Switch between trtllm-gen vs fa2/fa3 backends inside Attention wrappers #1493
Description
Switch between trtllm-gen vs fa2/fa3 backends inside Attention wrappers
Currently, vLLM uses the trtllm-gen attn kernels by directly calling the corresponding functions without going through the wrappers. Ideally, we should make the wrappers support all the features needed by trtllm-gen attn kernels so that vLLM only needs to call wrappers.
See: https://github.com/vllm-project/vllm/pull/21716/files#r2270261171