Minimax Dgx Spark

MiniMax DGX-Spark (Music3 + H3) via ComfyUI

Generate AI music (Music3) and AI video (H3) by submitting ComfyUI API workflows, then post-process outputs (MP3 / MP4). All paths/hosts are ENV-driven; nothing personal is hardcoded. This skill drives the sibling scripts in the MiniMax-DGX-Spark package (or standalone files on the node).

Prereqs

• A reachable ComfyUI instance.
  • Local: curl -s http://127.0.0.1:8188/system_stats
  • Remote node (e.g. a DGX Spark): curl -s http://<NODE_IP>:8188/system_stats
• ffmpeg, python3, and (for H3) no extra GPU beyond ComfyUI's.

Set the target once per session:

export COMFY_HOST="http://<NODE_IP>:8188"   # or leave unset for localhost:8188
export COMFY_OUTPUT_DIR="..."              # where ComfyUI writes outputs (node-local fs)

Speak in IPs, never hostnames

If remoting to a node, use its IP in COMFY_HOST. Hostnames in SSH config are connection aliases only. 127.0.0.1 works from inside the node.

Music3 (text-to-music)

Generate a track (MP3 default)

python3 music3/generate.py \
  --caption "Hard-hitting trap rap beat, aggressive 808 bass, tight hi-hats, melodic piano" \
  --lyrics "[intro]\nLet's go\n[verse 1]\nDigital rhythm sharp through the night\n[chorus]\nUnleash the spark, reaching new heights" \
  --duration 35.0 --prefix my_track --format mp3

• --format mp3 is the default. Add --format all for an MP4 waveform video.
• Keep --duration ≤ 35.0s when co-tenant with a large LLM.
• For max fidelity, switch unet_name to minimax_music3_dit_fp16.safetensors (default is int8).

Bypass the app (direct API, no app dependency)

If you only have the running ComfyUI and the checkpoint files, submit the raw API graph directly (see music3/workflows/*.json for the exact node wiring). The canonical pipeline: CLIPLoader (int8 TE) → UNETLoader (int8 DiT) → VAELoader (dav) → MiniMaxMusic3TextEncode (caption/lyrics/duration/cfg/topk) → EmptyMiniMaxMusic3LatentAudio → ConditioningZeroOut → KSampler (euler/simple, cfg 1.0) → VAEDecodeAudio → SaveAudio.

H3 (text/image-to-video)

Generate a short clip

python3 h3/generate-video.py --prompt "a calm ocean wave at dusk, photorealistic" \
  --duration 5 --prefix my_clip --host "$COMFY_HOST"

• --duration is rounded UP to H3's 17n+5 frame grid.
• Add --reference <img-in-comfy-input> (repeatable) for identity/stability across shots.
• Timings: a 5s 720p clip is ~7–8 min under heavy co-tenant load on GB10; much faster standalone.

Convert / Watchdog

• Convert a FLAC: python3 scripts/convert-media.py out.flac --mp3 (or --mp4).
• Long generations that could outlive an SSH session: submit, then run detached:

nohup python3 scripts/convert-watchdog.py --prompt-id <ID> --host "$COMFY_HOST" --format mp3 > /tmp/watchdog.log 2>&1 &

The watchdog polls /history/<ID> and converts when done, surviving disconnects.

⚠️ If you get a FLAC but no MP3/MP4, the generation SUCCEEDED — the post-processing step did not. Post-processing runs client-side, so a disconnected/timed-out session yields the ComfyUI output (FLAC) but the conversion never ran. Do NOT resubmit. Run the watchdog or convert manually: python3 scripts/convert-watchdog.py --prompt-id <ID> --format mp3. The prompt ID is printed on success (or fetch it from $COMFY_HOST/queue).

Pitfalls

1. Get the prompt ID — generate*.py prints it on success. If the client died before doing so, fetch curl -s "$COMFY_HOST/queue".
2. Do not start a second ComfyUI — check scripts/status-comfyui.sh / curl .../queue first; reuse a live instance.
3. Audio ownership — a <video> is muted; sound ships as separate <audio> (for HyperFrames muxing).
4. GP10 UMA: nvidia-smi CSV returns [N/A] for memory; read /proc/driver/nvidia/gpus/0/memory or system_stats vram_*.
5. Never interrupt a running ComfyUI job — wait for the queue to drain; kills orphan the job server-side.