v2.12.1
📦 pytorchView on GitHub →
🐛 3 fixes🔧 3 symbols
Summary
This release focuses on fixing regressions and silent correctness issues, primarily related to GPU operations using Triton and memory access bugs.
Migration Steps
- The fix for GPU issues required updating the internal Triton dependency to version 3.7.1.
🐛 Bug Fixes
- Fixed nondeterministic outputs in test_batch_invariance when using FLASH_ATTN on NVIDIA B200 GPUs.
- Fixed illegal memory access in the Triton convolution2d_bwd_weight kernel on B100/B200 (sm100) GPUs.
- Fixed fill_ operation on byte-dtype views that had a misaligned storage offset.