[Core] Add Flashinfer TRTLLM Backend for Flashinfer decode path (SM100). (#19825)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>
Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: shuw <shuw@nvidia.com>
Co-authored-by: mgoin <mgoin64@gmail.com>

vllm/engine/arg_utils.py

@@ -1424,6 +1424,8 @@ class EngineArgs:
                 from vllm.attention.utils.fa_utils import (
                     flash_attn_supports_fp8)
                 supported = flash_attn_supports_fp8()
+            elif envs.VLLM_USE_TRTLLM_DECODE_ATTENTION:
+                supported = True
             if not supported:
                 _raise_or_fallback(feature_name="--kv-cache-dtype",
                                    recommend_to_remove=False)