Skip Dynamo graph break for scalar-only bin_ops when tensorify is enabled (#2687)#2687
Open
nandesuka wants to merge 1 commit intopytorch:mainfrom
Open
Skip Dynamo graph break for scalar-only bin_ops when tensorify is enabled (#2687)#2687nandesuka wants to merge 1 commit intopytorch:mainfrom
nandesuka wants to merge 1 commit intopytorch:mainfrom
Conversation
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
|
@nandesuka has exported this pull request. If you are a Meta employee, you can view the originating Diff in D103055794. |
…bled (pytorch#2687) Summary: X-link: pytorch/pytorch#182026 When `tensorify_python_scalars` is enabled and `dynamic=True`, Dynamo lifts Python float/int arguments as 0-dim tensor placeholders followed by `.item()` calls, producing SymFloat/SymInt values. The tensorify pass (in AOTAutograd) then converts scalar ops back to tensor ops. However, `torch.add/sub/mul/div` on all-scalar args hit a graph break in TorchInGraphFunctionVariable ("Attempted to call torch in-graph function on only torch.SymInt arguments") that was added before the tensorify infrastructure existed. This prevents the graph from reaching AOTAutograd, so tensorify never runs. Fix: check `_is_tensorify_enabled()` to skip the graph break when tensorify can handle the scalar-to-tensor conversion downstream. This allows scalar-only binary ops to be compiled rather than graph-broken. The `_is_tensorify_enabled()` helper is moved to `torch/_dynamo/utils.py` so it is shared between the Dynamo tracing check (in `variables/torch.py`) and the tensorify FX pass itself (`torch/fx/passes/_tensorify_python_scalars.py`), eliminating the previous duplication. A comment at the Dynamo call site documents that this is an intentional abstraction violation: Dynamo peeks at a downstream pass's config to decide whether to graph-break, because when tensorify is enabled these scalar-only ops will be handled later in the pipeline. Changed files: - `torch/_dynamo/utils.py`: Add shared `_is_tensorify_enabled()` (cached env var + JustKnobs check) - `torch/_dynamo/variables/torch.py`: Import `_is_tensorify_enabled` from utils, add abstraction-violation comment, gate graph break on tensorify being disabled - `torch/fx/passes/_tensorify_python_scalars.py`: Replace inline knob logic with import of shared `_is_tensorify_enabled()` - `test/dynamo/test_misc.py`: Add `test_tensorify_scalar_only_bin_ops` and `test_tensorify_scalar_only_bin_ops_int` verifying scalar-only binary ops compile without graph break when tensorify is enabled Differential Revision: D103055794
nandesuka
added a commit
to nandesuka/pytorch
that referenced
this pull request
May 7, 2026
…bled (pytorch#182026) Summary: X-link: pytorch/benchmark#2687 When `tensorify_python_scalars` is enabled and `dynamic=True`, Dynamo lifts Python float/int arguments as 0-dim tensor placeholders followed by `.item()` calls, producing SymFloat/SymInt values. The tensorify pass (in AOTAutograd) then converts scalar ops back to tensor ops. However, `torch.add/sub/mul/div` on all-scalar args hit a graph break in TorchInGraphFunctionVariable ("Attempted to call torch in-graph function on only torch.SymInt arguments") that was added before the tensorify infrastructure existed. This prevents the graph from reaching AOTAutograd, so tensorify never runs. Fix: check `_is_tensorify_enabled()` to skip the graph break when tensorify can handle the scalar-to-tensor conversion downstream. This allows scalar-only binary ops to be compiled rather than graph-broken. The `_is_tensorify_enabled()` helper is moved to `torch/_dynamo/utils.py` so it is shared between the Dynamo tracing check (in `variables/torch.py`) and the tensorify FX pass itself (`torch/fx/passes/_tensorify_python_scalars.py`), eliminating the previous duplication. A comment at the Dynamo call site documents that this is an intentional abstraction violation: Dynamo peeks at a downstream pass's config to decide whether to graph-break, because when tensorify is enabled these scalar-only ops will be handled later in the pipeline. Changed files: - `torch/_dynamo/utils.py`: Add shared `_is_tensorify_enabled()` (cached env var + JustKnobs check) - `torch/_dynamo/variables/torch.py`: Import `_is_tensorify_enabled` from utils, add abstraction-violation comment, gate graph break on tensorify being disabled - `torch/fx/passes/_tensorify_python_scalars.py`: Replace inline knob logic with import of shared `_is_tensorify_enabled()` - `test/dynamo/test_misc.py`: Add `test_tensorify_scalar_only_bin_ops` and `test_tensorify_scalar_only_bin_ops_int` verifying scalar-only binary ops compile without graph break when tensorify is enabled Test Plan: ## Unit Tests Added two tests in `test_misc.py` (`DynamoOpPromotionTests`): - `test_tensorify_scalar_only_bin_ops`: Verifies `torch.add/sub/mul/div(float, float)` with `dynamic=True` compiles without graph break when tensorify is enabled - `test_tensorify_scalar_only_bin_ops_int`: Verifies `torch.add(int, int)` with `dynamic=True` compiles without graph break when tensorify is enabled Differential Revision: D103055794
meta-codesync
Bot
changed the title
nandesuka
force-pushed
the
export-D103055794
branch
from
d5fd2d5 to
9b7020e
Compare
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
nandesuka
had a problem deploying
to
docker-s3-upload
GitHub Actions
Failure
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
X-link: pytorch/pytorch#182026
When
tensorify_python_scalarsis enabled anddynamic=True, Dynamo lifts Python float/int arguments as 0-dim tensor placeholders followed by.item()calls, producing SymFloat/SymInt values. The tensorify pass (in AOTAutograd) then converts scalar ops back to tensor ops.However,
torch.add/sub/mul/divon all-scalar args hit a graph break in TorchInGraphFunctionVariable ("Attempted to call torch in-graph function on only torch.SymInt arguments") that was added before the tensorify infrastructure existed. This prevents the graph from reaching AOTAutograd, so tensorify never runs.Fix: check
_is_tensorify_enabled()to skip the graph break when tensorify can handle the scalar-to-tensor conversion downstream. This allows scalar-only binary ops to be compiled rather than graph-broken.The
_is_tensorify_enabled()helper is moved totorch/_dynamo/utils.pyso it is shared between the Dynamo tracing check (invariables/torch.py) and the tensorify FX pass itself (torch/fx/passes/_tensorify_python_scalars.py), eliminating the previous duplication.A comment at the Dynamo call site documents that this is an intentional abstraction violation: Dynamo peeks at a downstream pass's config to decide whether to graph-break, because when tensorify is enabled these scalar-only ops will be handled later in the pipeline.
Changed files:
torch/_dynamo/utils.py: Add shared_is_tensorify_enabled()(cached env var + JustKnobs check)torch/_dynamo/variables/torch.py: Import_is_tensorify_enabledfrom utils, add abstraction-violation comment, gate graph break on tensorify being disabledtorch/fx/passes/_tensorify_python_scalars.py: Replace inline knob logic with import of shared_is_tensorify_enabled()test/dynamo/test_misc.py: Addtest_tensorify_scalar_only_bin_opsandtest_tensorify_scalar_only_bin_ops_intverifying scalar-only binary ops compile without graph break when tensorify is enabledDifferential Revision: D103055794