checkpoint-200/training_config.yaml

aws_output_bucket: s3://sagemaker-us-east-1-107457652907/experiments/falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
train_file: /opt/ml/input/data/train/distant_path_v9.1_fix_no_shuffle.train.0.pkl
test_file: null
model:
  _target_: models.rw.RWForConditionalGenerationFlan.from_pretrained
  pad_token_id: 11
  use_peft: true
  lora_config:
    _recursive_: false
    _target_: models.rw.LoraConfig
    task_type: CAUSAL_LM
    inference_mode: false
    target_modules:
      _target_: models.rw.find_all_linear_names
      bits: 4
    r: 64
    lora_alpha: 16
    lora_dropout: 0.05
  gradient_checkpointing: true
  torch_dtype:
    _target_: general_util.training_utils.return_torch_dtype
    dtype: bfloat16
  quantization_config:
    _target_: transformers.utils.quantization_config.BitsAndBytesConfig
    load_in_4bit: true
    bnb_4bit_compute_dtype:
      _target_: general_util.training_utils.return_torch_dtype
      dtype: bfloat16
    bnb_4bit_use_double_quant: true
    bnb_4bit_quant_type: nf4
  device_map:
    _target_: models.rw.return_single_device_map
  load_in_4bit: true
  max_memory: true
read_tensor_train:
  _target_: data.wiki_entity_path_v9_1_2.convert_examples_into_features_seq2seq
  max_neg_num: 3
  aug_num: 3
  max_seq_length: 512
  shuffle_context: true
  min_rep_num: 5
  geo_p: 0.4
  deduct_ratio: 1.0
  context_ratio: 1.0
  noise_sent_ratio: 0.0
  num_workers: 128
extended_vocab: null
collator:
  _target_: data.collators.wiki_seq2seq_collator.WikiSeq2SeqCollatorWithCausalLM
  max_seq_length: 512
  tokenizer: ${model_name_or_path}
  causal_lm: true
  causal_lm_add_eos: false
  generative_mode: true
num_workers: 4
prefetch_factor: 2
do_preprocess: false
model_name_or_path: /tmp/falcon-40b
pretrain: null
exp_name: falcon.40b.q_lora.merit_v91_v91.seq2seq.v5.0.3aug.w16.adamw.500steps.NA100.0528.aws
exp_notes: null
output_dir: /tmp/${exp_name}
do_train: true
evaluate_during_training: false
do_eval: true
eval_sub_path: checkpoint-*
per_gpu_train_batch_size: 16
per_gpu_eval_batch_size: 8
learning_rate: 0.0005
gradient_accumulation_steps: 16
weight_decay: 0.0
adam_epsilon: 1.0e-06
adam_betas: (0.9, 0.99)
max_grad_norm: 0.3
num_train_epochs: 1
max_steps: -1
warmup_proportion: 0
warmup_steps: 50
optimizer: null
use_nvlamb: null
bit_training: null
logging_steps: 1
save_best: false
save_steps: 100
eval_steps: -1
ddp_eval: true
no_cuda: false
seed: 42
local_rank: 0
fp16: true
fp16_opt_level: O1
fp16_bfloat16: true
prediction_cfg:
  metric: acc
  measure: 1
  best_checkpoint: null
  best_result: null
eval_forward_fn:
  _target_: general_util.evaluator.DiscriminatorForwardFn
post_process: null
compile: false
fairscale_config: null
fsdp_config: null
ds_cfg:
  train_micro_batch_size_per_gpu: ${per_gpu_train_batch_size}
  gradient_accumulation_steps: ${gradient_accumulation_steps}
  optimizer:
    type: AdamW
    params:
      lr: ${learning_rate}
      betas:
      - 0.9
      - 0.999
      eps: ${adam_epsilon}
      weight_decay: ${weight_decay}
  scheduler:
    type: WarmupLR
    params:
      warmup_max_lr: ${learning_rate}
      warmup_num_steps: 50
      warmup_type: linear
  gradient_clipping: ${max_grad_norm}
  bf16:
    enabled: ${fp16}
  zero_optimization:
    stage: 1
    contiguous_gradients: true
    overlap_comm: true
    reduce_scatter: true
    reduce_bucket_size: 500000000.0
    allgather_bucket_size: 500000000.0
  steps_per_print: 1024
with_lightseq: false
summary_helper:
  _target_: general_util.tensorboard_helper.WandbWriter
  batch_index_or_keys: null
  outputs_index_or_keys:
    train/mlm_loss: mlm_loss
n_gpu: 1
device: cuda:0
train_batch_size: 16
eval_batch_size: null
world_size: 16