a100 deployment tweaks

infrastructure/gpus/base/vllm-servers/a100-vllm.yaml

@@ -1,19 +1,19 @@
 apiVersion: apps/v1
 kind: Deployment
 metadata:
-  name: openclaw-brain-vllm
+  name: a100-brain-vllm
   namespace: customer1
   labels:
-    app: openclaw-brain
+    app: a100-brain
 spec:
   replicas: 1
   selector:
     matchLabels:
-      app: openclaw-brain
+      app: a100-brain 
   template:
     metadata:
       labels:
-        app: openclaw-brain
+        app: a100-brain 
     spec:
       nodeSelector:
         cloud.google.com/gke-accelerator: "nvidia-a100-80gb"
@@ -32,35 +32,37 @@ spec:
             - name: TOKENIZER_MODE
               value: "hf"
           args:
-            - --model=Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16
+            - --model=sakamakismile/Huihui-Qwen3.6-27B-abliterated-NVFP4-MTP
             - --host=0.0.0.0
             - --port=8000
             - --tensor-parallel-size=1
             - --tokenizer-mode=hf 
             - --gpu-memory-utilization=0.95
-            - --kv-cache-dtype=fp8
+            - --kv-cache-dtype=fp8  
-            - --max-model-len=131072 
+            - --max-model-len=262144
             - --enable-auto-tool-choice 
             - --enable-chunked-prefill
-            - --max-num-batched-tokens=8192
+            - --max-num-batched-tokens=16384 
-            - --max-num-seqs=4
+            - --max-num-seqs=2
             - --trust-remote-code
             - --dtype=auto 
             - --enable-prefix-caching
             - --tool-call-parser=qwen3_xml
             - --reasoning-parser=qwen3
             - --disable-custom-all-reduce
+            - --attention-backend=flash_attn
+            - --quantization=modelopt 
           ports:
             - containerPort: 8000
           resources:
             limits:
               nvidia.com/gpu: 1
-              memory: "100Gi"
+              
-              cpu: "16"
+              
             requests:
               nvidia.com/gpu: 1
-              memory: "80Gi"
+              memory: "100Gi" 
-              cpu: "8"
+              cpu: "24" 
           
           startupProbe:
             httpGet:
@@ -99,12 +101,12 @@ spec:
       volumes:
         - name: model-cache
           persistentVolumeClaim:
-            claimName: vllm-model-qwen3.6-27b-uncensored
+            claimName: vllm-model-qwen3.6-27b-uncensored-4bit
 ---
 apiVersion: v1
 kind: PersistentVolumeClaim
 metadata:
-  name: vllm-model-qwen3.6-27b-uncensored
+  name: vllm-model-qwen3.6-27b-uncensored-4bit
   namespace: customer1
 spec:
   accessModes:
@@ -116,11 +118,11 @@ spec:
 apiVersion: v1
 kind: Service
 metadata:
-  name: openclaw-brain-service
+  name: a100-brain-service 
   namespace: customer1
 spec:
   selector:
-    app: openclaw-brain
+    app: a100-brain 
   ports:
     - protocol: TCP
       port: 8000

infrastructure/gpus/base/vllm-servers/kustomization.yaml

@@ -1,6 +1,6 @@
 apiVersion: kustomize.config.k8s.io/v1beta1
 kind: Kustomization
 resources:
-  #- a100-vllm.yaml
+  - a100-vllm.yaml
   - rtx6000-vllm.yaml
   #- vllm-l4.yaml