Accelerating MiniMax-H3 on Nvidia DGX Spark with ComfyUI and Sol-Engine

X postkeysTool

A public repository demonstrates a 1.54x end-to-end acceleration for the 33B MiniMax-H3 omni-DiT on a single Nvidia DGX Spark (GB10, sm_121) using ComfyUI. It ports Nvidia Sol-Engine techniques, applies Triton patches, and resolves critical Blackwell compilation issues.

Key points

  1. MiniMax-H3 (33B audio+video) render time drops from 312.7s to 202.5s for a 5-second 864x480 clip (1.54x end-to-end, 1.72x per denoising step).
  2. Fixes a silent miscompilation in Sol-Attn on sm_121 where flex_attention produced corrupted model outputs without errors.
  3. Ports Nvidia Sol-Engine FirstBlockCache and batched VAE tile decoding from multi-GPU stacks into single-GPU ComfyUI.
  4. Leverages kijai Triton INT8/TMA kernels to maximise quality and throughput on Blackwell architecture.
  5. Multi-node benchmarks confirmed that fabric limits (5.52 GB/s vs 503 GB/s NVLink) make CP=2 slower than a single tuned DGX Spark node.
  6. Provides a one-shot installer script (install.sh) and an agent-executable recipe (RECIPE.md) compatible with tools like Claude Code and Codex.

Why it matters

Running massive DiT models for video and audio generation locally requires heavy hardware optimization. This project resolves silent compiler bugs on modern GB10 chips and brings enterprise-grade multi-GPU optimizations down to single-GPU ComfyUI setups.

Analysis

Beyond performance gains, the project's primary value lies in identifying and resolving architecture-specific pitfalls on compute capability 12.1. By benchmarking each pipeline stage (Sol-Attn, FirstBlockCache, and VAE decode) and documenting why multi-node scaling fails on this fabric, the author provides clear architectural insights alongside turnkey configuration scripts.

Original post

keys 🧪

@u1tra_instinct · 5 August 2026

For those who are interested in Nvidia’s sol-engine and few speed integrations with the MinIMax H3 speed up on single dgx spark here is the repo to test it out just on comfyUI workflow optimization. My testing repo here: (may do More work tomorrow) t.co/jGaZM8ok8q

96 likes10 replies110 bookmarks15.6k views

View on X

Mentioned links

Links and tools from this entry

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • Repository:drowzeys/keys-SM121-Optimized-MiniMax-H3-Nvidia-Sol-Engine-Kijai-SolAttn_Triton-Single-DGX-Spark

    MiniMax-H3 Optimisation for NVIDIA DGX Spark (sm_121)

    An optimised ComfyUI setup for the MiniMax-H3 (33B) audio+video model on a single NVIDIA DGX Spark (GB10, sm_121). It fixes a silent miscompilation issue and accelerates end-to-end rendering by 1.54x.

    30starsPython

    AI & LLMs· Tool

  • Link:NVIDIA

    NVIDIA DGX Spark Developer Hub: Workflows for GB10 Grace Blackwell Superchip

    NVIDIA has launched the DGX Spark Developer Hub, offering playbooks and guides for running AI agent workflows on the NVIDIA GB10 Grace Blackwell Superchip. The hub features quickstarts for tools such as Open-WebUI, ComfyUI, and vLLM alongside multi-Spark networking setups.

    AI & LLMs· Resource

  • Repository:MiaAI-Lab/DeepSeek-v4-Flash-One-DGX-Spark

    Running DeepSeek-v4-Flash on a Single DGX Spark

    An open-source launcher and Docker recipe for running DeepSeek V4 Flash 0731 (EXL3) on a single NVIDIA DGX Spark with 128 GiB unified memory.

    328starsPython

    AI & LLMs· Tool

  • X post:Mia

    Running DeepSeek v4 Flash on a Single NVIDIA DGX Spark

    MiaAI-Lab published an open-source launcher allowing DeepSeek V4 Flash 0731 (EXL3) to run on a single NVIDIA DGX Spark node instead of requiring a multi-node setup.

    71bookmarks9.8kviews

    AI & LLMs· Tool

  • Repository:MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks

    Serving GLM-5.3 Flash EXL3 on Dual NVIDIA DGX Spark Kits

    Mia's AI Lab provides an optimised vLLM deployment overlay for GLM-5.3-Flash using 4 bpw EXL3 quantisation tailored for 2x NVIDIA GB10 systems.

    431starsPython

    AI & LLMs· Tool

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.