mm3-backend

Maps HOT-Step's native MiniMax-Music3 C++/GGML backend modules, endpoints, and parity infrastructure.

151|22|Updated Apr 19, 2026
One-click install
npx skills add https://github.com/scragnog/HOT-Step-CPP --skill mm3-backend-scragnog
Or copy as Structured Prompt for Agent
Please help me install this Agent Skill.
Skill: mm3-backend
Source: https://github.com/scragnog/HOT-Step-CPP/tree/main/.claude/skills/mm3-backend
Command: npx skills add https://github.com/scragnog/HOT-Step-CPP --skill mm3-backend-scragnog

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill includes references (resource) components.

What problem does it solve? Working on HOT-Step's MiniMax-Music3 backend requires navigating a large surface of engine modules, server routes, UI gating, model files, and parity fixtures, plus a long list of hard-won debugging traps that are expensive to rediscover. ## Core Features & Use Cases - Architecture map: Locates every MM3 piece — engine modules in engine/src/minimax/, server backend services, /mm3/* endpoints, UI capability gating, and the 5-way split GGUF model files. - Trap list and validation bar: Documents 18+ real debugging traps (tokenizer parity, scheduler sigmas, WAV interleaving, LoRA delta gating) and the forced-replay parity floors every change must meet. - Deep references on demand: Separate reference files cover feature subsystems (sampler plugins, caption composer, lyric timestamps), performance and caching (Plank, AR cache, streaming), and training (LM LoRA, codes export, runtime adapters). - Use Case: When debugging an MM3 generation that sounds wrong, consult the trap list to check known failure modes like the 2x time-stretch WAV bug or single-seed spectral misjudgment before changing code. ## Quick Start Ask the assistant to explain how the MM3 AR planner and flow DiT pipeline connect, or where to look when an MM3 generation produces unexpected audio.

Frequently Asked Questions about mm3-backend

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I debug a MiniMax-Music3 generation that sounds wrong?

Check the skill's trap list first: it documents real failures like the interleaved-vs-planar WAV bug causing 2x slow-motion audio, low step counts going thin rather than dull, and single-seed spectral judgments being meaningless. Verify preprocessing by decoding a target and listening.

How is the MM3 backend structured in HOT-Step?

Engine modules live in engine/src/minimax/ (loader, tokenizer, LM graph, AR loop, depth, condition, DiT, vocoder, pipeline, job queue, server). The Node server has backend services under server/src/services/backends/minimax/, and the UI gates features via capabilities in backendStore and BackendCapabilityGate.

What endpoints does the MM3 engine expose?

The engine serves /mm3/props, /mm3/warm, /mm3/unload, /mm3/synth for production generation, /mm3/job for progress, /mm3/stream for live audio, and /mm3/tokenize-check. Deprecated bring-up endpoints exist for parity work but must not be used in production paths.

Can MM3 models be mixed at different quantization levels?

Yes, the 5-way split (lm, depth, cond, dit, voc) supports per-role quant mixing, such as LM Q8_0 with DiT Q4_K_M. The cond and vocoder modules are never quantized and stay f16.

What is the validation bar for MM3 code changes?

Changes must pass forced-replay parity against fixtures, never sampled-path comparisons. Floors are 0.999 correlation per module versus bf16 dumps, or 0.9999 versus an fp32 CPU rerun; bit-neutral changes must be proven with deterministic seeds.

Does MM3 support LoRA adapter training?

Yes, via ace-train mm3-lm-train for LM LoRA against any installed quant base, with a practical VRAM floor near 10 GB. Runtime adapters load through mm3-lm-adapter.h with live per-group scales, and DiT LoRAs follow the ComfyUI MM3 format.