vLLM部署Qwen3.6-27B-FP8实战:FlashInfer后端、Chunked Prefill、Prefix Caching与MTP投机解码配置

vLLM 推理框架—Qwen3.6-27B-FP8部署

启动命令

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
CUDA_VISIBLE_DEVICES=0,1,2,3 \
python3 -m vllm.entrypoints.openai.api_server \
--model /data/models/Qwen3.6-27B-FP8/ \
--served-model-name Qwen/Qwen3.6-27B \
--dtype auto \
--disable-custom-all-reduce \
--host 0.0.0.0 \
--port 8080 \
--trust-remote-code \
--max-model-len 204800 \
--tensor-parallel-size 4 \
--gpu-memory-utilization 0.85 \
--max-num-batched-tokens 30720 \
--enable-chunked-prefill \
--enable-prefix-caching \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coder \
--reasoning-parser qwen3 \
--max-num-seqs 40 \
--attention-backend FLASHINFER \
--speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}' \
--api-key <YOUR_API_KEY>
本文结束 感谢您的阅读