pytorch / ao

PyTorch native quantization and sparsity for training and inference
BSD 3-Clause "New" or "Revised" License
1.62k stars 181 forks source link

[CI] CUDA nightly regression test is failing due to bnb + `triton.ops` #1338

Closed gau-nernst closed 3 days ago

gau-nernst commented 4 days ago

Failed CI: https://github.com/pytorch/ao/actions/runs/11997176417/job/33442516661?pr=1324 All other PRs will be affected too

bnb has a PR to remove triton.ops https://github.com/bitsandbytes-foundation/bitsandbytes/pull/1413. But even if it's merged, we are still stuck with the latest bnb version (without the removed triton.ops). Need to figure out a way to bypass this (e.g. skip bnb test in CUDA nightly, try to only import required bnb submodules so we don't indirectly import triton.ops). Open to other suggestions.