🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/147607

• 📄 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.

✅ No Failures

As of commit 6970267 with merge base aade4fb ():
💚 Looks good so far! There are no failures yet. 💚

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

Do not remove the support. 12.8 works just fine for these capabilities. Last night I compiled llama.cpp and ollama for sm_52.

Do not remove the support. 12.8 works just fine for these capabilities. Last night I compiled llama.cpp and ollama for sm_52.

If we can fix the linking of large binaries for libtorch with --relink, we can keep it.

sm_70 is deprecated in CUDA 12.8 and will be removed in future CUDA releases.
If you depend on older architectures, please use PyTorch binaries built for these (e.g. with CUDA 12.6.3).

If we can fix the linking of large binaries for libtorch with --relink, we can keep it.

It's not about a linking issue, but to drop deprecated architectures.

@pytorchbot rebase

@pytorchbot started a rebase job onto refs/remotes/origin/viable/strict. Check the current status here

Successfully rebased remove-sm70-cu128 onto refs/remotes/origin/viable/strict, please pull locally before adding more changes (for example, via git checkout remove-sm70-cu128 && git pull --rebase)

@pytorchmergebot merge

Merge started

Your change will be merged once all checks pass (ETA 0-4 Hours).

Learn more about merging in the wiki.

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

sm_70 is deprecated in CUDA 12.8 and will be removed in future CUDA releases.

This statement seems misleading. The official NVIDIA release notes for 12.8.1 say "Architecture support for Maxwell, Pascal, and Volta is considered feature-complete and will be frozen in an upcoming release." Frozen suggests to me that CUDA will continue to work on these architectures but no more fixes/features will come in the future.

The release notes for CUDA 12.9 clarify that sm_50-sm_70 support will remain for 12.x and it will be dropped for the next major release.

(none of this means the pytorch official build for CUDA >=12.8 has to continue to support sm_50-sm_70, I'm just clarifying that CUDA itself will continue to provide that support)