FireRedAudio 运行时控制 · T8star-Aix

Every generation node in this pack talks to a separate Python worker over local RPC - a deliberately isolated process running its own torch 2.8 / Transformers 5.8 environment. That isolation is what keeps FireRedAudio from nuking your ComfyUI environment, but it also means there's a process out there you can't see in the normal ComfyUI UI. RuntimeControl is the dashboard for that process: check whether the worker is alive, look at the reference-condition cache, cancel a stuck job, unload the model to free VRAM, or stop the managed worker entirely. One combo box, one status string, everything else is a button press.

The actions

• status (default) - returns manager + worker health as JSON: is the worker up, what model is loaded, real GPU and free-VRAM state. This is the first thing to run when a generation node misbehaves, because it tells you whether the worker even started.
• cache_status - audits the CPU LRU cache that reuses the RedAE/Patch Encoder condition for repeated reference clips (introduced in v0.7). Hits and current memory footprint are reported here, which is how you know the "reuse the reference, don't re-encode" optimization is actually working.
• clear_audio_cache - clears that audio-condition cache. Do this when you've changed a reference file in place and want to be sure the next run re-encodes instead of serving a stale condition.
• cancel_current - cancels the in-flight task in the worker. This is the manual interrupt: ComfyUI's own cancel usually propagates, but when it doesn't (or you queued a batch and want to stop the worker's current job), this is the lever.
• unload_model - releases the model from VRAM while keeping the worker process alive. The one you reach for when a generation finishes and you need the memory back for something else without paying worker restart costs.
• stop_worker - kills the managed worker process entirely. The worker restarts on the next model-using node, so it's not destructive - it's the "something is wedged, start fresh" option, or a shutdown step at the end of a session.

All of them need the model input (the runtime handle from the model loader), and all return a single 运行时状态 string. Wire it to a text node if you want to see it on the canvas, or just read it in the node output.

Where it fits

Tight-VRAM users live in this node. The pack's default memory mode is auto, which chooses sequential offloading when free VRAM drops below ~36 GiB - meaning the model trades between GPU and RAM over a run. When a run is done, unload_model is how you get that memory back before ComfyUI starts the next unrelated job. It's also the natural partner to the 模型/隔离运行时's release_after_run toggle: that one automates unload-per-run; this one lets you decide when.

Installing

Same pack install (ComfyUI Manager search comfyui-fireredaudio-T8, or clone + python scripts\setup_runtime.py). The one thing to internalize: this node can't tell you anything until the runtime is actually set up - if you skipped setup_runtime.py, status reports a worker that isn't there, which is your earliest possible diagnosis of "the pack isn't ready yet."