Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

scripts/install_server.sh — idempotent one-shot install for Linux servers. Clones the upstream repo, downloads both checkpoints, pins transformers to 4.49.0, installs MiMo's runtime deps, and builds flash-attn from source if needed. Configurable via env vars (COMFYUI_DIR, MODELS_DIR, REPOS_DIR, GPU_ARCH, MAX_JOBS).

Documentation overhaul based on real-world install pain.
- README now front-loads the three things that will break your install (transformers pin, flash-attn requirement, repo clone) before any step-by-step instructions.
- Each install step gets a copy-pasteable verification command.
- Added a "compatibility matrix" gut-check table.
INSTALL_NOTES.md rewritten as an exhaustive error → fix reference, covering every error message we observed in real installs:
- MiMo ASR repo path is empty
- Did not find src/mimo_audio/
- model path does not exist
- MimoAudio.__init__() got an unexpected keyword argument 'tokenizer_path'
- AttributeError: ... no attribute 'all_tied_weights_keys'
- TypeError: 'NoneType' object is not callable (flash-attn missing)
- flash-attn build failures (missing nvcc, OOM, wrong arch)
Documented the three install strategies for handling the transformers==4.49.0 requirement when it conflicts with other custom nodes (accept, dedicated ComfyUI instance, roll back).
Detailed flash-attn source-build guide with the exact env vars needed for an RTX 3090 host (TORCH_CUDA_ARCH_LIST="8.6", MAX_JOBS=4).

MiMoASRLoader now calls MimoAudio(model_path, tokenizer_path) with positional args. The HF model card shows kwargs, but the actual upstream __init__ signature doesn't accept tokenizer_path= as a keyword — run_mimo_asr.py confirms positional usage. Symptom was TypeError: MimoAudio.__init__() got an unexpected keyword argument 'tokenizer_path'.

MiMoASRLoader node — loads MiMo-V2.5-ASR + MiMo-Audio-Tokenizer with process-wide caching keyed by (model_path, tokenizer_path, repo_path).
MiMoASRTranscribe node — runs MimoAudio.asr_sft on a ComfyUI AUDIO input, writing it to a tempfile first; supports Auto / Chinese / English language tags.
cli/mimo_asr_cli.py — single-file CLI that uploads an audio file, builds the API-format workflow, runs it over WebSocket, and prints + optionally saves the transcript.
Example workflow JSON at example_workflows/mimo_asr_basic.json.
INSTALL_NOTES.md with the compatibility matrix and the transformers / flash-attn workarounds.