set fashinfer

clusters/k3s-dgx/nim-service/qwen.yaml

@@ -33,8 +33,6 @@ spec:
   env:
     - name: VLLM_CACHE_ROOT
       value: /model-store/vllm-cache
-    - name: VLLM_ATTENTION_BACKEND
-      value: FLASHINFER      # Faster attention on CUDA >= 12.1
     - name: VLLM_DISABLE_LOGGING
       value: "1"              # Remove logging overhead
   command:
@@ -64,6 +62,8 @@ spec:
     - --max-num-seqs
     - "254"                   # ↑ Allow more concurrent sequences
     - --enable-prefix-caching
+    - --attention-backend
+    - FLASHINFER
   authSecret: hf-api-secret
   storage:
     sharedMemorySizeLimit: 64Gi