vllm errors models size might be too big

apps/base/customer1/openclaw/vllm-gemma.yaml

@@ -32,15 +32,19 @@ spec:
             - name: VLLM_TOKENIZER_MODE
               value: "auto"
           args:
-            - "--model"
+            - --model=Jiunsong/supergemma4-26b-abliterated-multimodal
-            - "Jiunsong/supergemma4-26b-abliterated-multimodal" # Uncensored Abliterated Gemma 26B (Native)
+            - --host=0.0.0.0
-            - --dtype=auto
+            - --port=8000
-            - "bfloat16"
+            - --tensor-parallel-size=1
-            - --max-model-len=8192
+            - --gpu-memory-utilization=0.78          # ← Critical: lower than default 0.90
+            - --max-model-len=16384                  # Start conservative (you can raise to 32768 later)
+            - --max-num-batched-tokens=8192
+            - --max-num-seqs=4                       # Lower for single-user / low-concurrency
+            - --enforce-eager                        # Disables CUDA graphs → big memory saver at startup
+            - --disable-custom-all-reduce
             - --trust-remote-code
-            - --gpu-memory-utilization=0.80
+            - --dtype=auto
-            - --enforce-eager
+            - --limit-mm-per-prompt=image=1,video=0  # Optional: limit vision if not heavily using images yet 
-            
           ports:
             - containerPort: 8000
           resources: