高级训练配置

AdvancedTrainConfig is the "I know what I'm doing" drawer of this pack. Everything on it is optional - the node can sit with every field at its default and it does nothing. You wire its ADVANCED_TRAIN_CONFIG output into the advanced_config input on GeneralConfig only when the standard settings on that hub node aren't giving you what you want. If you're a first-timer, skip it entirely; it's there for the second or third run, when you actually know which knob is hurting you.

What the inputs actually do

Almost every field here is a zero-means-off control, so the defaults are deliberately inert. The ones worth knowing:

• max_steps (default 0) - caps training at a step count instead of running the full epochs from GeneralConfig. 0 means "ignore this, use epochs."
• lr_scheduler (default constant) - the learning-rate schedule. linear is the other option; if you switch to it, make sure you understand warmup (which GeneralConfig handles separately) or you can torch your loss curve.
• checkpoint_every_n_epochs (default 0) - the tooltip says it plainly: the author recommends enabling this, "otherwise you may lose part of your training progress." Diffusion-Pipe's own checkpointing writes a run directory you can resume from, and an occasional epoch checkpoint is cheap insurance. This one's easy to sleep on, and you'll regret it on a 20-hour run that OOMs at hour 19.
• compile (default off) - wraps the model in torch.compile. The author's own caveat: "not tested on all models." Try it on your setup with a short run first; when it works it's a free speedup, when it doesn't it's a confusing crash.
• map_num_proc (default 0) - how many parallel processes to use when caching the dataset. If you have many cores (or multiple GPUs), raising this can noticeably cut the one-time dataset-caching wall. 0 = let diffusion-pipe pick.
• partition_split - a manual layer-partition string like "10,20" meaning layers 0–9 on GPU 0, 10–19 on GPU 1, the rest on GPU 2. This is deep pipeline-parallel territory; most people never touch it.
• reentrant_activation_checkpointing - enable when pipeline_stages > 1 in GeneralConfig (i.e. you're splitting the model across GPUs). Off by default because it costs a bit of speed.

There are also a few niche ones: pseudo_huber_c (a robust-loss constant, only applies to models using the default loss function), save_every_n_steps / eval_every_n_steps (step-based alternatives to the epoch-based saves and evals on GeneralConfig), x_axis_examples (plot training progress against samples seen instead of steps in TensorBoard/WandB), force_constant_lr, and steps_per_print (log frequency).

How it fits in

Mechanically it's the same trick as every config node here: it serializes your choices into a dict and hands them to GeneralConfig, which folds everything into the TOML the real diffusion-pipe trainer consumes. The node itself never launches anything - the Train node does that. So nothing here is dangerous by itself; worst case you make a training run slower or worse, then revert.

Getting the pack installed

Same story as every node in this pack - it's a Linux/WSL2-only custom node, installed by cloning with submodules and running the requirements file (deepspeed is the heavyweight). The quick version:

cd ComfyUI/custom_nodes/
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
git submodule update
pip install -r Diffusion_pipe_in_ComfyUI/requirements.txt

Then restart ComfyUI and import the example workflow from example_workflows/ to see how AdvancedTrainConfig plugs into GeneralConfig.

Common gotchas

Because this node is inert by default, the #1 "problem" is people expecting it to do something and seeing no change - that's by design. When it does misbehave, it's usually compile on an unsupported model (turn it off and rerun) or max_steps fighting your epochs setting (whichever comes first wins, so set one of them to 0). And remember the WSL2 path convention from the pack: Windows drive-letter paths like F:/..., not /mnt/f/....