修复get_gcn_arch_name的导入bug

vllm/model_executor/layers/sparse_attn_indexer.py

@@ -98,7 +98,7 @@ def sparse_attn_indexer(
         # Get the full shared workspace buffers once (will allocate on first use)
         workspace_manager = current_workspace_manager()
         k_fp8_full, k_scale_full = workspace_manager.get_simultaneous(
-                ((total_seq_lens, head_dim), fp8_dtype if not current_platform.is_rocm() or get_gcn_arch_name == "gfx938" else k.dtype,),
+                ((total_seq_lens, head_dim), fp8_dtype if not current_platform.is_rocm() or get_gcn_arch_name() == "gfx938" else k.dtype,),
                 ((total_seq_lens, 4), torch.uint8),
             )
         for chunk in prefill_metadata.chunks: