Repository navigation
Drop CUDA 12.6 support - #24740
Merged
Merged
Drop CUDA 12.6 support#24740
Conversation
PyTorch no longer publishes cu126 nightlies (torchao's cu126 channel stops at dev20260907), so the 12.6 CI rows can no longer install their pinned deps. - install_utils: remove (12, 6) from SUPPORTED_CUDA_VERSIONS. - cuda.yml: drop 12.6 from the build matrix (13.0 and 13.4 remain, and 13.0 is used by every model e2e job). - vulkan.yml: move both real-GPU jobs from the 12.6 image to 13.2. - Remove the cu126 GPU arch lists, the 12.6 Windows cross-compile mapping, and 12.6 cases in CI unit tests. - Docs/comments: drop 12.6 mentions.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24740
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 Unclassified FailuresAs of commit e8dcbc0 with merge base bffccd1 ( UNCLASSIFIED FAILURES - DrCI could not classify the following jobs because the workflow did not run on the merge base. The failures may be pre-existing on trunk or introduced by this PR:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
github-actions
Bot
added
the
ciflow/docker
This PR needs a
|
SS-JIA
approved these changes
Nov 6, 2026
digantdesai
deployed
to
upload-benchmark-results
November 9, 2026 15:50 — with 
GitHub Actions
Active
GitHub Actions
Active
mergennachin
approved these changes
Nov 9, 2026
This branch was successfully deployed
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
PyTorch no longer publishes cu126 nightlies. The torchao cu126 channel stops at
0.19.0.dev20260907, so the 12.6 CI rows can't install the current pins. This removes CUDA 12.6 support.Support
install_utils.py: remove(12, 6)fromSUPPORTED_CUDA_VERSIONS. Supported versions are now 13.0, 13.2, 13.4.CI
.github/workflows/vulkan.yml: movetest-vulkan-nvidiaand the Vulkan LLM e2e job from thecuda12.6image tocuda13.2. Thepytorch/almalinux-builder:cuda13.2image exists and is current..github/workflows/cuda.yml: drop12.6from thetest-cuda-buildsmatrix. Coverage that remains:cuda-windows.yml) and the published wheel rows (filter_cuda_matrix.py: cu132 / cu134).ci/scripts/wheel/cuda_arch_list.sh: remove the cu126 GPU arch lists. A cu126 row now fails as an unknown train..ci/docker/common/install_cuda_windows_cross_compile.sh: remove the 12.6 entry..ci/scripts/tests/test_cuda_workflow.py,.ci/scripts/tests/test_cu134_dependencies.py: drop the 12.6 cases.Docs / comments
docs/source/backends/cuda/cuda-overview.md: the source build now lists CUDA 13.0 instead of 12.6.install_utils.py,setup.py,filter_cuda_matrix.py,test_shared_libraries.py,cuda_benchmark.py.Intentionally unchanged
backends/cuda/cuda_backend.py_setup_cuda_environment_for_fatbin: the CUDA 12 ptxas branch handles any user torch build on CUDA 12.x below 12.9, not just a 12.6 support row.setup.py/test_shared_libraries.pyCUDA 12 runtime package mapping (nvidia-cuda-runtime-cu12): this is keyed by CUDA major, not 12.6.Test plan
pytest .ci/scripts/tests/test_cu134_dependencies.py .ci/scripts/tests/test_cuda_workflow.py .ci/scripts/tests/test_filter_cuda_matrix.py: 49 passedexecutorch_cuda_arch_list: cu130/cu132/cu134 still resolve, cu126 now returns 64 (unknown train)12.6/cu126/(12, 6): two hits left, both kept on purpose: thenvidia-cuda-runtime-cu12==12.6.77example in a parsing comment, and theCUDA 12.6+comment on the unchanged ptxas branch incuda_backend.py