This PR fixes #157354
It fixes the issue in 'cmake/public/cuda.cmake' where a diagnostic message incorrectly showed an empty CUDA version when 'FindCUDA' and header-reported versions differed.
The problem was caused by this line:
set(${cuda_version_from_findcuda} ${CUDA_VERSION_STRING})
This incorrectly used the value of cuda_version_from_findcuda as a variable name. As a result the version string wasn't assigned and the error message omitted the version. This has been corrected to:
set(cuda_version_from_findcuda ${CUDA_VERSION_STRING})
Now the diagnostic message properly displays the CUDA version reported by FindCUDA.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157370
Approved by: https://github.com/soulitzer
Summary:
NumPy based tensor rebuilding from serialization has been deprecated by other backends (eg. [XLA](https://github.com/pytorch/pytorch/pull/137444)). The new flow has CPU storage being constructed with data from the file and then moved to the target backend device.
Furthermore, relying on numpy for serialization will fail loudly when torch.load flips weights_only.
Reviewed By: andyanwang
Differential Revision: D77843238
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157884
Approved by: https://github.com/albanD
Summary: Following implementation of the updated ATen Backend for mtia, and diffs enabling in tree view ops (D75266206, D75385411), we can remove custom logic from reducer to handle MTIA view operations.
Test Plan:
CI
Rollback Plan:
Reviewed By: egienvalue
Differential Revision: D77843212
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157882
Approved by: https://github.com/albanD, https://github.com/andyanwang
**Summary**
To enable use case where the input DTensor to `split` op has `Partial()` placement,
this PR treats `Partial()` in the same way with `Replicate()`. That means, `split` op
only unshards the `Shard(dim=x)` if `x == split_dim` and keep other placement
untouched.
**Test**
Added a new test because `test_dtensor_ops` doesn't test `Partial()` placement.
`pytest test/distributed/tensor/test_tensor_ops.py -s -k test_split_on_partial`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157991
Approved by: https://github.com/zpcore
Most of the work had already been done by @jeffdaily in #154680, but there was one remaining check that needed to be modified in order for `torch._scaled_mm` to use cuBLAS over CUTLASS when available.
I tested this change by rebuilding PyTorch locally with CUDA 12.9 and ran `torch._scaled_mm` under the profiler, and observed that the kernel being launched is called `nvjet_qqtst_128x128_128x6_1x1_h_bz_coopA_algo2_ovscale_TNT` (where `ovscale` stands for "outer vector scaling", I believe, which is how cuBLAS calls this scaling mode).
I then benchmarked the new kernels against the old CUTLASS ones on a standard 700W H100 GPU. I used the same approach as in #134781, and obtained these speed-ups:
We see that the two kernels perform very closely (I'm surprised, I would have expected cuBLAS to outperform CUTLASS across the board), with some thin/skewed shapes becoming worse but some very large shapes becoming better.
I guess the questions are whether we consider this a net-zero change (given that there's improvements _and_ degradations), and how large we consider the burden of maintaining our own CUTLASS kernels.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157905
Approved by: https://github.com/eqy, https://github.com/Skylion007, https://github.com/drisspg
# Motivation
Refactor `CUDAAllocatorConfig` to reuse `AcceleratorAllocatorConfig` and `ConfigTokenizer`. We would deprecate those option that overleap with `AcceleratorAllocatorConfig` in the following PR and keep them only for BC.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150312
Approved by: https://github.com/albanD
ghstack dependencies: #149601, #157908
**Summary**
Enable fp8 qconv on CPU. It's part of the plan to enable fp8 static quantization on CPU. This PR only adds FP8 support of the existing int8 qconv op. It does not add a new op nor does it affect frontend or quantization flow. The schema of the qconv op is not changed either.
So, the FP8 qconv shares the same op as INT8 qconv and the difference is that src/wei dtype is fp8 instead of int8. The output dtype can be fp8/float32/bfloat16. The implementation uses the oneDNN library.
Note:
OneDNN does not support quantized fp8 convolution until v3.9 but the version used in PyTorch is v3.7.2. So, the op goes to the reference kernel for now. And we have also update the oneDNN path so that it's compatible with the fp8 dtype. Once oneDNN is upgraded to v3.9 or newer, minimum changes are needed to enable the oneDNN path. And we have ensured that the behavior of the reference kernel is the same as the new oneDNN's implementation.
- oneDNN version < 3.9 (now)
- Always go to the reference kernel
- oneDNN version >= 3.9 (future)
- Go to reference kernel on old platforms (without AMX)
- Use oneDNN on new platforms (with AMX)
**Test plan**
```
pytest test/quantization/core/test_quantized_op.py -k "qconv and fp8"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157076
Approved by: https://github.com/leslie-fang-intel, https://github.com/jerryzh168
Refactor to allow TMA descriptors to be used in general codegen. TMA descriptors can only be generated if the conditions listed in the triton documentation for [make_tensor_descriptor](https://triton-lang.org/main/python-api/generated/triton.language.make_tensor_descriptor.html) are met.
Some implementation details:
- The `TMACompatibilityChecker` class holds and checks the conditions required for a load / store operation to be represented by a tma descriptor load / store
- The current TMA API requires that the innermost block size loads atleast 16 bytes of data. e.g. if the block shape is [YBLOCK, XBLOCK] and the tensor dtype is float32, this requires that XBLOCK >= 4. It is therefore required that the triton heuristics are aware of the minimum block sizes for the IO operations in the kernel. The minimum block sizes are determined in the `TMACompatibilityChecker` class and are passed to the triton heuristics when the block sizes are not static. The heuristic config options are then filtered to ensure that the minimum block size restriction is met.
Testing:
- Refactored test_torchinductor_strided_blocks.py to also test the `use_tensor_descriptor` option.
This requires an upgrade to Triton version 3.4.0: https://github.com/pytorch/pytorch/issues/154206
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157906
Approved by: https://github.com/jansel
Summary: In Pytorch 2.5 we added source code attribution to PT2 traces. Each Torch-Compiled Region will now have its frame id and frame compile id associated with it. Update the image in the doc and add a description of this in the doc itself
Test Plan:
{F1980179183}
Rollback Plan:
Differential Revision: D78118228
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158066
Approved by: https://github.com/aaronenyeshi
Users may want compile-related but customized logging info to dynamo_compile. One example is to logging the current training iteration index when recompilation happens. In general, current training iteration index is not available to compiler, since the same compiled function may be called multiple times in the same training iteration. The user could provide the training iteration index in a user hook where torch.compile logs it when recompilation happens.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157961
Approved by: https://github.com/masnesral
Fixes https://github.com/pytorch/pytorch/issues/154312
Fix logdet returning finite values for singular matrices on CUDA (https://github.com/pytorch/pytorch/issues/154312https://github.com/pytorch/pytorch/issues/154312)
PyTorch's logdet function returns mathematically incorrect finite values for
singular matrices on CUDA devices instead of the expected -inf. This occurs
because cuSOLVER and LAPACK produce tiny non-zero diagonal elements (~1e-16)
instead of exact zeros for singular matrices.
**Problem:**
Issue https://github.com/pytorch/pytorch/issues/154312 matrix returns finite values instead of -inf for singular matrices.
**Solution:**
Implemented NumPy-style two-tier singularity detection with GPU sync point removal:
1. **Primary detection**: Use LAPACK's built-in singularity detection via info parameter
2. **Backup detection**: Apply threshold-based detection for numerical edge cases
3. **Zero GPU sync points**: Eliminated all .item(), std::get<0>(), and scalar extractions
4. **Pure tensor operations**: All computations use tensor operations throughout
**Performance Impact:**
Based on comprehensive benchmarking across matrix sizes and data types:
- **Overall Impact**: 0.85× average speedup (+18.0% overhead)
- **CPU Performance**: 0.84× average speedup (+18.8% overhead)
- **CUDA Performance**: 0.85× average speedup (+17.3% overhead)
**Performance Trade-offs:**
- **Small matrices (16×16, 64×64)**: Higher overhead due to tensor operation setup costs
- **Large matrices (512×512, 2048×2048)**: Near-zero overhead, with some cases showing slight improvements
- **GPU sync elimination**: Removes expensive GPU→CPU synchronization bottlenecks
**Results:**
- ✅ All singular matrices now correctly return -inf on both CPU and CUDA
- ✅ Original issue https://github.com/pytorch/pytorch/issues/154312 matrix now works correctly
- ✅ Results match NumPy's slogdet behavior exactly
- ✅ Zero GPU synchronization points for improved performance
- ✅ Comprehensive edge case testing added
**Verification:**
Before: torch.linalg.slogdet(singular_matrix) → finite values (incorrect)
After: torch.linalg.slogdet(singular_matrix) → (sign=0, logabsdet=-inf) ✅
The implementation uses pure tensor operations to eliminate GPU sync points while
maintaining robust singularity detection through a two-tier approach.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-Authored-By: Claude <noreply@anthropic.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157910
Approved by: https://github.com/lezcano, https://github.com/IvanYashchuk, https://github.com/albanD
Co-authored-by: Claude <noreply@anthropic.com>
Differential Revision: D78089705
Previously to support overriding autotune configs for post fusion kernels in Inductor with a lookup table, we only keyed on the source code. However, the same source code could have multiple optimal configs, due to the input sizes. With this, we have many collisions in our lookup table, leading to subpar configs. A way around this is to add the size_hints to the lookup key as well
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158026
Approved by: https://github.com/jansel
Summary:
Replaced ASSERT_FLOAT_EQ which defaults to fixed kMaxUlps ( = 4-ULP , See gtest-internal.h) with ASSERT_NEAR which lets us set epsilon to 1e-3, (approximately 3 ULPs). This allows for slightly stricter and tunable comparison.
Test Plan:
**Before Fix**
✗ Fail:
qnnpack:pytorch_qnnpack_testApple - FULLY_CONNECTED_SPARSE_OP_8x1/unit_batch_dynamic_prepacked (0.0s)
'Expected equality of these values:
output_dynamic[i * outputChannels() + c]
Which is: 9.9160004
accumulators_float[i * outputChannels() + c]
Which is: 9.9159956
at 0, 17: reference = 9.9159955978393555, optimized = 9.9160003662109375
------------------------------
**After Fix**
Everything passes
Rollback Plan:
Differential Revision: D77911682
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157861
Approved by: https://github.com/kimishpatel, https://github.com/lucylq, https://github.com/malfet
## Overview
This PR adds a kwarg to the `table()` method of the profiler allowing users to specify a time unit to be used for all results in the profiling table. The available options are: `s`, `ms` and `us`. If an invalid unit or no unit is provided, then a time unit is selected based on the size of the value (current default behaviour).
## Testing
A unit test has been added to verify this works correctly.
## Documentation
I couldn't find any documentation specific to the `table()` function beyond doc strings which have been updated.
## Example Output
```
import torch
from torch.profiler import profile
with profile() as prof:
res = torch.mm(torch.rand(1024, 1024), torch.rand(1024, 1024))
print(prof.key_averages().table(time_unit="s"))
print(prof.key_averages().table(time_unit="ms"))
print(prof.key_averages().table(time_unit="us"))
print(prof.key_averages().table())
```
```
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
aten::rand 0.04% 0.000s 10.36% 0.014s 0.007s 2
aten::empty 0.04% 0.000s 0.04% 0.000s 0.000s 2
aten::uniform_ 10.27% 0.014s 10.27% 0.014s 0.007s 2
aten::mm 89.64% 0.119s 89.64% 0.119s 0.119s 1
aten::resolve_conj 0.00% 0.000s 0.00% 0.000s 0.000s 3
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 0.133s
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
aten::rand 0.04% 0.055ms 10.36% 13.735ms 6.868ms 2
aten::empty 0.04% 0.054ms 0.04% 0.054ms 0.027ms 2
aten::uniform_ 10.27% 13.626ms 10.27% 13.626ms 6.813ms 2
aten::mm 89.64% 118.892ms 89.64% 118.896ms 118.896ms 1
aten::resolve_conj 0.00% 0.004ms 0.00% 0.004ms 0.001ms 3
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 132.631ms
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
aten::rand 0.04% 55.495us 10.36% 13735.202us 6867.601us 2
aten::empty 0.04% 54.121us 0.04% 54.121us 27.061us 2
aten::uniform_ 10.27% 13625.586us 10.27% 13625.586us 6812.793us 2
aten::mm 89.64% 118892.284us 89.64% 118895.981us 118895.981us 1
aten::resolve_conj 0.00% 3.697us 0.00% 3.697us 1.232us 3
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 132631.183us
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Name Self CPU % Self CPU CPU total % CPU total CPU time avg # of Calls
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
aten::rand 0.04% 55.495us 10.36% 13.735ms 6.868ms 2
aten::empty 0.04% 54.121us 0.04% 54.121us 27.061us 2
aten::uniform_ 10.27% 13.626ms 10.27% 13.626ms 6.813ms 2
aten::mm 89.64% 118.892ms 89.64% 118.896ms 118.896ms 1
aten::resolve_conj 0.00% 3.697us 0.00% 3.697us 1.232us 3
---------------------- ------------ ------------ ------------ ------------ ------------ ------------
Self CPU time total: 132.631ms
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157913
Approved by: https://github.com/sraikund16
This adds new context manager based PG management to dist2. This allows for managing the active process group much in the same way as a stream
```py
with dist2.process_group(pg):
dist2.current_process_group().allreduce(...).wait()
```
matches
```py
with torch.cuda.stream(stream):
torch.cuda.current_stream().synchronize()
```
Test plan:
```
pytest test/distributed/test_dist2.py -k context
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157988
Approved by: https://github.com/fduwjj
DCE was incorrectly eliminating unused random operations like torch.rand() that have global RNG side effects, causing inconsistent results between eager and compiled execution modes.
**Root cause**: Python random functions (torch.rand, torch.randn, etc.) don't have the _nondeterministic_seeded attribute, so node.is_impure() returns False, allowing DCE to eliminate them despite advancing global RNG state.
**Solution**: Enhanced is_impure() in torch/fx/node.py to recognize Python random functions and mark them as impure when they use global RNG, regardless of the impure_random parameter setting. This ensures consistency between eager and compiled execution even when config.fallback_random=False.
**Key features**:
- Handles comprehensive list of random functions: rand, randn, randint, randperm, rand_like, randn_like, randint_like, normal, poisson, bernoulli, multinomial
- Generator optimization: Only marks as impure when using global RNG (no generator or generator=None). Operations with explicit generators don't affect global state and can be optimized.
- Works with both impure_random=True and impure_random=False cases
- Cleaner architecture: addresses root cause rather than working around it
**Tests**: Enhanced test_impure_random to verify both FX tracing and AOT compilation codepaths, ensuring random operations are preserved and eager/compiled execution consistency is maintained.
🤖 Generated with [Claude Code](https://claude.ai/code)
Fixes https://github.com/pytorch/pytorch/issues/151524
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157981
Approved by: https://github.com/mlazos
Co-authored-by: Claude <noreply@anthropic.com>
Summary:
We are testing enabling back autovectorization in some codepaths.
These resulted in crashes when compiling using clang17, we are now relying on clang19.
Test Plan:
buck2 build //caffe2/caffe2/fb/transforms:sigrid_interface
We are going to deploy it on ads workloads
Rollback Plan:
Differential Revision: D77448445
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157984
Approved by: https://github.com/Skylion007
Design doc: https://docs.google.com/document/d/1ncV7RpJ8xDwy8-_aCBfvZmpTTL824C-aoNPBLLVkOHM/edit?tab=t.0 (internal)
- Add codegen for static linkage
- refactor test code for test_compile_after_package tests
For now, the following options must be used together with `"aot_inductor.compile_standalone": True`.
"aot_inductor.package_cpp_only": True,
Will change `"aot_inductor.package_cpp_only"` to be automatically set to True in followup PR.
```
python test/inductor/test_aot_inductor_package.py -k test_compile_after_package
python test/inductor/test_aot_inductor_package.py -k test_run_static_linkage_model
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157129
Approved by: https://github.com/desertfire
We assumed that the output in an FX graph would always just be a
list[Tensor], even in the single tensor return case.
It is possible for the output to be a single Tensor. This can happen
by calling torch.fx.split_module on the module.
Test Plan:
- new test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157803
Approved by: https://github.com/oulgen
# Motivation
This PR aims to generalize `AllocatorConfig` to be device-agnostic. Introduce the class `AcceleratorAllocatorConfig` to clarify its scope as a configuration manager for accelerator backends (e.g., CUDA, XPU). The another name `AllocatorConfig` is now reserved for a potential future base class that can unify configuration handling for both CPU and accelerator allocators, should similar requirements arise for the CPU path.
# Design Rule
## Overall
This class configures memory allocation for both device and host memory. A single `AcceleratorAllocatorConfig` instance is shared across all accelerator backends, such as CUDA and XPU, under the assumption that relevant environment variables apply uniformly to all accelerators. Device-specific configuration extensions are supported via hooks (see `registerDeviceConfigParserHook`).
Introduce a new class `ConfigTokenizer` to help process the env variable config key-value pair
## Naming Convention:
- Public API names in `AcceleratorAllocatorConfig` should be device-generic.
- Members prefixed with `pinned_` are specific to the host/pinned allocator.
- Environment variable names should be generic across backends.
- Comma-separated key-value pairs in the format: `key:value`. Use square brackets `[]` for list values Example: `key1:123, key2:[val1,val2]`
## Environment Variables:
- The default environment variable for configuration is `PYTORCH_ALLOC_CONF`.
- For backward compatibility, `PYTORCH_CUDA_ALLOC_CONF` and `PYTORCH_HIP_ALLOC_CONF` are also supported with lower priority.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149601
Approved by: https://github.com/albanD
Introduces support for a new `OVERRIDEABLE` backend in the SDPA module, improves backend selection logic, and adds corresponding tests. In addition, a fallback mechanism was added when a specific backend is unavailable, enhancing user configurability.
### Backend Support and Selection Enhancements:
* Added `at::SDPBackend::overrideable` to the list of available SDPA backends in the `Context` class (`aten/src/ATen/Context.h`).
* Updated the backend selection logic in `select_sdp_backend_xpu` to include the `OVERRIDEABLE` backend and added a fallback mechanism for unsupported `FLASH_ATTENTION` on XPU.
* Adjusted error messaging in `_fused_sdp_choice_xpu` to reflect the inclusion of the `OVERRIDEABLE` backend. (`aten/src/ATen/native/mkldnn/xpu/Attention.cpp`)
### Test Additions for Backend Fallback and Selection:
* Added new unit tests to validate fallback behavior for `FLASH_ATTENTION` to `OVERRIDEABLE` and to verify correct backend selection when `MATH` is enabled. (`test/test_transformers.py`,)
### Codebase Updates for Backend Integration:
* Introduced `OVERRIDEABLE` as a new member of the `_SDPBackend` enum. (`torch/_C/__init__.pyi.in`)
* Extended `_backend_names` and updated related methods to handle the `OVERRIDEABLE` backend. (`torch/nn/attention/__init__.py`)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156669
Approved by: https://github.com/guangyey, https://github.com/drisspg
Summary:
`None` and `Ellipsis` in multi-dimensional indexing was previously not covered.
Moreover, we introduce a small optimization for `slice(None)` and a passthrough when symints do not appear in the indexing.
The remaining case is where indexing is by tensor, which is fairly complicated; we passthrough in that case.
Test Plan:
added tests
Rollback Plan:
Differential Revision: D77943929
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157821
Approved by: https://github.com/pianpwk
Summary:
This DIFF is to fix the following issue:
In python source code for CompiledFxGraph,the FX graph segment for the Triton kernel is broken. For example, the following function
def fn(a, b, c):
x = torch.nn.functional.linear(a, b)
x = x.sin()
x = x.t() + c
return x
Inductor compiled this FX graph into two nodes: the first one is mm, the second one is a triton kernel for sin + transpose + add. The FX graph segment for the triton kernel is like the following:
Graph fragment:
%add : [num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%permute_1, %arg2_1), kwargs = {})
Basically only "add" node in the FX graph.
The root cause is function caffe2/torch/_inductor/utils.py:gather_origins does not detect the realized node correctly.
To fix this issue, the IRNode is checked if it is one of the following IRNode:
ir.ComputedBuffer,
ir.InputsKernel,
ir.InputBuffer,
ir.ReinterpretView,
ir.TemplateBuffer,
If it is one of them, it is realized, otherwise, it is not.
Test Plan:
buck2 run mode/opt caffe2/test/inductor:provenance_tracing -- caffe2.test.inductor.test_provenance_tracing.TestProvenanceTracingArtifact.test_triton_kernel_to_post_grad_tracing_cuda
Rollback Plan:
Differential Revision: D77748371
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157578
Approved by: https://github.com/mlazos
@animesh pointed out using whitelist for strides can result in confusing graphs as follows
```
s60: "Sym(s60)", L_hidden_states_: "bf16[1, 4096, 3072][s60, 3072, 1]cuda:0"
```
We probably want to capture the relationship between sizes and strides anyways so let's make it so the whitelist only makes the sizes dynamic. That same graph now looks lik ethis
```
L_hidden_states_: "bf16[1, 4096, 64][262144, 64, 1]cuda:0"
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157960
Approved by: https://github.com/pianpwk
Summary:
- adding mmap-ing
- more efficient writing in larger chunks
latency from ~150s to ~6s for simple row-wise consolidation of a 7gb model sharded across 4 ranks
Test Plan:
ran consolidation with the following code:
```
from torch.distributed.checkpoint._consolidate_hf_safetensors import consolidate_safetensors_files
import time
start_time = time.time()
consolidate_safetensors_files(base_path, consolidated_path)
end_time = time.time()
print(f"Time taken: {end_time - start_time} seconds")
```
With the old code this was taking a couple minutes and this is now down to ~6s.
Internal users can find the tensor shards in the manifold path: manifold://ankita_test_bucket/tree/safetensors
Rollback Plan:
Differential Revision: D77960054
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157936
Approved by: https://github.com/teja-rao, https://github.com/pradeepfn
If the final output file is in remote storage, then create a local temp directory to write the files and upload the files to the remotes storage after they are written.
Add a new config to the storage writer, `enable_consolidation`, so we don't need to rely on the presence of the `consolidation_output_path` to decide if consolidation is enabled. If `enable_consolidation` is True and `consolidation_output_path` isn't provided, the consolidated safetensors will be added to the same path as the sharded ones.
Differential Revision: [D77554585](https://our.internmc.facebook.com/intern/diff/D77554585/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157371
Approved by: https://github.com/pradeepfn
Fixes#157195
### Summary:
Fixed Issue 157195 by adding a new error message for torch.binomial in **aten/src/ATen/native/Distributions.cpp**
### Explanation
According to the issue,
```
import torch
torch.binomial(torch.tensor([10]).long(), torch.tensor([0.5]))
```
`RuntimeError: Found dtype Float but expected Long`
It looks like we are getting a Tensor error rather than a binomial function error. Since the error is coming from **pytorch/aten/src/ATen/TensorIterator.cpp**, it seems like it is trying to align the tensor data to the same datatype for smooth tensor computations instead of giving a binomial function error.
I tried using both arguments as longs and both as ints and got the right binomial function error
```
torch.binomial(torch.tensor([10]).long(), torch.tensor([0.5]).long())
NotImplementedError: "binomial_cpu" not implemented for 'Long'
```
```
torch.binomial(torch.tensor([10.0]).int(), torch.tensor([0.5]).int())
NotImplementedError: "binomial_cpu" not implemented for 'Int'
```
But when I have both as different datatypes, the TensorIterator.cpp error comes back trying to align the datatypes.
`RuntimeError: Found dtype Float but expected Long`
I then tried finding where the NotImplementation Error was documented and found it in **pytorch/aten/src/ATen/Dispatch.h** in lines 193 - 211
```
#define AT_DISPATCH_SWITCH(TYPE, NAME, ...) \
[&] { \
const auto& the_type = TYPE; \
constexpr const char* at_dispatch_name = NAME; \
/* don't use TYPE again in case it is an expensive or side-effect op */ \
at::ScalarType _st = ::detail::scalar_type(the_type); \
RECORD_KERNEL_FUNCTION_DTYPE(at_dispatch_name, _st); \
switch (_st) { \
__VA_ARGS__ \
default: \
TORCH_CHECK_NOT_IMPLEMENTED( \
false, \
'"', \
at_dispatch_name, \
"\" not implemented for '", \
toString(_st), \
"'"); \
} \
}()
```
In the **AT_DISPATCH_SWITCH** function, it picks a tensor and its datatype and checks if the Tensor datatype matches the supported datatypes. If not we get the Not Implemented error. Unfortunately, I think the **AT_DISPATCH_SWITCH** function, uses the `common_dtype` from TensorIterator in order to run. So TensorIterator.cpp needs to happen before the AT_DISPATCH_SWITCH function.
### Summary: We are getting the wrong error message because **TensorIterator.cpp** gets called and errors out due to Tensor datatype mismatch before we can get the right error message in **Dispatch.h** for torch.binomial not supporting that datatype.
### Options for the Fix
**Option 1**: Make the error message in TensorIterator.cpp more general so it applies to torch.binomial. An error message along the lines
`RunTime Error : "Tensor Datatypes", op.target_dtype," and ", common_dtype_, "are different "`
**Option 2**: Add an error message for the binomial function datatype mismatch before the the TensorIterator.cpp error message gets called.
Although Option 1 seemed easier I think Option 2 might be better as it is more specific to the binomial function while Option1 would affect all Tensors with datatype mismatch.
**This PR applies the fix for Option 2**
After Fix :
```
torch.binomial(torch.tensor([10]).long(), torch.tensor([0.5]))
RuntimeError: Binomial function arguments count and prob must have same datatype of type Float, got: count = Long, prob = Float
```
```
torch.binomial(torch.tensor([10]).long(), torch.tensor([0.5]).long())
NotImplementedError: "binomial_cpu" not implemented for 'Long'
```
@malfet
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157658
Approved by: https://github.com/soulitzer
Previous to this PR, torch.onnx.export(..., dynamo=True, veriy=True, report=True) does not support symbolic arguments. Such examples are like follwing:
```python
class M(torch.nn.Module):
def forward(self, a, x):
return a + torch.tensor(1) + x
op = torch.onnx.export(M(), (1, torch.ones(2)),
dynamic_shapes=(torch.export.Dim.DYNAMIC, {0: torch.export.Dim.DYNAMIC}),
dynamo=True, report=True)
```
symbolic arguments are like constant arguments that they don't have tensor_meta wither. Besides, torch.export.export supports model inputs having constants, which is different from the legacy issue: https://github.com/pytorch/pytorch/issues/99534 where we tried to get the FX directly from dynamo export. Thus, `_remove_non_tensor` is deleted from args processing.
NOTE: If the ConstantArugment shows up in exported_program, it was kept to align the length of inputs to nn.Module, but it's irrelevant to the model graph, hwich is why in ONNX model the input is omitted.
The test `test_constant_argument_user_input_is_omitted_in_onnx_graph` needs #157719
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157734
Approved by: https://github.com/justinchuby
Dynamo was aggressively specializing on lazy VTs over `set_name_hint` in
`STORE_FAST`, etc., and `isinstance` in `LOAD_FAST_CHECK`. This causes
regional `torch.compile` from optimizing ComfyUI GGUF + LoRA to either
(1). exceed the recompialtion limit of 8, which results in suboptimal
performance, and (2). even if recompilation limit is increased, the
compilation time gets unnecessarily high (180s v.s. 20s for Flux).
This patch fixes the recompilation issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156891
Approved by: https://github.com/williamwen42, https://github.com/mlazos
Summary:
Change AOTI_RUNTIME_DEVICE_CHECK to the following depending on device:
AOTI_RUNTIME_CUDA_CHECK
AOTI_RUNTIME_XPU_CHECK
AOTI_RUNTIME_CPU_CHECK
Currently in the codebase, only `AOTI_RUNTIME_CUDA_CHECK` is used.
This shouldn't change anything as of now, but we do this to prepare for simultaneouly loading multiple backends (e..g CPU and CUDA) in AOTI standalone.
We don't want people writing `AOTI_RUNTIME_DEVICE_CHECK` for both CPU and CUDA checks. This could cause compilation problems when we statically link both CPU and CUDA models.
Test Plan:
CI
Rollback Plan:
Reviewed By: muchulee8
Differential Revision: D77742977
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157818
Approved by: https://github.com/jingsh
## Issue being addressed
`is_sparse` presents itself as determining if a tensor is sparse. HOWEVER, it only does checks against the tensor for `sparse_coo`. This has lead to confusion from developers as when non-coo sparse tensors are provided it return false, despite those tensors being sparse.
## Considered Remedy
Fixing this is do-able however would result in complexity as existing systems may depend on this behavior remaining consistent, and even inside of pytorch is_sparse is used by `bform` which states that it supports only `sparse_csr and sparse_coo` meaning additional work/thought would have to go into solving for `sparse_csc` and `sparse_bsr`
## Remedy provided in this PR
In lieu of these complications the lowest risk highest gain action was to add clear warning messaging to the function for now to avoid confusion to developers utilizing the function. The rest of the function behavior remains identical
## Issue content
Addresses issue number: #101385
Original issue: https://github.com/pytorch/pytorch/issues/101385
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157378
Approved by: https://github.com/soulitzer
Summary: We added the ability to make Annotating Global or Local based on an input flag in PyTorch but didn't add the args to the linter
Reviewed By: mzzchy
Differential Revision: D77959409
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157858
Approved by: https://github.com/mzzchy
Add env var AOT_INDUCTOR_ENABLE_LTO to enable clang's ThinLTO by setting AOT_INDUCTOR_ENABLE_LTO=1. The LTO is disabled by default because it may increase the build time.
Rollback Plan:
Differential Revision: D77899195
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157773
Approved by: https://github.com/desertfire
This is to unblock "dp2ep" Expert Parallel + TP integration in torchtitan https://github.com/pytorch/torchtitan/pull/1324.
It does two things:
1. Slightly modifies the glue code for FSDP/HSDP + TP to work with FSDP/HSDP + EP and FSDP/HSDP + EP + TP. I kept the name `FSDPParam._tp_spec` to make the change minimal. We can consider renaming it in the future if it confuses people, but I heard @wanchaol has a plan to rewrite DTensor strided sharding entirely.
2. Lifts the check of `_validate_tp_mesh_dim` for `torch.distributed.tensor.parallel.parallelize_module`, as in EP or EP+TP this check is too strict. In particular it assumes a DeviceMesh must have `mesh_dim_names` which is not always true. I'm also removing the file `torch/distributed/tensor/parallel/_utils.py` it belongs entirely, as the other check `_deprecate_warnings`, added two years ago, is not used any more.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157216
Approved by: https://github.com/wanchaol, https://github.com/weifengpy
I was confused about why the distributed tests weren't showing up quickly on HUD, its because the call of run_tests.py for distributed didn't include upload artifacts while running flag, so set it to default to IS_CI so I don't need to put the flag everywhere
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157868
Approved by: https://github.com/huydhn
Per title, as it fails with the following error if "+PTX" was used in `TORCH_CUDA_ARCH_LIST`:
```
File "/usr/local/lib/python3.12/dist-packages/torch/profiler/_pattern_matcher.py", line 313, in skip
has_tf32 = all(int(arch[3:]) >= 80 for arch in torch.cuda.get_arch_list())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/dist-packages/torch/profiler/_pattern_matcher.py", line 313, in <genexpr>
has_tf32 = all(int(arch[3:]) >= 80 for arch in torch.cuda.get_arch_list())
^^^^^^^^^^^^^
ValueError: invalid literal for int() with base 10: 'pute_120'
```
Because slicing `arch[3:]` will not end up on having only digits for `compute_120` element of `torch.cuda.get_arch_list()`:
```python
>>> torch.cuda.get_arch_list()
['sm_75', 'sm_80', 'sm_86', 'sm_90', 'sm_100', 'sm_120', 'compute_120']
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157711
Approved by: https://github.com/Skylion007, https://github.com/sraikund16
**Background:**
```Shell
[1376/2332] Building CUDA object caffe2/CMakeFiles/torch_...h/csrc/distributed/c10d/symm_mem/NCCLSymmetricMemory.cu.o
/root/Git.d/pytorch/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp(450): warning #68-D: integer conversion resulted in a change of sign
size_t numelIn_ = -1;
^
Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
/root/Git.d/pytorch/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp(451): warning #68-D: integer conversion resulted in a change of sign
size_t numelOut_ = -1;
^
/root/Git.d/pytorch/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp(450): warning #68-D: integer conversion resulted in a change of sign
size_t numelIn_ = -1;
^
Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
/root/Git.d/pytorch/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.hpp(451): warning #68-D: integer conversion resulted in a change of sign
size_t numelOut_ = -1;
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157889
Approved by: https://github.com/mlazos
The cpu inductor processes .to(torch.uint8) incorrectly, leading to numerical inconsistencies. The convert_float_to_int8 function may return incorrect results for negative inputs, such as -2.xx, when the data type is uint8_t, producing 0 instead of 255. This issue stems from the clamping logic; we should avoid converting min_val to uint8_t too early
Fixes https://github.com/pytorch/pytorch/issues/156788
@leslie-fang-intel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157267
Approved by: https://github.com/leslie-fang-intel
Summary: We found there's a special case in recent APS model where the input tensor has smaller size compared to the split size. It will be automatically truncated in split.Tensor thus we add extra condition check for split_with_sizes when do the normalization.
Test Plan:
### unit
```
buck2 test 'fbcode//mode/dev-nosan' fbcode//caffe2/test/inductor:split_cat_fx_aten_passes -- test_split_aten_normalization
```
Buck UI: https://www.internalfb.com/buck2/2ecd1ef8-8efe-4245-b4c8-282c23645b3c
Test UI: https://www.internalfb.com/intern/testinfra/testrun/7599824648585787
Network: Up: 3.9GiB Down: 9.2GiB (reSessionID-1396c91e-0dd2-457b-a49b-a6ab1f2a7d8f)
Loading targets. Remaining 0/5344 99617 dirs read, 1074949 targets declared
Analyzing targets. Remaining 0/123279 4988547 actions, 5966764 artifacts declared
Executing actions. Remaining 0/728058 209:52:59.9s exec time total
Command: test. Finished 12466 local, 209448 remote, 1226 cache (1% hit) 42:10.5s exec time cached (0%)
Time elapsed: 26:07.6s
Tests finished: Pass 2. Fail 0. Fatal 0. Skip 0. Build failure 0
### E2E
before fix:
aps-afoc_apop_pt2_v0-db2fe0449a
after fix:
aps-afoc_apop_pt2_v0-755ad0cdc6
Rollback Plan:
Differential Revision: D77961394
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157857
Approved by: https://github.com/anijain2305
Summary: When `compile_standalone` is True, we set `package_cpp_only` to True as well. We raise an error if `package_cpp_only` is explicitly set to False in config.
Test Plan:
```
buck2 run mode/dev-nosan fbcode//caffe2/test/inductor:test_aot_inductor -- -r TestAOTInductorConfig
```
Rollback Plan:
Differential Revision: D77889754
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157731
Approved by: https://github.com/desertfire
Fixes#157673
For the call trace:
```
......
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\codegen\common.py", line 2569, in reduction
return self.kernel.reduction(dtype, src_dtype, reduction_type, value)
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\codegen\cpp.py", line 2155, in reduction
self._gen_parallel_reduction_buffers(acc, acc_type, reduction_type, init_dtype)
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\codegen\cpp.py", line 1942, in _gen_parallel_reduction_buffers
reduction_prefix_array(
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\codegen\cpp.py", line 335, in reduction_prefix_array
if cpp_builder.is_msvc_cl()
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\cpp_builder.py", line 317, in is_msvc_cl
return _is_msvc_cl(get_cpp_compiler())
File "D:\Programs\Python\virtualenvs\torch_code-afvE469o\lib\site-packages\torch\_inductor\cpp_builder.py", line 240, in _is_msvc_cl
subprocess.check_output([cpp_compiler, "/help"], stderr=subprocess.STDOUT)
torch._inductor.exc.InductorError: UnicodeDecodeError: 'utf-8' codec can't decode byte 0xd3 in position 0: invalid continuation byte
```
On non-English language pack msvc environment, compiler path has raised `utf-8` issue. I add the `normalize_path_separator` to normalize the compiler path and avoid the issue.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157835
Approved by: https://github.com/jansel
Today, we always create and record an events in two places:
1) Upon seeing the first producer, we record an event on the producer, and we wait for this event in two places: (1) when the engine goes to run the consumer, the consumer stream waits for this event. (2) prior to doing accumulation, the accumulation stream waits for this event.
2) After doing accumulation, we record an event on the accumulation stream and wait for this event in a single place: when the engine goes to run the consumer.
We do not actually need to record the event in the cases where the 1st producer stream is the same as the consumer and as the accumulation stream, and where the accumulation stream is the same as the consumer stream.
Removing this unnecessary create + record event should save a few us for each instance avoided.
Fixes https://github.com/pytorch/pytorch/issues/157407
----
Manual test plan:
- [x] @eqy to confirm perf is restored
- [x] Running the repro originally reported before/after the patch
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157503
Approved by: https://github.com/eqy
ghstack dependencies: #155715
Previeusly, if users want to let pytorch determine the cuda arch when jit loading cuda extensions, they should left environment variable `TORCH_CUDA_ARCH_LIST` empty, but which will raise an warning. This commit add an option to set `TORCH_CUDA_ARCH_LIST=native`, to tell pytorch users want to use native cuda arch intentionally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156923
Approved by: https://github.com/ezyang
This implements a new `wait_stream` API in Work that matches how `wait` works for ProcessGroupNCCL for CPU based backends such as Gloo.
The idea is to support Gloo communication overlap in FSDPv2/HSDP with minimal changes to FSDP.
There was a previous attempt to make FSDPv2 use Work.wait but given the extensive stream semantics used it doesn't play nicely. https://github.com/pytorch/pytorch/pull/148780
This uses a "Baton" CUDA kernel which spinlocks on a pinned CPU tensor waiting for it to be set.
Test plan:
```
pytest test/distributed/test_c10d_gloo.py -v -k wait_stream
pytest test/distributed/test_c10d_nccl.py -v -k wait_stream
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156883
Approved by: https://github.com/kwen2501, https://github.com/fduwjj
# Feature
If a Triton kernel has a complicated indexing expression, Inductor may decide to precompute it on the host and pass it to the kernel as an argument. This happens in situations like broadcasts with dynamic shapes.
This PR adds support for this feature to Inductor's FX IR backend.
We generate FX IR for precomputed size args in 3 steps:
1. In `PythonWrapperCodegen`, this PR refactors the relevant code to use a `SymbolicCallArgLine` instead of raw Python strings. This stores a (symbol, expr) pair. (Prior to this PR, it was (str, expr), but changing this to a symbol makes it easier to do substitutions later on.)
2. In `WrapperFxCodegen`, keep a dict of {symbol: expr} arg defs which gets updated whenever we see a `SymbolicCallArgLine`.
3. When the FX backend sees a `KernelCallLine`, it uses this dict to replace symbolic call args with their definitions.
In the longer run, it might be desirable to emit FX nodes defining these symbolic call args. That way, we could reuse the size computation when the same kernel is called multiple times. However, I wasn't sure if there was an existing way to generate FX nodes from a sympy expression, and implementing that seemed like overkill for the present purposes.
# Test plan
Added a new CI test exercising this feature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157758
Approved by: https://github.com/jansel
Via google search I got to `torch.autograd.profiler` and implemented my code with it. Only to be taken by surprise finding `torch.profile.profiler`, which has a note saying the autograd one is legacy.
This just adds such note to `autograd.profiler` to avoid this confusion and waste of time to future people in my situation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157459
Approved by: https://github.com/sraikund16
Summary: These changes in D76442012 got reverted after the PR landed due to aps_models/ads/launchers/pearl/tests/ne/e2e_deterministic_tests:pearl_e2e_ne_tests failing with `Config not loaded due to no timely response from configerator. Likely configerator_proxy or falcon_proxy are not healthy`, but that test failing is definitely transient and unrelated to my changes, so re-creating the diff
Test Plan:
ensure tests pass
Rollback Plan:
Differential Revision: D77871099
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157715
Approved by: https://github.com/meetv18
We need to increase the tolerance slightly to ensure that certain models pass the accuracy check on the XPU device.
This pull request preserves the original tolerance threshold for CUDA/CPU devices and introduces a new key, higher_bf16_xpu, which only affects the XPU device.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156920
Approved by: https://github.com/soulitzer
This is to unblock "dp2ep" Expert Parallel + TP integration in torchtitan https://github.com/pytorch/torchtitan/pull/1324.
It does two things:
1. Slightly modifies the glue code for FSDP/HSDP + TP to work with FSDP/HSDP + EP and FSDP/HSDP + EP + TP. I kept the name `FSDPParam._tp_spec` to make the change minimal. We can consider renaming it in the future if it confuses people, but I heard @wanchaol has a plan to rewrite DTensor strided sharding entirely.
2. Lifts the check of `_validate_tp_mesh_dim` for `torch.distributed.tensor.parallel.parallelize_module`, as in EP or EP+TP this check is too strict. In particular it assumes a DeviceMesh must have `mesh_dim_names` which is not always true. I'm also removing the file `torch/distributed/tensor/parallel/_utils.py` it belongs entirely, as the other check `_deprecate_warnings`, added two years ago, is not used any more.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157216
Approved by: https://github.com/wanchaol, https://github.com/weifengpy
Summary:
I'm fairly sure the use of a custom metaclass is a holdover from pre-3.7 where Generic used a custom metaclass so we had to use multiple inheritance to avoid import-time failures.
At this point, `type(Generic)` is just `type` so it isn't needed, and we will get the least metaclass from our base classes, which means the `type(torch._C.Future)` isn't needed either, it will happen automatically just by inheritance.
Test Plan:
I'm fairly confident from local testing that this should be a no-op.
But also, Pytorch CI should give us pretty strong signal that this change doesn't break anything in case there's some edge case I missed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157757
Approved by: https://github.com/ezyang, https://github.com/Skylion007
This is useful for vLLM, which runs AOTAutograd directly on graphs after
they have been split.
I created a new flag for this instead of reusing
`keep_original_node_name` (please let me know if you think I should reuse this).
The reasoning is:
- The names of the placeholder nodes is different from the targets of
the placehoder nodes. The targets are the actual input names.
- Backwards compatibility: this API has been out for ~4 years, it
looks public, and it has extensive public use. For example, this change
would actually be BC-breaking to vLLM (they rely on the subgraph input
names being different at the moment).
Test Plan:
- new tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157733
Approved by: https://github.com/ezyang
Update s390x test marks
test_logs_out from test/dynamo/test_logging.py is updated
and no longer fails on s390x.
test_qengine from test/test_torch.py doesn't work on s390x:
no QEngine is available.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157541
Approved by: https://github.com/huydhn
In this PR, we are enabling `HPU` device-specific function calls for random operations. These calls will manage the setting and unsetting of the `context of Random Number Generator`.
While HPU devices typically utilize a `Mersenne-based RNG`, Dtensor-specific random operations employ an `offset-based (Philox) RNG tracker` which is specifically integrated with `CUDA` in scope.
To integrate a similar offset-based RNG tracker within the `HPU backend`, a backend-specific device handle function is necessary to identify the execution context of these random operations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156581
Approved by: https://github.com/jeromean, https://github.com/wanchaol
# Motivation
This PR aims to generalize `AllocatorConfig` to be device-agnostic. Introduce the class `AcceleratorAllocatorConfig` to clarify its scope as a configuration manager for accelerator backends (e.g., CUDA, XPU). The another name `AllocatorConfig` is now reserved for a potential future base class that can unify configuration handling for both CPU and accelerator allocators, should similar requirements arise for the CPU path.
# Design Rule
## Overall
This class configures memory allocation for both device and host memory. A single `AcceleratorAllocatorConfig` instance is shared across all accelerator backends, such as CUDA and XPU, under the assumption that relevant environment variables apply uniformly to all accelerators. Device-specific configuration extensions are supported via hooks (see `registerDeviceConfigParserHook`).
Introduce a new class `ConfigTokenizer` to help process the env variable config key-value pair
## Naming Convention:
- Public API names in `AcceleratorAllocatorConfig` should be device-generic.
- Members prefixed with `pinned_` are specific to the host/pinned allocator.
- Environment variable names should be generic across backends.
- Comma-separated key-value pairs in the format: `key:value`. Use square brackets `[]` for list values Example: `key1:123, key2:[val1,val2]`
## Environment Variables:
- The default environment variable for configuration is `PYTORCH_ALLOC_CONF`.
- For backward compatibility, `PYTORCH_CUDA_ALLOC_CONF` and `PYTORCH_HIP_ALLOC_CONF` are also supported with lower priority.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149601
Approved by: https://github.com/albanD
Target determination sorts the tests in a PR CI run based on heuristics about which tests are more relevant to the PR's changes. This can help provide faster CI signal as well as help alleviate capacity concerns as job durations should decrease due to catching failures earlier.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156545
Approved by: https://github.com/jeffdaily, https://github.com/clee2000
# Motivation
https://github.com/pytorch/pytorch/pull/155451 decoupled `torch._C._storage_Use_Count` from CUDA and introduced a corresponding unit test:
815545f2dd/test/test_torch.py (L257-L262)
However, this test fails when PyTorch is built with debug assertions enabled. @clee2000 disabled this UT in https://github.com/pytorch/pytorch/pull/156731. The root cause is that `_cdata` is obtained from an `intrusive_ptr`, not a `weak_intrusive_ptr`. As a result, calling `c10::weak_intrusive_ptr::use_count` on it triggers the internal assertion:
815545f2dd/c10/util/intrusive_ptr.h (L912-L917)
For example:
```python
a = torch.randn(10, device=device) # refcount=1, weakcount=1
prev_cf = torch._C._storage_Use_Count(a.untyped_storage()._cdata) # violate the assertation
```
This violates the expected invariant inside `weak_intrusive_ptr::use_count`, which assumes the pointer was originally constructed from a valid `weak_intrusive_ptr`. Actually, `storage_impl` is obtained from an `intrusive_ptr`.
815545f2dd/torch/csrc/Module.cpp (L2105-L2109)
# Solution
Use `c10::intrusive_ptr::use_count` instead.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157694
Approved by: https://github.com/albanD
Getting the build issue because of enablement of data type fp8 for onednn in qlinear and qlinear_prepack file after this commit c2185dc4a5626848df37cad214b73d5ae7dd4f17
Currrently cpuinfo is disable for power system because of that it is giving below error.
**Error:**
‘cpuinfo_has_x86_amx_int8’ was not declared in this scope
Made a required changes and now build issue got fixed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157469
Approved by: https://github.com/malfet
This PR uses `find_library` to replace `find_path`.
It also searches for NVSHMEM host lib and device lib separately.
Tested against system install location: /usr/local/lib and /usr/local/include.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157695
Approved by: https://github.com/Skylion007
ghstack dependencies: #157513
This PR adds a new config option, `caching_precompile`, and a `DynamoCache`, which loads and saves Dynamo Cache entries automatically. It also hooks up DynamoCache to PrecompileContext, so that we can save multiple cache entries.
When this configuration is turned on, we:
- Automatically create and initialize a CompilePackage on every torch.compile
- Automatically use BundledAutogradcache
- Automatically save the CompilePackage entry to DynamoCache after every compile
You can also use PrecompileContext.serialize() to manually serialize a full object.
I've added unit tests to exhibit this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155913
Approved by: https://github.com/zhxchen17
This PR uses `find_library` to replace `find_path`.
It also searches for NVSHMEM host lib and device lib separately.
Tested against system install location: /usr/local/lib and /usr/local/include.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157695
Approved by: https://github.com/Skylion007
ghstack dependencies: #157513
The structure is
```
torch/
__init__.py
version.py
```
When we import torch, only `torch/__init__.py` is executed by default.
The submodules like `version.py` are not automatically imported or attached to the torch module.
So without anything in `__init__.py`, `torch.version` may not be found. So in this PR, we make the import explicit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157584
Approved by: https://github.com/ezyang
Summary: We want to add versioning to DCP to the metadata so that whenever planner logic changes, we can use the version on save to determine how to load the data
Test Plan:
added a test
Rollback Plan:
Differential Revision: D76135887
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155343
Approved by: https://github.com/teja-rao
Note on backward precision over fp16:
A float16 number has 10 bits of mantissa, 5 bits of exponent, and 1 bit for the sign. If the sign bit is positive, then with a mantissa $m$ and exponent $e$ represented in base 10, the number that the float16 format represents is $(1 + m / 1024) \exp2(e)$. ([source](https://en.wikipedia.org/wiki/Half-precision_floating-point_format))
Consider adding two numbers $a$ and $b$ which have arbitrary mantissas, and say their exponents are $e_a = 1$ (so $2 \le a \lt 4$) and $e_b=-3$ (so $0.175 \le b \lt 0.25$). Assume that the result has the same exponent as $a$. Since the exponents differ by 4, we'll effectively need to truncate the 4 rightmost bits of $b$'s mantissa, which would introduce a maximum error on the order of $(2^4 / 1024) \exp2(-3) \approx 0.002$.
The error is nearly the same if $e_b = -2$ (so $0.25 \le b \lt 0.5$), where the 3 rightmost bits are truncated, giving a maximum error on the order of $(2^3 / 1024) \exp2(-2) \approx 0.002$. Same for $e_b=-1$.
So if we're adding up nine different numbers that all have exponents -3, -2, or -1, and they sum to a number with exponent 1, then we would expect a maximum error of several times greater than 0.002. In my comments above, summing those particular nine numbers in different ways gave results that ranged between 3.1816 and 3.1758, a difference of $0.0058 \approx 2.9 * 0.002$.
That's within the acceptable bounds, and we can safely just increase the error tolerance used in test_output_grad_match for the case of max_pool3d_backward with float16.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157498
Approved by: https://github.com/malfet
This PR fixes a minor typo in a comment in `torch/_dynamo/variables/torch.py`, changing 'paramter' to the correct spelling 'parameter'.
These small but meaningful changes help improve code readability and maintain the overall quality of the codebase.
Thanks for your time and review!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157651
Approved by: https://github.com/Skylion007
***SUMMARY***
The main function in this tests overrides that of the Gtest framework which contains it's `RUN_ALL_TESTS()` function. The main function in this test is called conditionally when conditions apply, in this case, when the C10_MOBILE directive is provided. This is wrong as we always want to call the `RUN_ALL_TEST()` function.
In this PR, we only make the test suite available for cases that apply, i.e if the C10_MOBILE directive exist which represents the caching allocator and is only exposed on mobile
***TEST PLAN***
This tests should run in modes where it applies which should be covered in the CI run.
Below shows a sample run in the dev-nosan mode which do not have the cache allocator
BEFORE
```
buck test fbcode//caffe2:cpu_caching_allocator_test
Discovered 0. Pass 0. Fail 0. Fatal 0. Skip 0. Timeout 0
⚠ Listing failed: caffe2:cpu_caching_allocator_test
Listing tests failed with error:
Failed to read from /data/users/ysuleiman/fbsource/buck-out/v2/test/buck-out/v2/test_discovery/fbcode/6dcc55a61c1b90b3/default/tpx_execution_dir/gtest_output_file.json. Listing process stdout: , stderr:
```
AFTER
```
buck test '@fbcode//mode/dev-nosan' fbcode//caffe2:cpu_caching_allocator_test
Analyzing targets. Remaining 0/46242 1871690 actions, 2251668 artifacts declared
Executing actions. Remaining 0/257870 83:28:24.4s exec time total
Command: test. Finished 10 remote, 112314 cache (99% hit) 83:22:43.5s exec time cached (99%)
Time elapsed: 2:57.7s
Tests finished: Pass 0. Fail 0. Fatal 0. Skip 0. Build failure 0
NO TESTS RAN
```
Rollback Plan:
steps:
- manual.note:
content: Revert this diff
Reviewed By: patskovn
Differential Revision: D77229077
Pull Request resolved: https://github.com/pytorch/pytorch/pull/156816
Approved by: https://github.com/kimishpatel
The process involves compiling thousands of files, and would take a long time. Fortunately, the compiled objects can be useful for your next build. When you modify some files, you only need to compile the changed files the next time.
@ -88,20 +88,19 @@ source venv/bin/activate # or `& .\venv\Scripts\Activate.ps1` on Windows
* If you want to have no-op incremental rebuilds (which are fast), see [Make no-op build fast](#make-no-op-build-fast) below.
* When installing with `python setup.py develop` (in contrast to `python setup.py install`) Python runtime will use
* When installing with `python -m pip install -e .` (in contrast to `python -m pip install .`) Python runtime will use
the current local source-tree when importing `torch` package. (This is done by creating [`.egg-link`](https://wiki.python.org/moin/PythonPackagingTerminology#egg-link) file in `site-packages` folder)
This way you do not need to repeatedly install after modifying Python files (`.py`).
However, you would need to reinstall if you modify Python interface (`.pyi`, `.pyi.in`) or
@ -785,7 +793,7 @@ If you are editing a single file and rebuilding in a tight loop, the time spent
Starting with CMake 3.29, you can specify the linker type using the [`CMAKE_LINKER_TYPE`](https://cmake.org/cmake/help/latest/variable/CMAKE_LINKER_TYPE.html) variable. For example, with `mold` installed:
Note on OpenMP: The desired OpenMP implementation is Intel OpenMP (iomp). In order to link against iomp, you'll need to manually download the library and set up the building environment by tweaking `CMAKE_INCLUDE_PATH` and `LIB`. The instruction [here](https://github.com/pytorch/pytorch/blob/main/docs/source/notes/windows.rst#building-from-source) is an example for setting up both MKL and Intel OpenMP. Without these configurations for CMake, Microsoft Visual C OpenMP runtime (vcomp) will be used.
@ -329,7 +335,6 @@ Additional libraries such as
You can refer to the [build_pytorch.bat](https://github.com/pytorch/pytorch/blob/main/.ci/pytorch/win-test-helpers/build_pytorch.bat) script for some other environment variables configurations
```cmd
cmd
@ -349,8 +354,7 @@ for /f "usebackq tokens=*" %i in (`"%ProgramFiles(x86)%\Microsoft Visual Studio\
:: [Optional] If you want to override the CUDA host compiler
set CUDAHOSTCXX=C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\bin\HostX64\x64\cl.exe
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.
