optimize get_kv_cache_torch_dtype (vllm-project#18531)

Signed-off-by: idellzheng <idellzheng@tencent.com>

Author: chunxiaozheng

vllm/utils.py

@@ -759,16 +759,15 @@ def get_kv_cache_torch_dtype(
         model_dtype: Optional[Union[str, torch.dtype]] = None) -> torch.dtype:
     if isinstance(cache_dtype, str):
         if cache_dtype == "auto":
-            if isinstance(model_dtype, str):
+            if isinstance(model_dtype,
+                          str) and model_dtype in STR_DTYPE_TO_TORCH_DTYPE:
                 torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[model_dtype]
             elif isinstance(model_dtype, torch.dtype):
                 torch_dtype = model_dtype
             else:
                 raise ValueError(f"Invalid model dtype: {model_dtype}")
-        elif cache_dtype in ["half", "bfloat16", "float"]:
+        elif cache_dtype in STR_DTYPE_TO_TORCH_DTYPE:
             torch_dtype = STR_DTYPE_TO_TORCH_DTYPE[cache_dtype]
-        elif cache_dtype == "fp8":
-            torch_dtype = torch.uint8
         else:
             raise ValueError(f"Invalid kv cache dtype: {cache_dtype}")
     elif isinstance(cache_dtype, torch.dtype):