Skip to content
Bud's Wiki

Sol-H3: Recursive Self-Improvement for MiniMax-H3 Inference Acceleration on Sol-Engine across Cloud and Edge

TL;DR

Section titled “TL;DR”

Sol-H3 is a full-stack inference pipeline for the 33B open-source MiniMax-H3 video diffusion model that targets both cloud (8×GB200) and edge (single DGX Spark) deployment through the same framework. It combines an algorithmic lever — a cross-resolution two-stage generation scheduler with a learned latent-to-latent mapping that eliminates the VAE decode/re-encode cycle between resolution stages — with a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts under joint latency+numerical-agreement evaluation. Headline: up to 30× E2E speedup, 20% lower memory, and a 5s 1344×768 video with audio at 3.5× real time on 8×GB200, with the same pipeline fully memory-resident on a single DGX Spark in under a minute.

Key claims

Section titled “Key claims”
  • MiniMax-H3’s 33B parameters and multi-step iterative denoising make production deployment latency-bound on cloud GPUs (NVIDIA-GB200) and memory-bound on edge devices (DGX-Spark), motivating a single pipeline that targets both regimes [Abstract].
  • A cross-resolution two-stage generation scheduler exploits the step-wise structure of diffusion: early low-resolution steps establish global layout, later high-resolution steps refine local/perceptual detail [Abstract].
  • A learned latent-to-latent mapping module connects the two resolution stages, eliminating the standard VAE decode → re-encode cycle otherwise required to transfer between resolutions [Abstract].
  • Operator implementation is driven by a Recursive Self-Improvement (RSI) loop that searches over kernel fusions and memory layouts, evaluating candidates jointly on latency and numerical agreement with a reference [Abstract].
  • End-to-end pipeline delivers up to 30× E2E speedup and 20% lower memory versus the unaccelerated H3 reference [Abstract].
  • A 5-second 1344×768 video with audio is generated at 3.5× faster than real time on an 8×GB200 node [Abstract].
  • The same pipeline fits fully memory-resident on a single DGX Spark, producing the same clip in under a minute — an explicit cloud→edge portability claim from one artifact [Abstract].

Method

Section titled “Method”

Sol-H3 attacks two independent bottlenecks. Algorithmically, generation is split into a coarse low-resolution denoising phase and a fine high-resolution refinement phase, connected by a small learned mapping in latent space rather than by the usual “decode to pixels, re-encode with the VAE at the higher resolution” bridge — sidestepping the compute cost of the VAE round-trip that dominates naive multi-resolution cascades. For the operator layer, the paper drops a Recursive Self-Improvement (RSI) loop on top of the standard kernel-tuning problem: candidate kernel fusions and memory layouts are generated and scored on latency together with numerical-agreement (against a reference implementation), and the loop iterates rather than committing to a fixed recipe. The framework is deployed against two hardware envelopes — 8×NVIDIA GB200 for cloud and a single DGX Spark for edge — with the same algorithmic scheduler and RSI-searched kernels.

Results

Section titled “Results”
  • 8×GB200 cloud: 5-second 1344×768 video with audio at 3.5× faster than real time [Abstract].
  • Single DGX Spark edge: same clip fully memory-resident, under one minute end-to-end [Abstract].
  • Up to 30× E2E speedup and 20% lower memory versus the reference stack [Abstract].
  • Per-lever ablation isolating cross-resolution scheduler vs latent mapping vs RSI-discovered kernels is not reported in the abstract available at ingest.

Why it’s interesting

Section titled “Why it’s interesting”

Sol-H3 is the third filed Sol-Engine artifact after Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation (agent-tuned five-lever framework across Cosmos3-Super / LTX-2.3 / SANA-Video) and Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification (the sparse-attention kernel behind Sol-Engine), and the first to target the MiniMax H3 base that has attracted a growing family of independent accelerators — FastVideo FastH3 V1 — 4-Step Sparse Distilled MiniMax H3 for 14x Blackwell Speedup (distillation-first), VideoDeltaNet on MiniMax H3 — Near-Lossless Hybrid Attention for Video Models (architecture retrofit + branch-parallel SP), TaoMate-H3 — Low-Latency Streaming LoRA Distillation of MiniMax H3 for Joint Audio-Video (LoRA + streaming), and Kamimoto — Jev-gated dynamic sparsity for MiniMax H3 video generation (41.7% wall-clock reduction on RTX 4070) (per-layer sparsity routing). Two things distinguish it in Diffusion serving optimization: (a) a cross-resolution two-stage scheduler with a learned latent bridge — a design point closest to Cross-Resolution Distribution Matching for Diffusion Distillation (RMD, which folds cascading into DMD distillation) but here composed with system-level acceleration rather than distillation; (b) a single pipeline that spans cloud GB200 and edge DGX Spark, extending the cloud↔edge portability story that Neodragon: Mobile Video Generation using Diffusion Transformer began on phone NPUs. The RSI-over-kernels framing also sharpens the AI-for-AI Research story: after Sol Video Inference Engine autoresearched which levers to compose, Sol-H3 autoresearches the kernels themselves under a numerical-agreement contract.

See also

Section titled “See also”