esh-pfi-infrastructure/stacks/char-rp-gguf/compose.yaml

# char-rp-gguf — dedicated GGUF character-RP seat on fv-ml1 GPU 0, REPLACING the
# broken ms32-24b-angel NVFP4 serve (garbage output — bad self-quant W4A4).
#
# Two co-located llama.cpp (llama-server) instances on GPU 0, served alongside the
# 35B-A3B heretic `gen` (qwen36-27b-aeon stack, :8015):
#
#   llama-charrp (:8016, gateway char-rp) — TheDrummer Magidonia-24B-v4.3 Q6_K.
#       Magistral (Mistral) dark-romantasy RP tune. NON-thinking PROSE seat: elite
#       literary prose, zero refusal, ~65 tok/s, precise POV/instruction adherence.
#
#   llama-charrp-reasoning (:8018, gateway char-rp-reasoning) — ArliAI QwQ-32B-RpR-v4 Q5_K_M.
#       QwQ reasoning RP tune whose reasoning DATA was generated with QwQ-ABLITERATED
#       → it does NOT re-censor in the think phase (the exact failure mode that killed
#       the Pantheon/DeepSeek-distilled reasoners: they reason themselves into refusals
#       inside <think>). llama.cpp MANAGES QwQ reasoning natively: --reasoning on
#       surfaces the trace in reasoning_content (clean prose in content, no <think>
#       leak), --reasoning-budget caps the chain-of-thought. ~50 tok/s @ Q5_K_M.
#
# WHY GGUF/llama.cpp (not vLLM NVFP4): sidesteps BOTH traps that killed the Angel serve
# — the vLLM NVFP4 self-quant breakage AND the Mistral-tokenizer/vision crash. llama.cpp
# handles Mistral + QwQ tokenizers natively. NEVER Ollama (banned fleet-wide).
#
# WHY TWO models (not one): no single dense 24-32B is BOTH an elite non-thinking prose
# seat AND a clean managed-reasoning seat on llama.cpp. Magidonia's Magistral [THINK]
# discipline is loose (won't reliably close [/THINK] on substantive reasoning → prose
# bleeds into reasoning_content, content empties); Cydonia-R1's <think> is emergent, so
# llama.cpp can't manage/cap it → runaway CoT that never reaches prose. QwQ's template
# opens <think> natively → llama.cpp manages+caps it. So: best-of-breed per seat.
# ONE-MODEL FALLBACK (consistent Mistral style, lighter reasoning): point both services
# at Magidonia via CHARRP_REASONING_MODEL in .env and blank CHARRP_REASONING_EXTRA_*.
#
# ALTERNATE prose model: PaintedFantasy-v4.1-24B (also Magistral, more literary flair
# but looser POV adherence) — set CHARRP_MODEL in .env. All candidate GGUFs are
# pre-pulled to /tank/aimodels/llm/rp/.
#
# VRAM (GPU 0, co-resident with gen ~38G): Magidonia Q6 ~19G + RpR-v4 Q5 ~23G + KV/
# compute ~6-8G = ~85-88G / 97G (~9-12G margin). Keep ctx modest; drop CHARRP_*_CTX
# to 8192 in .env if warmup bites. depends_on sequences char-rp first.
#
# API auth: blank (LAN-internal on the GPU host; matches API_KEY= in the AEON stack /
# gateway VLLM_API_KEY). llama-server ignores the gateway's api_key when none is set.
#
# All tunables live in .env — edit that, not this file.

name: char-rp-gguf

services:
  # ── PROSE seat — non-thinking. gateway char-rp. ──
  llama-charrp:
    image: ${LLAMA_IMAGE:-ghcr.io/mostlygeek/llama-swap:cuda}
    container_name: ${CHARRP_CONTAINER:-llama-charrp}
    restart: unless-stopped
    runtime: nvidia
    ports:
      - "${CHARRP_PORT:-8016}:8080"
    volumes:
      - ${MODELS_DIR:-/tank/aimodels/llm}:/models:ro
    environment:
      # Pin to GPU 0 (the on-demand large-model card; the always-on vLLM trio owns GPU 1).
      - NVIDIA_VISIBLE_DEVICES=${CHARRP_GPU_ID:-0}
    entrypoint: ["/app/llama-server"]
    command:
      - --model
      - /models/${CHARRP_MODEL:-rp/TheDrummer_Magidonia-24B-v4.3-Q6_K.gguf}
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --n-gpu-layers
      - "999"
      - --ctx-size
      - "${CHARRP_CTX:-98304}"
      - --flash-attn
      - on
      # q8_0 KV cache ~halves KV VRAM (8-bit, near-lossless) → ~2x the context per GB.
      # Mistral/Magistral handles q8 KV cleanly. Set f16 in .env to disable.
      - --cache-type-k
      - ${CHARRP_KV_TYPE:-q8_0}
      - --cache-type-v
      - ${CHARRP_KV_TYPE:-q8_0}
      - --jinja
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://localhost:8080/health >/dev/null || exit 1"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 240s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Inference
      - homepage.name=char-rp (Magidonia-24B GGUF)
      - homepage.icon=mdi-drama-masks
      - homepage.description=Dark-romantasy RP prose seat, non-thinking (llama.cpp, fv-ml1 GPU 0)
      - homepage.href=http://10.251.50.54:${CHARRP_PORT:-8016}

  # ── REASONING seat — NEO-CODE = Heretic2-Thinking (Qwen3.6-27B) managed thinking. gateway char-rp-reasoning. ──
  llama-charrp-reasoning:
    # ⚠️ CUSTOM llama.cpp build (master 6eddde0 + unmerged PR #25544). Needed for two reasons:
    # (1) recent master parses Qwen3.6's native qwen3_coder tool-call format (stock b8840 predates it —
    #     the <tool_call><function=..><parameter=..> XML is Qwen3.5/3.6-native, NOT an OpenHands quirk);
    # (2) PR #25544 multi-terminator reasoning-budget fix (Worldtree #355) — belt-and-suspenders now that
    #     NEO-CODE shows 0.0 runaway (R36 gate), but keep it. DO NOT revert to stock until #25544 merges.
    # Build recipe + why + rollback: ./llamacpp-custom/README.md.
    # Rollback: set LLAMA_REASONING_IMAGE=ghcr.io/mostlygeek/llama-swap:cuda in .env + recreate.
    image: ${LLAMA_REASONING_IMAGE:-llamacpp-charrp:custom-latest}
    container_name: ${CHARRP_REASONING_CONTAINER:-llama-charrp-reasoning}
    restart: unless-stopped
    runtime: nvidia
    # Sequence AFTER the prose seat is healthy so the two GPU-0 allocations don't race.
    depends_on:
      llama-charrp:
        condition: service_healthy
    ports:
      - "${CHARRP_REASONING_PORT:-8018}:8080"
    volumes:
      - ${MODELS_DIR:-/tank/aimodels/llm}:/models:ro
    environment:
      - NVIDIA_VISIBLE_DEVICES=${CHARRP_GPU_ID:-0}
    entrypoint: ["/app/llama-server"]
    command:
      - --model
      - /models/${CHARRP_REASONING_MODEL:-rp/Qwen3.6-27B-NEO-CODE-HERE-2T-OT-Q5_K_M.gguf}
      - --host
      - 0.0.0.0
      - --port
      - "8080"
      - --n-gpu-layers
      - "999"
      - --ctx-size
      - "${CHARRP_REASONING_CTX:-40960}"
      - --flash-attn
      - on
      # NEO-CODE = Qwen3.6-27B GDN-hybrid (16 of 64 layers cache KV → KV cheap); native ctx 262144
      # (256K). Full 256K @ q8_0 KV ≈ 8.6G, fits GPU0 w/ ~3.8G margin. q8_0 coherent; f16 in .env if gibberish.
      - --cache-type-k
      - ${CHARRP_REASONING_KV_TYPE:-q8_0}
      - --cache-type-v
      - ${CHARRP_REASONING_KV_TYPE:-q8_0}
      - --jinja
      # NEO-CODE's Qwen3.6 template natively opens <think> → llama.cpp manages the reasoning
      # (trace to reasoning_content, content stays clean prose); --reasoning-budget caps the CoT.
      # (R36 gate 2026-07-14: NEO-CODE composite 0.922 tool-calling + 0.0 runaway — beat Deckard
      # 0.08/0.80 and gen-reasoning 0.856. Budget held at 400: latency-coupled to soong's client timeout.)
      - --reasoning
      - on
      - --reasoning-format
      - deepseek
      - --reasoning-budget
      - "${CHARRP_REASONING_BUDGET:-400}"
      # Sampler defaults per the DavidAU/Qwen3.6 model card (thinking-mode, general tasks): temp 1.0,
      # top_p 0.95, top_k 20, min_p 0.0, no rep-penalty, no DRY (DRY was a QwQ/Deckard looping band-aid
      # NEO-CODE doesn't need). All tunable via .env. NOTE: 0.922 tool-gate was on the OLD Deckard
      # samplers (effective temp~0.8 + DRY); re-validate tools + slop on these card samplers.
      - --temp
      - "${CHARRP_REASONING_TEMP:-1.0}"
      - --top-p
      - "${CHARRP_REASONING_TOP_P:-0.95}"
      - --top-k
      - "${CHARRP_REASONING_TOP_K:-20}"
      - --min-p
      - "${CHARRP_REASONING_MIN_P:-0.0}"
    healthcheck:
      test: ["CMD-SHELL", "curl -fsS http://localhost:8080/health >/dev/null || exit 1"]
      interval: 30s
      timeout: 10s
      retries: 3
      start_period: 300s
    networks:
      - tnet
    labels:
      - homepage.group=AI - Dormant
      - homepage.name=char-rp-reasoning (QwQ-32B RpR-v4 GGUF)
      - homepage.icon=mdi-brain
      - homepage.description=Dark-romantasy RP reasoning seat, managed CoT (llama.cpp, fv-ml1 GPU 0)
      - homepage.href=http://10.251.50.54:${CHARRP_REASONING_PORT:-8018}

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