dequantize_qlora.py

"""Dequantize QLoRA checkpoint → fp32 Linear model (~970K params)."""
import os, sys, torch, torch.nn as nn
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
from scripts.model_tiny import TinyModel, QLoRALinear

def dequantize_model(qlora_ckpt_path, output_path):
    model = TinyModel()
    model.reset_weights()

    # Apply QLoRA to create matching architecture
    from scripts.model_tiny import apply_qlora
    model = apply_qlora(model, r=8, alpha=16, freeze_embeds=False)

    # Load QLoRA checkpoint
    ckpt = torch.load(qlora_ckpt_path, map_location="cpu", weights_only=True)
    model.load_state_dict(ckpt, strict=False)

    # Fix _has_bias flag (not in state dict, but bias buffer was loaded)
    for name, module in model.named_modules():
        if isinstance(module, QLoRALinear):
            bias_key = f"{name}.bias"
            if bias_key in ckpt:
                module._has_bias = True

    model.eval()

    # Dequantize: replace QLoRALinear → nn.Linear with dequantized weights
    for name, module in model.named_modules():
        if isinstance(module, QLoRALinear):
            weight = module._dequantized_weight() + module.lora_B @ module.lora_A * module.scaling
            linear = nn.Linear(module.in_features, module.out_features, bias=False)
            linear.weight = nn.Parameter(weight)

            # Replace in parent
            parent = model
            parts = name.split(".")
            for p in parts[:-1]:
                parent = getattr(parent, p)
            setattr(parent, parts[-1], linear)

    # Save fp32 state dict
    torch.save(model.state_dict(), output_path)
    n = sum(p.numel() for p in model.parameters())
    print(f"Dequantized: {n:,} params → {output_path}")

if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser()
    parser.add_argument("--input", default="best.pt")
    parser.add_argument("--output", default="best_fp32.pt")
    args = parser.parse_args()
    dequantize_model(args.input, args.output)