FireRedAudio 模型/隔离运行时 · T8star-Aix

Every FireRedAudio node in this pack takes a model input, and this is the node that produces it. It's where the pack's whole architecture becomes visible: it finds your model files, verifies them, and connects you to an isolated Python 3.10 / torch 2.8 / Transformers 5.8 worker so that FireRedAudio's dependency stack never touches your ComfyUI Python. That's the single most important thing to understand about this pack, because audio in ComfyUI has historically been a Transformers-version minefield - one pack's requirement upgrade breaks three other nodes. This one refuses to play that game.

What it does

First, model_name is a dropdown of discovered models (scanning ComfyUI/models/TTS/FireRedAudio/). If it says the default "no model found" string, run the downloader first:

cd ComfyUI\custom_nodes\comfyui-fireredaudio-T8
python scripts\download_models.py --profile full --variant int8-wo-safe-v1

--profile matters: full includes the RedAE decoder needed for generation and editing; lite is just the main model, enough for ASR/understanding and a lot less disk. int8-wo-safe-v1 is the recommended quantized variant - INT8 on the Qwen blocks, much smaller footprint, and the README is clear it's the stable default. GGUF is explicitly not supported.

The other inputs are the knobs you'll revisit:

• device - auto picks the first usable NVIDIA GPU.
• memory_mode - auto is smart: it reads the model's VRAM threshold and your actual free memory, and picks sequential offloading if there's less than ~36 GiB free, rather than trusting the card's sticker. On a 24 GB card you'll mostly live in sequential.
• acceleration_mode - auto_safe (default) uses a precompiled FlashAttention 2 wheel. off is the SDPA baseline for troubleshooting. DeepSpeed / FLA+Liger / torch_compile are experimental single-GPU modes; failures fall back loudly rather than silently.
• profile (full/lite) - validation profile.
• worker_mode - managed (the pack spawns and owns the worker) or external (point at a worker_url + worker_token from a desktop-integration worker you started yourself). External mode validates that URL and token are present.
• verify_hashes (off) - full SHA-256 check; slow, but the only way to be sure your download isn't corrupt.

Outputs: the runtime model handle (wire it everywhere) and model_info with root path, revisions, and what was verified.

The real install

Manager install (search comfyui-fireredaudio-T8) or:

cd ComfyUI\custom_nodes
git clone https://github.com/T8mars/comfyui-fireredaudio-T8.git
cd comfyui-fireredaudio-T8
python scripts\setup_runtime.py

setup_runtime.py builds the isolated venv with pinned uv - it's not optional, and Manager installs won't run it for you. The pack's own requirements.txt is deliberately empty; everything heavy lives in the worker.

Gotchas

Expect a cold start of a minute or more on first load - that's the model loading into the worker, not a hang. Don't touch requirements.txt or pip-install anything into ComfyUI hoping to speed it up; you'll only break the isolation that's the point. And the auto_safe default is opinionated for a reason - don't switch to DeepSpeed because a forum said it's faster; run the AccelerationBenchmark node and let your card vote. This loader is the least flashy node in the pack and the one everything depends on; treat its defaults as good advice.