feat: Autocast #3878
Open
feat: Autocast #3878
Conversation
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
github-actions
bot
added
component: lowering
zewenli98
commented
Oct 29, 2025
zewenli98
commented
Oct 29, 2025
Comment on lines
437
to
444
| enable_autocast: bool = _defaults.ENABLE_AUTOCAST, | ||
| low_precision_type: Optional[ | ||
| Union[torch.dtype, dtype] | ||
| ] = _defaults.LOW_PRECISION_TYPE, | ||
| nodes_to_exclude: Collection[str] = _defaults.NODES_TO_EXCLUDE, | ||
| targets_to_exclude: Collection[Target] = _defaults.TARGETS_TO_EXCLUDE, | ||
| data_max: float = _defaults.DATA_MAX, | ||
| max_depth_of_reduction: Optional[int] = _defaults.MAX_DEPTH_OF_REDUCTION, |
Collaborator
Author
There was a problem hiding this comment.
Before merging, these args should be added to other compile functions in this file.
zewenli98
commented
Oct 29, 2025
| ]: | ||
| # GEMM: A (M, K) @ B (K, N) = C (M, N) | ||
| self.reduction_depth = input_0_dims[-1] | ||
| # TODO: Add more reduction ops here |
Collaborator
Author
There was a problem hiding this comment.
Should any more reduction targets be added?
zewenli98
commented
Oct 29, 2025
Comment on lines
374
to
377
| assert ( | ||
| contiguous_inputs[i].dtype == self.input_dtypes[i] | ||
| ), f"Dtype mismatch for {i}th input({input_name}). Expect {self.input_dtypes[i]}, got {contiguous_inputs[i].dtype}." | ||
|
|
Collaborator
Author
There was a problem hiding this comment.
This precision check was removed because after autocasting, if the first layer runs in fp16 but the original input is fp32, input_dtypes will become fp16 but contiguous_inputs is still fp32.
Similarly, other runtimes also removed the check.
zewenli98
commented
Oct 30, 2025
Comment on lines
33
to
46
| # nodes = list(gm.graph.nodes) | ||
| # # insert enter autocast node in the beginning of the graph | ||
| # with gm.graph.inserting_before(nodes[0]): | ||
| # enter_autocast_node = gm.graph.call_function(torch.amp.autocast_mode._enter_autocast, args=("cuda", torch.float16, True, True)) | ||
| # enter_autocast_node.meta.update(getattr(nodes[0], "meta", {})) | ||
|
|
||
| # # insert exit autocast node before the return node, assuming the return node is the last node | ||
| # with gm.graph.inserting_before(nodes[-1]): | ||
| # exit_autocast_node = gm.graph.call_function(torch.amp.autocast_mode._exit_autocast, args=(enter_autocast_node,)) | ||
| # exit_autocast_node.meta.update(getattr(nodes[-1], "meta", {})) | ||
|
|
||
| # gm = clean_up_graph_after_modifications(gm) | ||
| # gm, new_signature = replace_autocast_with_hop_pass(gm, None) | ||
| # logger.debug("Graph after replace_autocast_with_hop_pass:\n%s", gm.graph) |
Collaborator
Author
There was a problem hiding this comment.
If using Pytorch autocast to wrap the whole model, pytorch will control the precision of each node per the doc and I didn't find a way to customize based on our ruleset.
peri044
reviewed
Oct 30, 2025
Collaborator
peri044
left a comment
peri044
left a comment
There was a problem hiding this comment.
Can you also update the documentation at https://github.com/pytorch/TensorRT/blob/main/docsrc/user_guide/mixed_precision.rst
Comment on lines
110
to
113
| auto expected_type = | ||
| util::TRTDataTypeToScalarType(compiled_engine->exec_ctx->getEngine().getTensorDataType(name.c_str())); | ||
| TORCHTRT_CHECK( | ||
| inputs[i].dtype() == expected_type, |
Collaborator
There was a problem hiding this comment.
Is this not necessary now ?
py/torch_tensorrt/dynamo/lowering/passes/rule_based_autocast.py
Outdated
Show resolved
Hide resolved
py/torch_tensorrt/dynamo/lowering/passes/rule_based_autocast.py
Outdated
Show resolved
Hide resolved
py/torch_tensorrt/dynamo/lowering/passes/rule_based_autocast.py
Outdated
Show resolved
Hide resolved
Collaborator
|
For Tests
L1 or L2 tests |
github-actions
bot
added
component: tests
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Description
Weak typing behavior in TensorRT is deprecated. However it is a good way to maximize performance. Therefore, we want to create similar PyTorch native system to use with Torch-TensorRT that recovers some of this behavior.
Fixes #3869
Type of change
Checklist: