FL MiniMax Music 3 Training Run

If you've ever lost a 12-hour training run to a power blip or a Ctrl+C, this node is the apology. It doesn't train anything - it opens the drawer where completed and interrupted training runs are persisted and hands you a handle to one of them. That handle is what you plug into the LoRA Trainer's resume_run input to continue from the latest full checkpoint instead of starting over.

The pack treats training as a durable artifact, not a queue event. Every run gets a directory under ComfyUI/user/fl_minimax_music3/runs/ with its logs, metrics, checkpoints, validation samples, interruption state, and resume metadata. This node is the read side of that system. In the graph it's a simple loader: pick a run from the dropdown, get back a training_run handle plus a human-readable summary.

What it gives you

• run_id - a dropdown populated from the runs actually on disk. Empty? It shows <no training runs found>, which is exactly what it sounds like.
• training_run - the handle the LoRA Trainer's resume_run optional input accepts.
• summary - a STRING snapshot: run id, status, current step vs total, last checkpoint, and whether an adapter has been exported. Useful as a quick status readout without digging through files.

Install

Part of the FL MiniMax Music 3 pack - ComfyUI Manager (search FL MiniMax Music 3) or:

cd ComfyUI/custom_nodes
git clone https://github.com/filliptm/ComfyUI-FL-MiniMaxMusic3.git
cd ComfyUI-FL-MiniMaxMusic3
pip install -r requirements.txt

Restart ComfyUI.

How resume actually behaves

Resume isn't a free-for-all, and that's deliberate. The trainer validates the new inputs against the run's original spec before continuing: you can't resume a run with a different dataset identity, a different LoRA rank/precision, or a different output name. Change any of those and it refuses. It will also refuse to resume a run with no full checkpoint on disk - the interruption has to have happened at a checkpoint boundary for clean recovery. And a completed run can't be resumed at all (it already exported its adapter; start a new run if you want to keep going).

One thing to internalize: this node is optional in a happy path. If training completes cleanly, the trainer hands you the lora_path directly and you never need the resume machinery. Where this node earns its keep is the failure case - interrupted training, or a run you want to keep extending in stages. That "train a little, check, resume" loop is exactly what the pack's durable-run design is for, and it's the workflow you'll actually want once a real multi-thousand-step training job is on the line.