FL VoxCPM Train Config

V1 LoRA training settings — what actually moves the needle

By filliptm·Created 7 months ago·Updated 4 months ago· 32

Train Config

◄learning_rate0.0001►

◄lora_rank32►

◄lora_alpha16►

◄lora_dropout0.00►

◄warmup_steps100►

◄grad_accum_steps1►

◄max_batch_tokens8192►

◄sample_rate44100►

◄weight_decay0.01►

◄enable_lm_loratrue►

◄enable_dit_loratrue►

◄enable_proj_lorafalse►

◄lr_schedulercosine►

This node is the settings sheet for training a voice LoRA on VoxCPM V1 (VoxCPM1.5 / 0.5B). It doesn't do anything on its own - it just packages a pile of hyperparameters into a VOXCPM_TRAIN_CONFIG object that the LoRA Trainer consumes. If you're new to training, it's the friendliest possible on-ramp: sensible defaults, tooltips on everything, and only a couple of fields you'll actually touch.

It's part of filliptm/ComfyUI-FL-VoxCPM, wired as Dataset Maker → Train Config → LoRA Trainer.

How it works

Every input is a named hyperparameter, passed straight through into a config dict the trainer reads. It's a data node, not a training node - the moment you'd otherwise be editing a YAML config file by hand is the moment this earns its place. The V2 sibling ships with OpenBMB's official defaults; this one is the V1 flavor, tuned for the 800M / 44.1kHz model.

Inputs that matter

For a beginner, three knobs dominate:

lora_rank (32) - the LoRA's capacity. This is the one the community actually pushes: people report "really impressive results" bumping rank to 128 for voice work. Higher rank = more faithful clone, more VRAM.
learning_rate (0.0001) - the default is the official 1e-4. If your training diverges or the loss goes wild, drop it a notch; if it underfits and sounds nothing like the target, that's usually a dataset problem, not an LR problem.
max_batch_tokens (8192) - the VRAM governor. Lower it if you OOM; the trainer sizes batches around it.

Worth knowing, not worth touching at first: lora_alpha (16, the scaling factor; alpha = rank gives scale 1.0), warmup_steps, grad_accum_steps (gradient accumulation for effective batch size), lr_scheduler (cosine vs constant), and the three LoRA target toggles - enable_lm_lora and enable_dit_lora both default on (Language Model + Diffusion Transformer), enable_proj_lora off. Leave those as-is unless you know why you're changing them.

One subtlety: sample_rate (44100) must match your training audio. VoxCPM1.5 is a 44.1kHz model, so your dataset should be too. If you're ripping audio from video at 48kHz, resample first.

Installation

Same pack as the rest - Manager search "FL VoxCPM", or clone + pip install -r requirements.txt as usual. This node itself has zero extra dependencies; the heavy argbind/datasets requirements are the trainer's problem, not this node's.

Common issues

Config mismatch errors at the trainer - the LoRA Trainer version-checks your config. A V1 config into a V2 model (or vice versa) raises instead of silently training garbage. Use this node only for V1 models and the V2 Train Config for VoxCPM2.
OOM during training - cut max_batch_tokens or lora_rank, not the dataset.
The output looks like a dict - it is. Train Config only has meaning when connected to the LoRA Trainer; it's not a file, it's a value.

CategoryFL/VoxCPM/Training

Inputs (13)

Name	Type	Default	Description
learning_rate	FLOAT	0.00010.000001–0.01	Learning rate for the optimizer.
lora_rank	INT	324–128	Rank (dimension) of the LoRA adapter.
lora_alpha	INT	161–128	Alpha scaling factor for LoRA.
lora_dropout	FLOAT	0.000–0.5	Dropout probability for LoRA layers.
warmup_steps	INT	1000–1000	Number of warmup steps for learning rate scheduler.
grad_accum_steps	INT	11–64	Number of steps to accumulate gradients before updating weights.
max_batch_tokens	INT	81921024–32768	Maximum number of tokens per batch to manage VRAM usage.
sample_rate	INT	4410016000–48000	Sample rate of the training audio.
weight_decay	FLOAT	0.010–0.1	Weight decay for regularization.
enable_lm_lora	BOOLEAN	true	Apply LoRA to the Language Model backbone.
enable_dit_lora	BOOLEAN	true	Apply LoRA to the Diffusion Transformer.
enable_proj_lora	BOOLEAN	false	Apply LoRA to projection layers.
lr_scheduler	COMBO	cosine	LR schedule: 'cosine' decays to 0 over training, 'constant' keeps LR fixed after warmup.

Outputs (1)

Name	Type	Description
Train Config	VOXCPM_TRAIN_CONFIG	—