[Bugfix] Switch bailout logic for kv-cache-dtype with SM100 Flashinfer (#20934)

Signed-off-by: Pavani Majety <pmajety@nvidia.com>

Author: pavanimajety

vllm/engine/arg_utils.py

@@ -1418,14 +1418,15 @@ def _is_v1_supported_oracle(self, model_config: ModelConfig) -> bool:
                 and not envs.is_set("VLLM_ATTENTION_BACKEND")
             ) or envs.VLLM_ATTENTION_BACKEND == "FLASH_ATTN_VLLM_V1"
             supported = False
-            if current_platform.is_rocm():
+            if current_platform.is_rocm() or (
+                    current_platform.is_cuda()
+                    and current_platform.is_device_capability(100)):
                 supported = True
             elif fp8_attention and will_use_fa:
                 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)