train_7b.py

"""
GIS-Coder 7B: Production QLoRA SFT Training Script
====================================================
Fine-tunes Qwen2.5-Coder-7B-Instruct for GIS code generation.

Hardware requirements:
  - Minimum: 1x A10G (24GB) or 1x RTX 4090 (24GB)
  - Recommended: 1x A100 (80GB) for faster training + larger batch
  - Also works on: H100, L40S, RTX 3090

Training recipe based on:
  - CFD fine-tuning (arxiv:2504.09602): QLoRA, r=16, 88.7% accuracy on domain tasks
  - MapCoder-Lite (arxiv:2509.17489): Qwen2.5-Coder-7B as best backbone for code LoRA
  - LoRA Without Regret: target all-linear layers, lr=2e-4 for LoRA

Usage:
  # Single GPU
  python train_7b.py

  # Multi-GPU with accelerate
  accelerate launch --num_processes 2 train_7b.py

  # With custom settings
  python train_7b.py --epochs 5 --lr 1e-4 --lora_r 32 --max_length 4096
"""

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

def parse_args():
    parser = argparse.ArgumentParser(description="Train GIS-Coder 7B")
    parser.add_argument("--model_id", type=str, default="Qwen/Qwen2.5-Coder-7B-Instruct")
    parser.add_argument("--dataset_id", type=str, default="RhodWeo/gis-code-instructions")
    parser.add_argument("--hub_model_id", type=str, default="RhodWeo/GIS-Coder-7B")
    parser.add_argument("--output_dir", type=str, default="./gis-coder-7b-output")
    
    # Training hyperparameters
    parser.add_argument("--epochs", type=int, default=3)
    parser.add_argument("--lr", type=float, default=2e-4, help="Learning rate (2e-4 for LoRA)")
    parser.add_argument("--batch_size", type=int, default=2, help="Per-device batch size")
    parser.add_argument("--grad_accum", type=int, default=8, help="Gradient accumulation steps")
    parser.add_argument("--max_length", type=int, default=4096, help="Max sequence length")
    parser.add_argument("--warmup_ratio", type=float, default=0.1)
    parser.add_argument("--weight_decay", type=float, default=0.01)
    parser.add_argument("--scheduler", type=str, default="cosine")
    
    # LoRA hyperparameters
    parser.add_argument("--lora_r", type=int, default=32, help="LoRA rank")
    parser.add_argument("--lora_alpha", type=int, default=16, help="LoRA alpha")
    parser.add_argument("--lora_dropout", type=float, default=0.05)
    parser.add_argument("--target_modules", type=str, default="all-linear",
                        help="Target modules (all-linear or comma-separated list)")
    
    # Quantization
    parser.add_argument("--no_quantize", action="store_true", help="Disable 4-bit quantization (full fp16)")
    parser.add_argument("--use_flash_attn", action="store_true", help="Use Flash Attention 2")
    
    # Tracking
    parser.add_argument("--use_trackio", action="store_true", help="Enable Trackio monitoring")
    parser.add_argument("--trackio_project", type=str, default="gis-coder-7b")
    
    return parser.parse_args()


def main():
    args = parse_args()
    
    # ─── Trackio (optional) ────────────────────────────────────────────────
    if args.use_trackio:
        import trackio
        trackio.init(
            project=args.trackio_project,
            config=vars(args),
        )
    
    # ─── Dataset ───────────────────────────────────────────────────────────
    print(f"Loading dataset: {args.dataset_id}")
    dataset = load_dataset(args.dataset_id, data_files="data/train.jsonl", split="train")
    print(f"  {len(dataset)} examples, columns: {dataset.column_names}")
    
    # ─── Model ─────────────────────────────────────────────────────────────
    print(f"Loading model: {args.model_id}")
    
    model_kwargs = {
        "trust_remote_code": True,
        "attn_implementation": "flash_attention_2" if args.use_flash_attn else "eager",
    }
    
    if not args.no_quantize:
        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_kwargs["quantization_config"] = bnb_config
        model_kwargs["dtype"] = torch.bfloat16
    else:
        model_kwargs["dtype"] = torch.bfloat16
    
    model = AutoModelForCausalLM.from_pretrained(
        args.model_id,
        device_map="auto",
        **model_kwargs,
    )
    
    tokenizer = AutoTokenizer.from_pretrained(args.model_id, trust_remote_code=True)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
        model.config.pad_token_id = tokenizer.eos_token_id
    
    if not args.no_quantize:
        model = prepare_model_for_kbit_training(model)
    
    print(f"  Parameters: {model.num_parameters()/1e9:.2f}B")
    
    # ─── LoRA ──────────────────────────────────────────────────────────────
    target = args.target_modules
    if target != "all-linear":
        target = target.split(",")
    
    peft_config = LoraConfig(
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        target_modules=target,
        lora_dropout=args.lora_dropout,
        bias="none",
        task_type="CAUSAL_LM",
    )
    print(f"  LoRA: r={args.lora_r}, alpha={args.lora_alpha}, targets={target}")
    
    # ─── Training Config ───────────────────────────────────────────────────
    training_args = SFTConfig(
        output_dir=args.output_dir,
        num_train_epochs=args.epochs,
        per_device_train_batch_size=args.batch_size,
        gradient_accumulation_steps=args.grad_accum,
        learning_rate=args.lr,
        lr_scheduler_type=args.scheduler,
        warmup_ratio=args.warmup_ratio,
        weight_decay=args.weight_decay,
        
        gradient_checkpointing=True,
        bf16=True,
        max_length=args.max_length,
        
        logging_steps=1,
        logging_first_step=True,
        logging_strategy="steps",
        disable_tqdm=True,
        report_to="trackio" if args.use_trackio else "none",
        
        save_strategy="epoch",
        save_total_limit=3,
        
        push_to_hub=True,
        hub_model_id=args.hub_model_id,
        hub_strategy="every_save",
        
        dataloader_num_workers=4,
        seed=42,
    )
    
    # ─── Trainer ───────────────────────────────────────────────────────────
    trainer = SFTTrainer(
        model=model,
        processing_class=tokenizer,
        args=training_args,
        train_dataset=dataset,
        peft_config=peft_config,
    )
    
    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    print(f"  Trainable: {trainable:,} ({trainable/total*100:.2f}%)")
    
    # ─── Train ─────────────────────────────────────────────────────────────
    eff_bs = args.batch_size * args.grad_accum
    print(f"\n{'='*60}")
    print(f"TRAINING: {args.model_id}")
    print(f"  Dataset: {len(dataset)} examples")
    print(f"  Method: {'QLoRA' if not args.no_quantize else 'LoRA'} (r={args.lora_r})")
    print(f"  LR: {args.lr}, Epochs: {args.epochs}, Eff. batch: {eff_bs}")
    print(f"  Max length: {args.max_length}")
    print(f"  Push to: {args.hub_model_id}")
    print(f"{'='*60}\n")
    
    result = trainer.train()
    
    # ─── Save ──────────────────────────────────────────────────────────────
    print("\nSaving final model...")
    trainer.save_model(os.path.join(args.output_dir, "final"))
    trainer.push_to_hub(commit_message="GIS-Coder 7B — final after training")
    
    m = result.metrics
    print(f"\nDone! Loss: {m.get('train_loss','?')}, Time: {m.get('train_runtime',0):.0f}s")
    print(f"Model: https://huggingface.co/{args.hub_model_id}")
    
    if args.use_trackio:
        import trackio
        trackio.log({"final_loss": m.get("train_loss", 0), "runtime": m.get("train_runtime", 0)})
        trackio.finish()


if __name__ == "__main__":
    main()