diff --git a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml index 6e105d6..32947b5 100644 --- a/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml +++ b/infrastructure/gpus/base/vllm-servers/rtx6000-vllm.yaml @@ -38,11 +38,13 @@ spec: - --tensor-parallel-size=1 - --tokenizer-mode=hf - --gpu-memory-utilization=0.95 + - --enforce-eager - --max-model-len=65536 - --enable-auto-tool-choice - --kv-cache-dtype=fp8 - --max-num-batched-tokens=65536 - - --max-num-seqs=32 + - --max-num-seqs=128 + - --block-size=32 - --enable-chunked-prefill - --trust-remote-code - --dtype=auto