🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/152814

• 📄 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.

❌ 10 New Failures

As of commit 3a18ffa with merge base 5796212 ():

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

Is this still needed after #148421 ? I'm seeing that on a fresh source build test_float8_rowwise_scaling_sanity_use_fast_accum_True_cuda test_float8_rowwise_scaling_sanity_use_fast_accum_False_cuda test_scaled_mm_vs_emulated_row_wise all pass

My bad, it is needed for sm_120, not sm_100.

@pytorchbot rebase

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

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

Hmm, is this because CUTLASS is missing those specializations? Or because we are missing something minor on our side that would unblock support? Like a missing template specialization, overly restrictive dispatch logic, or missing cmake are to build those kernels from SM120?

Doesn't it support it here? Are we missing dispatch logic?

Does just adding SM120 here fix it or is SM100 not compatible with SM120?

cc @eqy @malfet to review

Hmm, is this because CUTLASS is missing those specializations? Or because we are missing something minor on our side that would unblock support? Like a missing template specialization, overly restrictive dispatch logic, or missing cmake are to build those kernels from SM120?

No, it is not as trivial as adding sm_120 to those places. I have tried that, CUTLASS just fails as "uninitialized" (whatever that means).