The gfx928 architecture force to set VLLM_W8A8_BACKEND == 1

vllm/envs.py

@@ -8,6 +8,7 @@ import os
 import sys
 import tempfile
 import uuid
+import torch
 from collections.abc import Callable
 from typing import TYPE_CHECKING, Any, Literal
 
@@ -1896,7 +1897,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
     # cutlass: 2 (will remove in the future)
     # blaslt: 3 (default)
     # rocblas: others
-    "VLLM_W8A8_BACKEND": lambda: int(os.getenv("VLLM_W8A8_BACKEND", "3")),
+    "VLLM_W8A8_BACKEND": lambda: int(
+            1 if "gfx928" in torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] else os.getenv("VLLM_W8A8_BACKEND", "3")
+    ),
     # Capture MoE router logits for debugging/analysis.
     "VLLM_MOE_ROUTER_CAPTURE":
     lambda: (os.getenv("VLLM_MOE_ROUTER_CAPTURE", "0").lower() in ("true", "1")),