src/train.py

import os
import sys
import sqlite3
import pandas as pd
import torch
from datasets import load_dataset
from transformers import (
    AutoModelForCausalLM,
    AutoTokenizer,
    BitsAndBytesConfig,
    TrainingArguments,
    TrainerCallback
)
from peft import LoraConfig, get_peft_model, prepare_model_for_kbit_training
from trl import SFTTrainer, SFTConfig

class VRAMLoggerCallback(TrainerCallback):
    """Callback to print VRAM usage during training to monitor resource consumption."""
    def on_step_end(self, args, state, control, **kwargs):
        if torch.cuda.is_available():
            allocated = torch.cuda.memory_allocated(0) / (1024 ** 3)
            reserved = torch.cuda.memory_reserved(0) / (1024 ** 3)
            print(f" - [Step {state.global_step}] VRAM Allocated: {allocated:.2f} GB | Reserved: {reserved:.2f} GB")

def train():
    print("=" * 60)
    print("Starting Phi-3 Text-to-SQL QLoRA Fine-Tuning")
    print("=" * 60)
    
    # 1. Device Configuration and Hardware Checks
    if not torch.cuda.is_available():
        print("WARNING: CUDA is NOT available. Running on CPU is extremely slow and NOT recommended.")
        device = "cpu"
        use_bf16 = False
        use_fp16 = False
    else:
        device = "cuda"
        gpu_name = torch.cuda.get_device_name(0)
        print(f"CUDA Device Detected: {gpu_name}")
        
        # RTX 40-series (Ada Lovelace) natively supports bfloat16
        use_bf16 = torch.cuda.is_bf16_supported()
        use_fp16 = not use_bf16
        print(f"bfloat16 Supported: {use_bf16} | float16 Fallback: {use_fp16}")

    # 2. Paths
    dataset_dir = "data"
    train_file = os.path.join(dataset_dir, "train_dataset.jsonl")
    val_file = os.path.join(dataset_dir, "test_dataset.jsonl")
    output_dir = "models/phi3-text-to-sql"
    adapter_dir = "models/phi3-text-to-sql-adapter"
    
    if not os.path.exists(train_file) or not os.path.exists(val_file):
        print("ERROR: Dataset files not found. Run dataset.py first.")
        sys.exit(1)

    # 3. Quantization Config (Crucial for 6GB VRAM)
    print("Configuring 4-bit Quantization (QLoRA)...")
    bnb_config = BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_use_double_quant=True,
        bnb_4bit_compute_dtype=torch.bfloat16 if use_bf16 else torch.float16
    )

    # 4. Load Model and Tokenizer
    model_id = "microsoft/Phi-3-mini-4k-instruct"
    print(f"Loading base model: {model_id}...")
    
    model = AutoModelForCausalLM.from_pretrained(
        model_id,
        quantization_config=bnb_config,
        device_map="auto",
        trust_remote_code=False,
        attn_implementation="eager"
    )
    
    print("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=False)
    
    # Configure padding tokens for Phi-3 (which uses EOS for padding natively in standard chat SFT)
    tokenizer.pad_token = tokenizer.eos_token
    tokenizer.padding_side = "right"

    # 5. Prepare Model for Peft Training
    print("Preparing model for k-bit training and enabling gradient checkpointing...")
    model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=True)

    # 6. Configure LoRA
    # We target all linear layers to maximize standard model adaptation
    print("Configuring LoRA parameters...")
    peft_config = LoraConfig(
        r=8,  # Rank of adapter (8 is memory-efficient and very effective for code syntax)
        lora_alpha=16,
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj", 
            "gate_proj", "up_proj", "down_proj"
        ],
        lora_dropout=0.05,
        bias="none",
        task_type="CAUSAL_LM"
    )
    
    model = get_peft_model(model, peft_config)
    model.print_trainable_parameters()

    # 7. Load Datasets
    print("Loading formatted training datasets...")
    dataset = load_dataset("json", data_files={
        "train": train_file,
        "validation": val_file
    })
    
    # 8. Training Arguments
    # We optimize heavily for memory: batch_size=1, gradient_accumulation=4 (eff. batch size 4), paged_adamw optimizer
    print("Setting up memory-optimized SFT training arguments...")
    training_args = SFTConfig(
        output_dir=output_dir,
        num_train_epochs=3,
        per_device_train_batch_size=1,
        gradient_accumulation_steps=4,
        learning_rate=2e-4,
        weight_decay=0.01,
        lr_scheduler_type="cosine",
        warmup_ratio=0.05,
        logging_steps=1,
        eval_strategy="steps",
        eval_steps=5,
        save_strategy="steps",
        save_steps=10,
        save_total_limit=2,
        optim="paged_adamw_8bit" if device == "cuda" else "adamw_torch",
        bf16=use_bf16,
        fp16=use_fp16,
        gradient_checkpointing=True,
        max_grad_norm=0.3,
        report_to="none",
        ddp_find_unused_parameters=False,
        remove_unused_columns=False, # Crucial for TRL chat template processing
        max_length=512  # SFTConfig uses max_length instead of max_seq_length inside SFTTrainer constructor
    )

    # 9. Initialize SFTTrainer
    # SFTTrainer natively formats 'messages' using the model's chat template
    print("Initializing SFTTrainer...")
    trainer = SFTTrainer(
        model=model,
        train_dataset=dataset["train"],
        eval_dataset=dataset["validation"],
        peft_config=None,
        processing_class=tokenizer,
        args=training_args,
        callbacks=[VRAMLoggerCallback()] if device == "cuda" else []
    )

    # 10. Execute Fine-Tuning
    print("Launching Fine-Tuning loop...")
    trainer.train()
    
    # 11. Save the Fine-Tuned Adapter
    print(f"Fine-tuning complete! Saving adapter to {adapter_dir}...")
    trainer.model.save_pretrained(adapter_dir)
    tokenizer.save_pretrained(adapter_dir)
    print("Adapter saved successfully. Ready for evaluation and deployment!")

if __name__ == "__main__":
    train()