esh-pfi-infrastructure/stacks/vllm/compose.yaml

# vLLM — utility seats on fv-ml1: Skywork Reward-V2 classifier + the coder FIM seat.
#
# Originally created to replace the unmaintained Infinity stack (embed +
# rerank); generalized 2026-05-13 to host any vLLM-served model on fv-ml1,
# starting with the Skywork-Reward-V2-Llama-3.1-8B reward classifier
# (AWQ-quantized locally, served from /tank/aimodels/llm/).
#
# ⚠ Embedding + reranking LEFT this stack 2026-09-25 (Prime): they now run on TEI
# on esh-ml1 (stacks/embed-rerank), and TEI is the fleet's embed/rerank engine
# from now on. Do not re-add them here.
#
# All tunables live in .env — edit that, not this file.
#
# Skywork-Reward-V2-Llama-3.1-8B-AWQ is a locally-quantized model — lives at
# /tank/aimodels/llm/Skywork-Reward-V2-Llama-3.1-8B-AWQ on fv-ml1 and is
# bind-mounted into the reward service at /local-models. Not from HF Hub.

services:
  # vllm-embed (Qwen3-Embedding-0.6B, :8001) and vllm-rerank-a3
  # (bge-reranker-v2-m3, :8013) were RETIRED 2026-09-25 (Prime): embedding and
  # reranking moved to TEI on esh-ml1 (stacks/embed-rerank), and TEI is now the
  # fleet's embed/rerank engine. The gateway names `qwen3-embedding`, `reranker`
  # and `reranker-a3-bge-v2-m3` did not change. Their definitions are in git
  # history before this commit if a rollback is ever needed.

  vllm-reward:
    image: vllm/vllm-openai:${VLLM_VERSION}
    container_name: vllm-reward
    restart: unless-stopped
    ipc: host
    ports:
      - "${REWARD_PORT}:8000"
    volumes:
      # AWQ output lives in the legacy llama-swap models tree, not the HF cache
      # — bind-mount the LLM models dir read-only so the reward service can
      # load it as a local-path HF-format model.
      - /tank/aimodels/llm:/local-models:ro
    environment:
      - VLLM_API_KEY=${API_KEY:-}
    command:
      - /local-models/Skywork-Reward-V2-Llama-3.1-8B-AWQ
      - --served-model-name
      - Skywork/Skywork-Reward-V2-Llama-3.1-8B-AWQ
      # vLLM 0.19.1 deprecated --task in favor of --runner. The model's
      # config.json declares `LlamaForSequenceClassification` so the
      # pooling runner uses it as a classifier (single-label reward score)
      # without needing an explicit task flag.
      - --runner
      - pooling
      - --host
      - 0.0.0.0
      - --port
      - "8000"
      - --gpu-memory-utilization
      - ${REWARD_GPU_MEM_UTIL}
      - --max-model-len
      - ${REWARD_MAX_MODEL_LEN}
      - --dtype
      - auto
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids:
                - "${GPU_ID}"
              capabilities:
                - gpu
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 240s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Eval & Retrieval
      - homepage.name=vLLM Reward (Skywork)
      - homepage.icon=mdi-scale-balance
      - homepage.description=Skywork-Reward-V2 8B classifier via vLLM (fv-ml1)
      - homepage.href=http://10.251.50.54:${REWARD_PORT}/docs

  # vllm-granite (ibm-granite/granite-4.1-8b-fp8, :8004) — RETIRED 2026-08-12,
  # service block removed 2026-08-20. It was the fleet summarizer until the
  # `summarizer` and `classifier` aliases were repointed at the gen seat; the
  # container was stopped then and sat Exited for 8 days while this block still
  # claimed it as production.
  #
  # ⚠️ Retiring the SEAT did not retire its CONSUMERS, and nobody checked. The
  # `granite-4.1-8b` gateway alias was deleted at the same time, but `nevermore`
  # was still pinned to that alias by name — so its LLM summarization pass failed
  # 67 consecutive times over 8 days, 0 tokens, entirely silently, because a
  # failed gateway call still returns 200-shaped spend-log rows and nothing
  # alerts on status=failure. Found 2026-08-20 only because someone asked an
  # unrelated question about reranker VRAM.
  #
  # The rule this earns: RETIRING A MODEL IS A TWO-SIDED OPERATION. Grep every
  # consumer's config for the alias BEFORE deleting it, and prefer that consumers
  # pin stable ALIASES (`summarizer`) over model names (`granite-4.1-8b`) so the
  # gateway can repoint them without anyone editing a downstream .env.

  vllm-coder:
    image: vllm/vllm-openai:${VLLM_VERSION}
    container_name: vllm-coder
    restart: unless-stopped
    ipc: host
    ports:
      - "${CODER_PORT}:8000"
    volumes:
      - /tank/aimodels/huggingface:/hfcache
    environment:
      - HF_HOME=/hfcache
      - HF_HUB_CACHE=/hfcache/hub
      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
      - VLLM_API_KEY=${API_KEY:-}
    command:
      # Qwen2.5-Coder-1.5B (BASE) — FIM code-completion seat for Zed edit-predictions
      # (deep-research pick 2026-07-27). Native fill-in-the-middle: <|fim_prefix|> /
      # <|fim_suffix|> / <|fim_middle|> (IDs 151659/151660/151661); Zed sends the
      # FIM-formatted prompt to /v1/completions and vLLM passes it through (the FIM
      # special tokens live in the tokenizer). BASE not -Instruct (FIM is a
      # pretraining objective; base completions are cleaner). Apache-2.0. Runner-up =
      # Qwen2.5-Coder-3B (higher HumanEval-FIM but non-commercial Qwen-Research license).
      - ${CODER_MODEL}
      - --served-model-name
      - ${CODER_SERVED_NAME}
      - --host
      - 0.0.0.0
      - --port
      - "8000"
      - --gpu-memory-utilization
      - ${CODER_GPU_MEM_UTIL}
      - --max-model-len
      - ${CODER_MAX_MODEL_LEN}
      - --max-num-seqs
      - ${CODER_MAX_NUM_SEQS}
      - --dtype
      - auto
      - --kv-cache-dtype
      - ${CODER_KV_CACHE_DTYPE}
      - --enable-prefix-caching
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids:
                - "${CODER_GPU_ID}"
              capabilities:
                - gpu
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 300s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Inference
      - homepage.name=vLLM Qwen2.5-Coder 1.5B (FIM)
      - homepage.icon=mdi-code-braces
      - homepage.description=Qwen2.5-Coder-1.5B FIM code-completion (fv-ml1, Zed edit-predictions)
      - homepage.href=http://10.251.50.54:${CODER_PORT}/docs

  # vllm-lfm25 (LiquidAI/LFM2.5-2.6B, :8021) — RETIRED PERMANENTLY 2026-08-20 by
  # operator directive. It was an EVAL-ONLY bake-off seat against granite-4.1-8b
  # (brokkr R-target, 2026-08-10) that never received the operator ruling it was
  # pending. Its comparator is gone (granite retired from the roster 2026-08-15),
  # it was deliberately never wired into any default/fallback routing chain, and
  # LiteLLM spend logs showed 0 calls in the 4-day window ending 2026-08-21.
  # Freed 8,772 MiB on fv-ml1 GPU1. The `lfm2.5-2.6b` gateway alias was removed
  # in the same pass so the name 404s cleanly rather than erroring against a dead
  # backend. Weights remain in the shared HF cache; nothing was deleted from disk.

networks:
  tnet:
    name: traefik-net
    external: true