[Windows][ROCm] Fix missing native header includes causing DLL export…#179138
Closed
stsokolo wants to merge 1 commit intopytorch:mainfrom
Closed
[Windows][ROCm] Fix missing native header includes causing DLL export…#179138stsokolo wants to merge 1 commit intopytorch:mainfrom
stsokolo wants to merge 1 commit intopytorch:mainfrom
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/179138
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit f6b4087 with merge base 02521a0 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
jithunnair-amd
approved these changes
Apr 2, 2026
Contributor
|
Good catch! |
slayton58
approved these changes
Apr 2, 2026
jeffdaily
approved these changes
Apr 2, 2026
Collaborator
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Contributor
|
LGTM, though PR title is a bit misleading |
weifengpy
pushed a commit
that referenced
this pull request
Apr 7, 2026
#179138) Three operations — `torch._int_mm`,` torch._grouped_mm`, and `torch._scaled_mm_v2` — crash immediately on Windows ROCm builds with an access violation. The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it. Three source files were missing their corresponding headers: - Blas.cpp was missing _int_mm_native.h - GroupedBlas.cpp was missing _grouped_mm_native.h - ScaledBlas.cpp was missing _scaled_mm_v2_native.h Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash. Fixes: - ROCm/TheRock#4086 - ROCm/rocm-libraries#5205 - ROCm/TheRock#4079 The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now. Pull Request resolved: #179138 Approved by: https://github.com/jithunnair-amd, https://github.com/slayton58, https://github.com/jeffdaily
nklshy-aws
pushed a commit
to nklshy-aws/pytorch
that referenced
this pull request
Apr 7, 2026
pytorch#179138) Three operations — `torch._int_mm`,` torch._grouped_mm`, and `torch._scaled_mm_v2` — crash immediately on Windows ROCm builds with an access violation. The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it. Three source files were missing their corresponding headers: - Blas.cpp was missing _int_mm_native.h - GroupedBlas.cpp was missing _grouped_mm_native.h - ScaledBlas.cpp was missing _scaled_mm_v2_native.h Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash. Fixes: - ROCm/TheRock#4086 - ROCm/rocm-libraries#5205 - ROCm/TheRock#4079 The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now. Pull Request resolved: pytorch#179138 Approved by: https://github.com/jithunnair-amd, https://github.com/slayton58, https://github.com/jeffdaily
bobrenjc93
pushed a commit
to bobrenjc93/pytorch
that referenced
this pull request
Apr 9, 2026
pytorch#179138) Three operations — `torch._int_mm`,` torch._grouped_mm`, and `torch._scaled_mm_v2` — crash immediately on Windows ROCm builds with an access violation. The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it. Three source files were missing their corresponding headers: - Blas.cpp was missing _int_mm_native.h - GroupedBlas.cpp was missing _grouped_mm_native.h - ScaledBlas.cpp was missing _scaled_mm_v2_native.h Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash. Fixes: - ROCm/TheRock#4086 - ROCm/rocm-libraries#5205 - ROCm/TheRock#4079 The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now. Pull Request resolved: pytorch#179138 Approved by: https://github.com/jithunnair-amd, https://github.com/slayton58, https://github.com/jeffdaily
pytorch-bot Bot
pushed a commit
that referenced
this pull request
Apr 10, 2026
#179138) Three operations — `torch._int_mm`,` torch._grouped_mm`, and `torch._scaled_mm_v2` — crash immediately on Windows ROCm builds with an access violation. The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it. Three source files were missing their corresponding headers: - Blas.cpp was missing _int_mm_native.h - GroupedBlas.cpp was missing _grouped_mm_native.h - ScaledBlas.cpp was missing _scaled_mm_v2_native.h Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash. Fixes: - ROCm/TheRock#4086 - ROCm/rocm-libraries#5205 - ROCm/TheRock#4079 The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now. Pull Request resolved: #179138 Approved by: https://github.com/jithunnair-amd, https://github.com/slayton58, https://github.com/jeffdaily
2 tasks
tvukovic-amd
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Apr 16, 2026
pytorch#179138) Three operations — `torch._int_mm`,` torch._grouped_mm`, and `torch._scaled_mm_v2` — crash immediately on Windows ROCm builds with an access violation. The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it. Three source files were missing their corresponding headers: - Blas.cpp was missing _int_mm_native.h - GroupedBlas.cpp was missing _grouped_mm_native.h - ScaledBlas.cpp was missing _scaled_mm_v2_native.h Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash. Fixes: - ROCm/TheRock#4086 - ROCm/rocm-libraries#5205 - ROCm/TheRock#4079 The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now. Pull Request resolved: pytorch#179138 Approved by: https://github.com/jithunnair-amd, https://github.com/slayton58, https://github.com/jeffdaily
jeffdaily
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Apr 27, 2026
…3160) Cherry pick of pytorch#179138 Fixes: ROCm/TheRock#4086 ROCm/rocm-libraries#5205 ROCm/TheRock#4079 Co-authored-by: Stefan Sokolovic <stefan.sokolovic2@amd.com>
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
7 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.
Three operations —
torch._int_mm,torch._grouped_mm, andtorch._scaled_mm_v2— crash immediately on Windows ROCm builds with an access violation.The problem comes down to how Windows DLLs handle function visibility. Windows requires functions to be explicitly marked for export. In PyTorch, this happens through a somewhat indirect mechanism: when a source file includes a function's _native.h header, the compiler sees the TORCH_API declaration, recognizes it's defining that function locally, and automatically exports it.
Three source files were missing their corresponding headers:
Without these includes, the functions were compiled as internal symbols and never exported. Meanwhile, the auto-generated dispatch code tried to call them through import thunks that couldn't resolve — resulting in a jump to address zero and an immediate crash.
Fixes:
torch._grouped_mmaccess violation (0xC0000005) intorch_hip.dllROCm/TheRock#4086HIPBLAS_STATUS_NOT_SUPPORTEDROCm/rocm-libraries#52050xC0000005) intorch.nn.functional.scaled_mm(_scaled_mm_v2) ongfx1200ROCm/TheRock#4079The fix was tested on all reproduction examples mentioned in the issues above, and all of them are passing now.
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @jataylo @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang