isft_mistral.py

from datasets import load_dataset
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

import os


base_model_id = "mistralai/Mistral-7B-Instruct-v0.1" 

WORK = "vn_v2"
new_model_id = f"kmichiru/Nikaido-7B-mistral-instruct-v0.3-{WORK}"

# DSET = {
#     "train": f"dataset_iroseka/{WORK}_dataset.jsonl",
#     "eval": f"dataset_iroseka/{WORK}_validations.jsonl"
# }

DSET = {
    "train": f"dataset_iroseka/{WORK}_train.jsonl",
    "eval": f"dataset_iroseka/{WORK}_val.jsonl"
}


dataset = load_dataset("json", data_files=DSET)
# model = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype=torch.bfloat16)
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
# max_length = 1024
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

def dialogue(role, content):
    return {
        "role": role,
        "content": content
    }

def format_chat_history(example):
    user_msgs = []
    for msg in example["messages"]:
        if msg["role"] == "user":
            user_msgs.append(msg["content"])
    messages = [
        dialogue("user", "\n".join(user_msgs)), # join user messages together
        example["messages"][-1], # the last message is the bot's response
    ]
    encodeds = tokenizer.apply_chat_template(messages, tokenize=False)
    return encodeds

def prep_speaker(msg: str):
    msg = msg.replace("\u3000", " ") # replace full-width spaces
    speaker, content = msg.split(":", 1)
    speaker = speaker.strip()
    content = content.strip()
    if len(speaker) == 0:
        speaker = "傍白"

    return f"{speaker}: {content}"
    

def format_chat_history_v2(example):
    user_msg = []
    user_msg.append("<s>")
    for msg in example["messages"]:
        # [INST] What is your favourite condiment? [/INST]
        if msg["role"] != "system":
            user_msg.append(f"[INST] {prep_speaker(msg['content'])} [/INST]")
    # user_msg.append("</s>")
    return " ".join(user_msg)

# def format_chat_history_v2(example):
#     user_msgs = []
#     for msg in example["messages"]:
#         if msg["role"] == "user":
#             user_msgs.append(msg["content"])
#     messages = [
#         dialogue("user", "\n".join(user_msgs)), # join user messages together
#         example["messages"][-1], # the last message is the bot's response
#     ]
#     encodeds = tokenizer.apply_chat_template(messages, tokenize=False)
#     return encodeds

print(format_chat_history_v2(dataset['train'][0]))

def generate_and_tokenize_prompt(prompt, max_length=2048):
    result = tokenizer(
        format_chat_history_v2(prompt),
        truncation=True,
        max_length=max_length,
        padding="max_length",
    )
    result["labels"] = result["input_ids"]
    return result

tokenized_dataset = dataset.map(generate_and_tokenize_prompt)
print(tokenized_dataset['train'][0])

# # stats data length
# def plot_data_lengths(tokenized_dataset):
#     lengths = []
#     for split in tokenized_dataset:
#         lengths += [len(x['input_ids']) for x in tokenized_dataset[split]]
#     print(f"Max length: {max(lengths)}")
#     print(f"Min length: {min(lengths)}")
#     print(f"Mean length: {sum(lengths)/len(lengths)}")
#     print(f"Median length: {sorted(lengths)[len(lengths)//2]}")

# plot_data_lengths(tokenized_dataset)
print(tokenized_dataset['train'][0])

#Adding the adapters in the layers
from peft import LoraConfig, get_peft_model
def print_trainable_parameters(model):
    """
    Prints the number of trainable parameters in the model.
    """
    trainable_params = 0
    all_param = 0
    for _, param in model.named_parameters():
        all_param += param.numel()
        if param.requires_grad:
            trainable_params += param.numel()
    print(
        f"trainable params: {trainable_params:,} || all params: {all_param:,} || trainable%: {100 * trainable_params / all_param}"
    )
model = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype=torch.bfloat16)
# model = prepare_model_for_kbit_training(model)
peft_config = LoraConfig(
        r=64,
        lora_alpha=64,
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj","gate_proj", "up_proj", "down_proj"]
    )
model = get_peft_model(model, peft_config)
print_trainable_parameters(model)
print(model)

import wandb, os
# wandb.login()

wandb_project = "NikaidoLM"
if len(wandb_project) > 0:
    os.environ["WANDB_PROJECT"] = wandb_project

import transformers
from datetime import datetime

project = wandb_project
base_model_name = "mistral"
run_name = base_model_name + "-" + project
output_name = f"{run_name}-{WORK}-{datetime.now().strftime('%Y-%m-%d-%H-%M')}"
output_dir = "/scratch/generalvision/mowentao/mistral-out/" + output_name

trainer = transformers.Trainer(
    model=model,
    train_dataset=tokenized_dataset["train"],
    eval_dataset=tokenized_dataset["eval"],
    args=transformers.TrainingArguments(
        output_dir=output_dir,
        warmup_steps=500,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=2,
        num_train_epochs=3,
        weight_decay=5e-4,
        # max_steps=10_000,
        learning_rate=1e-4, # Want a small lr for finetuning
        bf16=True,
        optim="paged_adamw_32bit",
        logging_steps=100,              # When to start reporting loss
        logging_dir=output_dir,        # Directory for storing logs
        save_strategy="steps",       # Save the model checkpoint every logging step
        save_steps=500,                # Save checkpoints every 50 steps
        evaluation_strategy="steps", # Evaluate the model every logging step
        eval_steps=100,               # Evaluate and save checkpoints every 50 steps
        do_eval=True,                # Perform evaluation at the end of training
        report_to="wandb",           # Comment this out if you don't want to use weights & baises
        run_name=output_name,         # Name of the W&B run (optional)
        lr_scheduler_type="cosine",
    ),
    data_collator=transformers.DataCollatorForLanguageModeling(tokenizer, mlm=False),
)

model.config.use_cache = False  # silence the warnings. Please re-enable for inference!
trainer.train()
trainer.model.save_pretrained(new_model_id)
wandb.finish()