diff --git a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml index d95af39..abaab2b 100644 --- a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml +++ b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml @@ -51,7 +51,8 @@ spec: - --enable-prefix-caching - --tool-call-parser=qwen3_xml - --reasoning-parser=qwen3 - - --speculative-config "{"method": "mtp", "num_speculative_tokens": 2}" + - --speculative-config + - --'{"method": "mtp", "num_speculative_tokens": 2}' ports: - containerPort: 8000