[misc] 1. Add platform detection before using torch.cuda;

       2. Remove unnecessary detection of Navi4x platform;

vllm/model_executor/models/llama.py

@@ -95,9 +95,7 @@ def __init__(
         self.act_fn = SiluAndMul()
 
     def forward(self, x):
-        # Navi4x is diff from other HIP devices by using e4m3fn fp8 format
-        if is_hip() and not is_navi4x() \
-            and x.shape[0] == 1 and x.shape[1] == 1:
+        if is_hip() and x.shape[0] == 1 and x.shape[1] == 1:
             out = torch.empty(x.shape[0],
                               self.gate_up_proj.weight.shape[0] // 2,
                               dtype=x.dtype,

vllm/utils.py

@@ -428,7 +428,7 @@ def is_hip() -> bool:
 
 @lru_cache(maxsize=None)
 def is_navi4x() -> bool:
-    if not torch.cuda.is_available():
+    if not is_hip() or not torch.cuda.is_available():
         return False
     # All (visible) GPUs must be of the same type,
     # otherwise FP8 results can't be guaranteed.