2026-04-03 14:23:15 +08:00
|
|
|
|
pretrained_path: /path/to/VoxCPM2/
|
|
|
|
|
|
train_manifest: /path/to/train.jsonl
|
|
|
|
|
|
val_manifest: null
|
2026-04-07 22:57:42 +08:00
|
|
|
|
sample_rate: 16000 # AudioVAE encoder input rate; must match audio_vae_config.sample_rate
|
2026-04-08 18:15:17 +08:00
|
|
|
|
out_sample_rate: 48000 # AudioVAE decoder output rate; used for TensorBoard audio logging
|
2026-04-03 14:23:15 +08:00
|
|
|
|
batch_size: 2
|
|
|
|
|
|
grad_accum_steps: 8 # effective batch size = batch_size × grad_accum_steps = 16
|
|
|
|
|
|
num_workers: 8
|
|
|
|
|
|
num_iters: 1000
|
|
|
|
|
|
log_interval: 10
|
|
|
|
|
|
valid_interval: 500
|
|
|
|
|
|
save_interval: 500
|
|
|
|
|
|
learning_rate: 0.0001
|
|
|
|
|
|
weight_decay: 0.01
|
|
|
|
|
|
warmup_steps: 100
|
|
|
|
|
|
max_steps: 1000
|
|
|
|
|
|
max_batch_tokens: 8192
|
2026-04-08 18:15:17 +08:00
|
|
|
|
max_grad_norm: 1.0 # gradient clipping max norm; 0 = disabled
|
2026-04-03 14:23:15 +08:00
|
|
|
|
save_path: /path/to/checkpoints/finetune_lora
|
|
|
|
|
|
tensorboard: /path/to/logs/finetune_lora
|
|
|
|
|
|
lambdas:
|
|
|
|
|
|
loss/diff: 1.0
|
|
|
|
|
|
loss/stop: 1.0
|
|
|
|
|
|
|
|
|
|
|
|
# LoRA configuration
|
|
|
|
|
|
lora:
|
|
|
|
|
|
enable_lm: true
|
|
|
|
|
|
enable_dit: true
|
|
|
|
|
|
enable_proj: false
|
|
|
|
|
|
r: 32
|
|
|
|
|
|
alpha: 32
|
|
|
|
|
|
dropout: 0.0
|
|
|
|
|
|
|
|
|
|
|
|
# Distribution options (optional)
|
|
|
|
|
|
# - If distribute=false (default): save pretrained_path as base_model in lora_config.json
|
|
|
|
|
|
# - If distribute=true: save hf_model_id as base_model (hf_model_id is required)
|
|
|
|
|
|
# hf_model_id: "openbmb/VoxCPM2"
|
|
|
|
|
|
# distribute: true
|
