fast_div = FAST[torch.ops.aten.div.Tensor]
fast_div(fake_tensor, some_int)
```
Test Plan:
```
python test/test_fake_tensor.py -k test_fast_div
```
Differential Revision: D72667430
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150874
Approved by: https://github.com/angelayi
Summary:
Sometimes we get `MetadataMismatchError` in aoti compilation because draft export uses the flag below to infer the fake kernel when there’s a mismatch, but aoti doesn’t have this flag turned on.
https://fburl.com/code/9qzytl6q
torch._functorch.config.generate_fake_kernels_from_real_mismatches
If we set this flag to True, then aoti compilation would work.
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r aoti_runtime_asserts
```
Differential Revision: D72345085
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150651
Approved by: https://github.com/angelayi
The util converts a list of placements in the traditional DTensor format
(e.g. [_StridedShard(0), Shard(0)], where list position is mesh_dim and sharding
is always applied left-to-right (from dim 0 to higher dims))
to a more explicitly ordered format, also replacing '_StridedShard' with
simple 'Shard' placements in the process.
(e.g. the above becomes [(1, Shard(0)), (0, Shard(0)] where the first
item in the tuple is the mesh_dim and the ordering of the tuples is the
sharding order.
This is useful so far as a helper for fixing local shape computation for
strided sharding in the uneven shape case, in the following PR- but may
also be useful more broadly if we can use explicit orderings to simplify
other parts of DTensor logic.
This skips implementing some combinations of _StridedSharding that are
not currently used in the wild today, but could be supported easily.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150493
Approved by: https://github.com/wanchaol, https://github.com/XilunWu
This PR creates two utils for generating a schema for hops from example inputs and use base hop as an exmaple.
1. HopArgumentInfoGen creates an argument or an output schema with mutation information.
2. CFuncitonSchemaGen piece together the argument info of inputs and outputs and produces torch._C.FunctionSchema.
is_write attribute of argument info can be computed. Note that the is_write annotation only works when the inputs are flattened (e.g. cannot support mutation inside tuple). We need special handling the case where we have tuple inputs like cond.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149688
Approved by: https://github.com/zou3519
While looking at enabling FR analysis for coalesced collectives, I found that for the slow-path coalescing (cols which are not all-gather, all-reduce or reduce-scatter), we still record start event for them. This is wrong and we should do the same thing as endEvent recodring.
And I made the profiler title more visible when we pass in the opType for coalesced all-gather and reduce-scatter.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150863
Approved by: https://github.com/eqy, https://github.com/d4l3k, https://github.com/kwen2501
Currently for HPU device we don't have any support for _fused_sdp_choice_stub dispatcher function, so for `scaled_dot_product_attention` function by default selecting the `MATH Backend` using `_fused_sdp_choice_stub` for HPU device. With this PR we have enabled support for `_fused_sdp_choice_stub` dispatcher function, so that we can invoke any backend (for example math, flash_attention, efficient_attention, cudnn_attention, overrideable) according to user choice for HPU device.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149512
Approved by: https://github.com/drisspg
Observing failure in release workflow:
https://github.com/pytorch/pytorch/actions/runs/14346340202/job/40216804374
```
Traceback (most recent call last):
File "/opt/python/cp311-cp311/lib/python3.11/site-packages/wheel/bdist_wheel.py", line 11, in <module>
from setuptools.command.bdist_wheel import bdist_wheel as bdist_wheel
ModuleNotFoundError: No module named 'setuptools.command.bdist_wheel'
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/tmp/tmppwpqef_x/triton/python/setup.py", line 27, in <module>
from wheel.bdist_wheel import bdist_wheel
File "/opt/python/cp311-cp311/lib/python3.11/site-packages/wheel/bdist_wheel.py", line 13, in <module>
raise ImportError(ERROR) from exc
ImportError: The 'wheel.bdist_wheel' module has been removed.
Please update your setuptools to v70.1 or later.
If you're explicitly importing 'wheel.bdist_wheel', please update your import to point to 'setuptools.command.bdist_wheel' instead.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150931
Approved by: https://github.com/Skylion007
``modernize-use-default-member-init`` prefers initialisation in class members, that make more ``= default`` constructors possible. Some violations or modernize rules have been fixed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149046
Approved by: https://github.com/zou3519
Summary:
att
regular weight has the type of torch.nn.parameter.Parameter
buffer and tensor constant has the type of torch.Tensor
both types are valid.
Test Plan: CI
Differential Revision: D72657275
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150867
Approved by: https://github.com/zhxchen17
This bug was crazy hard to reproduce, so I can't seem to get a unit test written that isn't the internal one I used for debugging.
Here's a short TLDR of the bug:
- Due to D71983456(OSS: https://github.com/pytorch/pytorch/pull/149910), we cache CachingAutotuners in memory.
- Importantly: **Saving stuff in PyCodeCache in memory is not semantically equivalent to writing to disk**. By saving it in memory, CachingAutotuners do not reset global state.
- It's possible through recompiles for different dynamo frames to compile down to exactly the same inductor output code. This involves models that run multiple times, but differ very subtley, or in ways that cause a dynamo guard failure but not a different inductor output code.
- Because of this, we reuse CachingAutotuners for a second compile (with different example inputs, just the same triton kernel code)
- CachingAutotuners have a Coordinate Descent class on them, which has a cache: https://fburl.com/code/4igrsams (OSS: aafc4b6188/torch/_inductor/runtime/coordinate_descent_tuner.py (L69))
- Because we are caching these in memory and not on disk, this cache is **not cleared** between runs.
- However, this variable is *not* saved on the class, and is reinitialized every time we do autotuning: https://fburl.com/code/n2o8tmje
(OSS: aafc4b6188/torch/_inductor/runtime/triton_heuristics.py (L933))
- `config2launcher` is added when we call `benchmark_one_config`, but on a CoorDesc *cache hit*, we never call `benchmark_one_config`! So we end up returning None, and erroring with:
```
AttributeError: 'NoneType' object has no attribute 'store_cubin'
```
This fixes the problem for now by just recompiling the launcher. Technically, we might be able to save config2launcher on the class to avoid this, but I don't want to risk another weird cache safety bug here, so taking the simpler approach for now.
Note that this error only reproduces if:
- None of AOTAutogradCache, FXgraphCache hit on the second entry: otherwise, the CachingAutotuner will go through a pickling and then not be saved in memory
- We haven't spawned parallel compile workers. If there are parallel compile workers, we pickle the autotuner on the way from the worker to the parent process, once again resetting the Autotuner.
- The autotune cache doesn't already have the best config stored in it
So it was extraordinarily hard to debug/reproduce. Because of this, I have a complicated internal unit test but no OSS test that can trigger the exact problem. I'll work on a separate test later, but this needs to go in to fix a sev, so we're landing it based on an internal test only.
Differential Revision: [D72655382](https://our.internmc.facebook.com/intern/diff/D72655382/)
**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D72655382/)!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150860
Approved by: https://github.com/oulgen
To workaround limitation of 32-arguments per kernel and being able to eventually compile something like
```python
import torch
def foo(*args):
rc = torch.empty_like(args[0])
for arg in args:
rc += arg
return rc
tensors = torch.rand(100, 32, device='mps').unbind(0)
print(torch.compile(foo)(*tensors))
```
For now, introduce `at::native:🤘:get_tensor_gpu_address` and use it from both C++ test and compile_shader to convert list of tensors to list of pointers valid on GPU.
Initially this binding were done via `id< MTLArgumentEncoder>`, but according to [Improving CPU Performance by Using Argument Buffers](https://developer.apple.com/documentation/metal/improving-cpu-performance-by-using-argument-buffers?language=objc#Encode-Resources-into-Argument-Buffers) article, this is not necessary when targeting Tier2-only devices (which is true of all devices on MacOS-13 or newer):
> To directly encode the argument buffer resources on these Tier 2 devices, write the [MTLBuffer](https://developer.apple.com/documentation/metal/mtlbuffer?language=objc).[gpuAddress](https://developer.apple.com/documentation/metal/mtlbuffer/gpuaddress?language=objc) property — and for other resource types (samplers, textures, and acceleration structures), the [gpuResourceID](https://developer.apple.com/documentation/metal/mtlcomputepipelinestate/gpuresourceid?language=objc) property — into the corresponding structure member. To encode offsets, treat these property values as uint64 types and add the offset to them.
Add both C++ and PyThon unittests that validate that this works.
Please note, that using either ArgumentEncoder or directly encoding the data does not guarantee buffer will not be freed until shader execution is complete. On the other hand, this should already be guaranteed by MPSCachingAllocator that would only free the memory after all streams completed its execution.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150780
Approved by: https://github.com/dcci
This PR:
- cleans up some existing comments that don't make sense anymore
- hooks up the "custom_op_default_layout_constraint" back (that seems to
have broken)
- cleans up the "lazy registration path" which seems to never get hit
anymore
- adds dislike_padding to nodes that require exact strides
Test Plan:
- tests + CI
disable padding
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148104
Approved by: https://github.com/shunting314, https://github.com/eellison
ghstack dependencies: #150495
Summary: Added a field `protocol` to `ExternKernelNodes` and all the lowering pass will always use the oss schema to serialize external kernel nodes from now on.
Test Plan: CI
Differential Revision: D72020444
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150197
Approved by: https://github.com/zhxchen17
Includes ATen native transformers hipified sources in ROCm+Windows build. This was removed due to Trinton not being available on Windows, but this causes further linker errors. Setting `USE_FLASH_ATTENTION=0` and `USE_MEM_EFF_ATTENTION=0` during the build will mitigate the missing headers, but also not cause any linker errors, so we will use this approach for now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150521
Approved by: https://github.com/jeffdaily
I still don't really understand the original purpose of that env var, but it appears that its usage is completely disconnected from MemPools and from `ncclMemAlloc`/`Free`. In fact, when that env var is set, we invoke `ncclCommRegister` for _all_ NCCL communicators for _all_ the memory segments managed by the allocator (both the global ones, allocated with `cudaMalloc`, and the ones in private MemPools), and we do that both for the segments that already exist when the PG is initialized and for all segments that will be allocated later.
I'm reworking the code a bit, by using a few helper functions, whose name should make this behavior clearer.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150682
Approved by: https://github.com/kwen2501
ghstack dependencies: #150681
This consists mainly in two changes:
- ensure we can reliably obtain the device from a `NCCLComm` object (there was one constructor which didn't set the device)
- use a RAII pattern for acquiring the lock to the global dictionary of `NCCLComms` (which ensures the lock is released in case of exceptions)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150681
Approved by: https://github.com/kwen2501
This PR is to resolve issue reported in https://github.com/intel/torch-xpu-ops/issues/1478
There are two cases failing in our Windows CI enabling.
- **test_xpu.py::TestXpuXPU::test_lazy_init_xpu** Needs to add `if __name__ == '__main__':` for Windows when using multiprocess. Refer to https://stackoverflow.com/a/18205006
```
RuntimeError:
An attempt has been made to start a new process before the
current process has finished its bootstrapping phase.
This probably means that you are not using fork to start your
child processes and you have forgotten to use the proper idiom
in the main module:
if __name__ == '__main__':
freeze_support()
...
The "freeze_support()" line can be omitted if the program
is not going to be frozen to produce an executable.
Traceback (most recent call last):
File "C:\Users\sdp\lufengqing\torch-xpu-ops\test\xpu\xpu_test_utils.py", line 24, in <module>
test_multi_process(model, input)
File "C:\Users\sdp\lufengqing\torch-xpu-ops\test\xpu\xpu_test_utils.py", line 16, in test_multi_process
assert p.exitcode == 0
AssertionError
```
- **test_xpu.py::TestXpuXPU::test_wrong_xpu_fork_xpu** is a linux only test case, we should skip it on Windows. Refer to 248487f455/test/test_multiprocessing.py (L609)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150520
Approved by: https://github.com/guangyey, https://github.com/EikanWang
This PR extracts some test cases from TestPatternMatcher into a newly created TestPatternMatcherGeneric, and uses instantiate_device_type_tests to make them reusable across multiple devices.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150286
Approved by: https://github.com/jansel
# Changes over the previous PR
This reverts commit 61a1f09 and adds `__launch_bounds__` to the kernel.
Previously I merged 114d404 that did not work on Blackwell because it consumed too many registers. It got reverted in 61a1f09. For more context see: https://github.com/pytorch/pytorch/issues/150266.
This PR reverts the revert (i.e. reapplies the original diff), with one additional line with `__launch_bounds__` added:
```
git diff HEAD^
diff --git a/aten/src/ATen/native/cuda/layer_norm_kernel.cu b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
index 0d63a2f979c..3ce2c24c18e 100644
--- a/aten/src/ATen/native/cuda/layer_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
@@ -657,6 +657,7 @@ bool aligned_grid
>
__global__
void
+__launch_bounds__(block_dim_x * block_dim_y)
GammaBetaBackwardCUDAKernelTemplate(
int64_t M,
int64_t N,
```
I managed to get a Blackwell machine and verified that the fix works. The fix was verified using this repro that I got from @drisspg
<details>
<summary> Repro script that fails on Blackwell </summary>
```
import torch
from torch.nn import init
# from transformer_nuggets import init_logging
# from transformer_nuggets.utils.benchmark import profiler
# from pathlib import Path
# init_logging()
class PermuteModule(torch.nn.Module):
def __init__(self, permutation):
super(PermuteModule, self).__init__()
self.permutation = permutation
def forward(self, x:torch.Tensor) -> torch.Tensor:
assert len(x.shape) == len(self.permutation), f"Dimension mismatch! Unable to permute {len(x.shape)} dim input with a {len(self.permutation)} dim permutation!"
return x.permute(*self.permutation)
def test(n_layers:int, conv_stride:int):
_sequence = []
for _ in range(n_layers):
# Conv1d inputs are (N x C x L), LayerNorm expects (* x C). Dims must be permuted between modules.
_sequence += [
PermuteModule((0,2,1)),
torch.nn.Conv1d(in_channels=512, out_channels=512, groups=1, kernel_size=9, dilation=1, stride=conv_stride, padding=0, bias=False),
PermuteModule((0,2,1)),
torch.nn.LayerNorm(512),
torch.nn.ReLU()
]
model = torch.nn.Sequential(*_sequence).to(device="cuda")
data = torch.randn((100,2048,512), device="cuda")
out = model(data)
loss = torch.nn.functional.mse_loss(out, torch.rand_like(out))
loss.backward()
torch.autograd.set_detect_anomaly(True)
print(f"Torch version: {torch.__version__}")
# with profiler(Path("conv")):
# # print(f"layers=1, stride=1")
# # test(n_layers=1, conv_stride=1)
# # print(f"layers=2, stride=1")
# # test(n_layers=2, conv_stride=1)
# # print(f"layers=1, stride=2")
# # test(n_layers=1, conv_stride=2)
# print(f"layers=2, stride=2")
# test(n_layers=2, conv_stride=2)
print(f"layers=2, stride=2")
test(n_layers=2, conv_stride=2)
# we will not reach this print statement.
print("DONE.")
```
</details>
I also re-ran my performance benchmark and found no regressions over the previous PR.
# Full description of the old PR
Original PR: https://github.com/pytorch/pytorch/pull/148605
This PR adds a new kernel for producing gamma and beta values for the backward pass in a performant way.
To test the performance against the baseline, I measured the backward pass of layernorm while sweeping over the following variables:
1. dtype in {half, float}
2. M in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
3. N in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
4. Whether we flush the L2 cache before running the backward pass
Summary: The new code performs better than the old code, especially for powers of 2. For M >> N case, it performs very well (kernel itself can be 30x faster and the overall backward pass can be 5-10x faster).
In order to visualize results of the kernel when choosing different values of M, N and dtype, I wrote some code to generate a heatmap. The heatmap has N on the x-axis, M on the y-axis and color-coded points where green shows performance improvement and red shows regressions. For example, `m=32 n=2048 1.42x` in the heatmap would indicate the normalized shape had 32 elements. The leading dimensions' product was 2048 elements and the new kernel resulted in the *backward pass* being 1.42x faster than the old *backward pass*.
Important note: This heatmap shows the total backward pass time as seen by the user. The kernel time difference can be sometimes very large while the total backward pass time is not that high. For example, for dtype=torch.half, M=32 N=2048, flush_l2_cache=True case, the heatmap shows a speedup of 1.42x, while ncu tells me the new kernel is 2.5x faster than the old:
M=32 N=2048 dtype=half flush_l2=True Old Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.35
Elapsed Cycles cycle 27,526
Memory Throughput % 2.21
DRAM Throughput % 0.54
Duration us 20.42
L1/TEX Cache Throughput % 4.31
L2 Cache Throughput % 2.62
SM Active Cycles cycle 1,475.02
Compute (SM) Throughput % 0.29
----------------------- ----------- ------------
```
M=32 N=2048 dtype=half flush_l2=True New Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.34
Elapsed Cycles cycle 10,920
Memory Throughput % 5.64
DRAM Throughput % 1.35
Duration us 8.13
L1/TEX Cache Throughput % 1.92
L2 Cache Throughput % 6.89
SM Active Cycles cycle 3,554.41
Compute (SM) Throughput % 0.67
----------------------- ----------- ------------
```
Let's look at some rows from the heatmap. For dtype=float16 flush_l2_cache=True and when input shapes are powers of 2, we get the following:
<img width="1508" alt="image" src="https://github.com/user-attachments/assets/06179599-b2f0-4a45-8664-247a1067950b" />
There are 3 columns -- the first shows all data points, the second shows speedups only and the 3rd column shows regressions only. We can see that there are dramatic speedups for M >> N cases and the regressions are not that high (less than 1%, which could just be measurement noise). Here is a small guide I made:
For dtype=float32, we get a similar chart:
<img width="1499" alt="image" src="https://github.com/user-attachments/assets/c4d31a76-03b0-426c-9114-e1bfad29b530" />
The new code performs especially well for m >> n cases, and also where m and n are small. The m >> n case is special because we run 2 reduction kernels back to back and parallelize in the "M" dimension (the older kernel only parallelized in the "N" dimension).
The new code can sometimes have regressions for non-powers of 2. That is because the old code was using block sizes of {16, 32} while we have `threads.x = 32`. For example when N=33, the old code would have 3 blocks and we will have 2 blocks. I wrote some code to specialize for this case, but I think it will add complexity and @ngimel mentioned that non-powers of 2 are rare enough.
I am including the regressions here for completeness' sake:
<img width="1500" alt="image" src="https://github.com/user-attachments/assets/31c17cfb-ed9b-4106-b9c8-5c359751f530" />
To see this better:
1. Click the image
2. Right click the expanded image and open in a new tab
3. Go to that tab and left click once to zoom in
If you want to see the full data, here it is:
I also measured binary size and compile time since those are important for developers:
Binary size comparison
```
# Original
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
# This PR
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
```
The diff in bytes is 302kB which is about a 0.1% increase.
Compile time difference:
```
# Original
real 0m10.931s
user 0m9.676s
sys 0m1.004s
# this PR
real 0m16.720s
user 0m15.514s
sys 0m1.066s
# Command I ran
time /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DAT_PER_OPERATOR_HEADERS -DFLASHATTENTION_DISABLE_ALIBI -DFLASHATTENTION_DISABLE_SOFTCAP -DFLASH_NAMESPACE=pytorch_flash -DFMT_HEADER_ONLY=1 -DHAVE_MALLOC_USABLE_SIZE=1 -DHAVE_MMAP=1 -DHAVE_SHM_OPEN=1 -DHAVE_SHM_UNLINK=1 -DMINIZ_DISABLE_ZIP_READER_CRC32_CHECKS -DONNXIFI_ENABLE_EXT=1 -DONNX_ML=1 -DONNX_NAMESPACE=onnx_torch -DTORCH_CUDA_BUILD_MAIN_LIB -DTORCH_CUDA_USE_NVTX3 -DUNFUSE_FMA -DUSE_C10D_GLOO -DUSE_C10D_NCCL -DUSE_CUDA -DUSE_CUFILE -DUSE_DISTRIBUTED -DUSE_EXTERNAL_MZCRC -DUSE_FLASH_ATTENTION -DUSE_MEM_EFF_ATTENTION -DUSE_NCCL -DUSE_RPC -DUSE_TENSORPIPE -D_FILE_OFFSET_BITS=64 -Dtorch_cuda_EXPORTS -I/home/ahmads/personal/pytorch/build/aten/src -I/home/ahmads/personal/pytorch/aten/src -I/home/ahmads/personal/pytorch/build -I/home/ahmads/personal/pytorch -I/home/ahmads/personal/pytorch/cmake/../third_party/benchmark/include -I/home/ahmads/personal/pytorch/third_party/onnx -I/home/ahmads/personal/pytorch/build/third_party/onnx -I/home/ahmads/personal/pytorch/nlohmann -I/home/ahmads/personal/pytorch/third_party/flash-attention/csrc/flash_attn/src -I/home/ahmads/personal/pytorch/aten/src/THC -I/home/ahmads/personal/pytorch/aten/src/ATen/cuda -I/home/ahmads/personal/pytorch/third_party/fmt/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/tools/util/include -I/home/ahmads/personal/pytorch/build/caffe2/aten/src -I/home/ahmads/personal/pytorch/aten/src/ATen/.. -I/home/ahmads/personal/pytorch/build/nccl/include -I/home/ahmads/personal/pytorch/c10/cuda/../.. -I/home/ahmads/personal/pytorch/c10/.. -I/home/ahmads/personal/pytorch/third_party/tensorpipe -I/home/ahmads/personal/pytorch/build/third_party/tensorpipe -I/home/ahmads/personal/pytorch/third_party/tensorpipe/third_party/libnop/include -I/home/ahmads/personal/pytorch/torch/csrc/api -I/home/ahmads/personal/pytorch/torch/csrc/api/include -isystem /home/ahmads/personal/pytorch/build/third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/tensorpipe/third_party/libuv/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googletest/include -isystem /home/ahmads/personal/pytorch/third_party/protobuf/src -isystem /home/ahmads/personal/pytorch/third_party/XNNPACK/include -isystem /home/ahmads/personal/pytorch/third_party/ittapi/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/eigen -isystem /usr/local/cuda/include -isystem /home/ahmads/personal/pytorch/third_party/ideep/mkl-dnn/include/oneapi/dnnl -isystem /home/ahmads/personal/pytorch/third_party/ideep/include -isystem /home/ahmads/personal/pytorch/INTERFACE -isystem /home/ahmads/personal/pytorch/third_party/nlohmann/include -isystem /home/ahmads/personal/pytorch/third_party/NVTX/c/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/cudnn_frontend/include -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS -D_GLIBCXX_USE_CXX11_ABI=1 -Xfatbin -compress-all -DONNX_NAMESPACE=onnx_torch -gencode arch=compute_90,code=sm_90 -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=bad_friend_decl --expt-relaxed-constexpr --expt-extended-lambda -Wno-deprecated-gpu-targets --expt-extended-lambda -DCUB_WRAPPED_NAMESPACE=at_cuda_detail -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -O3 -DNDEBUG -std=c++17 -Xcompiler=-fPIC -DTORCH_USE_LIBUV -DCAFFE2_USE_GLOO -Xcompiler -Wall -Wextra -Wdeprecated -Wno-unused-parameter -Wno-missing-field-initializers -Wno-array-bounds -Wno-unknown-pragmas -Wno-strict-overflow -Wno-strict-aliasing -Wunused-function -Wunused-variable -Wunused-but-set-variable -Wno-maybe-uninitialized -MD -MT caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o -MF caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o.d -x cu -c /home/ahmads/personal/pytorch/aten/src/ATen/native/cuda/layer_norm_kernel.cu -o caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o
```
So the new PR is 6 seconds longer compile time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150625
Approved by: https://github.com/ngimel, https://github.com/atalman
hello,
I was going over the documentation to build pytorch from source.
Unfortunately, the first thing that come up is that you strongly recommend to use anaconda, which shouldn't be used because it's no longer free to use.
Could you please remove that from the doc?
I don't know if you are aware but anaconda is no longer free.
They changed their terms of service in 2020 to restrict commercial usage.
They changed their terms of service in 2024 to forbid downloading anaconda and forbid education and non-profit usage too.
The download is open and doesn't require any registration, but if you download anaconda they will sue you ^^
They started raining lawsuits against users since last year. You may have heard about anaconda vs intel in the news. They started another 5 or so in the last few months.
https://www.reuters.com/legal/litigation/intel-sued-copyright-infringement-over-ai-software-2024-08-09/
You may need to adjust more doc and adjust your build system. The free to use alternatives are miniforge with the conda-forge channel.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150619
Approved by: https://github.com/seemethere
Here are the following modifications made to cpp_extension.py- 1) Changed compiler flag to use --version.
2) Added a feature to convert alpha-numeric string to numeric string for the version string returned by compiler. This was the source of error as the parser was failing on parsing alpha-numeric version string.
Build with following pytorch extensions- Apex, TorchVision, TorchAudio & DeepSpeed.
Unit tested with following pytorch extensions- Apex, TorchVision.
(cherry picked from commit c873aeac35851a7d5000eb7f24561d3f56c2ffbd)
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150451
Approved by: https://github.com/jeffdaily
Detail of the issue:
If PyTorch issues send/recv to each 2 rank comm, and these comms are managed by a single ProcessGroupNCCL instance, then comms need to abort either in sequence or in group.
I.e. the following sequential abort will cause hang in NCCL. recv(..., comm0, stream);
send(..., comm1, stream);
abort(comm1);
abort(comm0);
Fixes#119797
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150690
Approved by: https://github.com/kwen2501
Summary: If there is only one safetensors file, we don't need users to have a metadata file and we can just construct it from the keys of that file. This is a use-case for some HuggingFace models, so adding support for it
Test Plan:
ensure existing tests pass
tested e2e in a notebook
Differential Revision: D72472490
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150701
Approved by: https://github.com/joecummings
Change loop unrolling strategy. Previously, the script only unrolls the inner loop over block_size when block size is multiple of vector length. This version instead unrolls the outer loop which reduces the number of load/store for accumulation into the output array and improves performance for cases when block size is not multiple of vector length.
Benchmarking script:
```python
# SPDX-FileCopyrightText: Copyright 2025 Arm Limited and/or its affiliate <open-source-office@arm.com>
# SPDX-License-Identifier: BSD-3-Clause
import torch
import torch.nn as nn
import numpy as np
import time
import sys
np.random.seed(0)
torch.manual_seed(0)
num_embeddings = 400000
embedding_dim = int(sys.argv[1])
multi_hot = 100
batch_size = 400
nrun = 1000
class SimpleEmbeddingBagModel(nn.Module):
def __init__(self, num_embeddings, embedding_dim):
super(SimpleEmbeddingBagModel, self).__init__()
weights = torch.from_numpy((np.random.random_sample((num_embeddings, embedding_dim)) + 1).astype(np.float32)).to(torch.float16)
# Defining the EmbeddingBag layer
self.embedding_bag = torch.nn.EmbeddingBag(num_embeddings, embedding_dim, _weight=weights,
mode='sum', include_last_offset=True, dtype=torch.float32)
def forward(self, input, offsets):
# Forward pass through the EmbeddingBag layer
result32 = self.embedding_bag(input, offsets, per_sample_weights=None)
return result32
# Instantiate the model
model = SimpleEmbeddingBagModel(num_embeddings=num_embeddings, embedding_dim=embedding_dim)
model.eval()
# Example input
input_tensor = torch.randint(0, num_embeddings, (batch_size * multi_hot,), dtype=torch.long)
offsets = torch.tensor(range(0, batch_size * multi_hot + 1, multi_hot))
with torch.no_grad():
# warm up
output32 = model(input_tensor, offsets)
ti = time.time_ns()
for i in range(nrun):
_ = model(input_tensor, offsets)
tf = time.time_ns()
print("{:3d} {:.3E}".format(embedding_dim, (tf-ti)/nrun/1.e6))
```
Speedup on NEOVERSEV1 with 1 thread
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150176
Approved by: https://github.com/digantdesai, https://github.com/malfet
Summary: Introduce barrier util in the DistWrapper for rank local checkpointing. This barrier will be used at the end of the rank local checkpointing to ensure all ranks synchronize.
Test Plan: UTs
Differential Revision: D72541431
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150748
Approved by: https://github.com/MeetVadakkanchery
When debugging FR missing dump and missing dump logs, I have couple initial findings:
1. On the same rank, if a second watchdog timeout triggers on a different PG(or subPG), that watchdog thread will immediately throw exception instead of sleeping. We want to fix that by still making the watchdog thread to wait for 1 min.
2. The FR dump takes about 900ms to 1200ms so, we are not checking the store frequently enough. But instead of changing the frequency from 1sec to 300ms, we finally decided to just let all ranks just sleep for 1 min universally rather than using a promise.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150652
Approved by: https://github.com/kwen2501
Summary:
Profiler side of memory snapshot.
1. Add API to actually do snapshot when client interface is called
2. Add ifdefs to builds so that kineto hooks snapshot correctly.
Design Philosophy: There is one interesting part of this implementation and it is during export. For export we are callign the python impl of the export rather than CPP even though we are already in CPP. This is because it is better to simply have one path of export rather than 2. Personally, I want there to be parity between auto-trace and on-demand so it if we can limit the side paths then we will have an easier time maintaining this relationship
Test Plan: {F1976563426}
Reviewed By: sanrise
Differential Revision: D70733247
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150559
Approved by: https://github.com/sanrise
Summary: Adding new ops, support for empty shards, and fixed initializations for downstream checkpointing.
Test Plan: buck2 run 'fbcode//mode/dev-nosan' fbcode//torchrec/distributed/tests:test_shards_wrapper
Differential Revision: D72271275
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150469
Approved by: https://github.com/XilunWu
Summary: To preserve global state guards we need to make the C++ type serialzable. Using json because it's easier to do and we don't have a lot of data in global state.
Test Plan: test_dynamo -k test_global_state_guard_serialization
Differential Revision: D72410611
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150636
Approved by: https://github.com/williamwen42
Summary:
`-Wambiguous-reversed-operator` warns about ambiguous reversed operators, e.g. `a < b` and `b > a` are both valid. Such operators are disallowed in C++20. This codemod fixes the warnings.
#buildsonlynotests - If this diff compiles, it works.
- If you approve of this diff, please use the "Accept & Ship" button :-)
Test Plan: Sandcastle
Differential Revision: D72535527
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150744
Approved by: https://github.com/drisspg
There is some sort of bug in `pytype` where if this function doesn't have type hints, `pytype` will spend 10 minutes inferring the types. Not that this matters much for a project not using `pytype`, but it led me to realize that this function could easily be type hinted and is not, so here is a PR adding some type hints.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150715
Approved by: https://github.com/Skylion007
I noticed that I couldn't use `vec::Vectorized` operations with scalars, even though there is an implicit conversion from `T` to `vec::Vectorized<T>`, so I made it work.
Test Plan: Added tests. Reverted vec_base.h, left the new tests in place, and confirmed that new tests don't compile in that state.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150380
Approved by: https://github.com/Skylion007
Summary: https://github.com/pytorch/pytorch/pull/149817 introduced an extra warmup run to compute AOTI memory compression ratio, but since weights are only loaded once in the AOTI run, the peak memory seen in the extra warmup won't include the weight, which causes an aritifically high memory compression ratio. This PR removes that extra warmup run, and calls reset_peak_memory_stats in the proper place instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150695
Approved by: https://github.com/yushangdi
Fixes#144196
Extends #144106 and #144110
## Open Problems:
- [ ] Annotating with `numbers.Number` is a bad idea, should consider using `float`, `SupportsFloat` or some `Procotol`. https://github.com/pytorch/pytorch/pull/144197#discussion_r1903324769
# Notes
- `beta.py`: needed to add `type: ignore` since `broadcast_all` is untyped.
- `categorical.py`: converted `else` branches of mutually exclusive arguments to `if` branch[^2].
- ~~`dirichlet.py`: replaced `axis` with `dim` arguments.~~ #144402
- `gemoetric.py`: converted `else` branches of mutually exclusive arguments to `if` branch[^2].
- ~~`independent.py`: fixed bug in `Independent.__init__` where `tuple[int, ...]` could be passed to `Distribution.__init__` instead of `torch.Size`.~~ **EDIT:** turns out the bug is related to typing of `torch.Size`. #144218
- `independent.py`: made `Independent` a generic class of its base distribution.
- `multivariate_normal.py`: converted `else` branches of mutually exclusive arguments to `if` branch[^2].
- `relaxed_bernoulli.py`: added class-level type hint for `base_dist`.
- `relaxed_categorical.py`: added class-level type hint for `base_dist`.
- ~~`transforms.py`: Added missing argument to docstring of `ReshapeTransform`~~ #144401
- ~~`transforms.py`: Fixed bug in `AffineTransform.sign` (could return `Tensor` instead of `int`).~~ #144400
- `transforms.py`: Added `type: ignore` comments to `AffineTransform.log_abs_det_jacobian`[^1]; replaced `torch.abs(scale)` with `scale.abs()`.
- `transforms.py`: Added `type: ignore` comments to `AffineTransform.__eq__`[^1].
- `transforms.py`: Fixed type hint on `CumulativeDistributionTransform.domain`. Note that this is still an LSP violation, because `Transform.domain` is defined as `Constraint`, but `Distribution.domain` is defined as `Optional[Constraint]`.
- skipped: `constraints.py`, `constraints_registry.py`, `kl.py`, `utils.py`, `exp_family.py`, `__init__.py`.
## Remark
`TransformedDistribution`: `__init__` uses the check `if reinterpreted_batch_ndims > 0:`, which can lead to the creation of `Independent` distributions with only 1 component. This results in awkward code like `base_dist.base_dist` in `LogisticNormal`.
```python
import torch
from torch.distributions import *
b1 = Normal(torch.tensor([0.0]), torch.tensor([1.0]))
b2 = MultivariateNormal(torch.tensor([0.0]), torch.eye(1))
t = StickBreakingTransform()
d1 = TransformedDistribution(b1, t)
d2 = TransformedDistribution(b2, t)
print(d1.base_dist) # Independent with 1 dimension
print(d2.base_dist) # MultivariateNormal
```
One could consider changing this to `if reinterpreted_batch_ndims > 1:`.
[^1]: Usage of `isinstance(value, numbers.Real)` leads to problems with static typing, as the `numbers` module is not supported by `mypy` (see <https://github.com/python/mypy/issues/3186>). This results in us having to add type-ignore comments in several places
[^2]: Otherwise, we would have to add a bunch of `type: ignore` comments to make `mypy` happy, as it isn't able to perform the type narrowing. Ideally, such code should be replaced with structural pattern matching once support for Python 3.9 is dropped.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144197
Approved by: https://github.com/malfet
Co-authored-by: Aaron Gokaslan <aaronGokaslan@gmail.com>
Fixes#142397
Basic implementation is done. What's left:
- [x] Different dtype/device tensors in the TensorList
- [x] fast path for grouping the foreach kernel
- [x] Tests
Regarding tests, I found some tests in `test/test_torch.py` for GradScaler but I couldn't figure out what is the best way to enable the test for MPS device.
By removing `@onlyNativeDeviceTypes`, one enables the tests for MPS but also enables tests for all other devices which are not included in the native device types. If I put:
`instantiate_device_type_tests(TestTorchDeviceType, globals(), allow_mps=True)`
This enables lots of tests in that class for MPS which were not(?) being tested before? This part needs some clarification
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150255
Approved by: https://github.com/malfet
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Enabled bf16 grouped gemm with an API similar to _scaled_group_gemm, except without scale and fast accum arguments. All transpose variants are enabled, unlike scaled gemm. Ideally we'd factor out a lot more code from scaled gemm, currently there's a lot of repetition between scaled and non-scaled versions. I factored out only a helper kernel that prepares arguments.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150374
Approved by: https://github.com/drisspg
When replacing placeholders with getattrs during constant folding, we can have an argument and parameter name mismatch. In fact, there is no guarantee that the parameter name is equivalent to the argument name used in the module call.
Differential Revision: D72415970
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150692
Approved by: https://github.com/jfix71
Summary:
We used RAIIAtenTensorHandle for ConstantMap, where RAIIAtenTensorHandle
is a unique_ptr, indicating that all memory handling is by the
AOTInductor internally.
In this PR, we introduce ConstantAtenTensorHandle which replaces
RAIIATenTensorHandle. This class holds a raw AtenTensorHandle, and also
owns a RAIIAtenTensorHandle if user decides to delegate memory
management to AOTInductor.
This is a prerequisite for user managed buffer, this PR, however only
introduces this class and make sure it works with existing AOTInductor
and has the default behavior identical as using RAIIAtenTensorHandle.
Test Plan:
Existing tests. No change should be introduced within this PR.
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150275
Approved by: https://github.com/chenyang78, https://github.com/desertfire
By using cooperative `simd_sum`/`simd_product` instead of a C-style for loop for threadgroup reductions. This also allows significantly reduce amount of shared memory needed to perform those reductions
Using such reduction increases the `torch.compile` performance for gpt-fast using `stories110M` from 29 tokens/sec to 630 tokens/sec on M4 and changes perf of torch.rand as follows:
|size| before | after |
|------------------------|------------|-------------|
| 512x512 | 202.1 | 131.8 |
| 1024x1024 | 780.6 | 176.9 |
| 2048x2048 | 1423.4 | 339.9 |
| 4096x4097 | 2982.2 | 1047.2 |
Unfortunately, none of the SIMDgroup operations are available for 64-bit integers, but one can simulate the behavior using using `simd_shuffle_down` of 64-bit values represented as `int2` types, that yields reduction in $log_2(threadgroup\\_size)$ steps. [`mlx/kernels/reduction/ops.h](86389bf970/mlx/backend/metal/kernels/reduction/ops.h (L15-L18)) contains an implementation of such algorithm, but alas it yields wrong results on M1/M2(and may be M3 machines) if not all threads in the simdgroup are active which could be observed by running
```python
import torch
lib=torch.mps.compile_shader("""
kernel void do_sum(device int* out, constant int* in, uint idx [[thread_position_in_grid]]) {
out[idx] = metal::simd_shuffle_down(in[idx], 8);
}
""")
x=torch.arange(22, device='mps', dtype=torch.int32)
y=torch.empty_like(x)
lib.do_sum(y, x)
print(y)
```
that returns following on M4
```
tensor([ 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 0, 0, 0, 0, 0, 0, 0, 0], device='mps:0', dtype=torch.int32)
```
but same kernel running on M1 returns
```
tensor([ 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 14, 15, 16, 17, 18, 19, 20, 21], device='mps:0', dtype=torch.int32)
```
This discrepancy in behavior can be addressed by using `simd_shuffle_and_fill_down`, but any kernels using simd_shuffle_and_fill_down cause an internal compiler error on MacOS-13.2. Considering that OS is to be EOL soon, skip the offending tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150566
Approved by: https://github.com/manuelcandales
ghstack dependencies: #150452, #150457
Summary: The previous diff broke a few tests that didn't run on internal or GH CI: T220169086, this fixes that issue. The {% if } block is only supposed to support autotuned parameters (constexpr), and should not be used for locals based on other examples.
Test Plan: buck test 'fbcode//mode/opt' fbcode//caffe2/test/inductor:fp8 -- --exact 'caffe2/test/inductor:fp8 - test_tensorwise_scaling_bfloat16_shape_16,32,32_has_bias_False_use_fast_accum_True_persistent_matmul_True (caffe2.test.inductor.test_fp8.TestFP8Lowering)'
Reviewed By: NikhilAPatel
Differential Revision: D72460516
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150686
Approved by: https://github.com/eellison, https://github.com/NikhilAPatel
Summary: add a param to specify to the storage writer how to save tensors. Write now the only options are safetensors and torch.save.
Test Plan:
(lintrunner) [ankitageorge@devgpu003.cco3 /data/users/ankitageorge/fbsource/fbcode/caffe2 (1d57cb27b)]$ buck2 test 'fbcode//mode/opt' fbcode//caffe2/test/distributed/checkpoint:test_hf_storage
File changed: fbcode//caffe2/torch/distributed/checkpoint/filesystem.py
Buck UI: https://www.internalfb.com/buck2/e80cc963-e34a-4876-b6f4-7ce2794e48dd
Test UI: https://www.internalfb.com/intern/testinfra/testrun/3659174965882569
Network: Up: 32KiB Down: 1.9KiB (reSessionID-ef9fa764-a40a-451b-ab58-08eabe7a9422)
Executing actions. Remaining 0/4 3.4s exec time total
Command: test. Finished 2 local
Time elapsed: 19.6s
Tests finished: Pass 4. Fail 0. Fatal 0. Skip 0. Build failure 0
Reviewed By: saumishr
Differential Revision: D70271943
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150025
Approved by: https://github.com/saumishr
Summary:
Splitting the type definition of ConstantType into a separate header because it's needed by Sigmoid OSS but the entire model.h header include cause the following compilation error:
```
2025-04-01T18:12:42.0391272Z FAILED: caffe2/CMakeFiles/torch_cpu.dir/__/torch/csrc/nativert/kernels/AOTICallDelegateKernel.cpp.o
2025-04-01T18:12:42.0417705Z /opt/cache/bin/sccache /opt/cache/bin/clang++ -DAT_PER_OPERATOR_HEADERS -DBUILD_ONEDNN_GRAPH -DCAFFE2_BUILD_MAIN_LIB -DCPUINFO_SUPPORTED_PLATFORM=1 -DFMT_HEADER_ONLY=1 -DFXDIV_USE_INLINE_ASSEMBLY=0 -DHAVE_MALLOC_USABLE_SIZE=1 -DHAVE_MMAP=1 -DHAVE_SHM_OPEN=1 -DHAVE_SHM_UNLINK=1 -DIDEEP_USE_MKL -DMINIZ_DISABLE_ZIP_READER_CRC32_CHECKS -DNNP_CONVOLUTION_ONLY=0 -DNNP_INFERENCE_ONLY=0 -DONNXIFI_ENABLE_EXT=1 -DONNX_ML=1 -DONNX_NAMESPACE=onnx_torch -DTORCH_ENABLE_LLVM -DUSE_C10D_GLOO -DUSE_DISTRIBUTED -DUSE_EXTERNAL_MZCRC -DUSE_RPC -DUSE_TENSORPIPE -DXNN_LOG_LEVEL=0 -D_FILE_OFFSET_BITS=64 -Dtorch_cpu_EXPORTS -I/var/lib/jenkins/workspace/build/aten/src -I/var/lib/jenkins/workspace/aten/src -I/var/lib/jenkins/workspace/build -I/var/lib/jenkins/workspace -I/var/lib/jenkins/workspace/cmake/../third_party/benchmark/include -I/opt/llvm/include -I/var/lib/jenkins/workspace/third_party/onnx -I/var/lib/jenkins/workspace/build/third_party/onnx -I/var/lib/jenkins/workspace/nlohmann -I/var/lib/jenkins/workspace/torch/csrc/api -I/var/lib/jenkins/workspace/torch/csrc/api/include -I/var/lib/jenkins/workspace/caffe2/aten/src/TH -I/var/lib/jenkins/workspace/build/caffe2/aten/src/TH -I/var/lib/jenkins/workspace/build/caffe2/aten/src -I/var/lib/jenkins/workspace/build/caffe2/../aten/src -I/var/lib/jenkins/workspace/torch/csrc -I/var/lib/jenkins/workspace/third_party/miniz-3.0.2 -I/var/lib/jenkins/workspace/third_party/kineto/libkineto/include -I/var/lib/jenkins/workspace/third_party/kineto/libkineto/src -I/var/lib/jenkins/workspace/third_party/cpp-httplib -I/var/lib/jenkins/workspace/aten/src/ATen/.. -I/var/lib/jenkins/workspace/third_party/FXdiv/include -I/var/lib/jenkins/workspace/c10/.. -I/var/lib/jenkins/workspace/third_party/pthreadpool/include -I/var/lib/jenkins/workspace/third_party/cpuinfo/include -I/var/lib/jenkins/workspace/aten/src/ATen/native/quantized/cpu/qnnpack/include -I/var/lib/jenkins/workspace/aten/src/ATen/native/quantized/cpu/qnnpack/src -I/var/lib/jenkins/workspace/aten/src/ATen/native/quantized/cpu/qnnpack/deps/clog/include -I/var/lib/jenkins/workspace/third_party/NNPACK/include -I/var/lib/jenkins/workspace/third_party/fbgemm/include -I/
2025-04-01T18:12:42.0444143Z In file included from /var/lib/jenkins/workspace/torch/csrc/nativert/kernels/AOTICallDelegateKernel.cpp:5:
2025-04-01T18:12:42.0445081Z In file included from /var/lib/jenkins/workspace/torch/csrc/nativert/executor/AOTIDelegateExecutor.h:6:
2025-04-01T18:12:42.0446002Z In file included from /var/lib/jenkins/workspace/torch/csrc/nativert/executor/AOTInductorModelImpl.h:5:
2025-04-01T18:12:42.0447549Z /var/lib/jenkins/workspace/torch/csrc/inductor/aoti_runtime/model.h:78:13: error: function 'RAII_cpuMalloc' is not needed and will not be emitted [-Werror,-Wunneeded-internal-declaration]
2025-04-01T18:12:42.0448656Z RAIIDataPtr RAII_cpuMalloc(size_t num_bytes) {
```
model.h defines RAII_malloc functions directly into anonymous namespace which seems pretty sad. we should do something about it but may not in the current diff.
Test Plan: CI
Differential Revision: D72320413
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150545
Approved by: https://github.com/desertfire
This PR is related to https://github.com/pytorch/pytorch/pull/145476 . That PR had two files (test_functions.py and test_misc.py) . test_functions was causing CI/rebase/merge issues and hence removed for now. This PR contains only test_misc.py.
This is a continuation of https://github.com/pytorch/pytorch/pull/144387 .
## MOTIVATION
We recently integrated support for Intel Gaudi devices (identified as 'hpu') into the common_device_type framework via the pull request at https://github.com/pytorch/pytorch/pull/126970. This integration allows tests to be automatically instantiated for Gaudi devices upon loading the relevant library. Building on this development, the current pull request extends the utility of these hooks by adapting selected CUDA tests to operate on Gaudi devices. Additionally, we have confirmed that these modifications do not interfere with the existing tests on CUDA devices.
Other accelerators can also extend the functionality by adding the device in the devices list. ( For eg: xpu )
## CHANGES
Create a separate class for test functions running on CUDA devices
Extend the functionality of these tests to include HPUs
Use instantiate_device_type_tests with targeted attributes to generate device-specific test instances within the new classes
Apply skipIfHPU decorator to bypass tests that are not yet compatible with HPU devices
PS: Most of these changes were initially part of https://github.com/pytorch/pytorch/pull/147609 , but closed that PR due to merge conflicts. The review comments were handled in this PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149499
Approved by: https://github.com/EikanWang, https://github.com/desertfire, https://github.com/cyyever
Fixes#129673
### Summary:
Modifying a tensor by reshaping in place (such as `unsqueeze_`) should cause a graph break; however, when accessed through `torch.Tensor` api as opposed to as self attribute caused the code to crash with an error (see attached issue)
Paths differed when traced due to the stack variable popped, as:
* `self.unsqueeze_` pops a `LazyVariableTracker` which gets resolved to `TensorVariable`, so when looking for the method, triggers the fn call `var_getattr` in `_dynamo/variables/tensor.py`; since this is an inplace view (metadata mutation) on graph input, it is not well supported so should fall back (see [L446](1017927c83/torch/_dynamo/variables/tensor.py (L446)) in that file)
* `torch.Tensor.unsqueeze` pops a `UserDefinedClassVariable` so when looking for the method, triggers the fn call `var_getattr` in `_dynamo/variables/user_defined.py` on [L273](a8f6b40e36/torch/_dynamo/variables/user_defined.py (L273)). This path tries to build a variable tracker from the obj popped, which resolves to a trace_rule , and as a Tensor method, is resolved to `TorchInGraphFunctionVariable` on [L3767](a8f6b40e36/torch/_dynamo/trace_rules.py (L3767))
So, one straightforward option is to check if the fn is an inplace_view on a input tensor in `torch.py` when we resolve the `__call__function` for the `TorchInGraphFunctionVariable` instead, which resolves the bug by providing a graph break
### Test
```
pytest test/dynamo/test_functions.py::FunctionTests::test_unsqueeze_inplace
```
Results in
```
Running 1 items in this shard
test/dynamo/test_functions.py . [100%]
=========================================================================================== 1 passed in 9.16s ==========================================================================================
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150573
Approved by: https://github.com/anijain2305
This PR makes it so that we don't crash due to logging if we invoke AOTAutogradCache/FXGraphCache without using dynamo. This is preparation for supporting certain VLLM use cases where they store graph modules and have special handling in conjunection with the caches.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150423
Approved by: https://github.com/oulgen
This PR fixes two race conditions that occur when UT tests are run:
- In a particular order within a single shard.
- Concurrently in multiple shards. Each test now gets a unique filename that depends on the test name.
There were two other minor improvements to the UTs:
- matmul_offline_mgpu could occasionally fail if run on 8 GPUs. Criteria was relaxed.
- bmm_tunableop_rocm checks that the rotating buffer is not zero. Otherwise, the test is not useful.
Additionally, several UTs took over 1 minute to run. Their duration was reduced by a combination of setting max tuning iterations to one, setting the rotating buffer size to zero, and/or reducing the matrix dimensions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150463
Approved by: https://github.com/jeffdaily
Summary: Add an experimental feature to defer pytorch library initialization cost to post startup. As noted this feature is not thread safe, it requires the client to maintain thread safety at library load time.
Reviewed By: zou3519
Differential Revision: D71917841
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150537
Approved by: https://github.com/zou3519
Allocations using cudaHostRegister should use corresponding cudaHostUnregister and similarly for cudaHostAlloc / cudaFreeHost. In test_cuda.py, the allocator config will change from test to test but the cache is not emptied prior to changing the config. This results in the wrong free being called later. Unit test sharding is avoiding this issue, but running the test_cuda.py with a single shard will fail.
The following reproducer demonstrates the problem.
```C++
int main(int argc, char **argv)
{
void *ptr;
assert(cudaSuccess == cudaHostAlloc(&ptr, 1024, cudaHostAllocDefault));
assert(cudaSuccess == cudaHostUnregister(ptr));
std::free(ptr);
return 0;
}
```
The above code results in the following failure because the ptr is an invalid argument to cudaHostUnregister.
```
a.out: test.cpp:53: int main(int, char**): Assertion `cudaSuccess == cudaHostUnregister(ptr)' failed.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146520
Approved by: https://github.com/ngimel
# Changes over the previous PR
This reverts commit 61a1f09 and adds `__launch_bounds__` to the kernel.
Previously I merged 114d404 that did not work on Blackwell because it consumed too many registers. It got reverted in 61a1f09. For more context see: https://github.com/pytorch/pytorch/issues/150266.
This PR reverts the revert (i.e. reapplies the original diff), with one additional line with `__launch_bounds__` added:
```
git diff HEAD^
diff --git a/aten/src/ATen/native/cuda/layer_norm_kernel.cu b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
index 0d63a2f979c..3ce2c24c18e 100644
--- a/aten/src/ATen/native/cuda/layer_norm_kernel.cu
+++ b/aten/src/ATen/native/cuda/layer_norm_kernel.cu
@@ -657,6 +657,7 @@ bool aligned_grid
>
__global__
void
+__launch_bounds__(block_dim_x * block_dim_y)
GammaBetaBackwardCUDAKernelTemplate(
int64_t M,
int64_t N,
```
I managed to get a Blackwell machine and verified that the fix works. The fix was verified using this repro that I got from @drisspg
<details>
<summary> Repro script that fails on Blackwell </summary>
```
import torch
from torch.nn import init
# from transformer_nuggets import init_logging
# from transformer_nuggets.utils.benchmark import profiler
# from pathlib import Path
# init_logging()
class PermuteModule(torch.nn.Module):
def __init__(self, permutation):
super(PermuteModule, self).__init__()
self.permutation = permutation
def forward(self, x:torch.Tensor) -> torch.Tensor:
assert len(x.shape) == len(self.permutation), f"Dimension mismatch! Unable to permute {len(x.shape)} dim input with a {len(self.permutation)} dim permutation!"
return x.permute(*self.permutation)
def test(n_layers:int, conv_stride:int):
_sequence = []
for _ in range(n_layers):
# Conv1d inputs are (N x C x L), LayerNorm expects (* x C). Dims must be permuted between modules.
_sequence += [
PermuteModule((0,2,1)),
torch.nn.Conv1d(in_channels=512, out_channels=512, groups=1, kernel_size=9, dilation=1, stride=conv_stride, padding=0, bias=False),
PermuteModule((0,2,1)),
torch.nn.LayerNorm(512),
torch.nn.ReLU()
]
model = torch.nn.Sequential(*_sequence).to(device="cuda")
data = torch.randn((100,2048,512), device="cuda")
out = model(data)
loss = torch.nn.functional.mse_loss(out, torch.rand_like(out))
loss.backward()
torch.autograd.set_detect_anomaly(True)
print(f"Torch version: {torch.__version__}")
# with profiler(Path("conv")):
# # print(f"layers=1, stride=1")
# # test(n_layers=1, conv_stride=1)
# # print(f"layers=2, stride=1")
# # test(n_layers=2, conv_stride=1)
# # print(f"layers=1, stride=2")
# # test(n_layers=1, conv_stride=2)
# print(f"layers=2, stride=2")
# test(n_layers=2, conv_stride=2)
print(f"layers=2, stride=2")
test(n_layers=2, conv_stride=2)
# we will not reach this print statement.
print("DONE.")
```
</details>
I also re-ran my performance benchmark and found no regressions over the previous PR.
# Full description of the old PR
Original PR: https://github.com/pytorch/pytorch/pull/148605
This PR adds a new kernel for producing gamma and beta values for the backward pass in a performant way.
To test the performance against the baseline, I measured the backward pass of layernorm while sweeping over the following variables:
1. dtype in {half, float}
2. M in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
3. N in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
4. Whether we flush the L2 cache before running the backward pass
Summary: The new code performs better than the old code, especially for powers of 2. For M >> N case, it performs very well (kernel itself can be 30x faster and the overall backward pass can be 5-10x faster).
In order to visualize results of the kernel when choosing different values of M, N and dtype, I wrote some code to generate a heatmap. The heatmap has N on the x-axis, M on the y-axis and color-coded points where green shows performance improvement and red shows regressions. For example, `m=32 n=2048 1.42x` in the heatmap would indicate the normalized shape had 32 elements. The leading dimensions' product was 2048 elements and the new kernel resulted in the *backward pass* being 1.42x faster than the old *backward pass*.
Important note: This heatmap shows the total backward pass time as seen by the user. The kernel time difference can be sometimes very large while the total backward pass time is not that high. For example, for dtype=torch.half, M=32 N=2048, flush_l2_cache=True case, the heatmap shows a speedup of 1.42x, while ncu tells me the new kernel is 2.5x faster than the old:
M=32 N=2048 dtype=half flush_l2=True Old Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.35
Elapsed Cycles cycle 27,526
Memory Throughput % 2.21
DRAM Throughput % 0.54
Duration us 20.42
L1/TEX Cache Throughput % 4.31
L2 Cache Throughput % 2.62
SM Active Cycles cycle 1,475.02
Compute (SM) Throughput % 0.29
----------------------- ----------- ------------
```
M=32 N=2048 dtype=half flush_l2=True New Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.34
Elapsed Cycles cycle 10,920
Memory Throughput % 5.64
DRAM Throughput % 1.35
Duration us 8.13
L1/TEX Cache Throughput % 1.92
L2 Cache Throughput % 6.89
SM Active Cycles cycle 3,554.41
Compute (SM) Throughput % 0.67
----------------------- ----------- ------------
```
Let's look at some rows from the heatmap. For dtype=float16 flush_l2_cache=True and when input shapes are powers of 2, we get the following:
<img width="1508" alt="image" src="https://github.com/user-attachments/assets/06179599-b2f0-4a45-8664-247a1067950b" />
There are 3 columns -- the first shows all data points, the second shows speedups only and the 3rd column shows regressions only. We can see that there are dramatic speedups for M >> N cases and the regressions are not that high (less than 1%, which could just be measurement noise). Here is a small guide I made:
For dtype=float32, we get a similar chart:
<img width="1499" alt="image" src="https://github.com/user-attachments/assets/c4d31a76-03b0-426c-9114-e1bfad29b530" />
The new code performs especially well for m >> n cases, and also where m and n are small. The m >> n case is special because we run 2 reduction kernels back to back and parallelize in the "M" dimension (the older kernel only parallelized in the "N" dimension).
The new code can sometimes have regressions for non-powers of 2. That is because the old code was using block sizes of {16, 32} while we have `threads.x = 32`. For example when N=33, the old code would have 3 blocks and we will have 2 blocks. I wrote some code to specialize for this case, but I think it will add complexity and @ngimel mentioned that non-powers of 2 are rare enough.
I am including the regressions here for completeness' sake:
<img width="1500" alt="image" src="https://github.com/user-attachments/assets/31c17cfb-ed9b-4106-b9c8-5c359751f530" />
To see this better:
1. Click the image
2. Right click the expanded image and open in a new tab
3. Go to that tab and left click once to zoom in
If you want to see the full data, here it is:
I also measured binary size and compile time since those are important for developers:
Binary size comparison
```
# Original
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
# This PR
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
```
The diff in bytes is 302kB which is about a 0.1% increase.
Compile time difference:
```
# Original
real 0m10.931s
user 0m9.676s
sys 0m1.004s
# this PR
real 0m16.720s
user 0m15.514s
sys 0m1.066s
# Command I ran
time /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DAT_PER_OPERATOR_HEADERS -DFLASHATTENTION_DISABLE_ALIBI -DFLASHATTENTION_DISABLE_SOFTCAP -DFLASH_NAMESPACE=pytorch_flash -DFMT_HEADER_ONLY=1 -DHAVE_MALLOC_USABLE_SIZE=1 -DHAVE_MMAP=1 -DHAVE_SHM_OPEN=1 -DHAVE_SHM_UNLINK=1 -DMINIZ_DISABLE_ZIP_READER_CRC32_CHECKS -DONNXIFI_ENABLE_EXT=1 -DONNX_ML=1 -DONNX_NAMESPACE=onnx_torch -DTORCH_CUDA_BUILD_MAIN_LIB -DTORCH_CUDA_USE_NVTX3 -DUNFUSE_FMA -DUSE_C10D_GLOO -DUSE_C10D_NCCL -DUSE_CUDA -DUSE_CUFILE -DUSE_DISTRIBUTED -DUSE_EXTERNAL_MZCRC -DUSE_FLASH_ATTENTION -DUSE_MEM_EFF_ATTENTION -DUSE_NCCL -DUSE_RPC -DUSE_TENSORPIPE -D_FILE_OFFSET_BITS=64 -Dtorch_cuda_EXPORTS -I/home/ahmads/personal/pytorch/build/aten/src -I/home/ahmads/personal/pytorch/aten/src -I/home/ahmads/personal/pytorch/build -I/home/ahmads/personal/pytorch -I/home/ahmads/personal/pytorch/cmake/../third_party/benchmark/include -I/home/ahmads/personal/pytorch/third_party/onnx -I/home/ahmads/personal/pytorch/build/third_party/onnx -I/home/ahmads/personal/pytorch/nlohmann -I/home/ahmads/personal/pytorch/third_party/flash-attention/csrc/flash_attn/src -I/home/ahmads/personal/pytorch/aten/src/THC -I/home/ahmads/personal/pytorch/aten/src/ATen/cuda -I/home/ahmads/personal/pytorch/third_party/fmt/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/tools/util/include -I/home/ahmads/personal/pytorch/build/caffe2/aten/src -I/home/ahmads/personal/pytorch/aten/src/ATen/.. -I/home/ahmads/personal/pytorch/build/nccl/include -I/home/ahmads/personal/pytorch/c10/cuda/../.. -I/home/ahmads/personal/pytorch/c10/.. -I/home/ahmads/personal/pytorch/third_party/tensorpipe -I/home/ahmads/personal/pytorch/build/third_party/tensorpipe -I/home/ahmads/personal/pytorch/third_party/tensorpipe/third_party/libnop/include -I/home/ahmads/personal/pytorch/torch/csrc/api -I/home/ahmads/personal/pytorch/torch/csrc/api/include -isystem /home/ahmads/personal/pytorch/build/third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/tensorpipe/third_party/libuv/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googletest/include -isystem /home/ahmads/personal/pytorch/third_party/protobuf/src -isystem /home/ahmads/personal/pytorch/third_party/XNNPACK/include -isystem /home/ahmads/personal/pytorch/third_party/ittapi/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/eigen -isystem /usr/local/cuda/include -isystem /home/ahmads/personal/pytorch/third_party/ideep/mkl-dnn/include/oneapi/dnnl -isystem /home/ahmads/personal/pytorch/third_party/ideep/include -isystem /home/ahmads/personal/pytorch/INTERFACE -isystem /home/ahmads/personal/pytorch/third_party/nlohmann/include -isystem /home/ahmads/personal/pytorch/third_party/NVTX/c/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/cudnn_frontend/include -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS -D_GLIBCXX_USE_CXX11_ABI=1 -Xfatbin -compress-all -DONNX_NAMESPACE=onnx_torch -gencode arch=compute_90,code=sm_90 -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=bad_friend_decl --expt-relaxed-constexpr --expt-extended-lambda -Wno-deprecated-gpu-targets --expt-extended-lambda -DCUB_WRAPPED_NAMESPACE=at_cuda_detail -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -O3 -DNDEBUG -std=c++17 -Xcompiler=-fPIC -DTORCH_USE_LIBUV -DCAFFE2_USE_GLOO -Xcompiler -Wall -Wextra -Wdeprecated -Wno-unused-parameter -Wno-missing-field-initializers -Wno-array-bounds -Wno-unknown-pragmas -Wno-strict-overflow -Wno-strict-aliasing -Wunused-function -Wunused-variable -Wunused-but-set-variable -Wno-maybe-uninitialized -MD -MT caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o -MF caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o.d -x cu -c /home/ahmads/personal/pytorch/aten/src/ATen/native/cuda/layer_norm_kernel.cu -o caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o
```
So the new PR is 6 seconds longer compile time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150625
Approved by: https://github.com/ngimel
Summary: Add support for caching of CUDA (nvcc) compilation errors to codecache.py
Test Plan: CI ( for example Cutlass backend unit tests )
Reviewed By: ColinPeppler
Differential Revision: D71562040
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149716
Approved by: https://github.com/ColinPeppler
This PR is the duplicated one for https://github.com/pytorch/pytorch/pull/139975.
This PR is to add torch._scaled_mm for CPU backend.
_scaled_mm_out_cpu and _scaled_mm_cpu are new added and included in torch._scaled_mm CPU dispatch. We also add _scaled_mm_out_cpu_emulated as a fallback function if the current platform cannot run FP8 matmul using oneDNN. And this PR also updates the various UTs related to FP8 to support CPU tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150410
Approved by: https://github.com/atalman
## Summary
remove the `tp_parallelize_plan` assignment that accidentally rewrites the previous assignments in `test_fsdp_dsd.py`.
## Test
`pytest test/distributed/checkpoint/fsdp/test_fsdp_dsd.py`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150354
Approved by: https://github.com/wconstab
Summary:
**Context**: https://github.com/pytorch/pytorch/pull/150122 (D71982587 - let's call this "the WS diff") introduces "bc/fc-breaking" cache changes.
In particular, it introduces `num_consumer_groups` and adds it to the cached config. In versions of torch that include the WS diff, `num_consumer_groups` is treated as a class variable on a triton.Config object (i.e. `triton.Config({..kwargs..}, num_consumer_groups=num_consumer_groups, ...`). And in versions of torch that don't include the WS diff, you generally don't expect to see this kwarg.
But if a program is run WS-torch (i.e. torch w/ the WS diff), and then later you run the same program with non-WS-torch, then non-WS-torch is going to find this autotune cache entry, and interpret `num_consumer_groups` as a kwarg, because there's no special handling for for num_consumer_groups in this version of torch. Then the program crashes with a triton failure message.
**The fix**: add the torch version / torch key into the hash, so that any changes to inductor will invalidate the cache (ensuring that other changes to triton_heuristics won't cause these bc/fc issues).
Test Plan: D72285868 (or https://gist.github.com/davidberard98/2ea697eb550c94d0d1948fedb5c5c7d8, but this doesn't repro in OSS because this version of warp specialization is not available in oss triton) can repro the failure, and the failure is fixed after this PR is patched.
Also, added a test in test/inductor/test_codecache.py which verifies that there's no cache hit if the torch_key changes (and verified that without the functional changes in this PR, the test fails).
Differential Revision: D72285303
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150494
Approved by: https://github.com/oulgen
An internal model was serialized in 2023, and is now breaking while loading with the following error:
```
File "<eval_with_key>.1675", line 4
def forward(self, arg1163_1, arg1164_1, , arg1166_1, , arg1168_1, arg1169_1, arg1170_1, , arg1172_1, arg1173_1, arg1174_1, arg1175_1, arg1176_1, arg1177_1, arg1178_1, arg1179_1, arg1180_1, arg1181_1, arg1182_1, arg1183_1, arg1184_1, arg1185_1, arg1186_1, arg1187_1, arg1188_1, arg1189_1, arg1190_1, arg1191_1, arg1192_1, arg1193_1, arg1194_1, arg1195_1, arg1196_1, arg1197_1, arg1198_1, arg1199_1, arg1200_1, arg1201_1, arg1202_1, arg1203_1, arg1204_1, arg1205_1, arg1206_1, arg1207_1, arg1208_1, arg1209_1, arg1210_1, arg1211_1, arg1212_1, arg1213_1, arg1214_1, arg1215_1, arg1216_1, , arg1218_1, arg1219_1, arg1220_1, arg1221_1, arg1222_1, arg1223_1, arg1224_1, , arg1226_1, arg1227_1, arg1228_1, , arg1230_1, , , , , , , , , , , , , , , ):
^
SyntaxError: invalid syntax
```
The syntax errors are due to inputs that are `None` when exporting. Prior to changes in https://github.com/pytorch/pytorch/pull/123590 (landed 4/2024), input specs for none inputs look like `InputSpec(userInput=UserInputSpec(arg=Argument(asNone=True)))`, and during deserialization when creating a node, we would just use a dummy name `arg`. After to those changes, the input specs for none inputs look like `InputSpec(constantInput=InputToConstantInputSpec(name='y', value=ConstantValue(asNone=True)))`, and when creating a node we would use the name `y` as the name. However the PR didn't handle the case if it's loading an old package which doesn't have this name, so ended up putting empty names in the placeholder nodes.
This error was uncovered after https://github.com/pytorch/pytorch/pull/149717, where we now use the GraphModule's python codegen to run the UnflattenedModule instead of going through the interpreter path. The placeholder nodes having empty names caused the python codegen to fail.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150515
Approved by: https://github.com/yushangdi
Summary: In the dashboard measurement script, AOTI needs to run Eager first to register the output pytree, so the peak memory compression ratio on the dashboard is always close to 1. Update AOTI run to use an extra warmup run, so the peak memory compression ratio measures the result at the run time instead of the compile time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150534
Approved by: https://github.com/yushangdi
# Motivation
Currently, in Pytorch XPU, `cudaStream_t` is mapped to `sycl::queue&`, so an implicit cast from `XPUStream` to `sycl::queue&` is provided just like `CUDAStream` has an implicit cast to `cudaStream_t`.
But on the SYCLomatic side, we migrate `cudaStream_t` to `sycl::queue*` but not `sycl::queue&` (One reason is that `cudaStream_t` is actually a pointer so users can do anything with that integer. Another reason is that the early `sycl::queue` was not impl-ed by a pointer, so copy by value is not desirable.)
Without this PR:
```
cudaStream_t a = getCurrentCUDAStream();
cudaStream_t b = getCurrentCUDAStream().stream();
```
need be migrated to:
```
queue_ptr a = &(sycl::queue&)getCurrentXPUStream();
queue_ptr b = &(getCurrentXPUStream().queue());
```
With this PR:
```
queue_ptr a = getCurrentXPUStream();
queue_ptr b = &(getCurrentXPUStream().queue());
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148646
Approved by: https://github.com/guangyey, https://github.com/EikanWang
Changes decomposition behavior of `aten.to` to respect the aliasing/non-aliasing behavior in eager, and to specialize to the input/conversion dtype & device.
Before change: we always decompose `aten.to` into `_to_copy`, regardless of aliasing behavior. This leads us to ban mutations on the result of `_to_copy` when aliased, since we can't guarantee correct program semantics. This meant users had to explicitly call `.clone()` before mutating. In the special cases where we don’t ban mutations (e.g. dtype conversion), we add runtime assertions on the input & conversion dtype/devices in the decomposed program (see https://github.com/pytorch/pytorch/pull/142420).
After change: we decompose to the aliasing/non-aliasing behavior that matches eager, allowing mutations in all cases. We also add dtype/device assertions for all `aten.to` ops, starting in the pre-dispatch graph, basically specializing the program to the dtype/devices.
Differential Revision: D71229547
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149235
Approved by: https://github.com/tugsbayasgalan
Summary: make a check function for each input to avoid too large to optimize error on `__check_inputs_outputs`
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:test_aot_inductor -- -r runtime_checks
```
Differential Revision: D72286280
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150553
Approved by: https://github.com/desertfire
Summary:
Codegen used to generate tmp_arg_{index} as temporary args, and index is the position of the caller.
We changed the logic of codegen such that we can reuse previous generated samples, and only delete after arg is no longer used. In this case, we need to make {index} unique, since different functions could reuse the same "tmp_arg_{index}" name string, but corresponds to different args.
Test Plan: `python test/inductor/test_aot_inductor.py -k test_autotuning_args_reuse`
Differential Revision: D72297084
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150522
Approved by: https://github.com/desertfire, https://github.com/22quinn
Summary:
My commandeer of https://github.com/pytorch/pytorch/pull/150102
Based on description of PR it seems that we need to add C calls for each starting python event with a callable such that when the tracing exits we will have a matching enter for any given exit. It adds some unnecessary events at worst but prevents segfaults/failures. My PR just cleans up some refcount impl and logging.
Contributors: @arjun-choudhry
Test Plan: Ran resnet test internally. Will check CI and ask reviewers to make sure it resolves their issues.
Differential Revision: D72207570
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150370
Approved by: https://github.com/aaronenyeshi
This patch effectively ignores traceable_tensor_subclasses, allowing
Dynamo to always try tracing into the `__torch_function__` of tensor
subclass. This helps us with 2 things:
1. allowing users to directly benefit from better compilation of tensor
subclass, by just upgrading pytorch, without having to change legacy
library code (see earlier patches in the stack for examples).
2. potentially exposing more issues in compiling tensor subclass, so we
can get signals and improve them.
As a consequence, it exposed and fixes 2 subtle bugs:
1. In `build_torch_function_fn`, we could get
`torch._C._disabled_torch_function_impl` because we have a
`Parameter` subclass without `__torch_function__` override or if we
have a tensor subclass with `__torch_dispatch__` override. We graph
break on this for now, and plan to add support -- the logic for
simulating `torch._C._disabled_torch_function_impl` is already in
`SuperVariable`, we just need to reuse it.
2. Sometimes we create `SyntheticLocalSource` and need to remove all the
guards installed on it, but we only removed the ones whose source
_is_ the created synthetic source `s`, but forgot about chained
source like `s.foo`, this showed up as
`SYNTHETIC_LOCAL['tmp_0'].__torch_function__.__func__`.
Differential Revision: [D71906141](https://our.internmc.facebook.com/intern/diff/D71906141)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149792
Approved by: https://github.com/jansel, https://github.com/mlazos
ghstack dependencies: #149482, #149483, #149484
This fixes most of the "torch.compile X tensor-subclass" issues
encountered in https://github.com/city96/ComfyUI-GGUF/issues/118. The
relevant tensor subclass definition is here:
298192ed60/ops.py (L18-L65).
A few things to note about the tensor subclass:
1. it overrides a lot of the `torch.Tensor` methods (e.g., `to`,
`clone`), so this patch updates `TensorWithTFOverrideVariable.var_getattr`
to support that.
2. it overrides the `shape` property, so this patch updates
`TensorWithTFOverrideVariable.var_getattr` to support property as well.
3. it has calls to `torch.Tensor.size`, which returns `torch.Size`,
which gets reconstructed in `torch.Tensor.__torch_function__`, so
this patch adds support for calling `torch.Size(...)` on non-constant
inputs.
Differential Revision: [D71906137](https://our.internmc.facebook.com/intern/diff/D71906137)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149484
Approved by: https://github.com/jansel, https://github.com/mlazos
ghstack dependencies: #149482, #149483
This fixes most of https://github.com/huggingface/diffusers/issues/10795,
except for `torch.Tensor._make_subclass`, which will be fixed in a
subsequent patch.
The relevant tensor subclass from the aforementioned issue is defined
here: fbf6b856cc/src/diffusers/quantizers/gguf/utils.py (L398-L435).
There are two things to note about the tensor subclass:
1. it calls `super().__torch_function__`, which is
`torch._C._disabled_torch_function_impl`, so this patch updates
`SuperVariable.call_method` to handle it (we can't do a simpler
polyfill due to some bug with `var_getattr` raising
`NotImplementedError`, which forgot to restore symbolic context).
2. it sets and reads attributes (`quant_type`), and
defines new methods (`as_data`), so this patch adds support for those.
3. it has a `__init__`, which Dynamo needs to trace through in
`TensorSubclassVariable.call_function`.
Differential Revision: [D71906140](https://our.internmc.facebook.com/intern/diff/D71906140)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149482
Approved by: https://github.com/jansel, https://github.com/mlazos
This was failing due to pybind being strict about their cmake version
requirements.
This resolves errors like:
```
652.1 Compatibility with CMake < 3.5 has been removed from CMake.
652.1
652.1 Update the VERSION argument <min> value. Or, use the <min>...<max> syntax
652.1 to tell CMake that the project requires at least <min> but has been updated
652.1 to work with policies introduced by <max> or earlier.
652.1
652.1 Or, add -DCMAKE_POLICY_VERSION_MINIMUM=3.5 to try configuring anyway.
652.1
652.1
652.1 -- Configuring incomplete, errors occurred!
```
Tested this locally with the following command:
```
./build.sh pytorch-linux-jammy-py3.12-halide -t 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/pytorch-linux-jammy-py3.12-halide:8a8989876ff1aa1d5b0e465177afebbc7a9da921
```
Closes https://github.com/pytorch/pytorch/issues/150420
Signed-off-by: Eli Uriegas <eliuriegas@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150560
Approved by: https://github.com/clee2000, https://github.com/ZainRizvi, https://github.com/atalman, https://github.com/malfet
Install nccl in the docker image (which is already being done in some docker images), and use USE_SYSTEM_NCCL=1 in CI builds
It takes some time to build nccl and doesn't happen in parallel, so theres less benefit in switching to a bigger runner and using more processes
The other changes in this PR are because there is an install_cuda script and an install_cuda_aarch64 script and they both build nccl from source and define their own pins for the nccl version. There is also a .ci/docker/nccl-cu11.txt and cu12.txt that define the pins, and this is an attempt to unify them. Unfortunately this leads to a lot of files needing to be copied to the docker build
Generally seems to increase docker pull times by <1 min, P1768456379 but its hard to tell what the real increase is
15761 mib -> 16221 [linux-focal-cuda11.8-py3.10-gcc9 / test (distributed](https://github.com/pytorch/pytorch/actions/runs/14114171729/job/39545500161#logs)
`jq '[.layers[].size, .config.size] | add / 1024 / 1024'`
Example 6eb3c2e282 (39520169577-box)
TODO:
* Figure out a way to verify that nccl was built + works properly when it is expected (this time i just checked torch.distributed.is_nccl_available)
* Merge the cusparse installation scripts
* Merge the cuda installation scripts
* Either split the nccl, cuda, and cusparse installations always, or make the always together in one bash script
distributed/test_distributed_spawn
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150226
Approved by: https://github.com/seemethere, https://github.com/atalman
1. Fixes Cmake update error: https://github.com/pytorch/pytorch/actions/runs/14223930697/job/39858632864
```
CMake Error at CMakeLists.txt:1 (cmake_minimum_required):
Compatibility with CMake < 3.5 has been removed from CMake.
Update the VERSION argument <min> value. Or, use the <min>...<max> syntax
to tell CMake that the project requires at least <min> but has been updated
to work with policies introduced by <max> or earlier.
Or, add -DCMAKE_POLICY_VERSION_MINIMUM=3.5 to try configuring anyway.
```
2. Removes deprecated CUDA 12.4 build
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150549
Approved by: https://github.com/clee2000
This patch effectively ignores traceable_tensor_subclasses, allowing
Dynamo to always try tracing into the `__torch_function__` of tensor
subclass. This helps us with 2 things:
1. allowing users to directly benefit from better compilation of tensor
subclass, by just upgrading pytorch, without having to change legacy
library code (see earlier patches in the stack for examples).
2. potentially exposing more issues in compiling tensor subclass, so we
can get signals and improve them.
As a consequence, it exposed and fixes 2 subtle bugs:
1. In `build_torch_function_fn`, we could get
`torch._C._disabled_torch_function_impl` because we have a
`Parameter` subclass without `__torch_function__` override or if we
have a tensor subclass with `__torch_dispatch__` override. We graph
break on this for now, and plan to add support -- the logic for
simulating `torch._C._disabled_torch_function_impl` is already in
`SuperVariable`, we just need to reuse it.
2. Sometimes we create `SyntheticLocalSource` and need to remove all the
guards installed on it, but we only removed the ones whose source
_is_ the created synthetic source `s`, but forgot about chained
source like `s.foo`, this showed up as
`SYNTHETIC_LOCAL['tmp_0'].__torch_function__.__func__`.
Differential Revision: [D71906141](https://our.internmc.facebook.com/intern/diff/D71906141)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149792
Approved by: https://github.com/jansel, https://github.com/mlazos
ghstack dependencies: #149482, #149483, #149484
This fixes most of the "torch.compile X tensor-subclass" issues
encountered in https://github.com/city96/ComfyUI-GGUF/issues/118. The
relevant tensor subclass definition is here:
298192ed60/ops.py (L18-L65).
A few things to note about the tensor subclass:
1. it overrides a lot of the `torch.Tensor` methods (e.g., `to`,
`clone`), so this patch updates `TensorWithTFOverrideVariable.var_getattr`
to support that.
2. it overrides the `shape` property, so this patch updates
`TensorWithTFOverrideVariable.var_getattr` to support property as well.
3. it has calls to `torch.Tensor.size`, which returns `torch.Size`,
which gets reconstructed in `torch.Tensor.__torch_function__`, so
this patch adds support for calling `torch.Size(...)` on non-constant
inputs.
Differential Revision: [D71906137](https://our.internmc.facebook.com/intern/diff/D71906137)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149484
Approved by: https://github.com/jansel, https://github.com/mlazos
ghstack dependencies: #149482, #149483
This fixes most of https://github.com/huggingface/diffusers/issues/10795,
except for `torch.Tensor._make_subclass`, which will be fixed in a
subsequent patch.
The relevant tensor subclass from the aforementioned issue is defined
here: fbf6b856cc/src/diffusers/quantizers/gguf/utils.py (L398-L435).
There are two things to note about the tensor subclass:
1. it calls `super().__torch_function__`, which is
`torch._C._disabled_torch_function_impl`, so this patch updates
`SuperVariable.call_method` to handle it (we can't do a simpler
polyfill due to some bug with `var_getattr` raising
`NotImplementedError`, which forgot to restore symbolic context).
2. it sets and reads attributes (`quant_type`), and
defines new methods (`as_data`), so this patch adds support for those.
3. it has a `__init__`, which Dynamo needs to trace through in
`TensorSubclassVariable.call_function`.
Differential Revision: [D71906140](https://our.internmc.facebook.com/intern/diff/D71906140)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149482
Approved by: https://github.com/jansel, https://github.com/mlazos
Mutable custom operators get wrapped into an auto_functionalized HOP, so
we need to store the arg_kwarg_vals on the auto_functionalized HOP
itself.
When Inductor does the re-inplacing, it'll use the pattern matcher to
decompose the auto_functionalized HOP back into the original op (and
0+ other view or clone operations). The pattern matcher uses the
arg_kwarg_vals to trace the subgraph to do the decomposition, so it
ultimately sets arg_kwarg_vals on the original op's node correctly.
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148091
Approved by: https://github.com/eellison
ghstack dependencies: #148046, #148063
Instead of always propagating arg_kwarg_vals in _COPY_META_FIELDS, we
special-case the pattern matcher to propagate arg_kwarg_vals when
it sees triton_kernel_wrapper_functional.
The strategy is:
1) trace out the replacement graph with arg_kwarg_vals (which have accurate eager-mode metadata)
2) trace out the replacement graph with vals (which have the accurate Inductor metadata)
3) Propagate the arg_kwarg_vals from the first graph to the second.
4) Use the second graph as the replacement graph.
The strategy is this because we want to extend this to handle
auto_functionalized later up in the stack.
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148046
Approved by: https://github.com/eellison
Add includes for torch.device, torch.dtype, torch.layout, and torch.memory_format to the cpp_wrapper common header, so that they get precompiled. Additionally, add move constructors and operator bool to RAIIPyObject.
Closes#142005.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149350
Approved by: https://github.com/desertfire
Previously, cudagraph is skipped if the graph contains any meta tensor. However, we should not skip since meta tensor does not have actual computation. This PR fixes the issue.
### Example
```python
import torch
def foobar(x, y):
return x * 2, y * 3
foo_c = torch.compile(mode="reduce-overhead")(foobar)
t = torch.empty((1, 16, 128, 128), device="meta")
y = torch.rand([64], device="cuda")
eager_out = foobar(t, y)
for _ in range(3):
compiled_out = foo_c(t, y)
```
Prior to this PR, above code leads to
```
skipping cudagraphs due to multiple devices: device(type='cuda', index=0), device(type='meta')
```
With this PR, we don't skip.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150478
Approved by: https://github.com/eellison
Adds an `Any` return type annotation to `__getattr__` methods in `torch/_ops.py` that return a union of types. Attribute access returning a union of types can cause issues downstream because consumers would need to handle all of the possible types to make the type checker happy. This doesn't seem to matter today for mypy, presumably because `Any` is always inferred when a return type annotation is missing, but it still makes explicit what mypy is already doing implicitly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150204
Approved by: https://github.com/malfet
operations
Summary:
Fix the test for memory tracking. This PR does:
(1) Add tracking before and after for all memory-related operations.
Make sure the operation do indeed captures memory both in CUDA and
torch's CUDACachAllocator Make sure the operation do indeed captures
consumed memory both in CUDA and torch's CUDACachAllocator.
(2) Keep track of memory being reserved by CUDACacheAllocator in
torch and it's relationship with global CUDA memory consumption.
Test Plan:
This PR is adding tests.
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150269
Approved by: https://github.com/jingsh, https://github.com/chenyang78, https://github.com/desertfire
Summary:
My commandeer of https://github.com/pytorch/pytorch/pull/150102
Based on description of PR it seems that we need to add C calls for each starting python event with a callable such that when the tracing exits we will have a matching enter for any given exit. It adds some unnecessary events at worst but prevents segfaults/failures. My PR just cleans up some refcount impl and logging.
Test Plan: Ran resnet test internally. Will check CI and ask reviewers to make sure it resolves their issues.
Differential Revision: D72207570
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150370
Approved by: https://github.com/aaronenyeshi
Summary:
Updates the meta registration for `torch._scaled_mm` to work for the
nvfp4 recipe.
Test Plan:
```bash
pytest test/test_matmul_cuda.py -s -k test_blockwise_nvfp4
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150462
Approved by: https://github.com/eellison
Summary:
X-link: https://github.com/facebookincubator/gloo/pull/423
This modifies `connectFullMesh` to take in a shared_ptr<IStore> instead of a reference. This is an API breaking change but fairly easy to work around.
To have backwards compatibility in PyTorch during the commit phase we add a new ifdef `GLOO_SHARED_STORE` which can provide backwards compatibility until we update the pinned Gloo version in pytorch OSS repo.
This also adds a new `wait_get` method to `IStore` which will allow us to do a more efficient operation in PyTorch TCPStore. PyTorch's `Store::get` automatically waits so we want to make sure we can avoid waiting twice to reduce network traffic.
This change will land simultaneously in PyTorch and Gloo repos.
Test Plan:
```
buck2 test //gloo/... //caffe2/caffe2/contrib/gloo:
```
Differential Revision: D72084111
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150230
Approved by: https://github.com/fduwjj
Summary: Update the GEMM template to include the necessary `tl.assume` annotations to enable bufferops with AMD.
Test Plan: Tested manually with a simple matmul run with torch.complie(f, mode="max-autotune") the environment variables TRITON_ALWAYS_COMPILE=1 AMDGCN_ENABLE_DUMP=1 AMDGCN_USE_BUFFER_OPS=1.
Inspecting the generated AMDGCN all loads/stores use bufferops.
Note: Since inductor is loading constants for many of the shape values assumes are generally not needed for the stride/shape information, but pid calculations are generally a gap in Triton's inference capability.
Differential Revision: D71922698
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150373
Approved by: https://github.com/eellison
The `reinplace_fsdp_all_gather` pass is currently only for Traceable FSDP2 and doesn't work together with SimpleFSDP. We should hide the pass behind `skip_fsdp_hooks` config which makes it only apply to Traceable FSDP2.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150436
Approved by: https://github.com/BoyuanFeng
Summary:
Instead of explicitly specifying dynamic shapes, it is possible to infer them from additional example inputs. Together with the example inputs provided to export, we can basically make any varying dim dynamic and keep any fixed dim static. This should be useful for prod scenarios that have access to tests and/or profiling data, yet are somewhat removed from the model authoring process.
However this alone is not satisfactory: the exported program by design has only one graph, representing one path through the model, and we cannot necessarily guarantee that this graph works for the additional example inputs because different guards might have been created if we had exported with them instead (corresponding to different traced paths). However, checking that the additional example inputs satisfy the guards created by the original export should be sufficient for generalization.
Now, while we don't preserve all guards in the exported program, we do check a subset of them as part of input matching. So we add a verification step at the end of export when such additional example inputs are provided. This should be enough for now.
Test Plan: added test (positive and negative cases)
Differential Revision: D72001771
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150144
Approved by: https://github.com/bobrenjc93
Smoke Test - disable pypi package validation for binaries that package cuda libs. These binaries do not install packages via pypi.
Should Resolve this from `linux-binary-manywheel / manywheel-py3_11-cuda12_6-full-test / test`:
```
Traceback (most recent call last):
File "/pytorch/.ci/pytorch/smoke_test/smoke_test.py", line 468, in <module>
main()
File "/pytorch/.ci/pytorch/smoke_test/smoke_test.py", line 462, in main
smoke_test_cuda(
File "/pytorch/.ci/pytorch/smoke_test/smoke_test.py", line 274, in smoke_test_cuda
compare_pypi_to_torch_versions(
File "/pytorch/.ci/pytorch/smoke_test/smoke_test.py", line 220, in compare_pypi_to_torch_versions
raise RuntimeError(f"Can't find {package} in PyPI for Torch: {torch_version}")
RuntimeError: Can't find cudnn in PyPI for Torch: 9.5.1
```
Link: https://github.com/pytorch/pytorch/actions/runs/14101221665/job/39505479587#step:15:982
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150194
Approved by: https://github.com/ZainRizvi
Needed this class for because `parallelize_module` takes a dict, which doesn't allow `PrepareModuleInput` and `PrepareModuleOutput` to be applied at the same time.
The `PrepareModuleInputOutput` in this PR initializes two variables `prepare_module_input` and `prepare_module_output` and uses them to process module / inputs / outputs.
I had another implementation which put all code in `PrepareModuleInputOutput` and let `PrepareModuleInput` and `PrepareModuleOutput` inherit the monolithic `PrepareModuleInputOutput`. But it is
1. less cleaner
2. conceptually abusing inheritance because `PrepareModuleInput` shouldn't be able to access class methods of `PrepareModuleOutput` and vice versa
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150372
Approved by: https://github.com/wanchaol
The default workspace for hipblaslt is larger than for cublas/cublaslt which requires a slight increase to the buffer needed.
Forward-fix for #150227 that broke ROCm distributed tests but wasn't part of initial CI signal.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150348
Approved by: https://github.com/jeffdaily
When torch.compile is applied to a module via `mod.compile(...)`, it's equivalent to `torch.compile(mod._call_impl)` which takes a different path than `OptimizedModule`. This PR ensures that the `wrap_top_frame` config can also take effect for the `torch.compile(mod._call_impl)` use case.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150209
Approved by: https://github.com/anijain2305
**Summary**
This PR adds a lowering pass for x86 backend
- Patterns of `dequantize -> conv/linear (-> quantize)` are fused to corresponding quantized onednn ops.
- Weights are prepacked ahead of time.
- Post ops of conv/linear are fused if supported.
- The pass returns a `GraphModule` with the modifications mentioned above.
**Test plan**
```
pytest test/quantization/pt2e/test_x86inductor_quantizer.py -k test_lowering_to_x86
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149708
Approved by: https://github.com/jerryzh168, https://github.com/leslie-fang-intel
Before this change:
```console
$ make setup-env-cuda PYTHON="${HOMEBREW_PREFIX}/bin/python3.12"
$ source venv/bin/activate
$ python3 -c 'import torch'
Traceback (most recent call last):
File "<string>", line 1, in <module>
File "/home/PanXuehai/Projects/pytorch/torch/__init__.py", line 379, in <module>
from torch._C import * # noqa: F403
^^^^^^^^^^^^^^^^^^^^^^
ImportError: libcudnn.so.9: cannot open shared object file: No such file or directory
```
This PR adds `site-packages/nvidia/**/lib` to `LD_LIBRARY_PATH` in `venv/bin/activate` script to let NVIDIA PyPI packages can be loaded correctly.
See also:
- #141837
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143262
Approved by: https://github.com/malfet
Altering the flag to use the correct streamType in CUDAPluggableAllocator class for ROCm gpu. The flag TORCH_HIP_VERSION does not work for ROCm as intended. This flag is replaced with USE_ROCM. This is impacting Distributed Fused Adam in Rocm/APEX when using nccl_ub feature. This has been tested with rocm/apex.
See PR https://github.com/ROCm/apex/pull/184
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150010
Approved by: https://github.com/jeffdaily
Relanding #148590 due to merge conflict.
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Squashed contents:
* [ptd][nccl] use current-stream as nccl-stream under async=False mode (#147820)
PTD current workflow:
- PTD creates its own dedicated `ncclStream` for comm operation
- it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective
such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us).
This diff:
- async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead
- async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready
- pass down async from c10d down to NCCL-PG
this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%**
* [PGNCCL] Make avoid-record-stream default
* [c10d] Add asyncOp argument to Ops
* Change python side wait
* Pass asyncOp at ProcessGroup level
* Watchdog unstashing tensors as a safety net
* Stash tensors for reduce_scatter_v and all_gather_v
Pull Request approved: https://github.com/pytorch/pytorch/pull/149753
* [c10d] Move unstashing from watchdog to main thread
Pull Request approved: https://github.com/pytorch/pytorch/pull/150079
* [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation
Pull Request approved: https://github.com/pytorch/pytorch/pull/150130
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150398
Approved by: https://github.com/atalman
In deep learning models, the tanh (hyperbolic tangent) function is a widely used activation function, primarily in feedforward networks, recurrent neural networks (RNNs), and various other architectures.
Also, the tanh (hyperbolic tangent) function is commonly used in **Physics-Informed Neural Networks (PINNs).** PINNs are a class of machine learning models designed to solve partial differential equations (PDEs) by incorporating the governing physics directly into the loss function, along with data-driven terms.
In PINNs, activation functions like tanh are used in the neural network architecture to enable the model to learn complex mappings between inputs (such as spatial and temporal coordinates) and outputs (such as field variables).
**Operator: tanh()**
**Current Implementation in OSS in ATen Backend:**
**SVE Flow:** Uses SVE sleef when available else std implementation.
**With this PR :**
**SVE Flow:** Uses SVE ACLE implementation. (Faster Implementation)
**Here are the performance improvements.**
**Single core perf numbers:**
**Metric:** CPU time avg time per iteration (In ms)
As you can see with both gcc and clang compilers, we see a significant performance gain with SVE ACLE implementation over current OSS Implementation (Sleef) and also Neon.
**Hardware:** m7g.8xlarge (Graviton 3 Instance)
**Script used in benchmarking:**
```python
import os
#os.environ["ATEN_CPU_CAPABILITY"] = "default"
os.environ["ATEN_CPU_CAPABILITY"] = "sve256"
import torch
import torch.nn as nn
#Set the random seed for reproducibility
torch.manual_seed(1)
#Create a tensor of shape (8521, 50)
x = torch.randn(8521, 50)
for i in range(10):
output = x.tanh()
#Perform the tanh operation 1000 times and profile the performance
print("### CPU tanh")
with torch.autograd.profiler.profile(record_shapes=True) as prof:
for i in range(1000):
output = x.tanh()
#Print the profiling results sorted by self CPU time
print(prof.key_averages().table(sort_by="self_cpu_time_total"))
#Optionally print the final output (if needed, uncomment the following line)
print(output)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143741
Approved by: https://github.com/malfet
Changes in this PR:
1. Add `is_structseq` and `is_structseq_class` functions to determine a object or a class is PyStructSequence.
2. Add a generic class `structseq` which can be used as the registration key for PyStructSequence types like `namedtuple` for Named Tuple types.
3. Change `is_namedtuple` to accept subclasses of namedtuple to be namedtuple. Before this PR, only namedtuple class directly created by `collections.namedtuple` or `typing.NamedTuple` were namedtuple classes while their subclasses were not. This PR makes `is_namedtuple` return true for subclasses of namedtuple class.
Resolves#75982. New tests are included in this PR.
- #75982
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113257
Approved by: https://github.com/zou3519
Summary:
dynamo_compile for the most part has been accounting for compile time except autotuning.
all_compilation_types had earlier been injected on fx_codegen_and_compile, which was incorrect.
Add autotuining to dynamo and deprcate all_compilation_types counter.
Differential Revision: D72145447
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150293
Approved by: https://github.com/masnesral, https://github.com/jamesjwu
Per title, this version uses symm mem input both as input source and as a work buffer, so input is modified after the end (similar to what fbgemm car reduction does). It is intended to be wrapped in an op that would first copy the real inputs to symm mem buffers that wouldn't be exposed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150153
Approved by: https://github.com/xw285cornell
Fix https://github.com/pytorch/pytorch/issues/148639.
Summary:
Optimize the heuristics of parallel reduction: When the number of steps of the first inner loop beyond the maximum parallel depth is much larger than the number of steps of all outer loops within the maximum parallel depth, change the starting depth of parallelism to the first inner loop and recalculate the maximum parallel depth. I ran the Inductor benchmark with this PR on CPU. A timm model poolformer_m36 BF16 has about 25% performance improvement, and no performance regression is seen.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149614
Approved by: https://github.com/jgong5, https://github.com/leslie-fang-intel, https://github.com/jansel
Fixes#143071
Operations performed on tensors with `requires_grad=True` such as
```python
import torch
x = torch.tensor(2.0, requires_grad=True)
y = x ** 3
```
and
```python
x = torch.tensor(2.0, requires_grad=True)
y = torch.pow(x,3)
```
are valid operations.
While an operation using `numpy` like
```python
import numpy as np
x = torch.tensor(2.0, requires_grad=True)
y = np.pow(x,3)
# > RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.
```
leads to an error.
However, an operation that uses `math` like
```python
import math
x = torch.tensor(2.0, requires_grad=True)
y = math.pow(x,3)
```
does not cause an error, and `y` is no longer a tensor with a gradient!
This represents a [footgun](https://en.wiktionary.org/wiki/footgun#Noun) for some users, like myself when training small, custom, non-neural network models.
To prevent future undesired behavior, I added a warning when converting tensors with `requires_grad=True` to scalars. Now, when using `math.pow` on a `tensor`, we get a single warning with:
```python
x = torch.tensor(2.0, requires_grad=True)
y = math.pow(x,3)
# > UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
# Consider using tensor.detach() first.
```
Please let me know if you have any questions 👍
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143261
Approved by: https://github.com/malfet
Co-authored-by: albanD <desmaison.alban@gmail.com>
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Previously, scaled_mm's (FP8 matmul) Triton lowering for inductor was in a separate template. This PR consolidates that lowering into the mm template, with an added epilogue to deal with multiplying the scales. This paves the way for future scaled variants of BMM, Grouped GEMM in inductor.
Currently, there is still a separate template for TMA+persistent version of scaled_mm. The current mm lowering has a separate template for TMA + Persistent version. Will hopefully consolidate the extra scaled_mm TMA+persistent template when the consolidation for the mm template is done.
TODO: Consolidate TMA+Persistent logic into 1 template and remove separate scaled_mm TMA template
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150045
Approved by: https://github.com/drisspg
As is the case with many inductor tests, this test adapts test criteria based on device type, where it should be adjusting for the backend registered for that device.
In this particular case, using the upstream triton CPU backend would lead to failures, as reference_in_float would be true as this is required for the C++/OpenMP backend which does not have float16 support. However most triton backends do, and as such should be tested in float16. Similarly a triton backend with a device not described as a GPU would get skipped from testing entirely.
A more generic solution would be ideal, but this would require a lot of work across many tests.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146911
Approved by: https://github.com/masnesral
`tuple[int]` means only a tuple of length 1, which is not what was intended.
```python
loss = torch.masked.mean(loss, mask=mask, dim=(-1, -2)) # Argument of type "tuple[Literal[-1], Literal[-2]]" cannot be assigned to parameter "dim" of type "DimOrDims"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149870
Approved by: https://github.com/Skylion007
This PR fixes a bug with how include directories with spaces are handled on Windows. I ran into an edge case with torch.compile() - it will error out with an exception on Windows. In particular, it will try to execute the following: `cl /I C:/Program Files/Python311/Include ...`, where `C:/Program` will be treated as separate from `Files/Python311/Include`.
I looked into using something like `shlex.quote` or `pathlib.Path`, but I didn't find those options to be suitable (shlex is POSIX shell only, pathlib.Path does not escape spaces).
There is another place in the function that also deals with escaping spaces. My fix follows the same style. 0ff2e6a85a/torch/_inductor/cpp_builder.py (L1464)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148271
Approved by: https://github.com/Skylion007
Co-authored-by: Aaron Gokaslan <aaronGokaslan@gmail.com>
Improvements to unit tests and warnings for unsupported cases in offline tuning. Here are more details:
- Previously we only compared the OpSig for the untuned vs. tuned entries. This was not strict enough so we now compare OpSig+ParamSig.
- The main offline and online UTs are now stricter to make sure we exercise the code paths for the four combinations of transA and transB.
- Offline tuning does not support some tensor shapes. Emit warning and skip tuning.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150142
Approved by: https://github.com/jeffdaily
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
Fixes#149876
## Stack
- [previous PR in stack] https://github.com/pytorch/pytorch/pull/149247
## TL;DR
This PR implements support in async TP for saving the reduce-scatter result for backward, which previously would break the torchtitan AC policies: no AC, per op SAC, and per layer SAC.
## Context
In torchtitan's LLama3 per op SAC policy, we want to save the output of `reduce_scatter` ops for backward, which is useful for TP. The reduce_scatter op is also saved for No AC (since all activations are saved) and per layer SAC (since we save the activations for N full layers, which do contain reduce-scatters for TP.
However, doing this causes incompatibility with Async TP for the AC policies above, for 2 reasons:
1) The graph pattern matching specifically only matches on reduce scatter nodes with 1 user, but reduce_scatter nodes saved for backwards will have 2 users (the 2nd one being the return/output node, which saves it for backward).
2) The subgraph replacement logic which replaces the users of the `wait_tensor` after the reduce-scatter with the new fused node has no mechanism to save the fused_node for backward instead of the reduce-scatter node. This means we cannot directly replace the subgraph, since we can't delete nodes which still have users (in this case, the output node is still using the reduce-scatter node).
To fix this, we do 2 things:
1) Add additional pattern matching logic to also match reduce-scatter nodes with 2 users, so we also perform fusion when reduce-scatter is saved for backward.
2) When replacing the subgraph with the fused node, detect if the reduce-scatter was saved for backward, and if so, save the result of the fused node for backward instead. This enables us to properly erase the subgraph and prevent the memory leak which occurred in #149876
## Other changes
- Continue to throw an error if we don't find any candidate all-gathers or reduce-scatters for fusion (since TP should have both) but DON'T throw an error if we don't fuse any matmul-reduce-scatters. This is because I've found there are actually valid graphs where we do fuse reduce scatters in the forward graph but not the backward graph (in the backward pass there are reduce-scatters but the producer op is an "add" not a mm/scaled_mm).
## Test plan
1. All unit tests are passing
2. Visualized the graphs and verified the fusion is occurring properly.
3. Verified via manual torchtitan runs there is no memory leak / OOM occurring anymore.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149946
Approved by: https://github.com/fegin
This PR adds CachingAutotuners that are statically launchable to FXGraphCache's cache entry.
Regular CachingAutotuners, with triton kernels attached to them, are not very good to cache: they are very large, and take huge amounts of space since they track all of the various binary files, along with various metadata. We could probably figure out what information we could delete from the kernel and have it still work, but with StaticCudaLauncher, we no longer have to. Instead, we can cache every compiled triton kernel that is statically launchable.
Because StaticTritonCompileResult is serializable, and designed to have a very small memory footprint, we can save it into FXGraphCache without increasing the cache size significantly. We store it as a part of CompiledFxGraph.triton_bundle.
Then, on load, we repopulate the CachingAutotuner into our CompiledTritonKernel cache.
The upsides of this are many:
- We no longer need to call into a separate process on cache hit
- We can *guarantee* that the triton kernel we got from our cache entry is the one we use to launch again, so no worries about triton's own caching logic
- Once we achieve feature parity and all torch.compiled triton kernels are statically launchable, we can clean up a bunch of TritonBundler code and simplify the cache hit logic.
Fixes#149449
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149054
Approved by: https://github.com/oulgen
Fixes#143071
Operations performed on tensors with `requires_grad=True` such as
```python
import torch
x = torch.tensor(2.0, requires_grad=True)
y = x ** 3
```
and
```python
x = torch.tensor(2.0, requires_grad=True)
y = torch.pow(x,3)
```
are valid operations.
While an operation using `numpy` like
```python
import numpy as np
x = torch.tensor(2.0, requires_grad=True)
y = np.pow(x,3)
# > RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.
```
leads to an error.
However, an operation that uses `math` like
```python
import math
x = torch.tensor(2.0, requires_grad=True)
y = math.pow(x,3)
```
does not cause an error, and `y` is no longer a tensor with a gradient!
This represents a [footgun](https://en.wiktionary.org/wiki/footgun#Noun) for some users, like myself when training small, custom, non-neural network models.
To prevent future undesired behavior, I added a warning when converting tensors with `requires_grad=True` to scalars. Now, when using `math.pow` on a `tensor`, we get a single warning with:
```python
x = torch.tensor(2.0, requires_grad=True)
y = math.pow(x,3)
# > UserWarning: Converting a tensor with requires_grad=True to a scalar may lead to unexpected behavior.
# Consider using tensor.detach() first.
```
Please let me know if you have any questions 👍
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143261
Approved by: https://github.com/albanD
Co-authored-by: albanD <desmaison.alban@gmail.com>
Summary:
This will log a wait counter with for backward compile and fixes weirdness with nested context managers.
Since the old wait counters added through dynamo_timed were never created with the nesting issue. I am also changing the key nomenclature from `pytorch.dynamo_timed` to `pytorch.wait_counter`. We want to use the same nomenclature, to make it easy to find keys.
Reviewed By: jamesjwu
Differential Revision: D72032055
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150235
Approved by: https://github.com/jamesjwu, https://github.com/masnesral
When generating reduction kernel, otherwise compiler can unroll loops too much that kernel could not be launched for the intended threadgroup size
Extend `c10:🤘:max` to accept different dtypes
Together this fixes `test_large_broadcast_reduction`
TODO:
- Explore different threadgroup_sizes for best perf
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150247
Approved by: https://github.com/jansel, https://github.com/dcci
ghstack dependencies: #150246
The intent of the existing code is to
> // Assign system TIDs to start events based on the system TID of the next
// observed event with the same Python TID.
However, if there are start events that don't share the same Python TID as later observed events, then they are left with the default initialization of DeviceAndResource and assigned values of `0`. This is problematic because Kineto uses `device=0, resource=0` for the first GPU (or other backend) device.
This PR maintains the previous logic of using TIDs from later events if any are present, but defaults to the current process and system thread IDs if there aren't later events to reference.
This issue was discovered while working to implement a custom backend and some CPU start events were appearing on the same process and thread as the device in the trace.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149757
Approved by: https://github.com/sraikund16
Remove the "special linking" that involves listing `BLAS_LIBRARIES` thrice if `TH_BINARY_BUILD` is set, as it should not be any different from listing it just once.
The code seems to date back to commit cfcf2af95f91a88ec61cbcac8b30a718e7332aa5. The original code already listed `BLAS_LIBRARIES` thrice, but it provided no explanation for doing that — and without `TH_BINARY_BUILD`, BLAS was not linked at all. The current version seems to originate in d6a8d28d6529a4f0b80a8c046ca9c36ca6c8b347 — and it already provided an `ELSE` clause listing `BLAS_LIBRARIES` only once. From this, I suspect that it is probably an unnecessary leftover.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145487
Approved by: https://github.com/malfet
#149548 Fixed the arbitrarily missing parallelism for NLL, but they also added an arbritrary #ifdef ROCM guard around this fix to prevent its use on CUDA gpus. There is also a problem with the way the kernel does the reduction from the intermediate shared memory, using only thread 0 walking linearly. This has been changed to a simple parallel reduction algorithm.
Tested changes with `python3 test/test_nn.py`
```
Ran 3551 tests in 200.554s
OK (skipped=998, expected failures=4)
```
Performance before and after with the script below with an RTX 3090, batch size x axis, time (sec) y axis. This GPU is also used for display graphics and such, so the measurements are pretty noisy, even with 100 samples.
## Before
## After ifdef removal
## After Parallel SMEM reduction
```python
import torch
from matplotlib import pyplot as plt
from torch.nn import functional as F
timing = []
batches= list(range(32, 4096, 32))
for batch in [32] + batches:
samples = []
for _ in range(100):
probs = torch.rand(batch, 10).cuda()
labels = torch.randint(0, 10, (batch,)).cuda()
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
F.nll_loss(probs, labels)
end.record()
torch.cuda.synchronize()
elapsed = start.elapsed_time(end)
samples.append(elapsed)
timing.append(sum(samples) / len(samples))
timing = timing[1:]
plt.plot(batches, timing)
plt.show()
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149779
Approved by: https://github.com/jeffdaily
Summary:
The original Diff D69500038 is reverted due to a false alarm on trunk health.
Implement torchbind support in OSSProxyExecutor.
Exactly the same as the implementation in FbProxyExecutor.
D69693697 - fbProxyExecutor
D69887230 - fbProxyExecutor but for torchbind method
D70746626 - Support None output type
Other changes:
- When generating the schema of the CallTrochBind HOP, the arg name of the torchbind object arg should be the same as the torchbind method's torchbind object arg (instead of `obj`).
- In `AOTIModelPackageLoader`, we extract everything in `data/constants` to `tmp_dir/data/aot_inductor/<model>/` folder, so the torchbind objs exist in the same folder as the rest of the files (e.g. cpp, so). This is to be consistent of how files are packaged internally (more details in internal Diff summary).
Note on using `filesystem`:
Seems like there'll be [issues](https://github.com/pytorch/pytorch/pull/137209) with using`filesystem` header in linux, so here I use string manipulation instead of `filesystem::path`.
Test Plan:
```
test/inductor:torchbind -- -r torchbind_aoti
test/inductor:torchbind -- -r aot_compile
```
Differential Revision: D72063691
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150196
Approved by: https://github.com/hl475, https://github.com/desertfire
Summary:
Currently only `num_warps` and `num_stages` are supported as one of the kernel options for inductor auto-tuning using `TritonTemplate`.
In order to allow warp-specialization kernel options should allow specifying `num_consumer_groups` and `num_buffers_warp_spec` as well.
NOTE: Currently gating changes to FBCODE using HAS_WARP_SPEC which is only available on triton/release-3.3.x
Test Plan:
## Unit test
Added tests for `test_triton_template_warp_specialization` to verify generated kenrnel contains configs for `num_consumer_groups` and `num_buffers_warp_spec`.
## Functional Testing
Specific to flexattention.
```
import torch
from torch.nn.attention.flex_attention import flex_attention
from triton.testing import do_bench
make_tensor = lambda: torch.rand(8, 16, 8192, 128, device="cuda", dtype=torch.bfloat16)
q, k, v = make_tensor(), make_tensor(), make_tensor()
flex_compiled = torch.compile(flex_attention, fullgraph=True)
print(do_bench(lambda: flex_compiled(q, k, v, kernel_options={"num_warps": 4})))
```
triton do_bench results:
- default compile: 15.176783561706543
- with warp-spec: 9.452800750732422
## Extra notes
- generated triton kernel using `TORCH_LOGS=output_code`: P1740612877
- TTGIR for fused kernel: P1740614685
Differential Revision: D71982587
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150122
Approved by: https://github.com/eellison, https://github.com/zou3519, https://github.com/jansel
Summary: Add extract_constant_map that allows users to inspect the constants being used by AOTInductor
Test Plan:
`python test/inductor/test_aot_inductor.py -k extract_constants_map`
`LD_LIBRARY_PATH=/data/users/$USER/pytorch/build/lib /data/users/$USER/pytorch/build/bin/test_aoti_inference`
Differential Revision: D72020400
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150163
Approved by: https://github.com/chenyang78
This supports `masterListenFd` which is required for full compatibility with the non-libuv TCPStore. The code was just missing a `uv_listen` call and now it works just fine.
This is required to migrate the last remaining uses of TCPStore off of the non-libuv backend.
Test plan:
```
pytest -v test/distributed/test_store.py -k test_take_over_listen_socket
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150215
Approved by: https://github.com/fduwjj
- Distinguish between conjugated/non_conjugated inputs by appending conjugation to the operator key
- For matmul or dot, add `conjugateWithTensor:name:` calls before running the op
- Enable testing for conjugated ops by passing `include_conjugated_inputs` to opinfo
- Filter `include_conjugated_inputs` argument from `sample_inputs_window` (probably should have landed as separate PR)
- Preserve conj property when gathering the views, that fixes `cov` operator
Fixes https://github.com/pytorch/pytorch/issues/148156
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150157
Approved by: https://github.com/dcci
Per title, this version uses symm mem input both as input source and as a work buffer, so input is modified after the end (similar to what fbgemm car reduction does). It is intended to be wrapped in an op that would first copy the real inputs to symm mem buffers that wouldn't be exposed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150153
Approved by: https://github.com/xw285cornell
Enable TD on distributed cpu, I think the only reason it's not is because I forgot to enable it
Get rid of some of the statements that are no ops:
* asan uses default shard
* nogpu got moved to periodic
* no windows cuda testing anymore
Only thing on pull and trunk that doesn't use TD is dynamo_wrapped but I think it's fast enough to be ok for now, we can take another look after this
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150028
Approved by: https://github.com/ZainRizvi
Sometimes I would find a log artifact that only has usage_logs.txt in it, even though there are other logs created by tests. I think this is somehow caused by output buffering with find. I don't understand how, but at the very least, I can see that all the jobs on this PR have the logs from the test runs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149577
Approved by: https://github.com/ZainRizvi
Sees to reduce docker pull times by ~3 min if triton is requested, some compressed docker sizes seems to have decreased by 1/3 ish
Also add check that triton is installed/not installed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149413
Approved by: https://github.com/malfet
I'm not sure if this is the right think to do, but cmake 4.0.0 got released on pypi and our builds are failing with it
Example:
aa70d62041 (39555975425-box)
I guess we have to go change all the cmake_minimum_required to >=3.5?
backwards compat still failing because its building with the base commit which this pr can't really change until it gets merged, but at least manywheel binary builds got past where they were originally failing
Also pin the conda installation, but the most recent version on conda is 3.31.2
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150158
Approved by: https://github.com/cyyever, https://github.com/malfet
The cpp contexts are only supported on x86 Linux.
The tests requiring them are skipped on non-Linux but not if the architecture is not x86.
In most places it is checked for ARM64 which is not enough as a check for x86 is required instead.
Fix the test decorators and factor out a common one in test_cuda.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148445
Approved by: https://github.com/eellison
This PR adds CachingAutotuners that are statically launchable to FXGraphCache's cache entry.
Regular CachingAutotuners, with triton kernels attached to them, are not very good to cache: they are very large, and take huge amounts of space since they track all of the various binary files, along with various metadata. We could probably figure out what information we could delete from the kernel and have it still work, but with StaticCudaLauncher, we no longer have to. Instead, we can cache every compiled triton kernel that is statically launchable.
Because StaticTritonCompileResult is serializable, and designed to have a very small memory footprint, we can save it into FXGraphCache without increasing the cache size significantly. We store it as a part of CompiledFxGraph.triton_bundle.
Then, on load, we repopulate the CachingAutotuner into our CompiledTritonKernel cache.
The upsides of this are many:
- We no longer need to call into a separate process on cache hit
- We can *guarantee* that the triton kernel we got from our cache entry is the one we use to launch again, so no worries about triton's own caching logic
- Once we achieve feature parity and all torch.compiled triton kernels are statically launchable, we can clean up a bunch of TritonBundler code and simplify the cache hit logic.
Fixes#149449
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149054
Approved by: https://github.com/oulgen
This PR was inspired by internal models that were cache missing due to PGO. At a high level the problem looks as follows
Run 1, Invocation 1: We do static compile, save some example values in PGO/automatic dynamic
Run 1, Invocation 2: We detect varying inputs, do dynamic compile, get a dynamic graph and save to PGO. Crucially what we save to PGO is actually a superset of what is actually dynamic. If we notice an input was varying, we mark it as dynamic in PGO even if later on that value gets specialized. When a value gets specialized, we actually remove the symbol from the graph. This results in an interesting conundrum where although we are producing the same isomorphic graph, PGO makes the second run cache miss. Let's see how....
Run 2, Invocation 1: We fetch the PGO, over-mark things as dynamic, get a fx graph, look it up in the cache and... whoops! cache miss! This is because of the aforementioned behavior where the PGO profile will cause us to over-allocate symbols. In practice this means we end up saving a graph in cache with symbols x:s1, y:s3 and on second attempt we cache miss with x:s1, y:s6 where symbols s3,s4,s5 were all optimistically marked dynamic by PGO and subsequently specialized.
We solve this problem by hashing the source names. This ensures somewhat stable assignment. To prevent catastrophic symbol collisions, we use linear probing to ensure no collisions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149665
Approved by: https://github.com/Mingming-Ding, https://github.com/laithsakka
This counter is designed to include all compilation pytorch does (triton +
dynamo_compile). However this wasn't including all of dynamo compilation, since
it was put in at the fx_codegen_and_compile spot.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149664
Approved by: https://github.com/masnesral
Summary: A couple follow-ups noted in review from https://github.com/pytorch/pytorch/pull/149700:
1. Make sure we correctly signal _all_ subproces to shutdown, even in the case where some processes are currently benchmarking.
2. Change how the pool singleton is created. That also allows us to fully initialize the object in the ctor and remove a bunch of asserts.
Test Plan: existing unit tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149890
Approved by: https://github.com/aorenste
ghstack dependencies: #149700
Summary: The primary change is to update the autotune-in-a-subproc implementation to avoid using multiprocessing spawn. Spawn (re)executes the toplevel script in the subproc, which can be problematic. The approach here is similar to Triton parallel compile: we Popen a subproc on a controlled entry point and communicate over pipes. That change drove a lot of refactoring in the TuningProcess class, so I took the opportunity to simplify some things, rename some methods, etc.
One other notable change is around the timeout / kill approach. After a timeout, we were previously attempting to stop the subproc in three steps (graceful shutdown, sigkill if graceful fails, sigterm if sigkill fails). I'm gonna argue think that's not useful: 1) The graceful shutdown is never going to work unless the subproc happens to have just completed its task and is ready to receive the next command. 2) If we're going to kill the subproc, let's just take the most aggressive approach and move on as quickly as possible to restarting it rather than waiting to see if previous shutdown attempts succeeded. The only downside that I can find find is maybe a little log spew?, e.g., ` ResourceWarning: subprocess 2987680 is still running`
List of changes:
* Use Popen instead of spawn for the autotuning subprocess.
* Introduced a new entry point `__autotune_main__.py`
* Renamed some TuningProcess methods. For example `shutdown` makes more sense than `terminate` because the latter implies a forced kill.
* Simplified the implementation around benchmarking timeout and how we kill the subproc after a timeout.
* Deprecated the unused timeout configs in `_inductor/config.py`
* Moved `get_ld_library_path` helper to a common utils file.
* Added more unit tests for subproc crashes / timeouts / exceptions, etc.
Test plan:
* New unit tests
* Also ran internally with all combinations of: build mode `opt` and `dev-nosan`, and `buck run` vs. executing the `.par` file directly.
* Made sure the functionality to parallelize autotuning across different GPUs is working (it wasn't clear to me this was behaving the way we wanted it to).
Differential Revision: [D71976971](https://our.internmc.facebook.com/intern/diff/D71976971)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149700
Approved by: https://github.com/aorenste, https://github.com/jansel, https://github.com/eellison
Summary:
Implement torchbind support in OSSProxyExecutor.
Exactly the same as the implementation in FbProxyExecutor.
D69693697 - fbProxyExecutor
D69887230 - fbProxyExecutor but for torchbind method
Other changes:
- When generating the schema of the CallTrochBind HOP, the arg name of the torchbind object arg should be the same as the torchbind method's torchbind object arg (instead of `obj`).
- In `AOTIModelPackageLoader`, we extract everything in `data/constants` to `tmp_dir/data/aot_inductor/<model>/` folder, so the torchbind objs exist in the same folder as the rest of the files (e.g. cpp, so). This is to be consistent of how files are packaged internally
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:torchbind -- -r torchbind_aoti
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:torchbind -- -r aot_compile
```
Differential Revision: D69500038
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149747
Approved by: https://github.com/desertfire
Fixes:
- https://github.com/pytorch/pytorch/issues/101138
**Description**
The PR enhances error handling in `_check_cuda_version` by verifying the existence of the `nvcc` executable before invoking `subprocess.check_output`. If `nvcc` is missing, a `FileNotFoundError` is raised with a clear message, guiding users to check their CUDA installation and path configuration.
**Testing**
Manually tested with and without `nvcc` present in the expected path.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148671
Approved by: https://github.com/malfet
Add includes for torch.device, torch.dtype, torch.layout, and torch.memory_format to the cpp_wrapper common header, so that they get precompiled. Additionally, add move constructors and operator bool to RAIIPyObject.
Closes#142005.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149350
Approved by: https://github.com/desertfire
ghstack dependencies: #147225
Originally, I excluded constant_pad_nd from fusing to be conservative on compilation time. But, on benchmarking, you do occasionally get speedups by fusing it. Also includes a fix for making single, contiguous dep for prologues.
For instance, the following benchmark gets a 7% speedup by fusing in the constant_pad_nd.
```
import torch
import torch.nn.functional as F
torch._inductor.config.force_disable_caches = True
padded_N = 2048
n_pad_rows = 100
K, N = 2048, 4096
tensor1 = torch.randn(padded_N - n_pad_rows, 4096, device="cuda").to(torch.bfloat16)
tensor2 = torch.randn(4096, 4096, device="cuda").to(torch.bfloat16)
@torch.compile(mode='max-autotune-no-cudagraphs')
def masked_linear(input, weight, n_pad_input_rows):
"""
Linear layer with input padded by `n_pad_input_rows` rows
"""
# Use constant_pad_nd to pad with zeros for the invalid rows
padded_input = F.pad(tensor1, (0, 0, 0, n_pad_input_rows), "constant", 0)
return F.linear(padded_input, weight)
# Invoke the function
masked_linear(tensor1, tensor2, n_pad_rows)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149947
Approved by: https://github.com/drisspg
During tracing it is possible for a `s1: VR[2, inf]` to be replaced by a `s0: VR[3, inf]` (note smaller range) by the shape env. But after export, unfortunately we'd previously record `range_constraints[s0] = VR[2, inf]` (note larger range), which is incorrect.
This is because we'd map `s1.node.expr` (`s0`) to the `var_to_range` of `s1.node._expr` (`s1`) when creating `range_constraints`. The comment surrounding this code suggests this predated `bound_sympy`, but now we can do better.
For users, this means that when using `Dim.DYNAMIC` previously they wouldn't get input constraints checked sufficiently, now they do (shifting errors early).
Differential Revision: D71962694
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150103
Approved by: https://github.com/zhxchen17
Summary: Given an explicit error when torchbind object is used as input to AoTI
Test Plan:
```
buck run fbcode//mode/dev-nosan //caffe2/test/inductor:torchbind -- -r test_torchbind_input
```
Differential Revision: D69490915
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149965
Approved by: https://github.com/desertfire
Summary:
When `a` and `b` have dtype `torch.float4_e2m1fn_x2` and `a_scale` and `b_scale` have dtype `torch.float8_e4m3fn`, makes
```python
c = torch._scaled_mm(a, b, a_scale, b_scale, out_dtype=torch.bfloat16)
```
call the cuBLAS fp4 gemm kernel, as specified in https://docs.nvidia.com/cuda/cublas/index.html?highlight=fp4#d-block-scaling-for-fp8-and-fp4-data-types
note: output scale (`scale_in_D` from the cuBLAS docs) is not tested in this PR - we can enable in a follow-up.
Test Plan:
```bash
pytest test/test_matmul_cuda.py -s -k mxfp8_nvfp4
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148792
Approved by: https://github.com/eqy
ghstack dependencies: #148791
This PR adds CachingAutotuners that are statically launchable to FXGraphCache's cache entry.
Regular CachingAutotuners, with triton kernels attached to them, are not very good to cache: they are very large, and take huge amounts of space since they track all of the various binary files, along with various metadata. We could probably figure out what information we could delete from the kernel and have it still work, but with StaticCudaLauncher, we no longer have to. Instead, we can cache every compiled triton kernel that is statically launchable.
Because StaticTritonCompileResult is serializable, and designed to have a very small memory footprint, we can save it into FXGraphCache without increasing the cache size significantly. We store it as a part of CompiledFxGraph.triton_bundle.
Then, on load, we repopulate the CachingAutotuner into our CompiledTritonKernel cache.
The upsides of this are many:
- We no longer need to call into a separate process on cache hit
- We can *guarantee* that the triton kernel we got from our cache entry is the one we use to launch again, so no worries about triton's own caching logic
- Once we achieve feature parity and all torch.compiled triton kernels are statically launchable, we can clean up a bunch of TritonBundler code and simplify the cache hit logic.
Fixes#149449
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149054
Approved by: https://github.com/oulgen
ghstack dependencies: #149657
Summary:
To align with thrift-python, we are adding the int base class for `non-Flag` enums. In order to not break production code, the annotation `python.NoIntBaseClassDeprecated` is added to opt-out some enums
After the related customer code logic changes, we can now safely remove the annotations that were added earlier.
Our ultimate goal is to unconditionally add the `int` base to `thrift-py3` enums.
Test Plan:
```
buck test 'fbcode//mode/opt' fbcode//caffe2/torch/fb/training_toolkit/applications/bulk_eval/tests:evaluator_test -- --exact 'caffe2/torch/fb/training_toolkit/applications/bulk_eval/tests:evaluator_test - test_setup_evaluation_utils (caffe2.torch.fb.training_toolkit.applications.bulk_eval.tests.evaluator_test.EvaluatorTest)'
```
Reviewed By: ahilger
Differential Revision: D71446522
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149744
Approved by: https://github.com/izaitsevfb, https://github.com/huydhn
Summary:
The _load_state_dict_from_keys method specifies that `Loads any key specified in this set. If no keys are specified, the entire checkpoint is loaded.`
But this isn't happening right now, because an empty keys arg is passed in as a set() to `_load_state_dict` and keys is expected to be None for it to actually be included in the state_dict https://fburl.com/code/l8yzojyx. So with the set() argument, the state_dict is always going to be empty
Test Plan: ensure existing tests pass
Differential Revision: D71930712
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150058
Approved by: https://github.com/saumishr
Summary:
Currently only `num_warps` and `num_stages` are supported as one of the kernel options for inductor auto-tuning using `TritonTemplate`. In order to allow warp-specialization kernel options should allow specifying `num_consumer_groups` and `num_buffers_warp_spec` as well.
Test Plan:
## Unit test
Added tests for `test_triton_template_warp_specialization` to verify generated kenrnel contains configs for `num_consumer_groups` and `num_buffers_warp_spec`.
## Functional Testing
Specific to flexattention.
```
import torch
from torch.nn.attention.flex_attention import flex_attention
from triton.testing import do_bench
make_tensor = lambda: torch.rand(8, 16, 8192, 128, device="cuda", dtype=torch.bfloat16)
q, k, v = make_tensor(), make_tensor(), make_tensor()
flex_compiled = torch.compile(flex_attention, fullgraph=True)
print(do_bench(lambda: flex_compiled(q, k, v, kernel_options={"num_warps": 4})))
```
triton do_bench results:
- default compile: 15.176783561706543
- with warp-spec: 9.452800750732422
## Extra notes
- generated triton kernel using `TORCH_LOGS=output_code`: P1740612877
- TTGIR for fused kernel: P1740614685
Differential Revision: D70212243
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148503
Approved by: https://github.com/eellison
Currently, as is the case with many inductor devices are assumed to be one of:
- CPU with Cpp coden, or
- GPU with triton codegen
This is not always the case, a CPU backend may be using the triton CPU backend, or some other codegen entirely. This goes some way to fixing it in the case where a CPU backend can use triton scheduling.
A more general solution could be implemented, but this would need to be quite robust, and is probably best done more centrally and by someone who can do more testing with CUDA devices.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146830
Approved by: https://github.com/eellison, https://github.com/albanD, https://github.com/guangyey
Co-authored-by: Xuehai Pan <XuehaiPan@outlook.com>
some context in this document:
https://docs.google.com/document/d/18nJsj-F2C_QXO7ClwzPcAUENQ-B440B43W7DdDnlDt4/edit?tab=t.0#heading=h.pgebnyi7pocj
But TLDR;
`guard_or_true`, `guard_or_false` are better than `guard_size_oblivious` due to :
- Easier to reason about what assumptions we are making while reading the code.
- Avoid size_oblivious complexity that is not needed.
- Avoid unsoundness that could make `guard_size_oblivious(a==1)` be true when its not true for some vaue `a` during runtime.
- Less data dependent errors for some cases: ex, when doing `guard_size_oblivious(a==1)` and we know `a` is a tensor size, if it's traced with `a=u1-u2` `guard_size_oblivious(a==1)` will throw a data dependent error but `guard_else_false` will just return `False`.
### How is it different from statically_known_true??
**`if(cond)`:** (normal guarding) will try to evaluate statically and guard on the condition, willing to restrict input space to evaluate cond. if it fails to evaluate due to data dependent error will throw an exception (that could be converted to graph break in some situations).
**`statically_known_true(cond)`:** would be used when you never want to add a guard (restrict your input space), but just want to do a best effort check to see if you can infer that something is true/false ONLY based on existing constraints.
**`guard_or_true(cond)`/`guard_or_false(cond)`:** Those would be used in situations you prefer to guard and know the result of the expression over not guarding, but in case you hit a data dependent error you are ok with just returning true or false.
Some reasons you might be ok with returning true/false instead could be:
1. It's an optimization I do not want to fail for not performing optimization.
2. I am willing to deviate from the normal semantics when I have unbacked for the benefit of not failing (See the doc above for more details).
**`definitely_true(cond)`**: same as `guard_or_false(cond)` except does not try to do static eval for unbacked (planning to deprecate it and replace uses with `guard_or_false` or make it alias to `guard_or_false`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148430
Approved by: https://github.com/bobrenjc93
internally.
Summary:
This diff allows freeing the usage of folded constants that's created by
AOTInductor through CUDACachingAllocator instead of the constant blob
from cudaMalloc directly.
Test Plan:
LD_LIBRARY_PATH=/data/users/$USER/pytorch/build/lib
/home/$USER/local/pytorch/build/bin/test_aoti_inference
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149825
Approved by: https://github.com/chenyang78, https://github.com/desertfire, https://github.com/jingsh
This PR was inspired by internal models that were cache missing due to PGO. At a high level the problem looks as follows
Run 1, Invocation 1: We do static compile, save some example values in PGO/automatic dynamic
Run 1, Invocation 2: We detect varying inputs, do dynamic compile, get a dynamic graph and save to PGO. Crucially what we save to PGO is actually a superset of what is actually dynamic. If we notice an input was varying, we mark it as dynamic in PGO even if later on that value gets specialized. When a value gets specialized, we actually remove the symbol from the graph. This results in an interesting conundrum where although we are producing the same isomorphic graph, PGO makes the second run cache miss. Let's see how....
Run 2, Invocation 1: We fetch the PGO, over-mark things as dynamic, get a fx graph, look it up in the cache and... whoops! cache miss! This is because of the aforementioned behavior where the PGO profile will cause us to over-allocate symbols. In practice this means we end up saving a graph in cache with symbols x:s1, y:s3 and on second attempt we cache miss with x:s1, y:s6 where symbols s3,s4,s5 were all optimistically marked dynamic by PGO and subsequently specialized.
We solve this problem by hashing the source names. This ensures somewhat stable assignment. To prevent catastrophic symbol collisions, we use linear probing to ensure no collisions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149665
Approved by: https://github.com/Mingming-Ding, https://github.com/laithsakka
Part of https://github.com/pytorch/torchtitan/issues/866
## Context
- Async TP needs to support the "reshape -> scaled_mm -> reshape" pattern because scaled mm only supports 2D input tensors and 2D scales.
- (a,b,c) => (a*b,c)
- (a\*b,c) @ (c,d) = (a\*b,d)
- (a\*b,d) => (a,b,d)
- Currently the implementation does not support scaled mm with rowwise scales **for all cases** of the reshape -> scaled_mm -> reshape pattern. The minimal example of this pattern is confirmed to work via this [unit test](00a2c68f67/test/distributed/tensor/parallel/test_micro_pipeline_tp.py (L406)), but more involved e2e examples in torchtitan fail silently (more context in final bullet point).
- Previously, the "A tensor" **node** referenced in the async TP graph manipulation code is the 3D+ node before the reshape, but the "A_scale" node is the 2d node from after the reshape, so they are incompatible.
- I previously implemented a simpler solution to this problem in https://github.com/pytorch/pytorch/pull/148001, with a [unit test](https://github.com/pytorch/pytorch/pull/148001/files#diff-115f1d0852382c9b58f22640d80999d879b33618e5f6c633fc9e4d0ca9781cecR406) confirming the fused node is indeed in the graph for the minimal example of the reshape->mm->reshape pattern. I also confirmed via manual e2e testing w/ torchtitan that the crash I was fixing no longer occurred. However, it turns out due to this [bug in torchtitan](https://github.com/pytorch/torchtitan/issues/866) it was causing async TP to fail silently and fall back to vanilla TP, hiding the fact that this original solution fixed the crash but the fusion would not occur for rowwise scales. Thus, more robust solution is needed to support all cases.
## Solution TL;DR
- Use the 2D 'A' tensor and corresponding 2D scales as input to the fused_matmul_reduce_scatter implementation, instead of the 3D+ tensor/scales.
- Track the "pre mm reshape" and "post mm reshape" separately, to be referenced in the `fused_scaled_matmul_reduce_scatter` implementation, to update the scatter dim through the pre-mm reshape, and apply the post-mm reshape before applying the reduce scatter and returning the output tensor.
- Separate the `fused_matmul_reduce_scatter` and the `fused_scaled_matmul_reduce_scatter` code paths, to simplify them both.
- By fixing the bug in torchtitan (PR https://github.com/pytorch/torchtitan/pull/965) and implementing support for rowwise scales in pytorch in this PR, together these changes will solve the problem of how to support rowwise scales with all types of AC.
## Additional details for reviewers
To use the 2D A tensor while also supporting the "reshape -> mm -> reshape" pattern, the following other changes were needed:
- Track the pre-mm reshape, as it will affect the scatter dim used in the fused_matmul_reduce_scatter impementation.
- Track the post-mm reshape, as it will affect the output shape used in the fused_matmul_reduce_scatter impementation
- Based on the pre-mm reshape and the original scatter dim, calculate the new scatter dim for the 2D tensor. This is needed because during the pipelined producer mm implementation, the scatter dim is moved to dim 0 (so it can be sharded along the first dim and then get chunks to do mm ops on by indexing into the first dim), then moved back to it's original place before the reduce-scatter.
- Use the tracked post-mm reshape to reshape the stacked partial 2D outputs of the mm ops into 3D outputs needed for 1) the reduce-scatter w/ the original scatter dim, and 2) the expected output shape to prevent shape errors with subsequent ops.
## Test plan
- All existing unit tests passing.
- Expand unit tests for rowwise scales to test more scatter dims
- Added unit tests enforcing that async TP fails fast / throws an error if it fails to perform any fusions. Previously it just "failed silently" (fell back to vanilla TP without the user knowing) which has led to confusion, so this will improve the UX.
- Compared loss curves of bf16 vs float8 w/ rowwise scales to confirm integrity of numerics
- Confirmed via manual testing with torchtitan and inspecting the compile graph that the fusion is working as intended for:
- bfloat16
- float8 with tensorwise scales
- float8 with rowwise scales
## Loss curves
Loss curves are virtually identical for bf16 + vanilla TP versus float8 with rowwise scales + async TP:
<img width="1017" alt="loss_async_tp" src="https://github.com/user-attachments/assets/4995db78-7012-490f-a370-f4fecc289a22" />
## Performance
#### Per op SAC
Performance benchmarks for torchtitan Llama3 8b training runs on 4 H100s with per op SAC, using FSDP degree=2, TP degree=2:
- bf16 (vanilla TP): TPS 5161.5, peak memory 50.53 GB
- bf16 (async TP): TPS 5229.5, peak memory 50.68 GB
- float8 tensorwise (vanilla TP): TPS: 5959.5, peak memory: 50.47 GB
- float8 tensorwise (async TP): TPS 5964.5, peak memory 50.47 GB
- float8 rowwise (vanilla TP): TPS: 4962.0, peak memory: 50.55 GB
- float8 rowwise (async TP): TPS 4966.5, peak memory 50.65 GB
#### Full AC
Llama3 70b training runs on 128 H100s with full AC, using FSDP=16, TP=8
- bf16 (vanilla TP): 598 TPS, peak memory 71.51 GB
- bf16 (async TP): TPS 673, peak memory 71.08 (+12.54% TPS vs vanilla TP)
- float8 tensorwise (vanilla TP): 820 TPS, peak memory 55.26 GB
- float8 tensorwise (async TP): 950 TPS, peak memory 55.91 GB (+15.85% TPS vs vanilla TP)
- float8 rowwise (vanilla TP): TPS: 540 TPS, peak memory 71.46 GB
- float8 rowwise (async TP): 560 TPS, peak memory 70.65 GB (+3.7% TPS vs vanilla TP but still unexpectedly lower than bf16)
As you can see, float8 rowwise is working but performance needs to be improved further.
## Other changes
- Added logging so the user will know why fusion failed if it does.
- Remove logic which inserted a reshape node targeting "A scale" to get it to be in 3D like the "A tensor" since it's no longer needed.
## Long term plan
- Add a `scaled_matmul` op in pytorch, which will natively support a 3D+ "A tensor" and allow us to simplify the async TP implementation by avoiding the reshape -> scaled_mm -> reshape pattern and the special handling for it.
## Visualizing fused nodes in graphs for torchtitan training runs
Below are examples of the visualized graph generated by torch compile for torchtitan llama3 8b training runs with per op SAC. These graphs provide additional evidence (beyond the new unit tests added) that the implementation is working correctly.
### bf16
<img width="900" alt="bf16-fusion" src="https://github.com/user-attachments/assets/a3bed917-28eb-4a56-8d6e-2d2bf498385c" />
### float8 with tensorwise scales
<img width="900" alt="tensorwise-node" src="https://github.com/user-attachments/assets/b212ec4a-1899-44de-a4de-18c74e1de68a" />
### float8 with rowwise scales
<img width="900" alt="rowwise" src="https://github.com/user-attachments/assets/ed3354a3-894b-4ec9-86d0-f80364bf3d83" />
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149247
Approved by: https://github.com/kwen2501
This PR adds a new kernel for producing gamma and beta values for the backward pass in a performant way.
To test the performance against the baseline, I measured the backward pass of layernorm while sweeping over the following variables:
1. dtype in {half, float}
2. M in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
3. N in `2**k, 2**k - 1, 2**k + 1 for k in range(...)`
4. Whether we flush the L2 cache before running the backward pass
Summary: The new code performs better than the old code, especially for powers of 2. For M >> N case, it performs very well (kernel itself can be 30x faster and the overall backward pass can be 5-10x faster).
In order to visualize results of the kernel when choosing different values of M, N and dtype, I wrote some code to generate a heatmap. The heatmap has N on the x-axis, M on the y-axis and color-coded points where green shows performance improvement and red shows regressions. For example, `m=32 n=2048 1.42x` in the heatmap would indicate the normalized shape had 32 elements. The leading dimensions' product was 2048 elements and the new kernel resulted in the *backward pass* being 1.42x faster than the old *backward pass*.
Important note: This heatmap shows the total backward pass time as seen by the user. The kernel time difference can be sometimes very large while the total backward pass time is not that high. For example, for dtype=torch.half, M=32 N=2048, flush_l2_cache=True case, the heatmap shows a speedup of 1.42x, while ncu tells me the new kernel is 2.5x faster than the old:
M=32 N=2048 dtype=half flush_l2=True Old Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.35
Elapsed Cycles cycle 27,526
Memory Throughput % 2.21
DRAM Throughput % 0.54
Duration us 20.42
L1/TEX Cache Throughput % 4.31
L2 Cache Throughput % 2.62
SM Active Cycles cycle 1,475.02
Compute (SM) Throughput % 0.29
----------------------- ----------- ------------
```
M=32 N=2048 dtype=half flush_l2=True New Kernel NCU summary:
```
----------------------- ----------- ------------
Metric Name Metric Unit Metric Value
----------------------- ----------- ------------
DRAM Frequency Ghz 1.59
SM Frequency Ghz 1.34
Elapsed Cycles cycle 10,920
Memory Throughput % 5.64
DRAM Throughput % 1.35
Duration us 8.13
L1/TEX Cache Throughput % 1.92
L2 Cache Throughput % 6.89
SM Active Cycles cycle 3,554.41
Compute (SM) Throughput % 0.67
----------------------- ----------- ------------
```
Let's look at some rows from the heatmap. For dtype=float16 flush_l2_cache=True and when input shapes are powers of 2, we get the following:
<img width="1508" alt="image" src="https://github.com/user-attachments/assets/06179599-b2f0-4a45-8664-247a1067950b" />
There are 3 columns -- the first shows all data points, the second shows speedups only and the 3rd column shows regressions only. We can see that there are dramatic speedups for M >> N cases and the regressions are not that high (less than 1%, which could just be measurement noise). Here is a small guide I made:
For dtype=float32, we get a similar chart:
<img width="1499" alt="image" src="https://github.com/user-attachments/assets/c4d31a76-03b0-426c-9114-e1bfad29b530" />
The new code performs especially well for m >> n cases, and also where m and n are small. The m >> n case is special because we run 2 reduction kernels back to back and parallelize in the "M" dimension (the older kernel only parallelized in the "N" dimension).
The new code can sometimes have regressions for non-powers of 2. That is because the old code was using block sizes of {16, 32} while we have `threads.x = 32`. For example when N=33, the old code would have 3 blocks and we will have 2 blocks. I wrote some code to specialize for this case, but I think it will add complexity and @ngimel mentioned that non-powers of 2 are rare enough.
I am including the regressions here for completeness' sake:
<img width="1500" alt="image" src="https://github.com/user-attachments/assets/31c17cfb-ed9b-4106-b9c8-5c359751f530" />
To see this better:
1. Click the image
2. Right click the expanded image and open in a new tab
3. Go to that tab and left click once to zoom in
If you want to see the full data, here it is:
I also measured binary size and compile time since those are important for developers:
Binary size comparison
```
# Original
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
# This PR
-rwxr-xr-x 1 ahmads users 307193112 Mar 6 08:46 ./torch/lib/libtorch_cuda.so
```
The diff in bytes is 302kB which is about a 0.1% increase.
Compile time difference:
```
# Original
real 0m10.931s
user 0m9.676s
sys 0m1.004s
# this PR
real 0m16.720s
user 0m15.514s
sys 0m1.066s
# Command I ran
time /usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DAT_PER_OPERATOR_HEADERS -DFLASHATTENTION_DISABLE_ALIBI -DFLASHATTENTION_DISABLE_SOFTCAP -DFLASH_NAMESPACE=pytorch_flash -DFMT_HEADER_ONLY=1 -DHAVE_MALLOC_USABLE_SIZE=1 -DHAVE_MMAP=1 -DHAVE_SHM_OPEN=1 -DHAVE_SHM_UNLINK=1 -DMINIZ_DISABLE_ZIP_READER_CRC32_CHECKS -DONNXIFI_ENABLE_EXT=1 -DONNX_ML=1 -DONNX_NAMESPACE=onnx_torch -DTORCH_CUDA_BUILD_MAIN_LIB -DTORCH_CUDA_USE_NVTX3 -DUNFUSE_FMA -DUSE_C10D_GLOO -DUSE_C10D_NCCL -DUSE_CUDA -DUSE_CUFILE -DUSE_DISTRIBUTED -DUSE_EXTERNAL_MZCRC -DUSE_FLASH_ATTENTION -DUSE_MEM_EFF_ATTENTION -DUSE_NCCL -DUSE_RPC -DUSE_TENSORPIPE -D_FILE_OFFSET_BITS=64 -Dtorch_cuda_EXPORTS -I/home/ahmads/personal/pytorch/build/aten/src -I/home/ahmads/personal/pytorch/aten/src -I/home/ahmads/personal/pytorch/build -I/home/ahmads/personal/pytorch -I/home/ahmads/personal/pytorch/cmake/../third_party/benchmark/include -I/home/ahmads/personal/pytorch/third_party/onnx -I/home/ahmads/personal/pytorch/build/third_party/onnx -I/home/ahmads/personal/pytorch/nlohmann -I/home/ahmads/personal/pytorch/third_party/flash-attention/csrc/flash_attn/src -I/home/ahmads/personal/pytorch/aten/src/THC -I/home/ahmads/personal/pytorch/aten/src/ATen/cuda -I/home/ahmads/personal/pytorch/third_party/fmt/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/tools/util/include -I/home/ahmads/personal/pytorch/build/caffe2/aten/src -I/home/ahmads/personal/pytorch/aten/src/ATen/.. -I/home/ahmads/personal/pytorch/build/nccl/include -I/home/ahmads/personal/pytorch/c10/cuda/../.. -I/home/ahmads/personal/pytorch/c10/.. -I/home/ahmads/personal/pytorch/third_party/tensorpipe -I/home/ahmads/personal/pytorch/build/third_party/tensorpipe -I/home/ahmads/personal/pytorch/third_party/tensorpipe/third_party/libnop/include -I/home/ahmads/personal/pytorch/torch/csrc/api -I/home/ahmads/personal/pytorch/torch/csrc/api/include -isystem /home/ahmads/personal/pytorch/build/third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/tensorpipe/third_party/libuv/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googletest/include -isystem /home/ahmads/personal/pytorch/third_party/protobuf/src -isystem /home/ahmads/personal/pytorch/third_party/XNNPACK/include -isystem /home/ahmads/personal/pytorch/third_party/ittapi/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/eigen -isystem /usr/local/cuda/include -isystem /home/ahmads/personal/pytorch/third_party/ideep/mkl-dnn/include/oneapi/dnnl -isystem /home/ahmads/personal/pytorch/third_party/ideep/include -isystem /home/ahmads/personal/pytorch/INTERFACE -isystem /home/ahmads/personal/pytorch/third_party/nlohmann/include -isystem /home/ahmads/personal/pytorch/third_party/NVTX/c/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/cudnn_frontend/include -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS -D_GLIBCXX_USE_CXX11_ABI=1 -Xfatbin -compress-all -DONNX_NAMESPACE=onnx_torch -gencode arch=compute_90,code=sm_90 -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=bad_friend_decl --expt-relaxed-constexpr --expt-extended-lambda -Wno-deprecated-gpu-targets --expt-extended-lambda -DCUB_WRAPPED_NAMESPACE=at_cuda_detail -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -O3 -DNDEBUG -std=c++17 -Xcompiler=-fPIC -DTORCH_USE_LIBUV -DCAFFE2_USE_GLOO -Xcompiler -Wall -Wextra -Wdeprecated -Wno-unused-parameter -Wno-missing-field-initializers -Wno-array-bounds -Wno-unknown-pragmas -Wno-strict-overflow -Wno-strict-aliasing -Wunused-function -Wunused-variable -Wunused-but-set-variable -Wno-maybe-uninitialized -MD -MT caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o -MF caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o.d -x cu -c /home/ahmads/personal/pytorch/aten/src/ATen/native/cuda/layer_norm_kernel.cu -o caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/layer_norm_kernel.cu.o
```
So the new PR is 6 seconds longer compile time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148605
Approved by: https://github.com/ngimel
Somehow the torch._dynamo.is_compiling is changed to torch.compiler.is_compiling(), which also checks whether we're exporting. This is not caught by cI because we don't have an export test for scan.
Changing to torch.compiler.is_dynamo_compiling and added a test.
edit: piggyback the re-tracing support in this PR. Related code in combine_fn_is_normalized.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149903
Approved by: https://github.com/zou3519
Summary:
Adds the meta registration logic for torch.compile to work with
`torch._scaled_mm` with mxfp8. Thanks to @eellison for the pointer to make inductor work with this.
Test Plan:
```
pytest test/test_matmul_cuda.py -k test_blockwise_mxfp8_compile -s
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148461
Approved by: https://github.com/drisspg, https://github.com/eellison
some context in this document:
https://docs.google.com/document/d/18nJsj-F2C_QXO7ClwzPcAUENQ-B440B43W7DdDnlDt4/edit?tab=t.0#heading=h.pgebnyi7pocj
But TLDR;
`guard_or_true`, `guard_or_false` are better than `guard_size_oblivious` due to :
- Easier to reason about what assumptions we are making while reading the code.
- Avoid size_oblivious complexity that is not needed.
- Avoid unsoundness that could make `guard_size_oblivious(a==1)` be true when its not true for some vaue `a` during runtime.
- Less data dependent errors for some cases: ex, when doing `guard_size_oblivious(a==1)` and we know `a` is a tensor size, if it's traced with `a=u1-u2` `guard_size_oblivious(a==1)` will throw a data dependent error but `guard_else_false` will just return `False`.
### How is it different from statically_known_true??
**`if(cond)`:** (normal guarding) will try to evaluate statically and guard on the condition, willing to restrict input space to evaluate cond. if it fails to evaluate due to data dependent error will throw an exception (that could be converted to graph break in some situations).
**`statically_known_true(cond)`:** would be used when you never want to add a guard (restrict your input space), but just want to do a best effort check to see if you can infer that something is true/false ONLY based on existing constraints.
**`guard_or_true(cond)`/`guard_or_false(cond)`:** Those would be used in situations you prefer to guard and know the result of the expression over not guarding, but in case you hit a data dependent error you are ok with just returning true or false.
Some reasons you might be ok with returning true/false instead could be:
1. It's an optimization I do not want to fail for not performing optimization.
2. I am willing to deviate from the normal semantics when I have unbacked for the benefit of not failing (See the doc above for more details).
**`definitely_true(cond)`**: same as `guard_or_false(cond)` except does not try to do static eval for unbacked (planning to deprecate it and replace uses with `guard_or_false` or make it alias to `guard_or_false`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148430
Approved by: https://github.com/bobrenjc93
Allows subclasses of `TritonTemplate` to override the kernel type, e.g.
```
class MyTritonTemplate(TritonTemplate):
kernel_type = MyTritonTemplateKernel
```
This means that all of the logic in `TritonTemplate` class doesn't need to be duplicated in subclasses if the only required change is the kernel type.
Note that there is precedent for doing this - see `SIMDScheduling` in `torch/_inductor/codegen/simd.py`:
```
class SIMDScheduling(BaseScheduling):
kernel_type: type[Any] = SIMDKernel # override in subclass
...
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150018
Approved by: https://github.com/jansel
I am splitting caching the loading of modules from the caching the codegen since its trivial and much easier.
Module loading is 50% of the cost, and codegen is 50% of maybe_append choice on full graph model. which is 40% of total compile time.
<img width="434" alt="Screenshot 2025-03-24 at 4 35 12 PM" src="https://github.com/user-attachments/assets/aa851c6a-bde9-43f8-b12d-e439504ef62c" />
running mm_loop benchmark,
before this change:
67947323682
after this change:
25845073249
2.6X faster.
it seems that the cache was there then got dropped. I added benchmark so it wont be dropped again by mistake.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149910
Approved by: https://github.com/eellison, https://github.com/aorenste
ghstack dependencies: #149932
Add `None` to type annotations of `torch.onnx.ops.symbolic*` ops and improve tests to test support for optional inputs. Previously it was omitted mistakenly even though the implementation supports it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150038
Approved by: https://github.com/titaiwangms
1. Add config selection for SM89.
2. Only build kernels if compiling for given arch.
3. Factor out CMake code to enforce compiling for needed archs for individual files into a function.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149978
Approved by: https://github.com/drisspg
Summary:
X-link: https://github.com/pytorch/executorch/pull/8703
Originally we created a bunch of empty `TARGETS` files to allow us to enable `BUCK` files in fbcode by hiding the existing BUCK file. These files were subsequently merged together using `non_fbcode_target` so these tombstones are no longer necessary.
This diff fixes all files that WOULD have had the useless tombstone merged into them. To create this diff, I just ran the merger script that Codemod Service is using and then deleted the "merged from" and tombstone lines with `sed`, `arc f` and reverted any lines that didn't make sense
Test Plan: CI
Differential Revision: D69994481
Pull Request resolved: https://github.com/pytorch/pytorch/pull/147897
Approved by: https://github.com/izaitsevfb
Summary: Triton-MTIA expects the codename of the device as the arch when querying the module map, not the compute capability. This diff gets rid of the following error: `No libdevice is provided for arch (0, 0)`
Test Plan: CI
Reviewed By: Myrthan
Differential Revision: D70072095
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149860
Approved by: https://github.com/jansel
This PR supports symbol inputs to graph partition functions. Before this PR, we rely on `node.read_writes` to get partition inputs. However, this does not cover symbol inputs.
In this PR, for each graph partition, we collect all symbol inputs which are required to be in scope to successfully perform codegen, including:
- free symbols used in partition nodes.
- free symbols in partition input/node shapes, strides, and offsets. This is needed for recording cudagraphs for tensors with dynamic shapes.
### Note1: MutationLayout
In this example, node.layout is MutationLayoutSHOULDREMOVE. The symint from index `n` does not appear in the size, offset, stridese of node.layout. This symint appear in node.layout.target. So we need extra handle for it.
```python
x = torch.zeros(7, device="cuda")
def fn(n, a):
a[n] = -1
return a
opt_fn = torch.compile(fn, fullgraph=True)
for n in range(2, x.shape[0]):
opt_fn(n, x)
```
### Note2: Composability with Padded Tensor Subclass
W/o graph partition, Padded Tensor subclass lifts outer shapes to input arguments (i.e., arg0_1 for s0, arg1_1 for s1) but does not lift inner shapes (i.e., s2 and s3). Since cudagraph cache relies on integer inputs, it will cache on outer shapes and ignore inner shapes, which is bad.
```
def call(args):
arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1 = args
args.clear()
s0 = arg0_1
s1 = arg1_1
arg2_1_size = arg2_1.size()
s2 = arg2_1_size[0]
s3 = arg2_1_size[1]
assert_size_stride(arg2_1, (s2, s3), (s3, 1))
with torch.cuda._DeviceGuard(0):
torch.cuda.set_device(0)
buf0 = empty_strided_cuda((s2, s3), (s3, 1), torch.float32)
# Topologically Sorted Source Nodes: [x1, mul], Original ATen: [aten.add, aten.mul]
triton_poi_fused_add_mul_0_xnumel = s2*s3
stream0 = get_raw_stream(0)
triton_poi_fused_add_mul_0.run(arg2_1, buf0, triton_poi_fused_add_mul_0_xnumel, stream=stream0)
del arg2_1
return (buf0, s0, s1, s1, )
```
w/ graph partition, the partition function only includes tensor and inner shapes as inputs, to make sure the cudagraph caching is correct. Full Comparison: [code](https://www.internalfb.com/intern/diffing/?paste_number=1761674743)
```python
def call(self, args):
arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1 = args
args.clear()
s0 = arg0_1
s1 = arg1_1
arg2_1_size = arg2_1.size()
s2 = arg2_1_size[0]
s3 = arg2_1_size[1]
assert_size_stride(arg2_1, (s2, s3), (s3, 1))
partition0_args = [arg2_1, s2, s3]
del arg2_1
(buf0,) = self.partitions[0](partition0_args)
del partition0_args
return (buf0, s0, s1, s1, )
```
The number of cudagraphs is validated below: (also added to test)
```python
import torch
from padded_tensor import PaddedTensor
# Turning off graph_partition leads to
# torch._inductor.cudagraph_trees.get_container(0).tree_manager.new_graph_id().id=6
# at the end, which is wrong.
# torch._inductor.config.graph_partition = False
# Turning on graph_partition leads to
# torch._inductor.cudagraph_trees.get_container(0).tree_manager.new_graph_id().id=4
# at the end, which is correct.
torch._inductor.config.graph_partition = True
def f(x):
x1 = x + 1
return x1 * 2
compiled_f = torch.compile(f, mode="reduce-overhead")
def run(shape):
x = torch.randn(*shape, device="cuda")
pad_x = PaddedTensor.from_tensor(x, multipliers={0:4, 1:4})
assert hasattr(pad_x, "multipliers"), breakpoint()
eager_out = f(pad_x)
for _ in range(3):
compiled_out = compiled_f(pad_x)
compiled_out = compiled_f(pad_x)
assert eager_out.shape == compiled_out.shape
assert eager_out.tensor.shape == compiled_out.tensor.shape
assert torch.allclose(eager_out.tensor, compiled_out.tensor)
# static shape. record a NEW cudagraph. 1 cudagraph in total now.
run((2,3))
# outer shape is dynamic, leading to a new dynamo graph
# this new dynamo graph forces a NEW cudagraph. 2 cudagraphs in total now
run((3,4))
# outer shape changed but inner shape does not change
# so NO new cudagraph is recorded
run((2,2))
# inner shape is dynamic now, leading to a new dynamo graph
# this new dynamo graph forces a NEW cudagraph. 3 cudagraphs in total now
run((5,6))
# does NOT record a new cudagraph
run((7,8))
# record a NEW cudagraph. 4 cudagraphs in total now
run((10,11))
assert torch._inductor.cudagraph_trees.get_container(0).tree_manager.new_graph_id().id == 4
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149458
Approved by: https://github.com/eellison
* When we try to install [libstdcxx-ng 12.3.0 from conda-forge](595293316d/.ci/docker/common/install_conda.sh (L65)), conda 24.7.1 updates the dependencies of that package, including libgcc-ng package to the following: `libgcc-ng-14.2.0 | h69a702a_2 52 KB conda-forge`
* However, conda updated their installer script on Feb 6 2025 to version 25.1.1, which behaves differently from previous versions when installing conda packages.
* conda 25.1.1 does *not* update any dependencies in the above step, and hence the same installation of libgcc-ng from "defaults" channel is present: `libgcc-ng pkgs/main/linux-64::libgcc-ng-11.2.0-h1234567_1`
* Adding the "--update-deps" flags to the conda install command installs a newer libgcc-ng package from the "conda-forge" conda channel: `libgcc-ng-12.3.0 | h77fa898_13 762 KB conda-forge`, which is compatible with the libstdcxx-ng 12.3.0 package
* Compare this [Feb 4 docker build](https://github.com/pytorch/pytorch/actions/runs/13148456164/job/36691412387#step:6:5179) to this [Feb 10 docker build](https://github.com/pytorch/pytorch/actions/runs/13247023578/job/36975931849#step:6:5451), which shows that the latter does *not* update libgcc-ng.
* This creates linking issues when trying to use a library, that was built with a newer libgcc_s.so.1 (from libcc-ng package), in the PyTorch conda environment. Eg. ONNX-RT:
```
[0;93m2025-02-13 10:18:38.492434704 [W:onnxruntime:Default, migraphx_execution_provider.cc:167 get_flags_from_env]
[MIGraphX EP] MIGraphX ENV Override Variables Set:[m
[1;31m2025-02-13 10:18:38.628064251 [E:onnxruntime:Default, provider_bridge_ort.cc:2028 TryGetProviderInfo_ROCM] /onnxruntime/onnxruntime/core/session/provider_bridge_ort.cc:1636 onnxruntime::Provider& onnxruntime::ProviderLibrary::Get() [ONNXRuntimeError] : 1 : FAIL : Failed to load library libonnxruntime_providers_rocm.so with error: /opt/conda/envs/py_3.10/bin/../lib/libgcc_s.so.1: version `GCC_12.0.0' not found (required by /opt/conda/envs/py_3.10/lib/python3.10/site-packages/onnxruntime/capi/libonnxruntime_providers_rocm.so)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149599
Approved by: https://github.com/malfet
Summary:
User defined Triton kernel sometimes rely on real inputs to determine
the path of execution. We need real inputs to invoke the correct
behavior of the user defined triton kernels (see example in test case,
where we have an early return for random inputs)
Test Plan:
Included in the commit.
python test/inductor/test_aot_inductor.py -k triton_autotuning
python test/inductor/test_aot_inductor.py -k triton_mutated_autotuning
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149553
Approved by: https://github.com/davidberard98, https://github.com/eellison
Summary: This diff adds the ability for HF reader/writer to read/write in a distributed way. We do this by sending all the tensors meant for the same file to the same rank.
Test Plan:
ensure existing tests pass
I also ran a full end to end test on my devserver to read/write from my HF repo
Differential Revision: D70096439
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148189
Approved by: https://github.com/joecummings, https://github.com/saumishr
By implementing `_cast_` flavors of both dense and strided ops. Add regression tests that tests `fmax`/`fmin` for mixed dtypes.
Been dreaded to write this PR for a while, as it end up to be pretty bulky:
- Adds 1C10_METAL_ALL_TYPES_FUNCTOR` and `c10:🤘:ScalarType` to `c10/metal/common.h` and test that its values always match `c10::ScalarType`
- Add `c10:🤘:cast_to` to `c10/metal/utils.h` which could be used to cast any scalar metal dtype to any other one, including complex values
- Implement `val_at_offs<T>(constant void *, long offs, ScalarType dtype)` that is used to dynamically cast types
- Add `binary_strided_cast` and `binary_dense_cast` that are invoked for output dtype and cast both inputs to that output before performing the op
Benchmark collected on M2Pro that runs fmax for 1 mln element tensors (Times are in microseconds.)
| | dense-dense | transp-transp | dense-transp | transp-dense | dense-scalar | dense-bcast |
|-------------------------|---------------|----------------|----------------|----------------|---------------|--------------- |
| fmax (torch.float16, torch.float16) | 160.9 | 159.9 | 270.5 | 270.9 | 236.6 | 293.0
| fmax (torch.float32, torch.float32) | 176.9 | 171.0 | 273.7 | 293.5 | 242.6 | 294.2
| fmax (torch.float32, torch.float16) | 171.4 | 170.9 | 283.6 | 303.0 | 253.7 | 302.3
| add (torch.float16, torch.float16) | 218.0 | 223.6 | 221.0 | 222.0 | 214.9 | 218.3
| add (torch.float32, torch.float32) | 227.4 | 233.9 | 228.8 | 231.9 | 218.9 | 221.4
| add (torch.float32, torch.float16) | 226.1 | 227.5 | 227.5 | 226.9 | 177.0 | 190.8
TODOS:
- Include input and output dtype in non-cast kernel name
- Make TensorFactory.h use `C10_METAL_ALL_TYPES_FUNCTOR`
- Extend mixed_dytpes testing via OpInfo
Fixes https://github.com/pytorch/pytorch/issues/149951
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149974
Approved by: https://github.com/manuelcandales
Summary:
- Add more tests for torchbind in aoti
**FallBackKernel**
- In FallbackKernel.find_device, do not check the device of torchbind obj because they don't have a fixed "device"
- If no device found for CallTorchBindObject, use cpu
- handle None output in `export_extern_kernel_node`
Test Plan:
```
buck run //sigmoid/inference/test:e2e_test_cpu -- -r CustomClassHolderConstantDynamic
```
Differential Revision: D70746626
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149749
Approved by: https://github.com/desertfire
This PR is cleanup only. There are no feature changes or bug fixes.
We create a TunableOp context manager for setting up and cleanup. We re-write TunableOp unit tests in terms of this context manager. Ultimately reduces the amount of copy-paste code.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149930
Approved by: https://github.com/jeffdaily
Fixes#148938
Context:
In triton 3.3, triton kernels expect a global scratch space arg to be passed in. This is fixed in #148051, which fixed most of the AOTI/cpp_wrapper failures; the fix is to inject a (null) global scratch space arg passed as an argument to all kernels.
But in the case of TMA, we need to call a non-triton-generated function - init1DTMADescriptor. The same `generate_args_decl` function used for calling triton kernels (and modified in #148051 to insert a global scratch space) is used to prepare the arguments to init1DTMADescriptor, and so it had an extra global scratch space arg. Then we'd get a null pointer passed into init1DTMADescriptor, resulting in an IMA later on when the TMA use kernel
This PR: adds an option to `generate_args_decl` to specify whether this is a triton kernel (in which case we should add the global scratch space arg) or not (when we shouldn't add the extra arg).
Note: this doesn't appear in CI because we don't run these tests with Hopper machines in CI.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149973
Approved by: https://github.com/drisspg
Summary:
Refine the error message if dlopen failed in AOTInductor.
The original error message was ominous, modified to recommend user to
rebuild AOTInductor if needed, otherwise it's fine.
Test Plan:
None. Error message change.
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149812
Approved by: https://github.com/chenyang78, https://github.com/jingsh
Add a couple of Jetson skips for oom tests in test/test_cuda.py due to failures in nvidia CI. Jetson not having full nvml support is a known issue so this is mostly a test side fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149587
Approved by: https://github.com/eqy
PYNVML related tests in test/test_cuda.py are failing in nvidia internal CI for Jetson devices because Jetson devices don't fully support nvml (it exists as a stub library). In addition to skipping PYNVML tests for Jetson, this PR also reworks the TEST_PYNVML logic a bit to be more consistent with the rest of TEST_{something} conditions in test/test_cuda.py
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149578
Approved by: https://github.com/janeyx99, https://github.com/eqy
Summary:
We might free the active buffer if we free the buffer twice.
Test Plan:
```
LD_LIBRARY_PATH=/data/users/$USER/pytorch/build/lib
/home/$USER/local/pytorch/build/bin/test_aoti_inference
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149810
Approved by: https://github.com/chenyang78
This will improve docker image build times by not having to rebuild magma rocm for unrelated changes. This PR is step 1 of 2. The next step is a second PR to modify the docker image builds to use the magma tarball that this PR will produce.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149902
Approved by: https://github.com/malfet
Fixes#139167
This PR:
* uses `named_buffers` to mark static
* Checks that `named_buffers` is of expected type (callable, iterator) before trying to iterate over; if not, we skip this pass
These changes fix the previous errors in dynamo causing to crash (as shown in issue above)
### Unit Test
```
python test/dynamo/test_buffers_override.py
```
Results in:
```
.
----------------------------------------------------------------------
Ran 2 tests in 5.344s
OK
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149882
Approved by: https://github.com/anijain2305
Summary: The primary change is to update the autotune-in-a-subproc implementation to avoid using multiprocessing spawn. Spawn (re)executes the toplevel script in the subproc, which can be problematic. The approach here is similar to Triton parallel compile: we Popen a subproc on a controlled entry point and communicate over pipes. That change drove a lot of refactoring in the TuningProcess class, so I took the opportunity to simplify some things, rename some methods, etc.
One other notable change is around the timeout / kill approach. After a timeout, we were previously attempting to stop the subproc in three steps (graceful shutdown, sigkill if graceful fails, sigterm if sigkill fails). I'm gonna argue think that's not useful: 1) The graceful shutdown is never going to work unless the subproc happens to have just completed its task and is ready to receive the next command. 2) If we're going to kill the subproc, let's just take the most aggressive approach and move on as quickly as possible to restarting it rather than waiting to see if previous shutdown attempts succeeded. The only downside that I can find find is maybe a little log spew?, e.g., ` ResourceWarning: subprocess 2987680 is still running`
List of changes:
* Use Popen instead of spawn for the autotuning subprocess.
* Introduced a new entry point `__autotune_main__.py`
* Renamed some TuningProcess methods. For example `shutdown` makes more sense than `terminate` because the latter implies a forced kill.
* Simplified the implementation around benchmarking timeout and how we kill the subproc after a timeout.
* Deprecated the unused timeout configs in `_inductor/config.py`
* Moved `get_ld_library_path` helper to a common utils file.
* Added more unit tests for subproc crashes / timeouts / exceptions, etc.
Test plan:
* New unit tests
* Also ran internally with all combinations of: build mode `opt` and `dev-nosan`, and `buck run` vs. executing the `.par` file directly.
* Made sure the functionality to parallelize autotuning across different GPUs is working (it wasn't clear to me this was behaving the way we wanted it to).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149700
Approved by: https://github.com/aorenste, https://github.com/jansel, https://github.com/eellison
We weren't handling `setattr(tensor_obj, "real", 42)` correctly, because
the attribute is a `GetSetDescriptorType` that has special setter logic.
See added test and comments for more explanations.
This patch makes it so that we graph break in those cases, rather than
resulting in silent incorrectness.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149791
Approved by: https://github.com/mlazos
ghstack dependencies: #149481
Patches over an issue where randomly generated example tensors can cause kernel autotuning to fail, when those tensors would not be possible outputs from previous kernels in the sequence. This fixes a failure in `test_torchinductor_opinfo.py` when run with compile-time autotuning, `test_comprehensive_nanquantile_cuda_float64`.
For clarity, the situation triggering this PR looks like kernels `A -> BCDE -> F` (`BCDE` is fused), where one of the outputs from `A` is a boolean tensor describing some of the input data. Previously, we randomly regenerated that boolean tensor and the input data before passing them to `BCDE`, so that they no longer matched. This caused a `tl.device_assert` call in `BCDE` to fail. With this PR, we reuse the random data input to `A` and the output Boolean tensor, such that they match and pass the device assertion in `BCDE`.
Fixes#147799.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146706
Approved by: https://github.com/desertfire
The "PendingUnbackedSymbolNotFound" error is when an unbacked symbol is created within a piece of code, but this symbol never appears in any of the outputs. I believe the original intention is to help catch incorrectly written meta kernels, where users might've unintentionally created an unbacked symbol but never used it anywhere, but in our case this is intentional. An example is the following test case:
```python
def test_pending_unbacked(self):
class M(torch.nn.Module):
@mark_compile_region
def gn(self, x):
u = x[0].item()
return x * u
def forward(self, x):
for _ in range(4):
x = self.gn(x)
return x
torch._dynamo.config.capture_scalar_outputs = True
torch.compile(M())(torch.randn(8))
```
This fails with the error:
```
torch._dynamo.exc.InternalTorchDynamoError: PendingUnbackedSymbolNotFound: Pending unbacked symbols {zuf1} not in returned outputs (FakeTensor(..., size=(8,)),) .
```
In this case, creating the unbacked symbol is intentional, so we can bypass this using `fake_mode.shape_env.ignore_fresh_unbakced_symbols()`.
Differential Revision: [D71298926](https://our.internmc.facebook.com/intern/diff/D71298926)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149297
Approved by: https://github.com/zou3519
ghstack dependencies: #149296
These were accidentally deleted in the refactor of DEVTOOLSET +
cxx11abi.
This happened because the `build_environment` variable wasn't aware of the `build_variant` for libtorch and subsequently overwrote the original file twice, leaving the last written as the actual workflow (which in this case was the debug builds).
One thing this has made me curious on is if we actually need `debug` builds for window at all? We don't release them for linux and I'd probably bet that they have low download numbers anyways so maybe it makes sense to cut them.
Adds a build_variant parameter to the dataclass so that we can extend
these easily in the future if we want.
Signed-off-by: Eli Uriegas <eliuriegas@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149863
Approved by: https://github.com/malfet, https://github.com/atalman
Summary:
Cache save plan metadata to reduce the collective overhead.
Global plan dedupe and metadata creation are the main overheads on Rank 0. This change saves all this cost for the subsequent saves if the plans do not change. A quick experiment with the 256 rank job, Global step overhead drops by ~99%, from 90s+ to mere 1.5s. 1.5s was mostly spent on creating the checkpoint module directories and near empty collective.
Differential Revision: D71631441
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149785
Approved by: https://github.com/MeetVadakkanchery
Summary: For Scalar variant resolution, we didn't handle a corner case of "Tensor_mode" variant (from aten::div). Adding the missing case to the graph pass.
Test Plan: buck test mode/opt caffe2/test:test_export -- -r test_operator_aten_tensor_mode_variant_cpp_runtime
Differential Revision: D71638433
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149755
Approved by: https://github.com/yushangdi
This adds a `reduce_scatter` implementation for ProcessGroupGloo. This is a pretty naive implementation as it does 1 allreduce per rank but may be useful for testing in FSDP etc. There was an existing implementation of reduce_scatter_tensor/reduce_scatter_tensor_coalesed that has a very similar implementation but requires a fixed tensor size per rank.
If users find these functions to be too slow we can address them as issues arise.
Gloo now supports all major distributed operations. Quite a few of these were added by @rohan-varma and @yifuwang but they didn't update the support chart. We also have `CUDAWork` variants of most operations so those were also added to the chart.
Test plan:
```
pytest -v test/distributed/test_c10d_gloo.py -k reduce_scatter
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149869
Approved by: https://github.com/fduwjj
Summary:
`--no-as-needed` is not available in ld64.lld
Applying this on all macos is potentially too broad? I am not sure if `fbcode//mode/mac` uses a different linker, but arvr mode for sure uses ld64.lld.
Test Plan: CI / used for a macOS build on top of the stack.
Differential Revision: D71315125
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149421
Approved by: https://github.com/colesbury
Summary: adding operator.truediv and operator.neg support to the runtime
Test Plan: buck run mode/opt caffe2/test:test_export -- -r test_sym_float_operators_cpp_runtime_nonstrict
Differential Revision: D71637267
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149754
Approved by: https://github.com/pianpwk
### Summary
When the block-size for `N` dimension is `48` for the AMX GEMM micro-kernel for int8 WoQ (BF16 activation, int8 statically quantized weights), the logic for handling the tail is incorrect - we can't always dequantize 32 elements of weights at a time because we may need to dequantize `32` followed by `16` when `block_n` is `48` (for each `K`).
This PR fixes that logic, which was initially exposed with `M=17, N=1024, K=1024`.
This PR also fixes the case of `block_n` being 16.
I had introduced [this bug ](ca9813ea14) after misreading GEMM blockings as `["block_m", "block_k", "block_n"]` instead of `["block_m", "block_n", "block_k"]` (so I had wrongly assumed that `block_n` was always 32).
### Future work
While this PR simply fixes a bug, it's possible to optimize the code pertaining to dequantizing & caching the B buffer - for `block_n` being `16` or `48`, `K` would always be a multiple of 2, so `K * block_n` will always be a multiple of 32. Since `dequantized_B_buf` stores rows contiguously, when `block_n` would be `16` or `48`, we could store 32 BF16 elements at a time instead of storing `16` at a time (when `block_n` is 16), or `32` followed by `16` at a time (when `block_n` is 48). Such an optimization would lower `register -> memory` data movements.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149359
Approved by: https://github.com/leslie-fang-intel, https://github.com/jgong5
This adds a new env var and flag,
autograd_cache_allow_custom_autograd_functions, (env var: `TORCHINDUCTOR_AUTOGRAD_CACHE_ALLOW_CUSTOM_AUTOGRAD`) which allows custom autograd functions into AOTAutogradCache.
@hirsheybar and I worked together to verify that the higher order op AutogradFunctionApply is pure with respect to the dynamo input being passed in, so this *should* be safe. I'm still putting it behind a flag and turning it on slowly, first on an internal model, though. Once we verify that it is correct on the internal model we can work to enable the flag by default.
Differential Revision: [D71633184](https://our.internmc.facebook.com/intern/diff/D71633184/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149751
Approved by: https://github.com/bdhirsh, https://github.com/zou3519
This patch updates existing `test_return_..._subclass` tests in
`test/dynamo/test_subclasses.py`, so that they end up invoking the
`__torch_function__` method of the newly constructed tensor subclass
instnaces.
This exposes a bug in `TensorVariable.method_as_subclass`, where it
forgot to grab the `__func__` out of `__torch_function__`, which led to
the an error down the line.
This patch fixes `TensorVariable.method_as_subclass` by centralizing how
we extract and wrap torch function, in `build_torch_function_fn`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149481
Approved by: https://github.com/jansel
This adds AVG support to ProcessGroupGloo to better support FSDP on CPU. I expect there will be more issues but this is easy enough to support in a naive fashion.
This applies to both reduce and allreduce.
This is a simple SUM + division and may not be the most numerically stable but that's expected. FSDP for low precision data types implements pre/post divide and uses SUM instead.
Test plan:
```
pytest -v test/distributed/test_c10d_gloo.py
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149781
Approved by: https://github.com/fduwjj
Which renders non-contiguous operations much faster for larger tensors, for example `fmax` of 1000x1000 strides tensors takes 270ms with new algorithm and 430ms with an old one, that needed additional tensor of 3e6 elements to function.
TODO: Add 64-bit indexing logic, as current implementation has the same limitation as `generateKernelDataOffsets`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149730
Approved by: https://github.com/dcci, https://github.com/manuelcandales
In ROCm 6.4 and newer, when building Triton in the Triton-ROCm wheel build flow, newer releases of ROCm no longer have **libamd_comgr.so.2** as the .so file has been updated to **libamd_comgr.so.3** in ROCm 6.4 and newer. We conditionalize on which ROCm the wheel build is for, and choose the .so accordingly.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149855
Approved by: https://github.com/Skylion007, https://github.com/jeffdaily
Lazos correctly pointed out this doesn't make sense for compile since
we graph break in compile. This results in tons of unwanted user log
spew. We do want this in export though since it's drastiaclly reduced
the support load for DDEs. This PR does the refactor to keep it in
export but remove it from compile
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149831
Approved by: https://github.com/mlazos
This preserves graph breaks in the case that one graph break directly causes another, e.g. graph breaks in generic context managers.
```python
import torch
class CtxMgr:
def __enter__(self):
return self
def __exit__(self, exc_type, exc_value, traceback):
pass
@torch.compile(backend="eager", fullgraph=True)
def fn():
with CtxMgr():
with CtxMgr():
pass
with CtxMgr():
with CtxMgr():
pass
torch._dynamo.graph_break()
fn()
```
Output:
```
torch._dynamo.exc.Unsupported: Call to `torch._dynamo.graph_break()`
Explanation: User-inserted graph break. Message: None
Hint: Remove the `torch._dynamo.graph_break()` call.
Developer debug context: Called `torch._dynamo.graph_break()` with args `[]`, kwargs `{}`
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/data/users/williamwen/pytorch/playground.py", line 23, in <module>
fn()
File "/data/users/williamwen/pytorch/torch/_dynamo/eval_frame.py", line 664, in _fn
raise e.with_traceback(None) from e.__cause__
torch._dynamo.exc.Unsupported: Graph break under GenericContextWrappingVariable
Explanation: Attempted to graph break in an active context manager(s) that doesn't support graph breaking.
Hint: Move the offending context manager(s) to outside the compiled region.
Hint: This graph break may have been caused by an earlier graph break. Resolving the earlier graph break may resolve this one.
Developer debug context: Active generic context managers: [GenericContextWrappingVariable(CtxMgr), GenericContextWrappingVariable(CtxMgr)]
from user code:
File "/data/users/williamwen/pytorch/playground.py", line 20, in fn
torch._dynamo.graph_break()
Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"
```
Note in particular that both graph breaks (torch._dynamo.graph_break and graph break in context manager) are present in the logs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149676
Approved by: https://github.com/jansel, https://github.com/zou3519, https://github.com/anijain2305
While we probably don't want to expand the set of default matmul tunings too much, this is the largest tile size usable by H100 and A100, and is usually the top performing tile size for large matmuls. E.g. on H100 adding this tile size improves perf of multiplying 8192-square matrices from 600->700 tflops. (cuBLAS 12.6 gets 780, so Triton still isn't SOTA, but closer)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149790
Approved by: https://github.com/jansel
Before, we would take the first argument with the largest number of shards, regardless if it had fewer dims than another arg with the same number of shards but more dimensions. This would lead to potentially fewer sharding options
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149721
Approved by: https://github.com/tianyu-l
Fixes#149450
This PR adds fallback support on StaticCudaLauncher for any number of kernel arguments. Above MAX_ARGS, we can do a heap allocation/malloc instead.
For 0 arguments, triton technically does some undefined behavior by allocating a 0 byte array and passing it to cuLaunchKernel. In reality, cuLaunchKernel never accesses the pointer if the singature of the cubin has no parameters, so we can just pass nullptr directly.
We could technically use `alloca` to stack allocate instead of heap allocate, though in my tests it didn't seem to affect runtime performance on benchmarks particularly impressively, and alloca has portability issues, so I'd rather just stick with something simpler for now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149442
Approved by: https://github.com/jansel
This PR threads through the correct boxed_forward_device_index from graph_kwargs to CompiledFXGraph.post_compile. This allows us to correctly update BoxedDeviceIndex from cache hits.
We don't actually need to save `boxed_forward_device_index` in CompiledFXGraph because its value is in the cache key, so it always matches to the ambient one anyway. On forward with cudagraphs enabled, derive `boxed_forward_device_index`'s value from `device_idxs`.
Testing:
```
python benchmarks/dynamo/cachebench.py --mode training --benchmark torchbench --model BERT_pytorch --device cuda --repeat 1 --dynamic --output="dynamic.json"
```
Now cache hits properly on FXGraphCache. AOTAutogradCache has a guard failure. Will look into that as a followup.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148130
Approved by: https://github.com/eellison
In the kernelBot leaderboard we support people competing with custom cuda extensions via `load_inline()`, however even on toy kernels this can result in cold starts of up to 90s - this feature is primarily responsible for us having to double our timeout values
I performed an investigation here https://github.com/msaroufim/load_inline_slow and the primary cause was that torch/extension.h and torch/types.h add in about 5,000 header files https://github.com/msaroufim/load_inline_slow/blob/main/header-analysis
So we introduce a mode `no_implicit_headers` which forces users to be explicit about exactly what they want to add. There's a proper test meant to be used in a CLI and a pytest test that's not terribly helpful
Then there's still an open question around what's the most minimal example implementation we can provide. For the baseline kernel we're showing here, it takes about 1 min to compile
1. There's using TensorBase.h (finicky to get right but can get compilation times down to 7s)
2. Just using Tensor.h (down to 15s)
3. Using Shim.h (did not try yet since the syntax is verbose relative to cuda)
This is my take so far https://gist.github.com/msaroufim/079a8d08ffebd0f91a1c2247eb0ce9e0 for a minimal implementation at 15s but @malfet has a simpler one at only 5s
There's more things I'd like to try moving forward like nvrtc and fancier compilation flags. Typical advice around using precompiled headers does not apply to us because we are mostly interested in cold starts where we tear down the machine after running a kernel
Also in a future PR I'd like to fix issue I've noticed with load_inline
1. It needs a force recompilation mode, I was using this quite a bit myself
2. The cache does not take into account changes in environment so the best way to force a recompilation is to change some string in the file
3. Instead of relying on pybind, can we use TORCH_LIBRARY instead
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149480
Approved by: https://github.com/malfet
As `cuBLAS` workspaces are already per-stream, there shouldn't be kernel execution overlap with `cuBLASLt` kernels.
This PR reuses `cuBLAS` workspaces for `cuBLASLt` for the following benefits:
+ caching (`cuBLAS` workspaces were already cached, so now we get that for `cuBLASLt`)
+ "free" workspace size bump for `cuBLASLt` `cuBLASLt` workspace sizes were previously smaller than those for `cuBLAS` by default which potentially hurts performance, and we encountered difficulty in increasing the size due to downstream OOMs , see also #120925
+ fixes behavior broken behavior with the memtracker; https://github.com/pytorch/pytorch/pull/139442 attempted to handle peaky allocation behavior that broke memtracker equivalence tests but it didn't seem to fully work, here the cached/reused `cuBLAS` workspace seems to fix it
+ one environment variable to rule them all: `CUBLAS_WORKSPACE_CONFIG` applies directly to `cuBLASLt` without a confusing `CUBLASLT_WORKSPACE_SIZE` that users would also need to consider
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145130
Approved by: https://github.com/ngimel
The main purpose of this PR is to fix offline tuning for ScaledGEMM. The previous UT passed because it was not strict enough. Additionally:
- All the offline tuning tests now do a comparison with the online results to ensure that ParamSignature match.
- We raise an error if submatrices are encountered as this is only supported in online tuning mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149677
Approved by: https://github.com/jeffdaily
Summary:
`use_triton_lce_replace_simple_LCE` and `use_triton_lce_replace_normal_LCE`
code is mostly the same, some minor changes to support aten IR
Test Plan:
```
scripts/aetk/aetk -L
%run ~/fbsource/fbcode/caffe2/test/inductor/fb/test_customized_triton_kernel_passes.py
```
will verify the qps after everything done in the stack
Reviewed By: frank-wei
Differential Revision: D68909857
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149702
Approved by: https://github.com/frank-wei
Similar to #140425, we are making the implementation usable via header-only code sharing.
Review note: #62546 by @yanbing-j removed expm1 usage from this path. I don't know why and expm1 should be more efficient, so I've put it back. Please let me know if there is a good reason I shouldn't.
Testing: existing correctness tests should cover.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149673
Approved by: https://github.com/cyyever, https://github.com/Skylion007
Today, if you run DTensor (or any tensor subclass) under __torch_dispatch__, you will start seeing `CompositeImplicitAutograd` ops show up in the torch_dispatch.
"handling" these ops is trivial: you can just tell them to decompose into their constituent ops. Normally this decomposing happens in autograd, above DTensor, but inference_mode turns autograd off, forcing the subclass to handle the op directly.
It looks like previously we manually added a few CompositeImplicitAutograd entries to DTensor (e.g. linear), but this PR tries to support these ops a bit more generically.
The main difference is that DTensor now needs to check if a given op is `CompositeImplicitAutograd` before attempting to run sharding prop. I ran a quick microbenchmark for the below code with `timeit`, which gave me overhead on the order of ~1us, which is hopefully not too bad for eager mode:
```
def fast_function():
return torch._C._dispatch_has_kernel_for_dispatch_key(op_call.name(), torch._C.DispatchKey.CompositeImplicitAutograd)
import timeit
time_taken = timeit.timeit(fast_function, number=1000)
# printed 0.12..., aka 1.2us
print(f'func={str(op_call)}, time={str(time_taken)}')
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149514
Approved by: https://github.com/kwen2501, https://github.com/albanD, https://github.com/wanchaol
Cudagraphs is careful to not allow any memory recorded to escape globally without having a reference to the tensor. This is because we may later reclaim that memory for a cudagraph recording and we need to mark the tensor as erroring on access. Very occasionally, a stray tensor will have been allocated locally but not yet cleaned up. In this case, we enter the slow path and try to gc.collect() to deallocate it. From a hard to repro internal use case, this was fixed by an additional `cuda.synchronize()`.
i also snuck in an outdated comment and a duplicate line removal.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149741
Approved by: https://github.com/BoyuanFeng, https://github.com/Skylion007
Split from #148186
The diff can be re-generated with the following code in the repo root directory on main branch:
```python
import re
from pathlib import Path
def replace(m: re.Match) -> str:
s = m.group()
if '\n' not in s:
return s
indent = m.group("indent")
varnames = s.removesuffix("None").replace("=", "").replace("(", "").replace(")", "").split()
return "\n".join(
[
f"{indent}(",
*(f"{indent} {varname}," for varname in varnames),
f"{indent}) = (None,) * {len(varnames)}",
]
)
file = Path('test/inductor/s429861_repro.py')
content = file.read_text(encoding='utf-8')
new_content = re.sub(
r"^(?P<indent> *)\w+ *=(\s*(\(\s*\w+\s*\)|\w+)\s*=\s*)+None$",
replace,
content,
flags=re.MULTILINE,
)
file.write_text(new_content, encoding='utf-8')
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148554
Approved by: https://github.com/jansel
Summary: When we call torch.inference_mode, we seem to skip Autograd key causing the custom op export uses to be not decomposed properly before subclass dispatching starts. We fix this by force desugaring this op at Python key
Test Plan: test
Differential Revision: D71599541
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149698
Approved by: https://github.com/bdhirsh
This is an attempt to fix#119698
I was unable to reproduce the original described problem on the latest trunk but the proposed fix makes sense. Instead of adding locks like the original (unlanded) fix I changed a few of the cache writes to be atomic file swaps (write to temp file, rename file) which should have the same effect without blocking reads.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149654
Approved by: https://github.com/eellison
Summary:
We need to properly fakify torchbind objects, including the ones in graph module attributes, so the resgitered fake implementation works properly.
- _fakify_script_objects in `compile_fx`
- Allow fake torchbind objects in `torchbind_constants`
Remove `node.meta["unbacked_bindings"]` for `aot_compile` in `compile_fx`. Otherwise `ShapeProp` will fail when trying to resolve the `unbacked_bindings` of `with_effect` tokens.
Update `sigrid_transforms_test` to use the latest `torch._inductor.aot_compile` API.
Add a test for `Fakify torchbind objects in compile_fx and add tests for SigridTransformsInstanceTorchBind` in `e2e_test`.
Test Plan:
```
buck run //caffe2/torch/fb/sparsenn:sigrid_test -- -r test_transform_torch_bind
buck run //sigmoid/inference/test:e2e_test_cpu -- -r SigridTransforms
buck2 run mode/dev-nosan sigmoid/inference/ts_migration:pt2i_readiness_main -- --model_id 545017754 --test_suite ads_all --mode test_preproc
```
Differential Revision: D70013257
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149529
Approved by: https://github.com/angelayi
Summary:
1\ The current write item structure does not contain the amount of data that needs to be written.
2\ the planner.item already has a size primitive 'tensor_storage_size'. https://fburl.com/code/7a0gsmw7 But only for tensors.
3\ Right now, the only way the writer layer get hold of this property (fro non tensor data)
first do a lookup in to the actual tensor/bytes
then calculate the nbytes.
This change introduce a way to capture non-tensor data size within a write-plan item.
Test Plan: Existing UT.
Differential Revision: D71599725
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149699
Approved by: https://github.com/MeetVadakkanchery
In similar vein as https://github.com/pytorch/pytorch/pull/149517
When we added the rocm-mi300.yml earlier this year, we had lower capacity and we were just pipecleaning the workflow, so we set the trigger to only respond to pushes to main branch. But now we have more stability as well as capacity, and we would really like to ensure that the release branch is being tested on MI300s as well.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149672
Approved by: https://github.com/jeffdaily
2025-03-21 16:02:03 +00:00
2787 changed files with 134392 additions and 70089 deletions
If you do not have upload permissions, please ping @seemethere or @soumith to gain access
## New versions
New ROCm versions can be added by creating a new make target with the next desired version. For ROCm version N.n, the target should be named `magma-rocmNn`.
Make sure to edit the appropriate environment variables (e.g., DESIRED_ROCM) in the `Makefile` accordingly. Remember also to check `build_magma.sh` to ensure the logic for copying over the files remains correct.
if ! python perf-tests/modules/test_cpu_torch.py "${ARGS[@]}";then
echo"To reproduce this regression, run \`cd .ci/pytorch/perf_test/ && bash ${FUNCNAME[0]}.sh\` on your local machine and compare the runtime before/after your code change."
if ! python perf-tests/modules/test_cpu_torch_tensor.py "${ARGS[@]}";then
echo"To reproduce this regression, run \`cd .ci/pytorch/perf_test/ && bash ${FUNCNAME[0]}.sh\` on your local machine and compare the runtime before/after your code change."
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
