[v1] Re-add fp32 support to v1 engine through FlexAttention (#19754)

Signed-off-by: Isotr0py <2037008807@qq.com>
Signed-off-by: Isotr0py <mozf@mail2.sysu.edu.cn>

vllm/engine/arg_utils.py

@@ -1393,13 +1393,6 @@ class EngineArgs:
                                recommend_to_remove=False)
             return False
 
-        # Only Fp16 and Bf16 dtypes since we only support FA.
-        V1_SUPPORTED_DTYPES = [torch.bfloat16, torch.float16]
-        if model_config.dtype not in V1_SUPPORTED_DTYPES:
-            _raise_or_fallback(feature_name=f"--dtype {model_config.dtype}",
-                               recommend_to_remove=False)
-            return False
-
         # No Mamba or Encoder-Decoder so far.
         if not model_config.is_v1_compatible:
             _raise_or_fallback(feature_name=model_config.architectures,