← Back to events
ActiveAIRelease0.6

trunk/160ef925057d994a806e29a078ded765a03d64c5: [ROCm] Enable test_cslt_sparse_mm_alpha for fp16 and bf16 (#199050)

Photo: PyTorch Releases

What happened

Summary Fixes #168851 Fixes #168852 This PR enables test_cslt_sparse_mm_alpha for float16 and bfloat16. The test passes now has both cuSPARSELt and hipSparseLt applying bias epilogue. The dense reference was only alpha * (A @ B), so it was off by 1 on every element. Add the bias to the reference on every platform and drop the CUDA skip. That skip called out cuSPARSELt v0.6.x, which CI no longer ships. Test plan pyth…

Summary assembled by rule from the sources below

Why it's spreading

Sources