[ROCm] gaurd on_gfx1250 call with rocm platform (#52625)

Signed-off-by: Kunshang Ji <kunshang.ji@intel.com>

vllm/model_executor/layers/quantization/utils/fp8_utils.py

@@ -881,9 +881,13 @@ def w8a8_triton_block_scaled_mm(
         torch.Tensor: The result of matmul.
     """
 
-    from vllm.platforms.rocm import on_gfx1250
+    _on_gfx1250 = False
+    if current_platform.is_rocm():
+        from vllm.platforms.rocm import on_gfx1250
 
-    if on_gfx1250():
+        _on_gfx1250 = on_gfx1250()
+
+    if _on_gfx1250:
         # Torch upcast reference: dequantize A,B to fp32 and matmul in fp32.
         # Avoids the gfx1250 native-fp8 block GEMM NaN bug. Correct but slow.
         _bn, _bk = block_size[0], block_size[1]