# Qwen3.6-35B-A3B-heretic NVFP4 + DFlash on DGX Spark
#
# Production config validated 2026-06-18 (vLLM 0.23.0 unified image).
#   - Image: ghcr.io/aeon-7/aeon-vllm-ultimate:latest (= :2026-06-18-v0.23.0-dflashfix)
#   - Model: AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4 (multimodal-preserved, ~22 GB)
#   - Drafter: z-lab/Qwen3.6-35B-A3B-DFlash (~905 MB; MUST be re-pulled if cloned before 2026-04-19)
#
# Measured performance (vLLM 0.23.0, aeon-vllm-ultimate:latest, DFlash n=11):
#   ~97 tok/s avg single-stream (120+ on math/code/reasoning); ~740 tok/s peak aggregate at c=64.
#   Scales to c=64 with no crash (the prior vllm-spark-omni-q36:v1.2 image died at c>=32 under DFlash).
#
# Default layout (override with QWEN36_MODELS_DIR=/your/path docker compose up -d):
#   /opt/qwen36/qwen36-nvfp4/    ← AEON-7/Qwen3.6-35B-A3B-heretic-NVFP4
#   /opt/qwen36/qwen36-dflash/   ← z-lab/Qwen3.6-35B-A3B-DFlash

services:
  vllm:
    image: ghcr.io/aeon-7/aeon-vllm-ultimate:latest
    container_name: vllm-qwen36-heretic
    restart: unless-stopped
    network_mode: host
    environment:
      - VLLM_ALLOW_LONG_MAX_MODEL_LEN=1
      - TORCH_MATMUL_PRECISION=high
      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
      - NVIDIA_DISABLE_REQUIRE=1
      # NOTE: the unified aeon-vllm-ultimate:latest (vLLM 0.23.0) image ships the
      # SM120-family CUTLASS NVFP4 kernels GB10 dispatches to and selects them by
      # default — the old v1/v1.2 VLLM_TEST_FORCE_FP8_MARLIN=1 MoE guard is no
      # longer required for this model and is intentionally omitted here. (Per
      # the DGX-Spark backend notes, the 256-expert x top-8 Qwen3.6 MoE shape
      # prefers CUTLASS; do NOT pin Marlin on this image.)
    volumes:
      - ${QWEN36_MODELS_DIR:-/opt/qwen36}/qwen36-nvfp4:/models/qwen36
      - ${QWEN36_MODELS_DIR:-/opt/qwen36}/qwen36-dflash:/models/qwen36-dflash
    command:
      - bash
      - -c
      - |
        exec vllm serve /models/qwen36 \
          --served-model-name qwen36-35b-heretic qwen36-fast qwen36-deep \
          --host 0.0.0.0 --port 8000 \
          --tensor-parallel-size 1 \
          --dtype auto \
          --quantization compressed-tensors \
          --max-model-len 262144 \
          --max-num-seqs 64 \
          --max-num-batched-tokens 65536 \
          --gpu-memory-utilization 0.70 \
          --enable-chunked-prefill \
          --enable-prefix-caching \
          --load-format safetensors \
          --trust-remote-code \
          --enable-auto-tool-choice \
          --tool-call-parser qwen3_coder \
          --reasoning-parser qwen3 \
          --speculative-config '{"method":"dflash","model":"/models/qwen36-dflash","num_speculative_tokens":11}' \
          --attention-backend flash_attn
          # NOTES (aeon-vllm-ultimate:latest, vLLM 0.23.0):
          #  - --enforce-eager is NOT required: the image aligns spec-decode CUDA-graph
          #    capture sizes (FULL graphs) and the latest z-lab drafter (re-pulled
          #    2026-04-19+) is long-context-stable, so FULL graphs give ~30% over eager.
          #  - DFlash num_speculative_tokens=11 is optimal for this model (acceptance /
          #    long-context fidelity drop as n climbs past ~11; n=15 wasted draft compute).
          #  - Leave the DFlash drafter's attention backend at its DEFAULT (do not add
          #    attention_backend to the spec config) and do NOT set --kv-cache-dtype:
          #    the non-causal drafter requires BF16 KV.
          #  - No --mamba-block-size override needed (this drafter is all-full-attention).
          #  - gpu-memory-utilization 0.70: GB10 shares one LPDDR5X pool; <=0.70 avoids thrash.
    ipc: host
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]