自部署Qwen3.8-27B，一句话生成我的世界

佬自部署的硬件是什么，想参考一下

name: qwen38-27b



services:

  sglang:

    # Verified upstream index: sha256:616a3e97...; linux/amd64 manifest: sha256:b91d664a...

    # The exact amd64 manifest is imported locally because Docker's proxy truncated the pull.

    image: lmsysorg/sglang:dev-qwen38-27b-dflash2

    container_name: sglang-qwen38-27b

    restart: unless-stopped



    gpus: all

    ipc: host

    shm_size: "32g"



    volumes:

      - "D:/sglang/cache/huggingface:/root/.cache/huggingface"



    # Local-only by default because the API has no authentication configured.

    ports:

      - "0.0.0.0:30000:30000"



    environment:

      HF_HOME: /root/.cache/huggingface

      HF_TOKEN: "${HF_TOKEN:-}"



    entrypoint: ["sglang", "serve"]

    command:

      - "--trust-remote-code"

      - "--model-path"

      # Default: official BF16 checkpoint. Set MODEL_PATH=Qwen/Qwen3.8-27B-FP8 for FP8.

      - "${MODEL_PATH:-Qwen/Qwen3.8-27B}"

      - "--kv-cache-dtype"

      - "fp8_e4m3"

      - "--mem-fraction-static"

      - "0.85"

      # Long-context profile: dedicate the token pool to one native 262K request.

      - "--context-length"

      - "262144"

      - "--max-total-tokens"

      - "262144"

      - "--max-running-requests"

      - "1"

      - "--attention-backend"

      - "flashinfer"

      - "--chunked-prefill-size"

      - "2048"

      - "--mamba-radix-cache-strategy"

      - "extra_buffer_lazy"

      # Keep headroom above the four active float32 slots required by one

      # extra_buffer_lazy request; four slots can be exhausted while caching

      # a long conversation prefix and crashes the scheduler.

      - "--max-mamba-cache-size"

      - "8"

      - "--mamba-ssm-dtype"

      - "float32"

      - "--reasoning-parser"

      - "qwen3"

      - "--tool-call-parser"

      - "qwen3_coder"

      - "--enable-cache-report"

      # DFlash2 accelerates decode while the target model remains BF16.

      - "--speculative-algorithm"

      - "DFLASH"

      - "--speculative-draft-model-path"

      - "incoai/Qwen3.8-27B-DFlash2"

      - "--speculative-num-draft-tokens"

      - "8"

      - "--host"

      - "0.0.0.0"

      - "--port"

      - "30000"



    healthcheck:

      test: ["CMD-SHELL", "curl -fsS http://127.0.0.1:30000/health || exit 1"]

      interval: 30s

      timeout: 10s

      retries: 10

      start_period: 30m