fix(rtx): add WAR to fall back grouped 3D deconvolutions to PyTorch#4188
Merged
zewenli98 merged 1 commit intopytorch:mainfrom Apr 17, 2026
Merged
fix(rtx): add WAR to fall back grouped 3D deconvolutions to PyTorch#4188zewenli98 merged 1 commit intopytorch:mainfrom
zewenli98 merged 1 commit intopytorch:mainfrom
Conversation
github-actions
bot
added
component: tests
Contributor
Author
|
This PR is based on #4178 and should be merged after it. |
tp5uiuc
force-pushed
the
grouped-deconv3d-war
branch
from
e0acfd4 to
b6f346b
Compare
tp5uiuc
commented
Apr 15, 2026
| bias=True, | ||
| output_padding=0, | ||
| ): | ||
| if groups > 1 and torch_tensorrt.ENABLED_FEATURES.tensorrt_rtx: |
Contributor
Author
There was a problem hiding this comment.
This test skips the capability validation and directly uses the TRTInterpreter, so I have skipped groups > 1 case for now. The newly added test_grouped_deconv3d_fallback in the test_models.py covers the fallback to pytorch route
tp5uiuc
force-pushed
the
grouped-deconv3d-war
branch
from
b6f346b to
7049608
Compare
Grouped 3D transposed convolutions (ConvTranspose3d with groups > 1) crash on TensorRT-RTX. This adds a convolution_capability_validator that detects these ops and rejects them from TRT conversion, causing the partitioner to keep them in PyTorch while other ops remain on TRT. Also renames depthwise_bf16_validator to convolution_capability_validator to reflect its broader scope, and removes the blanket skip on all 3D deconv tests — non-grouped cases now run through TRT on RTX. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
tp5uiuc
force-pushed
the
grouped-deconv3d-war
branch
from
7049608 to
2eb0734
Compare
zewenli98
approved these changes
Apr 17, 2026
Collaborator
zewenli98
left a comment
zewenli98
left a comment
There was a problem hiding this comment.
LGTM. feel free to merge
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
2 participants
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
Grouped 3D transposed convolutions (
ConvTranspose3dwithgroups > 1) crash on TensorRT-RTX. This PR adds a workaround that detects these ops at partitioning time and falls them back to PyTorch, while all other ops remain on TRT.Changes
aten_ops_converters.pydepthwise_bf16_validator→convolution_capability_validatorto reflect its broader scopetransposed=TrueANDgroups > 1AND input is 5D (3D spatial), reject from TRT on RTXis_grouped,is_transposed,is_3d,is_bf16) for readabilitytest_deconvolution_aten.py@unittest.skipIf(tensorrt_rtx)on all 3D deconv tests with a targetedself.skipTest()inside the test body for grouped cases onlytest_models.pytest_grouped_deconv3d_fallback: model-level test that verifies the fulltorch_tensorrt.compile→ partitioner → PyTorch fallback path with accuracy checks for grouped 3D deconvTest results (L40S, TRT-RTX nightly)
Type of change
Checklist: