Baharani Mix VoxCPM2 LoRA v1

Training config

• r=32, alpha=32, enable_lm=true, enable_dit=true, enable_proj=false
• lr=1e-4, 5000 steps, batch=2 × grad_accum=4 (effective 8)
• max_batch_tokens=8192
• RTX A6000 / cu124 / torch 2.5.1

Usage

Important: pass an explicit — VoxCPM's from_pretrained defaults to r=8 which silently breaks weight loading.

License