train.py

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from datasets import load_dataset
from peft import LoraConfig, TaskType
from trl import SFTTrainer, SFTConfig
import trackio

model_name = "./SmolLM3-3B-Base/"
dataset_path = "./MathInstruct/MathInstruct.json"
output_dir = "./SmolLMathematician-3B"
project_name = "SmolLMathematician-3B"
MAX_SEQ_LENGTH = 4096

trackio.init(project=project_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
    attn_implementation="flash_attention_2",
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
    model.config.pad_token_id = model.config.eos_token_id

with open("chat_template.jinja", "r") as f:
    chat_template = f.read()
tokenizer.chat_template = chat_template

model.gradient_checkpointing_enable()
dataset = load_dataset("json", data_files=dataset_path, split="train")

def formatInstructionWithTemplate(example: dict) -> str:
    messages = [
        {"role": "user", "content": example["instruction"]},
        {"role": "assistant", "content": example["output"]},
    ]
    return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)


def checkSequenceLength(example: dict) -> bool:
    formatted_text = formatInstructionWithTemplate(example)
    tokens = tokenizer(formatted_text)
    return len(tokens['input_ids']) <= MAX_SEQ_LENGTH

original_size = len(dataset)
train_dataset = dataset.filter(checkSequenceLength)
new_size = len(train_dataset)

print(f"Dataset: {original_size} → {new_size} samples (removed: {original_size - new_size})")

torch.cuda.empty_cache()
peft_config = LoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.1,
    target_modules=['q_proj', 'v_proj'],
    bias="none",
    task_type=TaskType.CAUSAL_LM,
)

training_args = SFTConfig(
    output_dir=output_dir,
    num_train_epochs=1,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    optim="paged_adamw_8bit",
    learning_rate=2e-5,
    weight_decay=0.01,
    adam_epsilon=1e-6,
    max_grad_norm=1.0,
    lr_scheduler_type="cosine",
    warmup_ratio=0.1,
    logging_steps=8,
    eval_strategy="no",
    save_strategy="steps",
    save_steps=32,
    save_total_limit=4,
    resume_from_checkpoint=True,
    report_to="trackio",
    bf16=True,
    packing=True,
    max_length=MAX_SEQ_LENGTH,
    dataloader_pin_memory=False,
    gradient_checkpointing_kwargs={"use_reentrant": False},
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    peft_config=peft_config,
    formatting_func=formatInstructionWithTemplate,
)

torch.cuda.empty_cache()
trainer.train()
torch.cuda.empty_cache()
trainer.save_model(output_dir)
print(f"LoRA adapter saved to {output_dir}")
trackio.finish()