MOSS 环境诊断与模型释放 · T8star-Aix

The first time your MOSS transcription dies mid-run, the question you'll ask is "what the hell is eating my VRAM?" This node is the answer sheet. It's a pure output node - one input, one JSON report - that tells you exactly what your environment looks like and, when you ask it to, unloads only the models this pack loaded.

It's part of the MOSS Transcribe Diarize pack, a set of ComfyUI V3 nodes that wrap the OpenMOSS MOSS-Transcribe-Diarize model (a ~0.9B audio model that does transcription and speaker diarization in one pass, no API key, all local). The Environment Release node is the pack's diagnostics corner. You tack it onto the end of a workflow and it gives you the report card.

What you actually get

One output: 环境报告 JSON (environment report JSON). Open it in a text node or just read it from the JSON viewer. It contains:

• Python version and platform
• torch version and the installed Transformers version, plus whether that Transformers version meets the pack's security minimum
• whether CUDA is available, the CUDA runtime version, and a per-GPU breakdown: name, total VRAM, BF16 support, and whether it clears the 12GB / 10GB baselines this pack cares about
• the model cache state - what this pack has loaded right now
• the pinned model revision, and the pack's own stated limitations

That per-GPU list is worth a look if you have heterogeneous cards, because the pack resolves BF16 support per target card, not just "the current one."

The one input that matters

The action combo has three options:

• report_only - just produce the report. Default, harmless.
• release_selected - unload one specific MOSS model. Needs a model wired in from the T8_MOSS_ModelLoader; if nothing's connected it raises an error.
• release_all_moss - free everything this pack has cached.

Note the phrasing: this pack's cached models. The design is deliberately scoped. It won't touch other custom nodes' loaded weights, and that's a feature - you never want a "free memory" button nuking the model another half of your graph still needs. If your OOM comes from somewhere else, the report tells you that too, because the cache state is right there in the JSON.

Installing it

Same as the rest of the pack:

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      # needs Transformers >= 5.5.0
python scripts/download_models.py --comfyui-root ..\..

Then restart ComfyUI. Full details live in the pack README; the model lands in ComfyUI/models/moss_transcribe_diarize/, not in custom_nodes.

Gotchas

• It's an output node (is_output_node), so it runs at the end and doesn't feed anything forward. Wire it up as the last thing in a diagnostic workflow, or alongside T8_MOSS_ModelLoader when you want to inspect a load.
• release_selected with no model connected throws. If you're seeing that error, you forgot the wire, not the pack.
• If the report says Transformers is below the 5.5.0 security minimum, that's the pack's loader refusing to run - run scripts/check_transformers.py and, if the rest of your environment is OK with it, pip install -r requirements-transformers-v5.txt to get to 5.15.1.
• Don't treat a high VRAM number as failure. A 10-minute transcribe on a 24GB card peaked around 11.6GB per the author's own tests. This node exists so you can see that number instead of guessing at it.