Move torch pin from the 2.11 to the 2026-04-09 nightly, and drop deprecated CUDA versions from CI#19072
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/19072
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 14 New Failures, 12 Cancelled Jobs, 17 Unrelated FailuresAs of commit b1e8cce with merge base cb94506 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOBS - The following jobs were cancelled. Please retry:
FLAKY - The following jobs failed but were likely due to flakiness present on trunk:
BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
Contributor
There was a problem hiding this comment.
Pull request overview
Updates ExecuTorch’s PyTorch dependency pin back to a Torch nightly (2.12) and aligns CI/install scripts and vendored header-only Torch/C10 code with that newer upstream.
Changes:
- Bump pinned Torch version to
2.12.0.dev20260410and update various install paths to use the nightly wheel index. - Update vendored
torch/headeronlymacros +BFloat16ROCm/CUDA conditionals to match newer upstream behavior. - Adjust a few example exporters’ decomposition tables and CI scripts to reflect the new pin.
Reviewed changes
Copilot reviewed 12 out of 12 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
torch_pin.py |
Updates the pinned Torch + nightly date used by tooling/scripts. |
install_requirements.py |
Switches to nightly wheel index and installs torch/nightly using the central pin values. |
runtime/core/portable_type/c10/torch/headeronly/util/BFloat16.h |
Updates CUDA/ROCm guards for bf16 support under newer ROCm/HIP versions. |
runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h |
Updates ROCm warp-size handling and old-compiler move-return macro conditions. |
examples/models/voxtral_realtime/export_voxtral_rt.py |
Narrows decompositions used for Metal lowering. |
examples/models/parakeet/export_parakeet_tdt.py |
Narrows decompositions used for Metal lowering; adds commented debug prints (should be removed). |
examples/models/moshi/mimi/install_requirements.sh |
Updates torchcodec install to nightly and switches index URL to nightly. |
.ci/scripts/utils.sh |
Reads torchaudio/vision pins from the checked-out PyTorch tree instead of hardcoded release branches. |
.ci/scripts/test_wheel_package_qnn.sh |
Installs pinned nightly torch in the wheel test flow (but currently uses --index-url, which can break dependency resolution). |
.ci/scripts/test_model_e2e.sh |
Updates torchcodec install to nightly and switches index URL to nightly. |
.ci/docker/common/install_pytorch.sh |
Reads torchaudio/vision pins from the checked-out PyTorch tree instead of hardcoded release branches. |
.ci/docker/ci_commit_pins/pytorch.txt |
Moves PyTorch source pin from a release branch to a specific commit hash. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
mergennachin
force-pushed
the
update_torch_nightly
branch
from
43a3edd to
b3931e0
Compare
manuelcandales
approved these changes
Apr 23, 2026
mergennachin
force-pushed
the
update_torch_nightly
branch
from
b3931e0 to
dce61c6
Compare
Contributor
There was a problem hiding this comment.
Pull request overview
Updates ExecuTorch’s PyTorch dependency pins back to a Torch nightly (2.12) and aligns several CI/scripts and vendored header-only Torch/C10 files with upstream changes needed for newer ROCm/CUDA behaviors.
Changes:
- Bump the pinned Torch nightly version to 2.12.0 +
dev20260411, and update install logic to construct nightly wheel versions fromtorch_pin.py. - Sync vendored header-only Torch/C10 code (warp size macros + BFloat16 guards) to accommodate newer ROCm/HIP behaviors.
- Update CI and example scripts to pull from the PyTorch nightly index and adjust related pins (e.g., torchcodec, PyTorch commit hash pin).
Reviewed changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
torch_pin.py |
Updates the pinned Torch version + nightly date suffix. |
install_requirements.py |
Switches to nightly base URL and constructs torch/vision/audio nightly versions from the pin. |
runtime/core/portable_type/c10/torch/headeronly/util/BFloat16.h |
Adjusts CUDA/HIP compile guards for bf16 handling. |
runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h |
Updates ROCm warp size handling and compiler-version logic. |
examples/models/moshi/mimi/install_requirements.sh |
Updates torchcodec install to a nightly build + nightly index. |
.ci/scripts/test_model_e2e.sh |
Updates torchcodec install to a nightly build + nightly index. |
.ci/scripts/utils.sh |
Updates how torchaudio/torchvision refs are sourced and includes a whitespace-only change. |
.ci/scripts/test_wheel_package_qnn.sh |
Updates torch install to use the pinned nightly version (but changes pip indexing behavior). |
.ci/docker/common/install_pytorch.sh |
Updates torchaudio/torchvision refs sourcing when building PyTorch from source in docker. |
.ci/docker/ci_commit_pins/pytorch.txt |
Switches PyTorch pin from release branch to a specific commit hash. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
mergennachin
changed the title
Contributor
There was a problem hiding this comment.
Pull request overview
Updates ExecuTorch’s PyTorch pin and related CI/scripts to track a newer nightly, alongside CI configuration changes for supported CUDA versions and synced header adjustments for ROCm/CUDA compatibility.
Changes:
- Bump the PyTorch version pin (and update scripts to install the pinned nightly wheel).
- Drop CUDA 12.8/12.9 from Linux CUDA CI matrix and move Windows CUDA CI to 13.0.
- Sync/update portable c10 header-only macros and BFloat16 ROCm/CUDA guards.
Reviewed changes
Copilot reviewed 14 out of 14 changed files in this pull request and generated 6 comments.
Show a summary per file
| File | Description |
|---|---|
torch_pin.py |
Updates pinned TORCH_VERSION / NIGHTLY_VERSION. |
install_requirements.py |
Uses pinned nightly versions and switches nightly base URL to /whl/nightly. |
runtime/core/portable_type/c10/torch/headeronly/util/BFloat16.h |
Adjusts CUDA/ROCm preprocessor guards for bf16 handling. |
runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h |
Updates ROCm warp-size handling and compiler feature guard. |
examples/models/moshi/mimi/install_requirements.sh |
Updates torchcodec install to nightly index/version. |
.ci/scripts/test_model_e2e.sh |
Updates torchcodec install to nightly index/version. |
.github/workflows/cuda.yml |
Removes CUDA 12.8/12.9 from matrix and updates messaging. |
.github/workflows/cuda-windows.yml |
Pins Windows CUDA CI to 13.0 (CUDA_HOME / expected version). |
.ci/scripts/utils.sh |
Changes how torchaudio/torchvision refs are sourced for domain installs. |
.ci/scripts/test_wheel_package_qnn.sh |
Installs pinned nightly torch in wheel test flow. |
.ci/docker/common/install_pytorch.sh |
Changes how torchaudio/torchvision refs are sourced for domain installs in docker. |
.ci/docker/common/install_cuda_windows_cross_compile.sh |
Updates CUDA installer mapping/filename logic for 13.0+. |
.ci/docker/ci_commit_pins/pytorch.txt |
Updates pinned PyTorch commit hash used in docker builds. |
.ci/docker/build.sh |
Updates CUDA version for the windows cross-compile docker image. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| else "torchvision" | ||
| ), | ||
| ( | ||
| f"torchaudio==2.11.0.{NIGHTLY_VERSION}" |
Comment on lines
143
to
147
| # Grab the pinned audio and vision commits from PyTorch | ||
| TORCHAUDIO_VERSION=release/2.11 | ||
| TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt) | ||
| export TORCHAUDIO_VERSION | ||
| TORCHVISION_VERSION=release/0.26 | ||
| TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt) | ||
| export TORCHVISION_VERSION |
Comment on lines
34
to
38
| # Grab the pinned audio and vision commits from PyTorch | ||
| TORCHAUDIO_VERSION=release/2.11 | ||
| TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt) | ||
| export TORCHAUDIO_VERSION | ||
| TORCHVISION_VERSION=release/0.26 | ||
| TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt) | ||
| export TORCHVISION_VERSION |
mergennachin
changed the title
mergennachin
changed the title
mergennachin
changed the title
Contributor
There was a problem hiding this comment.
Pull request overview
Updates ExecuTorch’s PyTorch/nightly pin and aligns CI/install scripts with the newer nightly and reduced CUDA test matrix.
Changes:
- Bump the pinned Torch version and nightly date, and update multiple install scripts to use the PyTorch nightly index.
- Reduce CUDA CI coverage to 12.6 and 13.0 (Windows CUDA set to 12.6).
- Adjust synced portable c10 headers/macros for updated compiler/CUDA/ROCm conditionals.
Reviewed changes
Copilot reviewed 14 out of 14 changed files in this pull request and generated 4 comments.
Show a summary per file
| File | Description |
|---|---|
torch_pin.py |
Updates pinned Torch base version and nightly date string. |
runtime/core/portable_type/c10/torch/headeronly/util/BFloat16.h |
Expands CUDA/ROCm gating for bfloat16 CUDA header usage and conversions. |
runtime/core/portable_type/c10/torch/headeronly/macros/Macros.h |
Refines “old compiler” detection for return-move optimization macro. |
install_requirements.py |
Switches to nightly wheel base URL and constructs pinned nightly versions for installs. |
examples/models/moshi/mimi/install_requirements.sh |
Updates torchcodec install to nightly index and version. |
.github/workflows/cuda.yml |
Drops CUDA 12.8/12.9 from Linux CUDA build matrix and messaging. |
.github/workflows/cuda-windows.yml |
Switches Windows CUDA expectations and paths from 12.8 to 12.6. |
.ci/scripts/utils.sh |
Changes torchaudio/torchvision pin source to external pin files (currently problematic). |
.ci/scripts/test_wheel_package_qnn.sh |
Installs pinned nightly torch (and reads NIGHTLY_VERSION) for QNN wheel tests. |
.ci/scripts/test_model_e2e.sh |
Updates torchcodec install to nightly index and version. |
.ci/docker/common/install_pytorch.sh |
Adds PyTorch build env override and changes torchaudio/vision pin source (currently problematic). |
.ci/docker/common/install_cuda_windows_cross_compile.sh |
Removes 12.8/12.9 mapping, adds CUDA 13.0 filename handling. |
.ci/docker/ci_commit_pins/pytorch.txt |
Pins PyTorch to a specific commit hash. |
.ci/docker/build.sh |
Sets CUDA cross-compile image CUDA version to 12.6. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
Comment on lines
39
to
43
| # Grab the pinned audio and vision commits from PyTorch | ||
| TORCHAUDIO_VERSION=release/2.11 | ||
| TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt) | ||
| export TORCHAUDIO_VERSION | ||
| TORCHVISION_VERSION=release/0.26 | ||
| TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt) | ||
| export TORCHVISION_VERSION |
| else "torchvision" | ||
| ), | ||
| ( | ||
| f"torchaudio==2.11.0.{NIGHTLY_VERSION}" |
| ) | ||
| echo "=== [$LABEL] Install torch==${TORCH_VERSION}.${NIGHTLY_VERSION} ===" | ||
|
|
||
| # Install torchao based on the pinned PyTorch version |
Comment on lines
143
to
147
| # Grab the pinned audio and vision commits from PyTorch | ||
| TORCHAUDIO_VERSION=release/2.11 | ||
| TORCHAUDIO_VERSION=$(cat .github/ci_commit_pins/audio.txt) | ||
| export TORCHAUDIO_VERSION | ||
| TORCHVISION_VERSION=release/0.26 | ||
| TORCHVISION_VERSION=$(cat .github/ci_commit_pins/vision.txt) | ||
| export TORCHVISION_VERSION |
This reverts commit b40d6fe.
Bumps torch/torchvision/torchaudio/torchcodec to the 2026-04-10 nightly, updates the PyTorch commit pin and grafts the corresponding c10/headeronly headers, and restores the guarding_hint_or_throw fallback in exir/sym_util.py so shape_env.size_hint is only used when the newer API is unavailable. Co-authored-by: Claude <noreply@anthropic.com>
CUDA 12.8 and 12.9 are no longer supported. Remove them from the Linux test-executorch-cuda-build matrix (leaving 12.6 and 13.0), and retarget the Windows cross-compile and E2E jobs from 12.8 to 13.0. The NVIDIA Windows installer filename format changed in 13.0 (no driver suffix), so the installer URL builder now omits that segment when empty. Authored with Claude.
Avoids pytorch c792f73a61f, which changed the AOTI shim signature aoti_torch_mps__scaled_dot_product_attention_math_for_mps from 10 to 11 arguments and broke the Metal SDPA e2e test with a SIGABRT. Authored with Claude.
Reverts the Windows side of 8757b66; Linux CUDA matrix is unchanged. The 12.6 entry in CUDA_DRIVER_MAP is already valid, so no change to install_cuda_windows_cross_compile.sh is required. Authored with Claude.
PyTorch's FindARM.cmake was recently strengthened to hard-fail when its SVE+BF16 compile check doesn't pass. The aarch64 CI image ships gcc-11, which can't compile the probe's mixed NEON/SVE/bfloat16 intrinsics, so `install_pytorch.sh` bails during CMake configure. Executorch's aarch64 runtime targets don't use SVE anyway, so set BUILD_IGNORE_SVE_UNAVAILABLE to keep the Docker build moving. Authored with Claude.
The 04-08 nightly is broken for iOS builds: vec128_convert.h's VecConvert specializations call convert_float_bfloat16 / convert_float_half, but those functions are gated behind !defined(C10_MOBILE) in the _neon.h headers. iOS defines C10_MOBILE, so the symbols are undeclared and the compile fails in kernels/portable/cpu/op_where.cpp. The offending upstream change (#177009) was reverted on 04-09, so move the pin one day forward. Refreshed the grafted c10/torch/headeronly/macros/Macros.h from the 04-09 tip via .github/scripts/update_pytorch_pin.py. Authored with Claude.
The gcc11 and zephyr-sdk docker-build jobs on linux.4xlarge OOM mid PyTorch source build (~step 5000/7393 of mkl-dnn), manifesting as "self-hosted runner lost communication with the server". Bumping to linux.12xlarge (96 GB) for just these two images keeps the other x86_64 from-source images on 4xlarge for now. Authored with Claude.
mergennachin
force-pushed
the
update_torch_nightly
branch
from
a5f3501 to
b1e8cce
Compare
mergennachin
changed the title
mergennachin
added a commit
that referenced
this pull request
Apr 27, 2026
…rop deprecated CUDA versions from CI" (#19160) Reverts #19072 Too many failures on https://hud.pytorch.org/hud/pytorch/executorch/main/1?per_page=50 Lots of AOTI/CUDA/Metal failures
zeel2104
pushed a commit
to zeel2104/executorch
that referenced
this pull request
May 5, 2026
…ecated CUDA versions from CI (pytorch#19072) - Move torch pin from the 2.11 to the 2026-04-09 nightly - Drop 12.8 and 12.9 from CI; now builds 12.6 and 13.0 only. Windows only 12.6 for now. --------- Co-authored-by: Claude <noreply@anthropic.com>
zeel2104
pushed a commit
to zeel2104/executorch
that referenced
this pull request
May 5, 2026
…rop deprecated CUDA versions from CI" (pytorch#19160) Reverts pytorch#19072 Too many failures on https://hud.pytorch.org/hud/pytorch/executorch/main/1?per_page=50 Lots of AOTI/CUDA/Metal failures
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.
Uh oh!
There was an error while loading. Please reload this page.