🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161257

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit the bot commands wiki or our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 3 New Failures, 2 Unrelated Failures

As of commit 2325f1b with merge base b2db293 ():

This comment was automatically generated by Dr. CI and updates every 15 minutes.

Is it possible to add that archs for next releases?
How much size is the wheel? #161378

Is it possible to add that archs for next releases? How much size is the wheel? #161378

This PR is part of the CUDA 13 upstream bringup #159779, adding sm80-120 support for CUDA 13 SBSA build.

You can check the size of the wheel in https://github.com/pytorch/pytorch/actions/runs/17186608287, 2.18 GB, with --compress-mode=size enabled. Comparing with 12.9 3.28 GB, that is 1.1 GB of savings and ~33.5% smaller.

#161378 seems dup as this PR will add support for Thor (sm_110), Spark (sm_120 compatible) and GB300 (sm_100 compatible). Thanks.

Is it possible to add that archs for next releases? How much size is the wheel? #161378

This PR is part of the CUDA 13 upstream bringup #159779, adding sm80-120 support for CUDA 13 SBSA build.

You can check the size of the wheel in https://github.com/pytorch/pytorch/actions/runs/17186608287, 2.18 GB, with --compress-mode=size enabled. Comparing with 12.9 3.28 GB, that is 1.1 GB of savings and ~33.5% smaller.

#161378 seems dup as this PR will add support for Thor (sm_110), Spark and GB300 (sm_120 compatible). Thanks.

Thanks! Very kind.
I’m very concerned because now, we have more devices available with ARM than x86 for generate wheels, so that increase wheels size. Majority of frameworks don’t want distribute for edge/small devices like jetson or spark, but it will be very beneficial if they can do it, due that compile in small devices is hard

After adding sm_110, build failure in nvshmem, which doesn't support building sm_110. https://github.com/pytorch/pytorch/actions/runs/17200303693/job/48789533255?pr=161257
nvlink error : Undefined reference to '_Z23nvshmemi_transfer_quietIL13threadgroup_t3EEvb' in 'caffe2/CMakeFiles/torch_nvshmem.dir/__/torch/csrc/distributed/c10d/symm_mem/nvshmem_extension.cu.o' (target: sm_110)

After adding sm_110, build failure in nvshmem, which doesn't support building sm_110. https://github.com/pytorch/pytorch/actions/runs/17200303693/job/48789533255?pr=161257 nvlink error : Undefined reference to '_Z23nvshmemi_transfer_quietIL13threadgroup_t3EEvb' in 'caffe2/CMakeFiles/torch_nvshmem.dir/__/torch/csrc/distributed/c10d/symm_mem/nvshmem_extension.cu.o' (target: sm_110)

What nvshem are you using?
It is only supported in the latest 3.3.24. I’ve got it in jetson-containers

https://pypi.jetson-ai-lab.io/sbsa/cu130/torch/2.9.0.dev20250823+cu130.g3e5b021

also, cusparselt for thor minimum version is 0.8.0

As mentioned by the NVSHMEM team, NVSHMEM by definition is intended for Datacenter GPUs (clusters), therefore not expected to support Thor. https://docs.nvidia.com/nvshmem/release-notes-install-guide/release-notes/release-3324.html
Skipping sm_110 for NVSHMEM on sbsa.

Previous build failures shows there are more places that call into undefined references than the header file
https://github.com/pytorch/pytorch/actions/runs/17232507556/job/48889683226

nvlink error   : Undefined reference to 'nvshmem_int64_p' in 'caffe2/CMakeFiles/torch_nvshmem.dir/__/torch/csrc/distributed/c10d/symm_mem/nvshmem_extension.cu.o' (target: sm_110)
nvlink error   : Undefined reference to 'nvshmemx_barrier_all_block' in 'caffe2/CMakeFiles/torch_nvshmem.dir/__/torch/csrc/distributed/c10d/symm_mem/nvshmem_extension.cu.o' (target: sm_110)
nvlink error   : Undefined reference to 'nvshmemx_getmem_block' in 'caffe2/CMakeFiles/torch_nvshmem.dir/__/torch/csrc/distributed/c10d/symm_mem/nvshmem_extension.cu.o' (target: sm_110)

Introduce _NVSHMEM_DEVICELIB_SUPPORTED as a helper to decide whether to set NVSHMEM_HOSTLIB_ONLY before including nvshmem.h, and guard whether to run the unsupported functions. 83d2623

@pytorchbot merge -i

Merge failed

Reason: Approvers from one of the following sets are needed:

• superuser (pytorch/metamates)
• Core Reviewers (mruberry, lezcano, Skylion007, ngimel, peterbell10, ...)
• Core Maintainers (soumith, gchanan, ezyang, dzhulgakov, malfet, ...)

@pytorchmergebot merge -f "All required signal look good"

Merge started

Your change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use -f as last resort and instead consider -i/--ignore-current to continue the merge ignoring current failures. This will allow currently pending tests to finish and report signal before the merge.

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team