Replace external vllm-flash-attn3 kernel with PyTorch cuDNN FA3 backend 1543d6f moose Claude Opus 4.7 (1M context) commited on May 13