This reverts commit 7743149b2be4a9eba7e0997ccdc6abe552bec266.
Reverts
* https://github.com/pytorch/pytorch/pull/135503
* https://github.com/pytorch/pytorch/pull/135502
* https://github.com/pytorch/pytorch/pull/135422
This passes this test. Earlier, the getitem would stay like a getitem in the Fx graph. But now the fake tensor propagations fails saying that .item is called. It seems that torch function is not getting triggered while fake tensor propagation.
```
import torch
from torch.nn.attention.flex_attention import BlockMask, _mask_mod_signature, _score_mod_signature, flex_attention
from torch._inductor.lowering import make_pointwise, register_lowering
from torch._inductor.virtualized import ops
from torch.nn.attention.flex_attention import create_block_mask
torch.set_default_device('cuda')
flex_attention = torch.compile(flex_attention, dynamic=False)
prefix_lengths = torch.arange(8)
def prefix_lm(b, h, q, kv):
return prefix_lengths[b] >= kv
mask = create_block_mask(prefix_lm, 8, None, 512, 512, _compile=True)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136590
Approved by: https://github.com/Chillee
Summary:
X-link: https://github.com/pytorch/benchmark/pull/2454
This adds structured logging overhead at a per compile basis to compilation metrics.
To do so, we track the frame_id_frame_compile_id that trace_structured uses to categorize compiles, and use that as the key in our timing table.
Implementation notes:
- If there's times we call trace_structured without a compile id, the time won't be measured. Not really a good way around that today given the compile id framework of compilation metrics. Strobelight is still the best way to measure on a per job basis.
- We don't actually measure the time it takes to log the compilation metrics itself. Fundamentally, it's not possible to log this properly if we're storing the logging number *in* compilation metrics, since there's no way to measure it before we do it(unless we want discrepancies between dynamo_compile and tlparse, which seems suboptimal). Hopefully for a large job, the cost of structured_logging compilation metrics itself is small.
- I wanted to use frame_phase_timing here, but there's a bunch of ids to iron out, and I don't really want to deal with that headache. compilation_time_metrics is sort of what I want, but that isn't by frame/compile id, so it's also a bit off. Putting it into torch.logging as a separate thing so logging tracks its own overhead seems fine, though.
Test Plan:
Run benchmarks/nanogpt and staging logger. See that the new compilation metric is logged to the staged dynamo_compile table:
https://fburl.com/scuba/logger_staging_jjwu_30582a48f1ff9cf5f4ac50a4c40af/xazjg5xq
Note that the sum(structured_logging_overhead_s) / sum(entire_frame_compile_time) = 8.387 / 124.278 = 6%, which seems reasonable as the overhead for a small compilation like this.
You can also look at samples for a more detailed log of this.
Reviewed By: oulgen
Differential Revision: D62643611
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136142
Approved by: https://github.com/bobrenjc93
Summary:
Record remote cache time saved via frame_phase_timing
We add to the "phase" when remote cache hits and saves us time, so that we have a 1:1 correspondence between a frame and time saved.
Test Plan:
Internally run benchmark, see that it's populated in sandbox table after previous diff lands and logger config is actualized.
Show that column exists in table:
https://fburl.com/scuba/logger_staging_jjwu_30582a48f1ff9cf5f4ac50a4c40af/fp2te0ff
Note that an earlier version of D62105258 had the column as a string so the staging table is a bit messed up. But you can see the most recent samples have the column populates as a float.
Reviewed By: aorenste
Differential Revision: D62106921
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135490
Approved by: https://github.com/aorenste
Fixes https://github.com/pytorch/pytorch/issues/133858
Details: Previously Dynamo would treat dataclasses as UserDefinedVariables. This was non-desirable if we would like to proxy the value into the graph, which is needed for TensorSubclassMetadata. To rectify this, frozen dataclasses are now able to be proxied similarly to NamedTuples. We require the object to be frozen, because if arbitrary mutation were allowed, we would need to replay those mutations in the graph after construction of the object.
For tracing construction of the variable, the generated `__init__` for the dataclass uses `object.__setattr__` because frozen dataclasses throw errors on the usual `__setattr__` invocation. With this treatment, no special handling is needed in dynamo for frozen dataclass construction.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134846
Approved by: https://github.com/bdhirsh, https://github.com/anijain2305
Restart the work from PR https://github.com/pytorch/pytorch/pull/100331 in this new PR since it's hard to rebase. It would be expected that some code is copy/pasted from the previous PR and main idea is the same.
Previously we see relatively large compilation time increase due to too many loop orders being considered. This PR tries to continue the work by doing pruning and only considering loop orders that we know for sure are relevant (i.e. do it on demand).
Some manually created cases that loop ordering matters are added as unit tests. The PR can make sure inductor does not miss fusion opportunities for them.
This PR should solve the not-able to fusion problem in https://github.com/pytorch/pytorch/issues/130015
Right now there is still significant increase of compilation time. I'll disable the feature by default. Later on after the compilation time issue is resolved, I'll enable it by default.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126254
Approved by: https://github.com/jansel
PYTHONPATH=$(pwd) python benchmarks/update_hint_benchmark.py out
as of this diff, compile_time_instruction_count counts the number of instruction from within
convert_frame.compile_inner
```
update_hint_regression,compile_time_instruction_count,10522459165
```
will add result from CI once populated.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133834
Approved by: https://github.com/aorenste
Summary:
We should always emit an end event in a finally block so that if a unit test or job fails, the stack is still correct.
Also, we use thread local storage for the stack, so that in multithreaded scenarios the stack will still be correctly added.
Test Plan:
Run benchmark and see that everything still works
Run
```
TORCH_LOGS=dynamo buck run test/functorch:test_aotdispatch -- -r test_backward_mutation_on_grad_out
```
With some extra logging to see that start events with the correct stack are emitted, and the end events are also emitted even though the test fails at runtime.
Differential Revision: D61682556
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134279
Approved by: https://github.com/aorenste
Part of #134054.
This corresponds to the pytorch mypy changes from D61493706. Updating takes so
long and touches so many files that it's impossible to land as a whole without conflicting with some other intermediate change.
So landing these 'type: ignore' for pytorch in advance of them actually being needed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134202
Approved by: https://github.com/Skylion007
Summary:
This diff implements a bunch of views for internal scuba viewing.
TODOS that I might punt to another diff:
- Saving cache stats via counter is definitely sus here, but there's not really a good way to track "fx graph cache hit for this compile phase" right now. Will think about this more.
- We should definitely log frame id, compile id, etc
- We should definitely be logging configs. That way, we can A/B test based on whether a config is turned on.
- idk what I'm doing with compile_uuid yet, but it's useful when you want to look at samples for a single run. I think if we had mast job info this field is not needed, but it's nice to be able to drill down to a single run and get its chrome trace view or icicle view, so idk
Test Plan:
All of the above views are run with nanogpt benchmark:
```
buck run mode/opt caffe2/benchmarks/dynamo:torchbench -- --training --backend=inductor --only nanogpt --performance
```
Differential Revision: D61603243
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134118
Approved by: https://github.com/oulgen
This PR adds support for tracing `torch._C._pop_torch_function_stack()` without graph breaking and in order to verify the state change also adds replay of mutations to the torch function mode stack via side_effects appending supplemental bytecode as we do for other python mutable objects.
Details:
To represent the torch function mode stack symbolically a deque field is added to the instruction translator. When the InstructionTranslator is initialized, all modes are read from the current torch function mode stack, and stashed in a global weak ref for later access (using existing sources) without needing to push/pop the python/cpp torch function mode stack.
During tracing, when `_pop_torch_function_stack` is encountered a value is popped from this deque and the variable tracker representing the mode is returned. To ensure the true torch function mode stack matches this state, `TorchFunctionModeStackVariable`, a singleton, is marked as mutated, this adds it to side effects, where during final codegen, side effects will codegen a call to a python helper which will update the python torch function mode stack.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133131
Approved by: https://github.com/jansel
ghstack dependencies: #133130, #133729
This PR adds a guard on the torch function mode stack state at the beginning of tracing. The way this is implemented is via a new leaf guard which is passed the initial stack state at construction and compares it to the stack state at the time the guard is run.
Details:
The stack state is extracted via popping all modes, appending them to a list, and pushing all modes back. This list is stored on the output graph and read during guard construction to pass to the stack mode guard. There the length and types of the modes are recorded. Next time the guard is run it compares this recorded state to the current mode stack state.
To implement this in python a helper function was added to utils.py and this is used if cpp guards are not enabled.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133130
Approved by: https://github.com/anijain2305
`torch.cuda.Event` objects are different from `torch.cuda.Stream` in that events are not pooled, meaning we can't look up a previously created CUDA event object by ID. This prevents CUDA event object created outside of the Dynamo graph from being used within the graph (since Dynamo needs a way to emit a `call_function` line in the graph that does the retrieval of the event object for downstream op use). This PR adds a simple object pool within Dynamo utility, to support looking up CUDA event object by ID from within the Dynamo graph.
After this PR, if a user creates a CUDA event object outside of the graph and use that event within the graph, the behavior will exactly match eager.
Test commands:
- `pytest -rA test/dynamo/test_ctx_manager.py::CtxManagerTests::test_cuda_event_created_outside_of_graph`
- `pytest -rA test/dynamo/test_ctx_manager.py::CtxManagerTests::test_cuda_event_across_graph_break`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133635
Approved by: https://github.com/yifuwang
ghstack dependencies: #133532, #133531, #133636
