# spark-control launch recipes
#
# These are NOT the dashboard menu. The menu is whatever is actually downloaded
# on the Sparks — Spark Control scans the Hugging Face cache on each load and
# shows what it finds. These entries are launch *recipes*: matched to an on-disk
# model by `repo`, they say HOW to launch it. A downloaded model with no recipe
# here shows up as "needs setup", and the dashboard infers + saves one on first
# use (from the model's own config.json). Add a recipe to make a known model
# launch correctly the moment it's downloaded, with no setup prompt.
#
# Each recipe produces this command on Spark 1:
#   cd ~/spark-vllm-docker
#   ./launch-cluster.sh [--solo] -d exec vllm serve <repo> \
#     --port=<defaults.port> --host=<defaults.host> <vllm_args...>

defaults:
  port: 8888
  host: 0.0.0.0

models:
  qwen3-vl:
    display_name: "Qwen3-VL 235B (vision)"
    description: >-
      Qwen's flagship multimodal model. 235B total parameters with ~22B
      active per token (Mixture-of-Experts). Handles text, images, and
      many languages. The most capable model in this catalog — also the
      slowest to load because it splits across both Sparks.
    repo: RedHatAI/Qwen3-VL-235B-A22B-Instruct-NVFP4
    size_gb: 135
    mode: cluster
    capabilities: [vision, multilingual]
    expected_ready_seconds: 300
    vllm_args:
      - --gpu-memory-utilization=0.7
      - -tp=2
      - --distributed-executor-backend=ray
      - --max-model-len=32768
      - --max-num-batched-tokens=16384

  gemma4:
    display_name: "Gemma 4 31B"
    description: >-
      Google's mid-size reasoning model. 31B dense parameters with built-in
      thinking mode and function-calling. Strong on math, logic, and
      structured outputs; also supports vision input. Runs solo on one Spark.
    repo: RedHatAI/gemma-4-31B-it-NVFP4
    size_gb: 23
    mode: solo
    capabilities: [vision, reasoning, tools]
    expected_ready_seconds: 240
    vllm_args:
      - --gpu-memory-utilization=0.8
      - --max-model-len=32768
      - --max-num-batched-tokens=16384
      - --reasoning-parser=gemma4
      - --tool-call-parser=gemma4
      - --enable-auto-tool-choice
      - --load-format=fastsafetensors
      - --enable-prefix-caching
      - --kv-cache-dtype=fp8

  gemma4-26b:
    display_name: "Gemma 4 26B-A4B (vision, light)"
    description: >-
      Lighter, faster sibling of the Gemma 4 31B above: a Mixture-of-Experts
      model with 26B total parameters but only ~4B active per token, so it
      generates quickly. Takes images as well as text (good for tasks like
      reading a business card into structured text). Reasoning is a bit
      shallower than the dense 31B. Runs solo on one Spark.
    repo: nvidia/Gemma-4-26B-A4B-NVFP4
    size_gb: 17
    mode: solo
    capabilities: [vision, reasoning, tools]
    expected_ready_seconds: 240
    vllm_args:
      - --gpu-memory-utilization=0.8
      - --max-model-len=32768
      - --max-num-batched-tokens=16384
      - --reasoning-parser=gemma4
      - --tool-call-parser=gemma4
      - --enable-auto-tool-choice
      # MoE backend: research found this model's expert layers fall back to
      # 'marlin' on GB10 (the fast flashinfer_cutlass path errors on sm_121).
      # If a swap fails to start, this flag is the first thing to flip.
      - --moe_backend=marlin
      - --load-format=fastsafetensors
      - --enable-prefix-caching
      - --kv-cache-dtype=fp8

  qwen36:
    display_name: "Qwen3.6 35B-A3B (daily driver)"
    description: >-
      Qwen's latest fast Mixture-of-Experts model: 35B total parameters but
      only ~3B active per token, making inference quick. Long 64K-token
      context window. A good default for everyday chat and longer documents.
    repo: RedHatAI/Qwen3.6-35B-A3B-NVFP4
    size_gb: 20
    mode: solo
    # Qwen3.6-35B-A3B is natively multimodal (Qwen3_5MoeForConditionalGeneration,
    # vision tower ships in the checkpoint). Confirmed reading a business card
    # cleanly on this cluster — use the "Vision check" button on the live card.
    capabilities: [vision, reasoning]
    expected_ready_seconds: 300
    vllm_args:
      - --gpu-memory-utilization=0.85
      - --max-model-len=65536
      - --max-num-batched-tokens=16384
      - --reasoning-parser=qwen3
      - --moe_backend=flashinfer_cutlass
      - --load-format=fastsafetensors
      - --enable-prefix-caching
      - --kv-cache-dtype=fp8
      # Cap image resolution: a large phone photo (e.g. 12MP) otherwise expands
      # to ~11.8k vision tokens, blowing past vLLM's ~4096-image-token limit and
      # getting rejected with a 400. ~2MP auto-downscales big images server-side
      # (so every /v1 consumer is covered) while staying sharp enough for OCR.
      - '--mm-processor-kwargs={"max_pixels": 2000000}'