try befier model

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

@@ -43,40 +43,32 @@ spec:
 apiVersion: apps.nvidia.com/v1alpha1
 kind: NIMService
 metadata:
-  name: qwen3-32b-instruct
+  name: qwen36-27b-fp8
   namespace: nim-service
 spec:
-  env:
+  args:
-    - name: GPU_MEMORY_UTILIZATION
+    - "--gpu-memory-utilization=0.85"
-      value: "0.90"
+    - "--served-model-name=qwen36"
-    - name: NIM_MAX_NUM_SEQS
+    - "--max-model-len=256K"
-      value: "32"
+    - "--language-model-only"
-    - name: NIM_MAX_NUM_BATCHED_TOKENS
+    - "--reasoning-parser=qwen3"
-      value: "16384"
+    - "--enable-auto-tool-choice"
-    - name: ENABLE_AUTO_TOOL_CHOICE
+    - "--tool-call-parser=qwen3_coder"
-      value: "true"
+    - "--enable-chunked-prefill"
-    - name: ENABLE_PREFIX_CACHING
+    - "--max-num-batched-tokens=32768"
-      value: "true"
+    - "--max-num-seqs=10"
-    - name: TRUNCATION_SIDE
+    - "--enable-prefix-caching"
-      value: "left"
+    - '--speculative-config={"method":"mtp","num_speculative_tokens":2}'
-    - name: VLLM_LOGGING_LEVEL
-      value: "INFO"
-    - name: VLLM_ALLOW_LONG_MAX_MODEL_LEN
-      value: "true"
-    - name: NIM_MAX_MODEL_LEN
-      value: "32768"
-    - name: NIM_ENABLE_KV_CACHE_REUSE
-      value: "1"
   image:
-    repository: nvcr.io/nim/qwen/qwen3-32b-dgx-spark
+    repository: nvcr.io/nim/nvidia/llm-nim
-    tag: "1.1.0-variant"
+    tag: "1.12"
     pullPolicy: IfNotPresent
     pullSecrets:
       - ngc-secret
   authSecret: ngc-api-secret
   storage:
     nimCache:
-      name: qwen3-32b-instruct
+      name: qwen36-27b-fp8
   replicas: 1
   resources:
     limits: