Release 2.7.1 validations checklist and cherry-picks

[atalman]

🐛 Describe the bug

Similar to: #150628

Cherry Picks to validate on final RC:

• modded-nanogpt flaky NCCL hang starting 3/30 nightly #152623 @kwen2501
• [c10d] Fix extra CUDA context created by barrier #152834 @kwen2501
• [c10d] Turn off default non-blocking API mode to work around hang in NCCL 2.26 #154085 @kwen2501
• [FlexAttention] Remove Old Constraint on lastdim strides #151959 @drisspg
• [FlexAttention] Remove old constraint that was causing assert failure #151521 @drisspg
• [FlexAttention] Remove Old Constraint on lastdim strides #151959 @drisspg
• [FlexAttention] explicilty create grad_q w/ strides #152641 @drisspg
• [dynamo][super variable] Fix bug to use correct source #151154 @anijain2305
• [cudagraphs] Fix issue in collecting static_input_idxs #152287 @anijain2305
• [dynamo][super variable] Fix bug to use correct source #152774 @anijain2305
• [cudagraphs][HF][torch 2.7] Excessive cudagraph re-recording for HF LLM models #152275 @anijain2305
• Add device guard for xpu conv on multi device #153067 @chuanqi129
• [CD] Fix the libgomp twice load issue #150084 @chuanqi129
• Add device guard for xpu conv on multi device #153345 @chuanqi129
• [CD] Fix the libgomp twice load issue (#150084) #153518 @chuanqi129
• XPU inference output abnormal with device 'XPU:1' #153022 @chuanqi129
• Pip-installed pytorch limits threads to 1 when setting GOMP_CPU_AFFINITY (likely due to bundled GOMP) #149422 @chuanqi129
• Illegal Instruction Caused by grid_sample Under Windows #152385 @xuhancn
• [ATen][CUDA] Optimize 128 bit vectorization #148320 @atalman - New Cherry-Pick done for this issue
• Remove cuda dependencies from non cuda buids #152333 @atalman
• [binary builds] Linux aarch64 CUDA builds. Make sure tag is set correctly #154136 @atalman
• [release only] Bump triton version to 3.3.1 #153554 @atalman
• [BUG] einops is unsupported and break dynamo graph with torch 2.7 #153476 @ZainRizvi
• [Dynamo] Exception raised inside torch.autocast causes crash AttributeError: 'NoneType' object has no attribute 'is_python_constant #152012 @ZainRizvi
• [MKLDNN] Check that strides are positive #151848 @malfet
• Fix tensorpipe compilation with clang-17 #151344 @malfet
• [vec128] Fix fmsub NEON defintion #152075 @malfet
• [Cherry-pick] Fix copysign + scalar correctness issue #153098 @malfet
• Inductor doesn't support tensor.view(dtype).copy_(...) #151156 @atalman
• Mark auto_functionalized HOPs as cacheable (#151194) #153304 @zou3519
• [ONNX] Update decomposition logic to loop over onnx registry #153168 @titaiwangms
• Only print dde partial fx graph for export #153218 @StrongerXi
• [dynamo] replace unimplemented with unimplemented_v2 in variables/functions.py #153533 @StrongerXi
• [ROCm] Update CUDAPluggableAllocator.h (#1984) #153974 @jithunnair-amd

Additional validation checks - @atalman :

• Validate Linux aarch64 CUDA builds with triton (Please note all CUDA Aarch64 builds where validated by Nvidia)

• Python 3.13 and 3.13t wheel validate - https://github.com/pytorch/test-infra/actions/runs/15303152347/job/43059048408

• Amazon Linux 2023 Test + torch.compile + no numpy installed: https://github.com/pytorch/test-infra/actions/runs/15303152347/job/43059048408

• Validate Metadata section of wheels - make sure python versions are set

• PyTorch 2.7.1 exposes statically linked libstdc++ CXX11 ABI symbols : PyTorch 2.5.0 exposes statically linked libstdc++ CXX11 ABI symbols. #133437

  • Tested on macOS by running and verified and no matches
  • (release2.7) ~/test/release2.7/.venv/lib/python3.12/site-packages/torch/lib nm -gU libtorch_cpu.dylib | grep "recursive_directory_iterator"

• CUDA

  • pypi binaries with slimmed dependencies are usable in standard AWS containers 2023 regression in 1.13 - https://github.com/pytorch/test-infra/actions/runs/15352480877/job/43203837080 - TODO REMOVE NEXT ITERATION. AUTOMATED
  • Check cuda 1.12.1 update issue: torch.linalg.eigh fails on GPU #94772 with small wheels . Passes on GPU but failing on CPU, new issue: torch.linalg.eigh fails on CPU #145801

• torch.compile

  • Basic test works (for example see test mentioned in Search for libdevice relative to shared library triton-lang/triton#1176 ) in PyTorch docker container
  • torch.compile raises an error if used on Windows. Test (part of torchvision): https://github.com/pytorch/test-infra/actions/runs/14182325015/job/39731076931#step:9:447
  • torch.compile works on 3.13 : Test: https://github.com/pytorch/test-infra/actions/runs/14315674885/job/40121143490#step:15:3483
  • torch.compile raises error on 3.13t: Validated : RuntimeError: torch.compile is not supported on Python built with GIL disabled

• MPS

  • Resnet is usable out of the box (https://github.com/pytorch/test-infra/actions/runs/14315674885/job/40121143490#step:15:3469)
  • Is torchvision usable? True German shepherd (cpu): 37.6% German shepherd (mps): 34.1%

• Validate docker release builds

Versions

2.7.1

[atalman]

Instructions to install runners:

DOWNLOAD INSTRUCTIONS FOR TESTING FINAL RC

PIP CPU

Windows/Linux/MacOS:
pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cpu
PIP CUDA 11.8, 12.6, 12.8
Windows/Linux:

pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu118
pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu126
pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cu128
PIP ROCM 6.2.4

pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/rocm6.2.4
PIP ROCM 6.3

pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/rocm6.3
PIP XPU

Linux/Windows:
pip3 install torch==2.7.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/xpu
PIP Linux Aarch6

pip3 install torch==2.7.1 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/test/cpu
PIP Linux CUDA 12.8 Aarch64

pip3 install torch==2.7.1 torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/test/cu128
Libtorch

CPU Linux:

(cxx11 ABI): https://download.pytorch.org/libtorch/test/cpu/libtorch-cxx11-abi-shared-with-deps-latest.zip
CPU Windows:

Download here (Release version): https://download.pytorch.org/libtorch/test/cpu/libtorch-win-shared-with-deps-latest.zip
Download here (Debug version): https://download.pytorch.org/libtorch/test/cpu/libtorch-win-shared-with-deps-debug-latest.zip
CPU MacOS:

https://download.pytorch.org/libtorch/nightly/cpu/libtorch-macos-arm64-latest.zip
CUDA 11.8, 12.6, 12.8 Linux

(cxx11 ABI):

https://download.pytorch.org/libtorch/test/cu118/libtorch-cxx11-abi-shared-with-deps-latest.zip
[https://download.pytorch.org/libtorch/test/cu126/libtorch-cxx11-abi-shared-with-deps-latest.zip](https://download.pytorch.org/libtorch/test/cu121/libtorch-cxx11-abi-shared-with-deps-latest.zip)
[https://download.pytorch.org/libtorch/test/cu128/libtorch-cxx11-abi-shared-with-deps-latest.zip](https://download.pytorch.org/libtorch/test/cu124/libtorch-cxx11-abi-shared-with-deps-latest.zip)
Windows CUDA 11.8, 12.6, 12.8

(Release version):

https://download.pytorch.org/libtorch/test/cu118/libtorch-win-shared-with-deps-latest.zip
[https://download.pytorch.org/libtorch/test/cu126/libtorch-win-shared-with-deps-latest.zip](https://download.pytorch.org/libtorch/test/cu121/libtorch-win-shared-with-deps-latest.zip)
[https://download.pytorch.org/libtorch/test/cu128/libtorch-win-shared-with-deps-latest.zip](https://download.pytorch.org/libtorch/test/cu124/libtorch-win-shared-with-deps-latest.zip)
(Debug version):

https://download.pytorch.org/libtorch/test/cu118/libtorch-win-shared-with-deps-debug-latest.zip
[https://download.pytorch.org/libtorch/test/cu126/libtorch-win-shared-with-deps-debug-latest.zip](https://download.pytorch.org/libtorch/test/cu121/libtorch-win-shared-with-deps-debug-latest.zip)
[https://download.pytorch.org/libtorch/test/cu128/libtorch-win-shared-with-deps-debug-latest.zip](https://download.pytorch.org/libtorch/test/cu124/libtorch-win-shared-with-deps-debug-latest.zip)
Docker images CUDA 11.8, 12.6, 12.8
Devel:

docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda11.8-cudnn9-devel
docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda12.6-cudnn9-devel
docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda12.8-cudnn9-devel
Runtime:

docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda11.8-cudnn9-runtime
docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda12.6-cudnn9-runtime
docker pull ghcr.io/pytorch/pytorch-test:2.7.1-cuda12.8-cudnn9-runtime