[AMD] [Quantization] Add override flag for attention dtype instead of using kv_cache_dtype trigger (#17331)

Signed-off-by: Randall Smith <Randall.Smith@amd.com>

vllm/config.py

@@ -417,6 +417,8 @@ class ModelConfig:
     available.\n
     - "vllm" will use the vLLM model implementation.\n
     - "transformers" will use the Transformers model implementation."""
+    override_attention_dtype: Optional[str] = None
+    """Override dtype for attention"""
 
     def compute_hash(self) -> str:
         """
@@ -517,6 +519,12 @@ class ModelConfig:
 
         from vllm.platforms import current_platform
 
+        if (self.override_attention_dtype is not None
+                and not current_platform.is_rocm()):
+            warnings.warn(
+                "override-attention-dtype is set but not using ROCm platform",
+                stacklevel=2)
+
         if (self.enable_sleep_mode
                 and not current_platform.is_sleep_mode_available()):
             raise ValueError(