code/spec_rag/llm_lora.py

from __future__ import annotations

from dataclasses import dataclass
from typing import Optional


@dataclass
class LoRAConfigSpec:
    r: int = 16
    alpha: int = 32
    dropout: float = 0.05
    target_modules: Optional[list[str]] = None


def build_lora_model(
    model_name: str,
    lora_cfg: LoRAConfigSpec,
    load_in_4bit: bool = True,
    device_map: str = "auto",
):
    import warnings
    import os
    
    # Suppress bitsandbytes warnings if it fails
    os.environ.setdefault('BITSANDBYTES_NOWELCOME', '1')
    
    from transformers import AutoModelForCausalLM
    from peft import LoraConfig, get_peft_model

    kwargs = {"device_map": device_map}
    if load_in_4bit:
        # Check if bitsandbytes is available and working
        bitsandbytes_available = False
        try:
            # Try to import bitsandbytes (this may raise RuntimeError if CUDA setup fails)
            import bitsandbytes as bnb
            # If import succeeds, try to use BitsAndBytesConfig
            try:
                from transformers import BitsAndBytesConfig
                # Create config to test if it works
                test_config = BitsAndBytesConfig(load_in_4bit=True)
                bitsandbytes_available = True
            except Exception as e:
                print(f"⚠️  Warning: bitsandbytes configuration failed: {str(e)[:100]}")
                print("   Falling back to full precision training.")
                bitsandbytes_available = False
        except (ImportError, RuntimeError, Exception) as e:
            error_msg = str(e)
            if "CUDA Setup failed" in error_msg or "libcudart" in error_msg or "libstdc++" in error_msg:
                print("⚠️  Warning: bitsandbytes CUDA setup failed (missing CUDA libraries).")
                print("   Falling back to full precision training.")
                print("   To fix: Install CUDA libraries or use full precision (remove --load-in-4bit)")
            else:
                print(f"⚠️  Warning: bitsandbytes not available: {error_msg[:100]}")
                print("   Falling back to full precision training.")
            bitsandbytes_available = False
        
        if bitsandbytes_available:
            try:
                from transformers import BitsAndBytesConfig
                # Use BitsAndBytesConfig for better control
                quantization_config = BitsAndBytesConfig(
                    load_in_4bit=True,
                    bnb_4bit_quant_type="nf4",  # Use NF4 for training (recommended by QLoRA paper)
                    bnb_4bit_compute_dtype="float16",
                    bnb_4bit_use_double_quant=True,  # Double quantization for better memory efficiency
                )
                kwargs["quantization_config"] = quantization_config
                print("✓ Using 4-bit quantization with bitsandbytes")
            except Exception as e:
                print(f"⚠️  Warning: Failed to configure 4-bit quantization: {str(e)[:100]}")
                print("   Falling back to full precision training.")
                load_in_4bit = False
        else:
            load_in_4bit = False
    
    if not load_in_4bit:
        print("ℹ️  Training in full precision (FP16/BF16). This requires more GPU memory.")
        print("   If you run out of memory, try reducing --batch-size or install/fix bitsandbytes.")
    
    # Suppress bitsandbytes warnings during model loading
    import warnings
    import torch
    
    # Set default torch_dtype to float16 or bfloat16 for memory efficiency
    kwargs.setdefault("torch_dtype", torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16)

    with warnings.catch_warnings():
        warnings.filterwarnings("ignore", category=UserWarning, module="bitsandbytes")
        model = AutoModelForCausalLM.from_pretrained(model_name, **kwargs)
    
    target_modules = lora_cfg.target_modules or ["q_proj", "k_proj", "v_proj", "o_proj"]
    peft_cfg = LoraConfig(
        r=lora_cfg.r,
        lora_alpha=lora_cfg.alpha,
        lora_dropout=lora_cfg.dropout,
        bias="none",
        task_type="CAUSAL_LM",
        target_modules=target_modules,
    )
    
    try:
        return get_peft_model(model, peft_cfg)
    except Exception as e:
        print(f"⚠️  Error creating PEFT model: {e}")
        print("   This might be due to bitsandbytes issues. Try removing --load-in-4bit.")
        raise