training.sh

#!/bin/bash
set -eux

export PIP_NO_CACHE_DIR=1
export CUDA_VISIBLE_DEVICES=0

export HF_HOME=/workspace/.cache/huggingface
export HF_DATASETS_CACHE=/workspace/.cache/huggingface/datasets
export HF_HUB_DOWNLOAD_TIMEOUT=600
export HF_HUB_DOWNLOAD_RETRY=10

export TORCH_HOME=/workspace/.cache/torch
export XDG_CACHE_HOME=/workspace/.cache

export WANDB_DIR=/workspace/wandb
export PIP_CACHE_DIR=/workspace/.cache/pip


export HUGGINGFACE_HUB_TOKEN=$HF_TOKEN

rm -rf /tmp/*

echo "📦 Editable install"
apt-get update && apt-get install -y git


echo "📦 Installing Python deps (NO torch)"
pip install --no-cache-dir -r requirements.txt

echo "📦 Editable install"
pip install --no-cache-dir -e  .

wandb login 4549b8694e56e01e56c6e24350a275ebbe271e6e
export WANDB_PROJECT="vibevoice-finetune-baseten"
echo "🚀 Starting VibeVoice fine-tuning"

python src/finetune_vibevoice_lora.py \
  --model_name_or_path vibevoice/VibeVoice-7B \
  --processor_name_or_path src/vibevoice/processor \
  --dataset_name PharynxAI/merged_multilingual_tts_6k_each \
  --text_column_name text \
  --audio_column_name audio \
  --output_dir /workspace/output/lora_hi_only_v1 \
  --per_device_train_batch_size 1 \
  --gradient_accumulation_steps 1 \
  --learning_rate 2.5e-5 \
  --max_steps 30000 \
  --logging_steps 10 \
  --save_steps 1000 \
  --save_total_limit 3 \
  --eval_split_size 0.00 \
  --remove_unused_columns False \
  --do_train \
  --bf16 True \
  --gradient_clipping \
  --gradient_checkpointing True \
  --dataloader_pin_memory True \
  --dataloader_num_workers 4 \
  --ddpm_batch_mul 1 \
  --diffusion_loss_weight 1.0 \
  --ce_loss_weight 0.02 \
  --voice_prompt_drop_rate 0.2 \
  --lora_target_modules q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj \
  --lr_scheduler_type cosine \
  --warmup_ratio 0.03 \
  --max_grad_norm 0.8 \
  --optim adamw_torch \
  --push_to_hub True \
  --hub_model_id Diksha2001/vibevoice-multiligual-6k-v1 \
  --hub_strategy every_save