train.py

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "transformers>=4.44.2",
#     "accelerate>=0.24.0",
#     "bitsandbytes>=0.41.0",
#     "datasets",
#     "scipy",
#     "hf_transfer",
#     "rich",
#     "trackio",
# ]
# ///

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

# === CONFIGURATION ===
MODEL_NAME = "Qwen/Qwen3-8B"           # Base model — fits locally on M2 Pro 16GB after fine-tuning
DATASET_NAME = "ceperaltab/diamond-vision-dataset"
OUTPUT_DIR = "diamond-vision-expert"
HF_USERNAME = "ceperaltab"

def main():
    print("=" * 60)
    print("Diamond Vision Expert — QLoRA Fine-tuning")
    print(f"Base model : {MODEL_NAME}")
    print(f"Dataset    : {DATASET_NAME}")
    print("=" * 60)

    # Load dataset
    print(f"📦 Loading dataset: {DATASET_NAME}...")
    dataset = load_dataset(DATASET_NAME, split="train")
    print(f"✅ Dataset loaded: {len(dataset)} examples")

    # Train / eval split
    print("🔀 Creating train/eval split...")
    dataset_split = dataset.train_test_split(test_size=0.05, seed=42)
    train_dataset = dataset_split["train"]
    eval_dataset  = dataset_split["test"]
    print(f"   Train: {len(train_dataset)} | Eval: {len(eval_dataset)}")

    # Training config
    config = SFTConfig(
        output_dir=OUTPUT_DIR,
        push_to_hub=True,
        hub_model_id=f"{HF_USERNAME}/{OUTPUT_DIR}",
        hub_strategy="every_save",

        # Training
        num_train_epochs=1,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=8,
        learning_rate=2e-4,
        # NOTE: max_seq_length is NOT supported in SFTConfig (trl>=0.12.0) — removed

        # Logging & checkpointing
        logging_steps=10,
        save_strategy="steps",
        save_steps=500,
        save_total_limit=2,

        # Evaluation
        eval_strategy="steps",
        eval_steps=500,

        # Optimization
        warmup_ratio=0.03,
        lr_scheduler_type="cosine",
        gradient_checkpointing=True,
        bf16=True,                     # A10G supports bf16

        # Monitoring
        report_to="trackio",
        project="diamond-vision-training",
        run_name="diamond-vision-qwen3-8b-v1",
    )

    # LoRA
    peft_config = LoraConfig(
        r=64,
        lora_alpha=16,
        lora_dropout=0.1,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj",
            "gate_proj", "up_proj", "down_proj",
        ],
    )

    # 4-bit QLoRA quantization
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_use_double_quant=True,
    )

    # Load model
    print(f"🔄 Loading base model: {MODEL_NAME}...")
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=True,
    )

    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    # Train
    print("🎯 Initializing trainer...")
    trainer = SFTTrainer(
        model=model,
        processing_class=tokenizer,
        train_dataset=train_dataset,
        eval_dataset=eval_dataset,
        args=config,
        peft_config=peft_config,
    )

    print("🚀 Starting training...")
    trainer.train()

    print("💾 Pushing final adapter to Hub...")
    trainer.push_to_hub()

    trackio.finish()
    print("✅ Done! Adapter pushed to:", f"https://huggingface.co/{HF_USERNAME}/{OUTPUT_DIR}")

if __name__ == "__main__":
    main()