> 2026 06 28 更新 这是经过测试 可以正常使用的，注意此物理机 硬件为 K100 AI * 8，1.5T DDR5 内存。 /data/docker/API/01_llm/Qwen3.5/transformers 是最新的 git 仓库 ```yml services: qwen3-5-35b-a3b-tool: image: image.sourcefind.cn:5000/dcu/admin/base/custom:pytorch2.9.1-ubuntu22.04-dtk26.04-0130-py3.10-20260204-qwen3_5 container_name: dcu-qwen3-5-35b-a3b-tool privileged: true devices: - /dev/kfd - /dev/dri - /dev/mkfd group_add: - video ipc: host user: root # 对应 -u root cap_add: - SYS_PTRACE security_opt: - seccomp:unconfined environment: - HIP_VISIBLE_DEVICES=0,1,2,3 - VLLM_USE_V1=0 - OMP_NUM_THREADS=256 - NCCL_MIN_NCHANNELS=16 - NCCL_MAX_NCHANNELS=16 - VLLM_USE_FLASH_ATTN_PA=0 - VLLM_USE_FUSED_RMS_ROPE=1 - VLLM_CUSTOM_CACHE=0 volumes: - my-pip-cache:/root/.cache/pip - /data/models/Qwen3.5-35B-A3B:/data/models/Qwen3.5-35B-A3B - /opt/hyhal:/opt/hyhal:ro - /sys/kernel/debug:/sys/kernel/debug:ro - /data/docker/API/01_llm/Qwen3.5/transformers:/data/transformers shm_size: '512G' # command: bash # tty: true # stdin_open: true command: bash -c "cd /data/transformers && pip install . && \ pip install numpy==1.25.0 -i https://pypi.tuna.tsinghua.edu.cn/simple && \ export ALLREDUCE_STREAM_WITH_COMPUTE=1 && \ unset NCCL_ALGO && \ export NCCL_MIN_NCHANNELS=16 && \ export NCCL_MAX_NCHANNELS=16 && \ export VLLM_USE_V1=0 && \ export VLLM_CUSTOM_CACHE=0 && \ export NCCL_NET_GDR_READ=1 && \ export HIP_VISIBLE_DEVICES=0,1,2,3 && \ export VLLM_SPEC_DECODE_EAGER=1 && \ export VLLM_MLA_DISABLE=0 && \ export VLLM_USE_FLASH_MLA=1 && \ export VLLM_RPC_TIMEOUT=1800000 && \ export VLLM_NUMA_BIND=1 && \ export VLLM_RANK0_NUMA=0 && \ export VLLM_RANK1_NUMA=1 && \ export VLLM_RANK2_NUMA=2 && \ export VLLM_RANK3_NUMA=3 && \ export VLLM_RANK4_NUMA=4 && \ export VLLM_RANK5_NUMA=5 && \ export VLLM_RANK6_NUMA=6 && \ export VLLM_RANK7_NUMA=7 && \ vllm serve /data/models/Qwen3.5-35B-A3B \ --served-model-name Qwen3.5-35B-A3B \ --port 8000 \ --host 0.0.0.0 \ --reasoning-parser qwen3 \ --gpu-memory-utilization 0.85 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice \ --disable-custom-all-reduce \ --tensor-parallel-size 4 \ --enable-chunked-prefill \ --enable-prefix-caching \ --trust-remote-code --api-key sk-Wp1S****************************pPZD86" ports: - "18000:8000" ulimits: nproc: 65535 nofile: soft: 65535 hard: 65535 volumes: my-pip-cache: ```