v2.12.0分支用例拉起#137
Conversation
kerer-ai
commented
May 20, 2026
- Update pytorch version defaults from 2.7.1 to 2.12.0
- Update branch refs from v2.7.1_dev_fast to v2.12.0_dev
- Keep docker images at pytorch_2.7.1 for now
- Add .github/actions, workflows, scripts and test_upstream directory
- Update pytorch version defaults from 2.7.1 to 2.12.0 - Update branch refs from v2.7.1_dev_fast to v2.12.0_dev - Keep docker images at pytorch_2.7.1 for now - Add .github/actions, workflows, scripts and test_upstream directory Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
- Replace pytorch_version input with docker_image parameterization - Update all workflow refs from v2.7.1_dev to v2.12.0_dev - Add ascend-pytorch-github artifact packaging in prepare job - Increase artifact retention-days to 60, test timeout to 1200s - Add build-docker-images.yml for builder/test image CI pipeline - Add .ci/docker with builder (manylinux_2_28) and test (ubuntu:22.04) Dockerfiles - Upgrade requirements-test.txt deps to match upstream v2.12.0 - Fix hardcoded pytorch-version in report.yml Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
Upstream renamed tags from cpu-2.7 to cpu-v2.12.0-rc9 convention. - x86_64: pytorch/manylinux2_28-builder:cpu-v2.12.0-rc9 - aarch64: pytorch/manylinux2_28_aarch64-builder:cpu-aarch64-v2.12.0-rc9 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
Replace hardcoded cpython-X.Y.Z paths with dynamic discovery via /opt/_internal/cpython-3.*/ glob. The v2.12.0 base image has different Python patch versions and may lack Python 3.9. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…iles Dockerfile RUN \ line continuations join lines without adding command separators. The for-loop body commands were concatenated without ; causing "unexpected end of file" syntax errors. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…in test Dockerfiles PyG wheel index (data.pyg.org) has no torch-2.12.0+cpu.html yet. These packages build from source and need torch at build time, which fails in PEP 517 isolated build environments. Install them with --no-build-isolation after torch is already installed. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
The v2.12.0-rc9 manylinux image includes Python 3.14t (nogil/free-threaded). numpy 1.26.4 cannot build on this ABI, and torch 2.12 has no wheels for it. Only install torch on Python 3.10-3.13. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
- numpy: use version markers (1.x for <3.13, 2.x for >=3.13) matching upstream pattern; numpy 1.26.4 has no cp313 wheels causing source builds - torchvision: fix 0.27.1 -> 0.27.0 (0.27.1 doesn't exist); install separately with --extra-index-url to get CPU wheel instead of CUDA - numpy in test: 1.26.2 -> 1.23.2 to match upstream's Python 3.10 pin Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…ements.txt These versions come from torch-npu's own test/requirements.txt. Install with --no-build-isolation after torch since PyG has no pre-built wheels for torch 2.12 on data.pyg.org. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
… git commit Upstream PyTorch installs torchvision from source using a pinned commit (.github/ci_commit_pins/vision.txt) rather than a pre-built wheel. Use the same commit 601776dc from v2.12.0. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
torchvision v0.27.0 was released 2026-05-13 with full pre-built wheels on download.pytorch.org/whl/cpu. No need to build from git source. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
torch_geometric 2.5.3 is a pure Python wheel (py3-none-any.whl), no source build needed. Only torch-scatter requires --no-build-isolation (C++ extension with no torch 2.12 pre-built wheel on data.pyg.org). Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The hand-edited hunk had line count mismatch and missing space prefix on blank context lines. Regenerated using GNU diff -u to ensure proper unified diff format. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…d update context - Remove BOM (byte order mark) from patch headers - Fix trailing whitespace and newline issues in patches - Update patch context lines for correctness - 895 patch files regenerated Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
Resolve conflict in test_freezing.py.patch — keep full SHA format Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…stream Upstream test/dynamo/test_functions.py refactored from ~5400 to 2724 lines. The DefaultsTests code this patch was modifying no longer exists. Patch application (hunk Ascend#2 at line 5466) fails in CI. Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…2.12.0 Upstream (ascend/v2.12.0) deleted ~70% of source files under test/ and torch/. Patches targeting non-existent files fail in CI with 'hunk FAILED' errors. Bulk-removed all patches whose target source no longer exists upstream. Reduced patch count from 847 to 149. Co-Authored-By: Claude <noreply@anthropic.com>
…stream v2.12.0" This reverts commit b7a01eb.
…moved upstream" This reverts commit fed8aa0.
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…against upstream v2.12.0 - test_functions.py.patch: fix hunk Ascend#2 indentation (was 7 spaces, now 8) - test_install_free_tensors.py.patch: add missing trailing newline Patches now apply cleanly against PyTorch v2.12.0 tag. Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
- Remove UTF-8 BOM (ef bb bf) from 5 patch files - Add missing trailing newlines to 25 patch files - Regenerate test_misc.py.patch (fix wrong paths + BOM) against upstream v2.12.0 BOM and missing newlines cause CI failures: - 'can't find file to patch at input line' when BOM corrupts path - 'patch unexpectedly ends in middle of line' for missing trailing newline Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
…e count Hunk header said +1,9 but actual content had only 8 lines. Regenerated against upstream pytorch v2.12.0. Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
All 905 test_upstream patches now verified to apply cleanly against upstream pytorch v2.12.0 tag. Fixed patches (context mismatch / malformed / truncated): - test/nn/test_packed_sequence.py.patch - test/test_dataloader.py.patch - test/test_functionalization.py.patch - test/test_nn.py.patch - test/test_python_dispatch.py.patch - torch/cuda/__init__.py.patch - torch/testing/_internal/common_cuda.py.patch - torch/testing/_internal/common_device_type.py.patch - torch/testing/_internal/opinfo/definitions/special.py.patch Co-Authored-By: Claude <noreply@anthropic.com>
CLA Signature Passkerer-ai, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Pass05mengmeizi, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
1 similar comment
CLA Signature Pass05mengmeizi, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Pass05mengmeizi, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Pass05mengmeizi, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |
CLA Signature Pass05mengmeizi, thanks for your pull request. All authors of the commits have signed the CLA. 👍 |