mamba-sft-qlora.yaml

# Model arguments
model_name_or_path: state-spaces/mamba-2.8b-hf
model_revision: main
torch_dtype: float32

# LoRA arguments
load_in_4bit: true
use_peft: true
lora_r: 16
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules:
- x_proj
- embeddings
- in_proj
- out_proj

# Data training arguments
chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}"
dataset_mixer:
  HuggingFaceH4/ultrachat_200k: 1.0
dataset_splits:
- train_sft
- test_sft
preprocessing_num_workers: 12

# SFT trainer config
# set per_device_train_batch_size such that the total batchsize for all used GPUs = 64 for SFT and 32 for DPO 
bf16: false
do_eval: false
evaluation_strategy: epoch
gradient_accumulation_steps: 2
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
hub_model_id: mamba-2.8b-sft-qlora
hub_strategy: every_save
learning_rate: 2.0e-03
log_level: info
logging_steps: 10  
logging_strategy: steps
lr_scheduler_type: cosine
max_seq_length: 2048
max_steps: -1
num_train_epochs: 1
optim: adamw_torch
adam_beta1: 0.9
adam_beta2: 0.95
weight_decay: 0.1
max_grad_norm: 1
output_dir: /mnt/fast/nobackup/users/nt00601/mamba-alignment-handbook/data/mamba-2.8b-sft-qlora
overwrite_output_dir: true
per_device_eval_batch_size: 8
per_device_train_batch_size: 64
push_to_hub: true
report_to:
- tensorboard
save_strategy: "steps"
save_steps: 100
save_total_limit: 1
seed: 42
warmup_ratio: 0.1