Tentatively eliminate graph break overhead #3741
Draft
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: api [Python]
narendasan
reviewed
Aug 1, 2025
peri044
reviewed
Aug 15, 2025
narendasan
reviewed
Aug 19, 2025
narendasan
reviewed
Aug 19, 2025
| @@ -174,6 +173,8 @@ def __init__( | |||
| self.cudagraph: Optional[torch.cuda.CUDAGraph] = None | |||
| self._caller_stream: Optional[torch.cuda.Stream] = None | |||
| self._engine_stream: Optional[torch.cuda.Stream] = None | |||
| self.output_tensors: Optional[List[torch.Tensor]] = None | |||
| self.sync_stream = True | |||
There was a problem hiding this comment.
Just inherit stream from PyTorch / input tensors
narendasan
reviewed
Aug 19, 2025
| @@ -381,16 +405,17 @@ def setup_input_tensors( | |||
|
|
|||
| # For shape tensors, we use CPU pointers and for data tensors, we use GPU pointers | |||
| # as per TensorRT requirements | |||
| if self.engine.is_shape_inference_io(input_name): | |||
| if self.is_shape_inference_io[i]: | |||
There was a problem hiding this comment.
Probably better to make this a dictionary and key on names, instead of implicitly relying on input order to stay the same over time
narendasan
reviewed
Aug 19, 2025
| input_name, tuple(contiguous_inputs[i].shape) | ||
| ) | ||
| if shape_changed: | ||
| self.context.set_input_shape( |
There was a problem hiding this comment.
Can we safely assume execution context holds shape between inference calls?
narendasan
reviewed
Aug 19, 2025
| @@ -994,6 +994,10 @@ def preserve_module_specs( | |||
| ) as f: | |||
| f.write(trt_module.get_layer_info()) | |||
|
|
|||
| # Only set the requires_unique_output flag for the last TRT Module when user has access to the output tensor | |||
| if trt_module and settings.use_python_runtime: | |||
| trt_module.set_requires_unique_output(True) | |||
There was a problem hiding this comment.
How is this going to work with serialization in C++?
Also make the name clearer like trt_module.module_is_output_operator or trt_module.requires_unowned_output_tensor
Yeah once we think all changes in pytorch is valid and I can make changes accordingly |
cehongwang
force-pushed
the
graph-break
branch
from
b2ef228 to
a9a27b1
Compare
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
Please include a summary of the change and which issue is fixed. Please also include relevant motivation and context. List any dependencies that are required for this change.
Fixes # (issue)
Type of change
Please delete options that are not relevant and/or add your own.
Checklist: