code/qlora.py

import torch, os, wandb, uuid, json
import bitsandbytes as bnb
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer, BitsAndBytesConfig, TrainerCallback
from peft import prepare_model_for_kbit_training, LoraConfig, get_peft_model
from accelerate import Accelerator
from accelerate.utils import set_seed
from datasets import load_dataset, DatasetDict, Dataset,load_from_disk
from functools import partial

set_seed(42)

accelerator = Accelerator()
run_id = str(uuid.uuid4())
modelpath="microsoft/phi-2"
dataset_name="teknium/OpenHermes-2.5"
lr=0.00002
bs=10            # batch size
bs_eval=16      # batch size for evals
ga_steps=4     # gradient acc. steps
epochs=1
max_length=1024
output_dir=f"out_{run_id}"

# Load model
model = AutoModelForCausalLM.from_pretrained(
    modelpath,    
    device_map={"": accelerator.process_index},
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_quant_type="nf4",
    ),
    torch_dtype=torch.bfloat16,
    # does not work yet
    # attn_implementation="flash_attention_2",          
)

# Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(modelpath, use_fast=False)    # fast tokenizer sometimes ignores the added tokens

# Add tokens <|im_start|> and <|im_end|>, latter is special eos token, 
tokenizer.add_tokens(["<|im_start|>", "<PAD>"])
tokenizer.pad_token = "<PAD>"
tokenizer.add_special_tokens(dict(eos_token="<|im_end|>"))
model.config.eos_token_id = tokenizer.eos_token_id

# Add adapters to model
model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True) 

lora_config = LoraConfig(
    r=32, 
    lora_alpha=32, 
    target_modules = [ "q_proj", "k_proj", "v_proj", "dense" ],
    modules_to_save = ["lm_head", "embed_tokens"],
    lora_dropout=0.1, 
    bias="none", 
    task_type="CAUSAL_LM",
)
model = get_peft_model(model, lora_config)

model.config.use_cache = False

# Print stats
if accelerator.is_main_process:
    model.print_trainable_parameters()

# Load dataset
with accelerator.main_process_first():
    dataset = load_dataset(dataset_name)
    dataset = dataset["train"].train_test_split(test_size=0.1)

    # Format (chatML) and tokenize dataset
    templates= {
        "system": "<|im_start|>system\n{msg}<|im_end|>",
        "human": "<|im_start|>user\n{msg}<|im_end|>",
        "gpt": "<|im_start|>assistant\n{msg}<|im_end|>",
    }
    IGNORE_INDEX=-100

    def tokenize(input, max_length):
        input_ids, attention_mask, labels = [], [], []

        for i,msg in enumerate(input["conversations"]):
            msg_role=msg["from"]
            msg_content=msg["value"]
            isHuman=msg_role=="human"
            if not msg_role in templates: return  # this will break it
            msg_chatml=templates[msg_role].format(msg=msg_content)
            msg_tokenized=tokenizer(msg_chatml, truncation=False, add_special_tokens=False)

            input_ids+=msg_tokenized["input_ids"]
            attention_mask+=msg_tokenized["attention_mask"]
            labels+=[IGNORE_INDEX]*len(msg_tokenized["input_ids"]) if isHuman else msg_tokenized["input_ids"]

        return {
            "input_ids": input_ids[:max_length],
            "attention_mask": attention_mask[:max_length],
            "labels": labels[:max_length],
        }

    dataset_tokenized = dataset.map(
        partial(tokenize, max_length=max_length), 
        batched=False, 
        # num_proc=os.cpu_count()//accelerator.num_processes,    # multithreaded
        num_proc=os.cpu_count(),    # multithreaded
        remove_columns=dataset["train"].column_names  # don't need this anymore, we have tokens from here on
    )

# collate function - to transform list of dictionaries [ {input_ids: [123, ..]}, {.. ] to single batch dictionary { input_ids: [..], labels: [..], attention_mask: [..] }
def collate(elements):
    tokens=[e["input_ids"] for e in elements]
    tokens_maxlen=max([len(t) for t in tokens])

    for i,sample in enumerate(elements):
        input_ids=sample["input_ids"]
        labels=sample["labels"]
        attention_mask=sample["attention_mask"]

        pad_len=tokens_maxlen-len(input_ids)

        input_ids.extend( pad_len * [tokenizer.pad_token_id] )   
        labels.extend( pad_len * [IGNORE_INDEX] )    
        attention_mask.extend( pad_len * [0] ) 

    batch={
        "input_ids": torch.tensor( [e["input_ids"] for e in elements] ),
        "labels": torch.tensor( [e["labels"] for e in elements] ),
        "attention_mask": torch.tensor( [e["attention_mask"] for e in elements] ),
    }

    return batch
 
steps_per_epoch=len(dataset_tokenized["train"])//(accelerator.num_processes*bs*ga_steps)

args = TrainingArguments(
    output_dir=output_dir,
    per_device_train_batch_size=bs,
    per_device_eval_batch_size=bs_eval,
    evaluation_strategy="steps",
    logging_steps=1,
    eval_steps=steps_per_epoch//3,    # 2 evals per epoch
    save_steps=steps_per_epoch//3,     # save once per epoch
    gradient_accumulation_steps=ga_steps,
    num_train_epochs=epochs,
    lr_scheduler_type="constant",
    optim="paged_adamw_32bit",      # val_loss will go nan with paged_adamw_8bit
    learning_rate=lr,
    group_by_length=False,
    bf16=True,        
    ddp_find_unused_parameters=False,
)

trainer = Trainer(
    model=model,
    tokenizer=tokenizer,
    args=args,
    data_collator=collate,
    train_dataset=dataset_tokenized["train"],
    eval_dataset=dataset_tokenized["test"],
)

if accelerator.is_main_process:
    run = wandb.init(
        project="phi2-teknium1",
        name=modelpath+"_"+dataset_name+f"_bs-{bs}_LR-{lr}_GPUs-{accelerator.num_processes}_maxlen-{max_length}_{run_id}",
        config={
            "model_name": modelpath,
            "run_id": run_id,
            "dataset": dataset_name,
            "output_dir": output_dir,
            "lr": lr,
            "max_length": max_length,
            "train_batch_size": bs,
            "validation_batch_size": bs,
            "ga_steps": ga_steps,
            "lora_config": lora_config, 
            "training_args": args,
            "GPUs": accelerator.num_processes,
        }
    )
    run.log_code()

trainer.train()