Experiments set I: baselines

Tensorboard: https://huggingface.co/bigscience/dechonk-logs-1/tensorboard

Run 0: continue training 6b3

Using bloom-6b3 at 156K steps Model version: https://huggingface.co/bigscience/bloom-6b3/commit/7e0ac14c5fcba85ad009953ab243eef4624f1503 Re-warming-up the learning rate to value 8e-4 (the last recorded LR in https://huggingface.co/bigscience/tr11f-6B3-logs/tensorboard ) Training params:

deepspeed --num_gpus 8 ./run_clm.py --do_train --do_eval \
    --model_name $INITIAL_MODEL_PATH --tokenizer_name $INITIAL_MODEL_PATH \
    --dataset_name $DATASET_NAME_OR_PATH --dataset_config_name $DATASET_CONFIG_NAME --run_name $RUN_NAME \
    --block_size 2048 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 16 \
    --learning_rate 0.00008 --max_grad_norm 1.0 --lr_scheduler_type cosine --max_steps 200000 --warmup_steps 1000 \
    --adam_epsilon 1e-8 --weight_decay 0.1 --adam_beta1 0.9 --adam_beta2 0.95 --fp16=True --seed 42 \
    --cache_dir $INPUT_PATH/data/cache --output_dir $SNAPSHOT_PATH --overwrite_output_dir=True \
    --logging_dir $LOGS_PATH --report_to tensorboard --logging_first_step --logging_steps 10 \
    --evaluation_strategy steps --eval_steps 100 --prediction_loss_only --eval_subset_size 512 \
    --save_steps 100 --save_total_limit 1 --dataloader_num_workers 4 --deepspeed ds_config.json

Run 1: train 6b3 from scratch

Why: to get an early checkpoint with which we can test Iz's proposal. Model: same as above, but re-initialize weights using HF BLOOM initializer.

Training params:

deepspeed --num_gpus 8 ./run_clm.py --do_train --do_eval \
    --config_name $MODEL_CONFIG_PATH --tokenizer_name $TOKENIZER_PATH \
    --dataset_name $DATASET_NAME_OR_PATH --dataset_config_name $DATASET_CONFIG_NAME --run_name $RUN_NAME \
    --block_size 2048 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 16 \
    --learning_rate 0.00012 --max_grad_norm 1.0 --lr_scheduler_type cosine --max_steps 350000 --warmup_steps 1000 \
    --adam_epsilon 1e-8 --weight_decay 0.1 --adam_beta1 0.9 --adam_beta2 0.95 --fp16=True --seed 42 \
    --cache_dir $INPUT_PATH/data/cache --output_dir $SNAPSHOT_PATH --overwrite_output_dir=True \
    --logging_dir $LOGS_PATH --report_to tensorboard --logging_first_step --logging_steps 10 \
    --evaluation_strategy steps --eval_steps 100 --prediction_loss_only --eval_subset_size 512 \
    --save_steps 100 --save_total_limit 1 --dataloader_num_workers 4 --deepspeed ds_config.json

Run 2: train student-sized model from scratch

model_config.json:

{
  "apply_residual_connection_post_layernorm": false,
  "attention_dropout": 0.0,
  "attention_softmax_in_fp32": true,
  "bias_dropout_fusion": true,
  "bos_token_id": 1,
  "dtype": "float16",
  "eos_token_id": 2,
  "pad_token_id": 3,
  "unk_token_id": 0,
  "hidden_dropout": 0.0,
  "initializer_range": 0.02,
  "layer_norm_epsilon": 1e-05,
  "masked_softmax_fusion": true,
  "model_type": "bloom",
  "n_embed": 2048,
  "n_inner": null,
  "n_layer": 30,
  "num_attention_heads": 16,
  "offset_alibi": 100,
  "pretraining_tp": 4,
  "seq_length": 2048,
  "skip_bias_add": true,
  "skip_bias_add_qkv": false,
  "transformers_version": "4.20.0.dev0",
  "use_cache": false,
  "vocab_size": 250880
}

Training params:

deepspeed --num_gpus 8 ./run_clm.py --do_train --do_eval \
    --config_name $MODEL_CONFIG_PATH --tokenizer_name $TOKENIZER_PATH \
    --dataset_name $DATASET_NAME_OR_PATH --dataset_config_name $DATASET_CONFIG_NAME --run_name $RUN_NAME \
    --block_size 2048 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 16 \
    --learning_rate 0.0002 --max_grad_norm 1.0 --lr_scheduler_type cosine --max_steps 350000 --warmup_steps 1000 \
    --adam_epsilon 1e-8 --weight_decay 0.1 --adam_beta1 0.9 --adam_beta2 0.95 --fp16=True --seed 42 \
    --cache_dir $INPUT_PATH/data/cache --output_dir $SNAPSHOT_PATH --overwrite_output_dir=True \
    --logging_dir $LOGS_PATH --report_to tensorboard --logging_first_step --logging_steps 10 \
    --evaluation_strategy steps --eval_steps 100 --prediction_loss_only --eval_subset_size 512 \
    --save_steps 100 --save_total_limit 5 --dataloader_num_workers 4 --deepspeed ds_config.json