MOSS 远程推理连接 · T8star-Aix

The MOSS Transcribe Diarize model that this pack wraps is genuinely good - transcribing and telling speakers apart in one local pass - but it wants a real GPU. Twelve gigs of VRAM is the pack's stated floor, and the 10-minute stress test that peaks at 11.6 GB happened on a 24 GB RTX 5090. If your card is smaller, or you'd rather keep the heavy lifting on one beefy box and use ComfyUI as the front end, T8_MOSS_RemoteModelLoader is the escape hatch. It swaps the local model handle for a connection to an OpenAI-compatible /v1/audio/transcriptions endpoint served by SGLang Omni or vLLM, and the rest of the graph keeps working exactly as if a local model were loaded.

What this node is not is the usual sketchy "paste your API key" cloud wrapper. Read the loader's behavior the way you'd read a security-conscious pack, because it's the anti-LLMVISION design: the audio leaves your machine (that's the whole point, and the loader tells you so in its info JSON with "audio_leaves_comfyui_host": true), but the pack does everything it can to keep that honest. You cannot accidentally upload - allow_remote_upload defaults to false and the loader hard-errors if you run without flipping it. Any optional Bearer key is read only from the MOSS_TRANSCRIBE_API_KEY environment variable, never stored in the workflow file. And the URL is sanitized hard: non-loopback HTTP is rejected outright (it refuses to send your audio in plaintext), and URLs can't carry usernames, passwords, query tokens, or fragments. A node that's supposed to phone home is exactly the shape of thing that's been weaponized in this ecosystem once, so this level of belt-and-suspenders is worth calling out.

The four inputs are the whole interface:

• endpoint_url - root address or a full /v1/audio/transcriptions path. The default http://127.0.0.1:8000 is a sensible localhost default; anything not loopback must be HTTPS.
• remote_model - the model ID sent in the request body, defaulting to OpenMOSS-Team/MOSS-Transcribe-Diarize.
• timeout_seconds - request timeout, 10 to 7200, default 300.
• allow_remote_upload - the explicit consent switch. Leave it off until you've actually reviewed the endpoint.

It outputs a T8_MOSS_TRANSCRIBE_MODEL handle plus a JSON info string describing the backend, endpoint, timeout, whether a key is configured, and the audio-leaves-host flag. The loader itself is lazy - it builds the handle and does the network call later, in the transcription node, using multipart WAV upload with response_format=verbose_json and temperature=0.

Installation is the pack's standard dance (see below), but note this node needs no local model - you're skipping download_models.py's heavy step if you only ever use remote. The one workflow file you want is example_workflows/api/03_remote_transcribe.json; it exists to show you exactly what to flip.

Where people get burned: forgetting allow_remote_upload and staring at a "远程音频上传未授权" error, or pointing at an http:// address on another machine and watching it refuse. Also remember that responses are capped at 16 MiB and progress/cancellation during an in-flight request depend entirely on the server implementation - vLLM and SGLang handle that differently, so a long file can feel unresponsive even when it's working.

Installing the pack

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-MOSS-Transcribe-Diarize-T8.git comfyui-MOSS-Transcribe-Diarize-T8

Then, in the ComfyUI Python environment (use ..\..\python_embeded\python.exe on Windows Portable - don't pollute system Python): pip install -r requirements.txt, run scripts/check_transformers.py (the pack needs Transformers >= 5.5.0, < 6 and will reject older versions with a security error), and restart ComfyUI. ComfyUI Manager users can just search comfyui-moss-transcribe-diarize-t8. Note the node UI defaults to Chinese and follows ComfyUI's own language setting - there's an English locale if you want it.

One last tip: because the remote node only makes sense when you actually own the serving box, it's really for people who already run SGLang or vLLM. If you're just trying to get subtitles out of a video and you have 12 GB+ VRAM, the local loaders are the path of least resistance - remote is a power tool, not a starting point.