update moe configs name

vllm/utils/__init__.py

@@ -84,9 +84,9 @@ DEFAULT_MAX_NUM_BATCHED_TOKENS = 2048
 POOLING_MODEL_MAX_NUM_BATCHED_TOKENS = 32768
 MULTIMODAL_MODEL_MAX_NUM_BATCHED_TOKENS = 5120
 
-gpuname = torch.cuda.get_device_properties(torch.cuda.current_device()).name
-is_kme = gpuname.startswith('K100_AI') or gpuname.startswith('K500SM_AI')
-SUPPORT_TC = gpuname.startswith('K100_AI') or gpuname.startswith('K500SM_AI') or gpuname.startswith('BW')
+GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
+is_kme = any(arch in GPU_ARCH for arch in ["gfx928"])
+SUPPORT_TC = any(arch in GPU_ARCH for arch in ["gfx928", "gfx936"])
 
 def _generate_random_int8(
     tensor: torch.Tensor,