cd /Users/z/work/zen/gym
conda create -n zen-gym python=3.10
conda activate zen-gym
pip install -r requirements.txt

# For FlashAttention-2 (recommended)
pip install flash-attn --no-build-isolation

# For Unsloth acceleration
pip install "unsloth[colab-new] @ git+https://github.com/unslothai/unsloth.git"

# Fine-tune zen-nano with LoRA
llamafactory-cli train \
    --stage sft \
    --do_train \
    --model_name_or_path Qwen/Qwen3-0.6B \
    --dataset your_dataset \
    --template qwen3 \
    --finetuning_type lora \
    --lora_target all \
    --output_dir ./zen-nano-lora \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --num_train_epochs 3 \
    --lr_scheduler_type cosine \
    --learning_rate 5e-5 \
    --save_steps 100 \
    --logging_steps 10 \
    --flash_attn fa2 \
    --use_unsloth true

# GRPO training for zen-eco
llamafactory-cli train \
    --stage grpo \
    --do_train \
    --model_name_or_path Qwen/Qwen3-4B \
    --dataset your_preference_dataset \
    --template qwen3 \
    --finetuning_type lora \
    --output_dir ./zen-eco-grpo \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 8 \
    --num_train_epochs 1 \
    --learning_rate 1e-5 \
    --save_steps 50

# QLoRA 4-bit training
llamafactory-cli train \
    --stage sft \
    --do_train \
    --model_name_or_path Qwen/Qwen3-4B \
    --dataset your_dataset \
    --template qwen3 \
    --finetuning_type lora \
    --quantization_bit 4 \
    --output_dir ./zen-eco-qlora \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --num_train_epochs 3 \
    --learning_rate 2e-4

# Launch Zen Gym web interface
llamafactory-cli webui

# Export trained model to GGUF for llama.cpp
llamafactory-cli export \
    --model_name_or_path ./zen-eco-lora \
    --adapter_name_or_path ./zen-eco-lora \
    --template qwen3 \
    --export_dir ./zen-eco-gguf \
    --export_size 4 \
    --export_quantization_bit 4 \
    --export_legacy_format false

--finetuning_type lora \
--lora_rank 64 \
--lora_alpha 32 \
--lora_dropout 0.1 \
--lora_target all

--finetuning_type lora \
--quantization_bit 4 \
--lora_rank 64 \
--lora_alpha 32

--stage grpo \
--finetuning_type lora \
--learning_rate 1e-5

--stage gspo \
--finetuning_type lora \
--learning_rate 1e-5

cd /Users/z/work/zen/zen-musician
# Train with zen-gym
llamafactory-cli train \
    --config /Users/z/work/zen/gym/configs/zen_musician_lora.yaml

cd /Users/z/work/zen/zen-nano
# Train with zen-gym
llamafactory-cli train \
    --config /Users/z/work/zen/gym/configs/zen_nano_lora.yaml

cd /Users/z/work/zen/zen-eco
# Train with zen-gym
llamafactory-cli train \
    --config /Users/z/work/zen/gym/configs/zen_eco_lora.yaml

# Enable WandB logging
export WANDB_PROJECT=zen-models
llamafactory-cli train --report_to wandb --config your_config.yaml

# View TensorBoard
tensorboard --logdir ./output

# Deploy with vLLM
llamafactory-cli api \
    --model_name_or_path ./zen-eco-lora \
    --template qwen3 \
    --infer_backend vllm

# Deploy with SGLang
llamafactory-cli api \
    --model_name_or_path ./zen-eco-lora \
    --template qwen3 \
    --infer_backend sglang

llamafactory-cli chat \
    --model_name_or_path ./zen-eco-lora \
    --template qwen3

# Q4_K_M (recommended for most use cases)
llamafactory-cli export \
    --model_name_or_path ./model \
    --export_dir ./gguf \
    --export_quantization_bit 4

# Q8_0 (higher quality)
llamafactory-cli export \
    --model_name_or_path ./model \
    --export_dir ./gguf \
    --export_quantization_bit 8

# Q2_K (maximum compression)
llamafactory-cli export \
    --model_name_or_path ./model \
    --export_dir ./gguf \
    --export_quantization_bit 2

# Convert to MLX format
python -m mlx_lm.convert \
    --hf-path ./model \
    --mlx-path ./mlx \
    --quantize

# Reduce batch size
--per_device_train_batch_size 1

# Increase gradient accumulation
--gradient_accumulation_steps 32

# Enable gradient checkpointing
--gradient_checkpointing true

# Use QLoRA
--quantization_bit 4

# Enable FlashAttention-2
--flash_attn fa2

# Enable Unsloth
--use_unsloth true

# Enable Liger Kernel
--enable_liger_kernel true

# Lower learning rate
--learning_rate 1e-5

# Add warmup
--warmup_ratio 0.1

# Use cosine scheduler
--lr_scheduler_type cosine