Skip to content

fix: Improve input weight handling to acc_ops convolution layers in FX#1886

Merged
gs-olive merged 1 commit intopytorch:mainfrom
gs-olive:acc_conv_bugfix
May 10, 2023
Merged

fix: Improve input weight handling to acc_ops convolution layers in FX#1886
gs-olive merged 1 commit intopytorch:mainfrom
gs-olive:acc_conv_bugfix

Conversation

@gs-olive
Copy link
Copy Markdown
Contributor

@gs-olive gs-olive commented May 4, 2023

Description

  • In line with recent deprecation of the kEXPLICIT_PRECISION flag in TRT
  • Add improved logic for weight handling in conv layers to bring all accelerated conv layers into agreement and fix errors in the Dynamo path arising from TRTTensor inputs
  • Fix minor typos in error messages

Fixes #1885

Type of change

  • Bug fix (non-breaking change which fixes an issue)

Checklist:

  • [ x ] My code follows the style guidelines of this project (You can use the linters)
  • [ x ] I have performed a self-review of my own code
  • [ x ] I have commented my code, particularly in hard-to-understand areas and hacks
  • [ x ] I have made corresponding changes to the documentation
  • [ x ] I have added tests to verify my fix or my feature
  • [ x ] New and existing unit tests pass locally with my changes
  • [ x ] I have added the relevant labels to my PR in so that relevant reviewers are notified

@gs-olive
fix: Input weight handling to accelerated conv
a3b654f 
- In line with recent deprecation of the kEXPLICIT_PRECISION flag in TRT
- Add improved logic for weight handling in `conv` layers to bring all
accelerated conv layers into agreement and fix errors in the Dynamo path
arising from TRTTensor inputs
- Fix minor typos
@gs-olive gs-olive added priority: high release: v1.4 Tagged to be included in v1.4 labels May 4, 2023
@gs-olive gs-olive requested review from frank-wei and wushirong May 4, 2023 22:56
@gs-olive gs-olive self-assigned this May 4, 2023
@github-actions github-actions bot requested a review from narendasan May 4, 2023 22:56
gs-olive
gs-olive commented May 4, 2023
View reviewed changes
py/torch_tensorrt/fx/converters/acc_ops_converters.py
bias = to_numpy(kwargs["bias"]) # type: ignore[arg-type]

if network.has_explicit_precision:
if network.has_explicit_precision or isinstance(kwargs["weight"], TRTTensor):
Copy link
Copy Markdown
Contributor Author

@gs-olive gs-olive May 4, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Mirrors the existing code for acc_ops.conv1d

@gs-olive gs-olive added the component: dynamo Issues relating to the `torch.compile` or `torch._dynamo.export` paths label May 4, 2023
@gs-olive gs-olive mentioned this pull request May 5, 2023
Closed
frank-wei
frank-wei approved these changes May 10, 2023
View reviewed changes
Copy link
Copy Markdown
Contributor

@frank-wei frank-wei left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM!

@gs-olive gs-olive merged commit 39585b1 into pytorch:main May 10, 2023
@gs-olive gs-olive deleted the acc_conv_bugfix branch May 10, 2023 22:21
gs-olive added a commit that referenced this pull request May 10, 2023
@gs-olive
fix: Improve input weight handling to acc_ops convolution layers in…
9d2bd91 
… FX (#1886)
@gs-olive gs-olive mentioned this pull request May 10, 2023
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@frank-wei frank-wei frank-wei approved these changes

@wushirong wushirong Awaiting requested review from wushirong

@narendasan narendasan Awaiting requested review from narendasan

Assignees

@gs-olive gs-olive

Labels

cla signed component: api [Python] Issues re: Python API component: dynamo Issues relating to the `torch.compile` or `torch._dynamo.export` paths component: fx fx priority: high release: v1.4 Tagged to be included in v1.4

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

🐛 [Bug] Encountered bug when compiling model with acc_ops conv converter

3 participants

@gs-olive @frank-wei @facebook-github-bot