Commit History
3 days ago
-
Use BF16 state at every CUDA batch and stabilize physical-ring attention (#1)
- Use BF16 state at every CUDA batch and stabilize physical-ring attention (623a14eb5f8b3b9ac5ab9227975c6cb2ff0043ca) Co-authored-by: Aaron Xu <hkust-aaron@users.noreply.huggingface.co>49ab3cd
5 days ago
-
Use BF16 inference and align figures and results with the technical report
Verified signature799fa1f