Finetune_Llama3.py

from geobleu.Report import report_geobleu_dtw_gpt
from unsloth import FastLanguageModel
import torch
import wandb
import os
os.environ["WANDB_PROJECT"] = "HuMob2024cityA"
os.environ["WANDB_LOG_MODEL"] = "checkpoint"
os.environ["WANDB_MODE"] = "offline"


max_seq_length = 50000 # Choose any! We auto support RoPE Scaling internally!
dtype = None # None for auto detection. Float16 for Tesla T4, V100, Bfloat16 for Ampere+
load_in_4bit = True # Use 4bit quantization to reduce memory usage. Can be False.

# 4bit pre quantized models we support for 4x faster downloading + no OOMs.
fourbit_models = [
    "unsloth/mistral-7b-v0.3-bnb-4bit",      # New Mistral v3 2x faster!
    "unsloth/mistral-7b-instruct-v0.3-bnb-4bit",
    "unsloth/llama-3-8b-bnb-4bit",           # Llama-3 15 trillion tokens model 2x faster!
    "unsloth/llama-3-8b-Instruct-bnb-4bit",
    "unsloth/llama-3-70b-bnb-4bit",
    "unsloth/Phi-3-mini-4k-instruct",        # Phi-3 2x faster!
    "unsloth/Phi-3-medium-4k-instruct",
    "unsloth/mistral-7b-bnb-4bit",
    "unsloth/gemma-7b-bnb-4bit",             # Gemma 2.2x faster!
] # More models at https://huggingface.co/unsloth

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/llama-3-8b-Instruct-bnb-4bit", # Choose ANY! eg teknium/OpenHermes-2.5-Mistral-7B
    max_seq_length = max_seq_length,
    dtype = dtype,
    load_in_4bit = load_in_4bit,
    # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
)

"""We now add LoRA adapters so we only need to update 1 to 10% of all parameters!"""

model = FastLanguageModel.get_peft_model(
    model,
    r = 16, # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 16,
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = "unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,  # We support rank stabilized LoRA
    loftq_config = None, # And LoftQ
)

import json
from datasets import Dataset
from unsloth.chat_templates import get_chat_template

def load_custom_dataset(file_path):
    with open(file_path, 'r', encoding='utf-8') as f:
        data = json.load(f)
    
    # Create a dataset structure suitable for the datasets library
    formatted_data = {
        "conversations": [
            convo["messages"] for convo in data
        ]
    }
    
    # Convert the dataset into the Hugging Face datasets format
    dataset = Dataset.from_dict(formatted_data)
    return dataset

def format_conversations(conversation):
    formatted_convo = []
    for message in conversation:
        formatted_message = {
            "from": message["role"],
            "value": message["content"]
        }
        formatted_convo.append(formatted_message)
    return formatted_convo

def formatting_prompts_func(examples):
    convos = examples["conversations"]
    texts = []
    for convo in convos:
        formatted_convo = format_conversations(convo)
        try:
            text = tokenizer.apply_chat_template(formatted_convo, tokenize=False, add_generation_prompt=False)
            texts.append(text)
        except Exception as e:
            print(f"Error processing conversation: {convo}")
            raise e
    return {"text": texts}

# Initialize the tokenizer with the correct chat template
tokenizer = get_chat_template(
    tokenizer,
    chat_template="llama-3",
    mapping={"role": "from", "content": "value", "user": "human", "assistant": "gpt"},
)

# Load and format the custom dataset
train_dataset = load_custom_dataset("datasets/datasetB_train_0-17599.json")
# train_dataset = train_dataset.shuffle(seed=42)
# train_dataset = train_dataset.select(range(2000, 2100))
train_dataset = train_dataset.map(formatting_prompts_func, batched=True)
val_dataset = load_custom_dataset("datasets/datasetB_eval_17600-21999.json")
val_dataset = val_dataset.select(range(100))
val_dataset = val_dataset.map(formatting_prompts_func, batched=True)

"""<a name="Train"></a>
### Train the model
Now let's use Huggingface TRL's `SFTTrainer`! More docs here: [TRL SFT docs](https://huggingface.co/docs/trl/sft_trainer). We do 60 steps to speed things up, but you can set `num_train_epochs=1` for a full run, and turn off `max_steps=None`. We also support TRL's `DPOTrainer`!
"""

from trl import SFTTrainer
from transformers import TrainingArguments, EarlyStoppingCallback
from unsloth import is_bfloat16_supported

early_stopping_callback = EarlyStoppingCallback(
    early_stopping_patience=3,
)

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = train_dataset,
    eval_dataset= val_dataset,
    dataset_text_field = "text",
    max_seq_length = max_seq_length,
    dataset_num_proc = 8,
    packing = False, # Can make training 5x faster for short sequences.
    args = TrainingArguments(
        per_device_train_batch_size = 1,
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        num_train_epochs = 3,
        learning_rate = 2e-3,
        fp16 = not is_bfloat16_supported(),
        bf16 = is_bfloat16_supported(),
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "cosine_with_restarts",
        seed = 3407,
        output_dir = "outputs",
        report_to = "wandb",
        logging_steps = 1, # Change if needed
        save_steps = 100, # Change if needed
        run_name = "B_eval_finetune", # (Optional)
        eval_strategy="steps",
        eval_steps=100,
        per_device_eval_batch_size=1,
        load_best_model_at_end=True,
        greater_is_better=False,
    ),
    # callbacks=[early_stopping_callback],
)

#@title Show current memory stats
gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"GPU = {gpu_stats.name}. Max memory = {max_memory} GB.")
print(f"{start_gpu_memory} GB of memory reserved.")

# Initial finetuning
trainer_stats = trainer.train()

# If you want to continue finetuning from checkpoints
# run = wandb.init(name="B_finetune")
# artifact = run.use_artifact('user/HuMob2024cityA/model-B_eval_finetune:epoch_3.0', type='model')
# artifact_dir = artifact.download()
# trainer.train(resume_from_checkpoint=artifact_dir)

model.save_pretrained("lora_model") # Local saving
# model.push_to_hub("your_name/lora_model", token = "...") # Online saving