esh-pfi-infrastructure/stacks/mistral-small-4/compose.yaml

# mistral-small-4 — Mistral-Small-4-119B-2603 (official NVFP4) on fv-ml1 GPU 0.
#
# Mistral Small 4 is a 119B-total / 6.5B-active MoE (128 experts, 4 active),
# 256K context, multimodal, Apache-2.0 (released 2026-03). This serves the
# OFFICIAL NVFP4 checkpoint (mistralai/Mistral-Small-4-119B-2603-NVFP4) — 74.4 GB
# of compressed-tensors (llm-compressor, a vLLM + Red Hat collaboration, day-0
# vLLM support). It is the GPU-0 tenant (the slot formerly reserved for a
# creative-writing pick — operator reassigned 2026-06-15; tune-for-creative-
# writing comes after base-characteristic probing).
#
# WHY NVFP4 (not FP8/bf16): on a SINGLE 96 GB card, NVFP4 (74.4 GB weights) is
# the only variant that fits at TP=1 — FP8 (~119 GB) and bf16 (~238 GB) need both
# GPUs. The card is Blackwell (sm_120) with FP4 tensor cores, so NVFP4 gets a real
# speedup, not just a VRAM save. NOTE: this is the COMPRESSED-TENSORS NVFP4 path
# (vendor-shipped, vLLM-tested) — distinct from the nvidia-ModelOpt NVFP4 MoE
# loader that broke on Qwen3.6 (#44081); different code path, day-0 supported.
#
# WHY TP=1 here: Mistral's official card uses --tensor-parallel-size 2 (their
# reference 80 GB cards can't fit 74.4 GB + context on one). The 96 GB Blackwell
# flips that to single-card: 74.4 GB weights + ~5 GB overhead leaves ~17 GB for
# KV. Mistral Small 4 uses MLA attention (TRITON_MLA) so KV is compressed/cheap —
# big context stays affordable even on a constrained KV pool. We serve the FULL
# native 256K (max-model-len 262144) — the KV pool is util-bound (~862K tokens)
# so 256K costs no extra VRAM, it just lets one request use up to 256K (max
# concurrency 3.29x at full length). For novel-length consistency-checking.
#
# vLLM PIN: v0.22.0 (in .env) — the last release with WORKING Mistral vision
# (#44911 fetch_images regression hit 0.22.1+/0.23.0). See the .env header.
#
# Serve flags mirror Mistral's official command (cited in README), adapted for
# single-card: TP 2->1, util 0.8->0.93, max-num-seqs 128->32 (32 keeps the
# 256K warmup transient flat). All tunables live in .env — edit that, not this file.

name: mistral-small-4

services:
  vllm-mistral4:
    image: ${MISTRAL_IMAGE}
    container_name: ${MISTRAL_CONTAINER_NAME}
    restart: unless-stopped
    ipc: host
    ports:
      - "${MISTRAL_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:
      - ${MISTRAL_MODEL}
      # Pre-quantized NVFP4 (compressed-tensors) — vLLM auto-detects the quant;
      # no --quantization flag.
      - --served-model-name
      - mistral-small-4
      - --host
      - 0.0.0.0
      - --port
      - "8000"
      - --tensor-parallel-size
      - "1"
      - --gpu-memory-utilization
      - ${MISTRAL_GPU_MEM_UTIL}
      - --max-model-len
      - ${MISTRAL_MAX_MODEL_LEN}
      # MLA attention backend (DeepSeek-style latent KV → compressed, cheap KV).
      - --attention-backend
      - TRITON_MLA
      # Mistral tool-calling + configurable reasoning (per the official card).
      - --tool-call-parser
      - mistral
      - --enable-auto-tool-choice
      - --reasoning-parser
      - mistral
      - --max-num-seqs
      - ${MISTRAL_MAX_NUM_SEQS}
      # VISION ENABLED. vLLM is pinned to v0.22.0 in .env — the last release BEFORE
      # the Mistral multimodal regression (#44911, `MistralCommonImageProcessor has
      # no attribute fetch_images`, landed ~0.22.1+; 0.23.0 is affected). v0.22.0
      # still has Mistral-Small-4 arch + compressed-tensors NVFP4 support (the
      # #44081 ModelOpt-NVFP4 bug on 0.22.0 is a DIFFERENT quant path, doesn't touch
      # this compressed-tensors checkpoint). Gives a verified working vision tower
      # as the abliteration/tuning baseline. (qwen36 stays on 0.23.0 — separate
      # container; it NEEDS 0.23.0 for its ModelOpt NVFP4.)
      - --dtype
      - auto
      - --enable-prefix-caching
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              device_ids:
                - "${MISTRAL_GPU_ID}"
              capabilities:
                - gpu
    healthcheck:
      test: ["CMD", "curl", "-f", "http://localhost:8000/health"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 600s
    networks:
      - tnet
    labels:
      - homepage.group=AI Systems
      - homepage.name=Mistral Small 4 (NVFP4)
      - homepage.icon=mdi-creation
      - homepage.description=Mistral-Small-4-119B-2603 MoE (NVFP4) via vLLM (fv-ml1 GPU 0)
      - homepage.href=http://10.251.50.54:${MISTRAL_PORT}/docs

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