training.py

import sys
import logging

import datasets
from datasets import load_dataset
import torch
import transformers
from trl import SFTTrainer
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, BitsAndBytesConfig
from typing import Dict, List

logger = logging.getLogger(__name__)


###################
# Hyper-parameters
###################

training_config = {
    "bf16": True,
    "do_eval": False,
    "learning_rate": 1e-04,
    "log_level": "info",
    "logging_steps": 20,
    "logging_strategy": "steps",
    "lr_scheduler_type": "cosine",
    "num_train_epochs": 3,
    "max_steps": -1,
    "output_dir": "./tulu_sft",
    "overwrite_output_dir": True,
    "per_device_eval_batch_size": 4,
    "per_device_train_batch_size": 4,
    "remove_unused_columns": True,
    "save_steps": 1000,
    "save_total_limit": 1,
    "seed": 0,
    "gradient_checkpointing": True,
    "gradient_checkpointing_kwargs":{"use_reentrant": False},
    "gradient_accumulation_steps": 4, # You may use a bigger batch size, and manually normalize the loss with gradient_accumulation_steps in modeling_ddllama.py
    "warmup_ratio": 0.03,
    "ddp_find_unused_parameters": True,
    }
train_conf = TrainingArguments(**training_config)


###############
# Setup logging
###############
logging.basicConfig(
    format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
    datefmt="%Y-%m-%d %H:%M:%S",
    handlers=[logging.StreamHandler(sys.stdout)],
)
log_level = train_conf.get_process_log_level()
logger.setLevel(log_level)
datasets.utils.logging.set_verbosity(log_level)
transformers.utils.logging.set_verbosity(log_level)
transformers.utils.logging.enable_default_handler()
transformers.utils.logging.enable_explicit_format()

# Log on each process a small summary
logger.warning(
    f"Process rank: {train_conf.local_rank}, device: {train_conf.device}, n_gpu: {train_conf.n_gpu}"
    + f" distributed training: {bool(train_conf.local_rank != -1)}, 16-bits training: {train_conf.fp16}"
)
logger.info(f"Training/evaluation parameters {train_conf}")


################
# Model Loading
################

checkpoint_path = "./"
model_kwargs = dict(
    use_cache=False,
    trust_remote_code=True,
    attn_implementation="flash_attention_2",
    torch_dtype=torch.bfloat16,
    device_map=None
)
model = AutoModelForCausalLM.from_pretrained(checkpoint_path, **model_kwargs)
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Meta-Llama-3-8B-Instruct")
tokenizer.model_max_length = 2048
tokenizer.pad_token = "<|reserved_special_token_0|>"
tokenizer.pad_token_id = 128002
tokenizer.padding_side = 'right'


##################
# Data Processing
##################
def apply_chat_template(
    example,
    tokenizer,
):
    messages = example["messages"]
    example["text"] = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=False)
    return example

raw_dataset = load_dataset("allenai/tulu-v2-sft-mixture")
train_dataset = raw_dataset["train"]
column_names = list(train_dataset.features)

processed_dataset = train_dataset.map(
    apply_chat_template,
    fn_kwargs={"tokenizer": tokenizer},
    num_proc=64,
    remove_columns=column_names,
    desc="Applying chat template to train_sft",
)


###########
# Freeze Transformer
###########
for param in model.parameters():
    param.requires_grad = False

for name, param in model.named_parameters():
    if 'router' in name.lower():
        param.requires_grad = True

###########
# Training
###########
trainer = SFTTrainer(
    model=model,
    args=train_conf,
    peft_config=None,
    train_dataset=processed_dataset,
    eval_dataset=None,
    max_seq_length=2048,
    dataset_text_field="text",
    tokenizer=tokenizer,
    packing=True
)

train_result = trainer.train()
metrics = train_result.metrics
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()

# ############
# # Save model
# ############
trainer.save_model(train_conf.output_dir)