Frosty40/h3-spark

MiniMax H3 — single DGX Spark (GB10) serving package

Complete serving package for running MiniMax H3 video generation on a single NVIDIA DGX Spark, at serving speed. Two parts:

Repository hub: the canonical source tree, install script, demo footage, and issue tracker live on GitHub: https://github.com/newjordan/h3-spark — start there to actually run the system. This HF repo mirrors the same source and hosts the serving configuration alongside the weights.

1. weights/ — pruned + NVFP4-quantized checkpoints (modified derivatives of the MiniMax H3 weights; see weights/MODEL_CARD.md and weights/LICENSE-MiniMax-H3-Community.txt).
2. the serving config at the repository root — ComfyUI integration source, native kernels, qualification receipts, and the validated serving environment (config/h3_spark_serving.env).

Demo

Four-shot segment generated and served from one DGX Spark (drone orbit on the wizard -> knight riding solo -> riding together -> wide shot with a dragon silhouette), with narration voice-over and score mix. Watch it in the GitHub README (it embeds the video):

https://github.com/newjordan/h3-spark

What this config does

• H3-scoped SageAttention 2++ quality attention with real-activation calibration (H3_ATTENTION=sage2-quality).
• Exact segmented modulation fusion (H3_MODULATION_FUSION=exact).
• Native NVFP4 packing for the Omni-Transformer linears (H3_SWIGLU_NVFP4_FUSION=auto, H3_RMS_ADALN_NVFP4_FUSION=auto) — 200 NVFP4 linears per evaluation.
• Promoted Q RMSNorm + rot96 RoPE → Sage Q-INT8 fusion (H3_Q_RMS_ROPE_INT8_FUSION=auto), byte-exact on default and non-default streams.
• VAE fast path: same-256px tile-batched decode/encode, SwiGLU silu(gate)*up native kernel, GPU fp16 host transfer, persistent host decode buffer, NHWC staging (COMFY_MINIMAX_H3_VAE_*).
• Step-invariant caches: rope/cond/mask, audio sinc-filter, ViT3D grid caches; identical-reference encode/vision/pack reuse.
• Optional static NVFP4 activation scales with fail-closed promotion (see labs/nvfp4_static/).

Measured on one DGX Spark (dgx-spark-host, GB10, sm120)

Warm fixed-reference workload 20-step render — 864×480, 124 frames at 24 fps, six references, S=20,423, boosting GB10, SM clock recorded in the same sentence as the time:

Config	Sampler	% faster	s/step	Executor	% faster	Wall	SM
Stock Comfy (BF16 ops)	243 s	—	12.1	~278 s	—	—	boosting
This config, dynamic NVFP4 (2026-08-15)	153.6 s	36.8%	7.68	171.7 s	38.2%	180.2 s	2300–2330 MHz
This config, promoted static NVFP4 scales (2026-08-15, 200/200 layers)	148.4–149.4 s	38.5–38.9%	7.42–7.47	166.4–167.6 s	39.7–40.1%	170.0–170.1 s	2242–2255 MHz

The 2026-08-15 static artifact passed the fail-closed promotion gate (verify_promotion.py exit 0): 200/200 static layers, zero excluded layers, held-out max ratio 0.9873. Static scales are shape-bound and workload-bound — dynamic NVFP4 is the always-valid default.

Install

Prereqs: ComfyUI (GPL-3.0), comfy-kitchen ≥ 0.2.27 (Apache-2.0), SageAttention ≥ 2.2.0 (Apache-2.0), PyTorch 2.13.0+cu130, Triton 3.7.x, NVIDIA driver ≥ 580, CUDA 13.0. Target: NVIDIA DGX Spark (sm120) or other Blackwell parts (B200 validation in progress).

# 1. get the source (canonical clone — GitHub)
git clone https://github.com/newjordan/h3-spark
cd h3-spark

# 2. weights: official upstream MiniMax H3 checkpoints
hf download MiniMaxAI/MiniMax-H3 --local-dir models --include "Ref2VA/*" "FL2VA/*"
mkdir -p models/ComfyUI/models
cp -r models/Ref2VA   models/ComfyUI/models/
cp -r models/FL2VA    models/ComfyUI/models/

# 3. overlay: install the serving source overlay into ComfyUI
bash scripts/install_comfyui_overlay.sh /path/to/ComfyUI

# 4. env: export the validated serving defaults
set -a; source config/h3_spark_serving.env; set +a

Our pruned + NVFP4-quantized derivatives (the numbers in the table below) ship as a companion weights package — status and download in the GitHub README.

Start ComfyUI with:

python main.py --listen 127.0.0.1 --port 18188 \
  --models-directory /path/to/models/ComfyUI/models \
  --disable-pinned-memory --fp16-intermediates --highvram \
  --disable-dynamic-vram --fast fp16_accumulation cublas_ops autotune

License and attribution

• Model weights in weights/ are modified derivatives of MiniMax H3 and are licensed under the MiniMax H3 Community License Agreement (full text in weights/LICENSE-MiniMax-H3-Community.txt, required NOTICE file alongside). Read it, including the Applicable Territory definition and Sections III and V, before use or redistribution.
• Qwen3-VL-32B text encoder: Apache-2.0.
• ComfyUI portions of the serving source: GPL-3.0. comfy-kitchen and SageAttention portions: Apache-2.0. See third_party_licenses/.
• Original serving-config work: see docs/PROVENANCE.md.