MOSS 转写说话人模型加载器 · T8star-Aix

Every workflow in the MOSS Transcribe Diarize pack starts here. T8_MOSS_ModelLoader finds the pinned MOSS-Transcribe-Diarize model on disk, verifies it's intact, and hands a model handle to the transcribe nodes. It's the pack's gatekeeper - and unlike most model loaders you've used, this one is lazy: the weights only get pulled into VRAM when transcription actually starts. The loader node itself is nearly free to run, which matters when you're chaining a loader into several transcribe paths.

The model is the OpenMOSS-Team MOSS-Transcribe-Diarize checkpoint (~0.9B params, roughly a 1.8GB safetensors file), pinned to a fixed Hugging Face revision. The pack is an independent integration - it's not an official OpenMOSS release, and the README says so plainly.

The inputs that matter

• model_name - a dropdown of models found under ComfyUI/models/moss_transcribe_diarize (plus any paths registered in extra_model_paths.yaml). If you haven't downloaded the model yet, it shows [未找到] 请将模型放入 models/moss_transcribe_diarize. That's your cue to run the download script, not to panic.
• device and precision - auto does the sensible thing: BF16 on a CUDA card that supports it, FP16 otherwise on CUDA, FP32 on CPU. You can force bfloat16/float16/float32 if you're experimenting, but auto is fine for 99% of cases.
• memory_policy - keep (resident, fastest), release_under_pressure (frees when free VRAM drops below 2GB or 20%), or release_after_run (most frugal). This is the knob to touch if you're sharing a card with other nodes.
• attention_implementation - auto tries FlashAttention-2, then SDPA, then eager, in that order, and records each skip or failure. The pinned upstream code fixes a Transformers path that could silently pick eager attention and blow up VRAM quadratically on long audio. If you explicitly pick a backend and it fails, it errors loudly instead of quietly swapping - no silent fallback, by design.
• verify_hashes - the loader always uses full-file SHA-256 for model identity; flipping this on also compares every file against the pinned manifest's expected digests. Slower, but it's the difference between "probably fine" and "proven intact."
• custom_model_path - absolute path for models living outside the standard tree.

Outputs: MOSS 模型 (the handle you wire into T8_MOSS_TranscribeDiarize or T8_MOSS_SmartLongAudio) and 模型信息 (a string with resolved device, actual dtype, memory policy, attention backend, and the revision - useful as a sanity check via a ShowText node).

Installing and getting the model

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-MOSS-Transcribe-Diarize-T8.git comfyui-MOSS-Transcribe-Diarize-T8
cd comfyui-MOSS-Transcribe-Diarize-T8
pip install -r requirements.txt          # Windows Portable: ..\..\python_embeded\python.exe -m pip install -r requirements.txt
python scripts/check_transformers.py      # enforces Transformers >= 5.5.0, < 6
python scripts/download_models.py --comfyui-root ..\..

The download script prints its target directory before starting, and refuses to guess if it can't find your ComfyUI root - it won't silently drop weights into custom_nodes. Requirements install av, librosa, soundfile, soxr, webrtcvad-wheels and friends, but deliberately do not touch torch or Transformers.

Where people get burned

• The model list is empty. You cloned the pack but never ran download_models.py. The script is step three for a reason.
• The loader refuses to run on an old Transformers. This pack enforces a 5.5.0 security minimum and rejects anything older; run scripts/check_transformers.py to see exactly where you stand. Fix with pip install -r requirements-transformers-v5.txt (5.15.1) once you've checked it won't break your other nodes.
• VRAM pressure mid-workflow. The loader is lazy, but once transcription starts, a 10-minute clip can peak around 11.6GB. The 12GB baseline in the model-info string is the author telling you the truth: 8–10GB cards are only a short-audio compatibility tier. If you're squeezed, release_under_pressure is your friend.