From 2e1ab2ea2d172fe6736e55e459730bd4596f9411 Mon Sep 17 00:00:00 2001
From: HaimKortovich <haimkortovich88@gmail.com>
Date: Thu, 7 May 2026 14:54:30 -0500
Subject: [PATCH] try befier model

---
 clusters/k3s-dgx/nim-service/qwen.yaml | 42 +++++++++++---------------
 1 file changed, 17 insertions(+), 25 deletions(-)

diff --git a/clusters/k3s-dgx/nim-service/qwen.yaml b/clusters/k3s-dgx/nim-service/qwen.yaml
index 4835811..4e68ddf 100644
--- a/clusters/k3s-dgx/nim-service/qwen.yaml
+++ b/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:
-    - name: GPU_MEMORY_UTILIZATION
-      value: "0.90"
-    - name: NIM_MAX_NUM_SEQS
-      value: "32"
-    - name: NIM_MAX_NUM_BATCHED_TOKENS
-      value: "16384"
-    - name: ENABLE_AUTO_TOOL_CHOICE
-      value: "true"
-    - name: ENABLE_PREFIX_CACHING
-      value: "true"
-    - name: TRUNCATION_SIDE
-      value: "left"
-    - 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"
+  args:
+    - "--gpu-memory-utilization=0.85"
+    - "--served-model-name=qwen36"
+    - "--max-model-len=256K"
+    - "--language-model-only"
+    - "--reasoning-parser=qwen3"
+    - "--enable-auto-tool-choice"
+    - "--tool-call-parser=qwen3_coder"
+    - "--enable-chunked-prefill"
+    - "--max-num-batched-tokens=32768"
+    - "--max-num-seqs=10"
+    - "--enable-prefix-caching"
+    - '--speculative-config={"method":"mtp","num_speculative_tokens":2}'
   image:
-    repository: nvcr.io/nim/qwen/qwen3-32b-dgx-spark
-    tag: "1.1.0-variant"
+    repository: nvcr.io/nim/nvidia/llm-nim
+    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: