train.py

import json
import torch
from pathlib import Path
from datasets import load_dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    DataCollatorForLanguageModeling,
    TrainingArguments,
    Trainer,
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training, PeftModel

project = Path("/home/zeus/btl-1")
base_model_name = "Qwen/Qwen2.5-7B-Instruct"
max_seq_length = 4096
train_batch_size = 8
grad_accum = 2
epochs = 1
learning_rate = 2e-4

print("Loading dataset...")
train_ds = load_dataset("json", data_files=str(project / "data" / "final" / "train.jsonl"), split="train")
eval_ds = load_dataset("json", data_files=str(project / "data" / "final" / "eval.jsonl"), split="train")
print(f"train: {len(train_ds)}, eval: {len(eval_ds)}")

print("Loading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True, use_fast=True)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "right"

print("Loading model (QLoRA 4-bit)...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)
model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    trust_remote_code=True,
    quantization_config=bnb_config,
    device_map="auto",
    attn_implementation="sdpa",
)
model = prepare_model_for_kbit_training(model)
model.config.use_cache = False
model.gradient_checkpointing_enable()

lora_config = LoraConfig(
    r=64,
    lora_alpha=128,
    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, lora_config)
model.print_trainable_parameters()

print("Tokenizing...")
def render_messages(messages):
    return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)

def to_text(batch):
    return {"text": [render_messages(m) for m in batch["messages"]]}

train_text = train_ds.map(to_text, batched=True, remove_columns=train_ds.column_names)
eval_text = eval_ds.select(range(min(500, len(eval_ds)))).map(to_text, batched=True, remove_columns=["messages"])

def tokenize_batch(batch):
    return tokenizer(batch["text"], truncation=True, max_length=max_seq_length, padding=False)

train_tok = train_text.map(tokenize_batch, batched=True, remove_columns=train_text.column_names)
eval_tok = eval_text.map(tokenize_batch, batched=True, remove_columns=eval_text.column_names)

collator = DataCollatorForLanguageModeling(tokenizer=tokenizer, mlm=False)

training_args = TrainingArguments(
    output_dir="/home/zeus/btl-1/checkpoints",
    num_train_epochs=epochs,
    per_device_train_batch_size=train_batch_size,
    per_device_eval_batch_size=32,
    gradient_accumulation_steps=grad_accum,
    eval_strategy="steps",
    eval_steps=500,
    save_strategy="steps",
    save_steps=500,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    learning_rate=learning_rate,
    warmup_ratio=0.03,
    lr_scheduler_type="cosine",
    logging_steps=10,
    save_total_limit=2,
    bf16=torch.cuda.is_available(),
    fp16=False,
    optim="paged_adamw_8bit",
    report_to="none",
    gradient_checkpointing=True,
    remove_unused_columns=False,
)

trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_tok,
    eval_dataset=eval_tok,
    data_collator=collator,
)

print("Starting training...")
train_result = trainer.train()
trainer.save_state()
print(f"Training complete: {train_result}")

print("Saving adapter...")
adapter_dir = project / "artifacts" / "qlora-adapter"
adapter_dir.mkdir(parents=True, exist_ok=True)
model.save_pretrained(adapter_dir)
tokenizer.save_pretrained(adapter_dir)
print(f"Adapter saved to {adapter_dir}")

print("Loading best checkpoint...")
best = trainer.state.best_model_checkpoint
if best:
    print(f"Best checkpoint: {best}")
    reloaded_base = AutoModelForCausalLM.from_pretrained(
        base_model_name,
        trust_remote_code=True,
        quantization_config=bnb_config,
        device_map="auto",
    )
    reloaded = PeftModel.from_pretrained(reloaded_base, best)
    reloaded.save_pretrained(adapter_dir / "best")
    tokenizer.save_pretrained(adapter_dir / "best")
    print(f"Best adapter saved to {adapter_dir / 'best'}")