name: llama
model:
  pretrained_model_name_or_path: 'mistralai/Mistral-7B-v0.1'
  cache_dir: '/juice/scr/scr110/scr/nlp/data/neo/hub/'
  return_dict: true
  quantization: false
  device_map: auto  # null
  low_cpu_mem_usage: true  # false
  torch_dtype: bfloat16
  attn_implementation: eager  # so we can load attention weights
  rope_theta: 10000.0

attention:
  attention_type: hedgehog_llama
  feature_map: softmax_dim
  feature_map_kwargs:
    input_dim: 128
    eps: 1e-12
    # mlp: null  # to set
    fullspace: true
  layer_idx: null  # to set
  learned_kernel: untied_head
  learned_kernel_kwargs:
    feature_dim: 128
    skip_connection: false
    bias: false
    zero_init: false
  tie_qk_kernels: false
  train_qk: true
  peft:
    method: lora
    kwargs:
      r: 8  # 256
      lora_alpha: 16  # 512
      lora_dropout: 0.1  # 0.05
      target_modules: ['self_attn.q_proj', 'self_attn.k_proj']

dataset:
  name: alpaca_clean
  dataset_config:
    name: alpaca
    path: yahma/alpaca-cleaned
    chunk_size: 1024  # 2048
    concat_data: true
    cache_dir: '/u/scr/nlp/data/alpaca'
  pretrained_model_config:
    pretrained_model_name_or_path: 'mistralai/Mistral-7B-v0.1'
    cache_dir: '/juice/scr/scr110/scr/nlp/data/neo/hub/'
  preprocess_config: null

dataloader:
  batch_size: 1
  num_workers: 2
  drop_last: false
  pin_memory: true

optimizer:
  optim: adamw_torch_fused
  lr: 0.001
  weight_decay: 0.0

lr_scheduler:
  lr_scheduler_type: reduce_lr_on_plateau
  mode: min
  factor: 0.1
  patience: 10
  min_lr: 0.00001

trainer:  # HuggingFace Trainer-like arguments  
  name: distill_attention
  token_reduce: true
  bottom_attention_only: false
  reverse_kl: false
  
  bf16: true
  train_split: train
  val_split: validation
  num_train_epochs: 2
  gradient_accumulation_steps: 8
  seed: 42
  batch_size: 1
  load_best_model_at_end: true
  greater_is_better: false
  metric_for_best_model: distill/eval/loss
  logging_steps: 100
  evaluation_strategy: steps
  max_steps: -1
  eval_steps: 100
  max_eval_batches: null

dataset:
  name: alpaca_clean
  dataset_config:
    name: alpaca
    path: yahma/alpaca-cleaned
    chunk_size: 1024  # 2048
    concat_data: true
    cache_dir: '/u/scr/nlp/data/alpaca'
  pretrained_model_config:
    pretrained_model_name_or_path: 'mistralai/Mistral-7B-v0.1'
    cache_dir: '/juice/scr/scr110/scr/nlp/data/neo/hub/'
  preprocess_config: null    

dataloader:
  batch_size: 1
  num_workers: 2
  drop_last: false
  pin_memory: true

optimizer:
  optim: adamw_torch_fused
  lr: 1e-4
  weight_decay: 0.0

lr_scheduler:
  lr_scheduler_type: reduce_lr_on_plateau
  mode: min
  factor: 0.1
  patience: 10
  min_lr: 0.00001

trainer:  # HuggingFace Trainer-like arguments  
  name: default
  bf16: true
  train_split: train
  val_split: validation
  num_train_epochs: 2
  gradient_accumulation_steps: 8
  seed: 42
  batch_size: 1
  load_best_model_at_end: true
  greater_is_better: false
  metric_for_best_model: eval/loss  # eval/rouge/geometric_mean
  logging_steps: 100
  evaluation_strategy: steps
  max_steps: -1
  eval_steps: 100
  max_eval_batches: null

finetune:
  method: lora
  kwargs:
    r: 8
    lora_alpha: 16  # 32
    lora_dropout: 0  # 0.05
    target_modules: ['q_proj', 'k_proj', 'v_proj', 'o_proj']