lora_sft_train.py

import argparse
import gc
import math
import os
import time
from pathlib import Path

import torch
import torch.nn as nn
import yaml
from huggingface_hub import hf_hub_download
from torch.amp import GradScaler, autocast

from sft_train import LUNAModel, SFTDataset, cosine_lr, probe_hardware, run_eval_prompts


SEP = "=" * 72


class LoRALinear(nn.Module):
    def __init__(self, base_layer, rank=16, alpha=32, dropout=0.05):
        super().__init__()
        if not isinstance(base_layer, nn.Linear):
            raise TypeError("LoRALinear expects a torch.nn.Linear base layer")
        self.base = base_layer
        self.rank = rank
        self.alpha = alpha
        self.scale = alpha / max(rank, 1)
        self.dropout = nn.Dropout(dropout)
        self.lora_a = nn.Linear(base_layer.in_features, rank, bias=False)
        self.lora_b = nn.Linear(rank, base_layer.out_features, bias=False)
        nn.init.kaiming_uniform_(self.lora_a.weight, a=math.sqrt(5))
        nn.init.zeros_(self.lora_b.weight)

        for parameter in self.base.parameters():
            parameter.requires_grad = False

    def forward(self, x):
        base_out = self.base(x)
        lora_out = self.lora_b(self.lora_a(self.dropout(x))) * self.scale
        return base_out + lora_out


def load_config(config_path):
    with open(config_path, encoding="utf-8") as handle:
        raw = yaml.safe_load(handle)

    cfg = {
        "auto_config": raw.get("auto_config", True),
        "hf_model_repo": raw.get("hf_model_repo", "ASTERIZER/LUNA-100M"),
        "hf_model_file": raw.get("hf_model_file", "sft_v1/final/model.pth"),
        "pretrained_ckpt": raw.get("pretrained_ckpt", "Base/out/input_models/luna_sft_v1/model.pth"),
        "train_json": raw.get("train_json", "Base/Datasets/rag_mcp_sft/train.json"),
        "val_json": raw.get("val_json", "Base/Datasets/rag_mcp_sft/val.json"),
        "out_dir": raw.get("out_dir", "Base/out/sft/rag_mcp_lora"),
        "tokenizer_dir": raw.get("tokenizer_dir", "Base/checkpoints/EleutherAI/pythia-160m"),
        "vocab_size": raw["model"]["vocab_size"],
        "seq_len": raw["model"]["seq_len"],
        "n_layer": raw["model"]["n_layer"],
        "n_embd": raw["model"]["n_embd"],
        "n_head": raw["model"]["n_head"],
        "epochs": raw["train"]["epochs"],
        "lr_warmup_steps": raw["train"]["lr_warmup_steps"],
        "save_interval": raw["train"]["save_interval"],
        "log_interval": raw["train"]["log_interval"],
        "eval_interval": raw["train"]["eval_interval"],
        "max_norm": raw["train"]["max_norm"],
        "lr": raw["optimizer"]["lr"],
        "min_lr": raw["optimizer"]["min_lr"],
        "weight_decay": raw["optimizer"]["weight_decay"],
        "betas": tuple(raw["optimizer"]["betas"]),
        "eps": raw["optimizer"]["eps"],
        "global_batch": raw["batch"]["global_batch"],
        "micro_batch": raw["batch"]["micro_batch"],
        "grad_accum": raw["batch"]["grad_accum"],
        "auto_probe_batch": raw["batch"].get("auto_probe_batch", True),
        "probe_safety": raw["batch"].get("probe_safety", 0.94),
        "num_workers": raw["dataloader"]["num_workers"],
        "pin_memory": raw["dataloader"]["pin_memory"],
        "precision": raw["hardware"]["precision"],
        "eval_prompts": raw.get("eval_prompts", []),
        "lora_rank": raw["lora"]["rank"],
        "lora_alpha": raw["lora"]["alpha"],
        "lora_dropout": raw["lora"]["dropout"],
        "target_modules": list(raw["lora"]["target_modules"]),
    }
    return cfg


def resolve_checkpoint(cfg):
    ckpt_path = Path(cfg["pretrained_ckpt"])
    if ckpt_path.exists():
        return ckpt_path

    ckpt_path.parent.mkdir(parents=True, exist_ok=True)
    hf_hub_download(
        repo_id=cfg["hf_model_repo"],
        filename=cfg["hf_model_file"],
        local_dir=str(ckpt_path.parent),
        token=os.environ.get("HF_TOKEN"),
    )
    downloaded = ckpt_path.parent / cfg["hf_model_file"]
    if not downloaded.exists():
        raise FileNotFoundError(f"Expected downloaded checkpoint at {downloaded}")
    return downloaded


def inject_lora(model, target_modules, rank, alpha, dropout):
    replaced = []
    for module_name, module in list(model.named_modules()):
        if not isinstance(module, nn.Linear):
            continue
        if not any(module_name.endswith(target) for target in target_modules):
            continue
        parent_name, _, child_name = module_name.rpartition(".")
        parent_module = model.get_submodule(parent_name) if parent_name else model
        wrapped = LoRALinear(module, rank=rank, alpha=alpha, dropout=dropout)
        wrapped = wrapped.to(device=module.weight.device, dtype=module.weight.dtype)
        setattr(parent_module, child_name, wrapped)
        replaced.append(module_name)
    if not replaced:
        raise RuntimeError("No target modules matched for LoRA injection")
    return replaced


def get_lora_state_dict(model):
    state_dict = model.state_dict()
    return {
        name: tensor.cpu()
        for name, tensor in state_dict.items()
        if "lora_a.weight" in name or "lora_b.weight" in name
    }


def count_trainable_parameters(model):
    return sum(parameter.numel() for parameter in model.parameters() if parameter.requires_grad)


def probe_max_micro_batch_lora(model, trainable_parameters, device, dtype, seq_len, vocab_size, safety=0.94, grad_accum_sim=2):
    if device.type != "cuda":
        return 1

    optimizer = torch.optim.AdamW(trainable_parameters, lr=1e-4)
    lo, hi, best = 1, 512, 1

    while lo <= hi:
        mid = (lo + hi) // 2
        try:
            torch.cuda.empty_cache()
            gc.collect()
            optimizer.zero_grad(set_to_none=True)

            for _ in range(grad_accum_sim):
                input_ids = torch.randint(0, vocab_size, (mid, seq_len), device=device)
                loss_mask = torch.ones_like(input_ids)
                with autocast(device_type="cuda", dtype=dtype):
                    _, loss = model(input_ids, targets=input_ids, loss_mask=loss_mask, return_logits=False)
                    loss = loss / grad_accum_sim
                loss.backward()
                del input_ids, loss_mask, loss

            optimizer.step()
            optimizer.zero_grad(set_to_none=True)
            best = mid
            lo = mid + 1
        except (torch.cuda.OutOfMemoryError, RuntimeError) as error:
            if "out of memory" not in str(error).lower() and not isinstance(error, torch.cuda.OutOfMemoryError):
                raise
            optimizer.zero_grad(set_to_none=True)
            torch.cuda.empty_cache()
            gc.collect()
            hi = mid - 1

    del optimizer
    torch.cuda.empty_cache()
    gc.collect()

    safe = max(1, int(best * safety))
    print(f"  LoRA batch probe: max_micro_batch={best}, using {safe} ({int(safety * 100)}% safety)")
    return safe


def load_base_weights(model, checkpoint_path, device):
    checkpoint = torch.load(checkpoint_path, map_location=device, weights_only=True)
    state_dict = checkpoint["model"] if isinstance(checkpoint, dict) and "model" in checkpoint else checkpoint
    model.load_state_dict(state_dict, strict=True)


def train(cfg):
    hw = probe_hardware()
    device = torch.device(hw["device"])
    dtype = hw.get("dtype", torch.float32) if cfg["auto_config"] else {
        "bf16": torch.bfloat16,
        "fp16": torch.float16,
        "fp32": torch.float32,
    }.get(cfg["precision"], torch.float32)

    from transformers import AutoTokenizer

    tokenizer = AutoTokenizer.from_pretrained(cfg["tokenizer_dir"])
    ckpt_path = resolve_checkpoint(cfg)

    model = LUNAModel(
        vocab_size=cfg["vocab_size"],
        block_size=cfg["seq_len"],
        n_layer=cfg["n_layer"],
        n_embd=cfg["n_embd"],
        n_head=cfg["n_head"],
    ).to(device)
    load_base_weights(model, ckpt_path, device)

    for parameter in model.parameters():
        parameter.requires_grad = False

    replaced = inject_lora(
        model,
        target_modules=cfg["target_modules"],
        rank=cfg["lora_rank"],
        alpha=cfg["lora_alpha"],
        dropout=cfg["lora_dropout"],
    )
    trainable_params = count_trainable_parameters(model)
    total_params = sum(parameter.numel() for parameter in model.parameters())
    trainable_parameters = [parameter for parameter in model.parameters() if parameter.requires_grad]

    if cfg["auto_config"] and device.type == "cuda" and cfg["auto_probe_batch"]:
        print("  Probing LoRA micro_batch against available VRAM...")
        cfg["micro_batch"] = probe_max_micro_batch_lora(
            model,
            trainable_parameters=trainable_parameters,
            device=device,
            dtype=dtype,
            seq_len=cfg["seq_len"],
            vocab_size=cfg["vocab_size"],
            safety=cfg["probe_safety"],
        )
        cfg["grad_accum"] = max(1, math.ceil(cfg["global_batch"] / cfg["micro_batch"]))
        torch.cuda.reset_peak_memory_stats(device)

    effective_batch = cfg["micro_batch"] * cfg["grad_accum"]

    train_dataset = SFTDataset(cfg["train_json"], tokenizer, max_len=cfg["seq_len"])
    val_dataset = SFTDataset(cfg["val_json"], tokenizer, max_len=cfg["seq_len"]) if Path(cfg["val_json"]).exists() else None

    train_loader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=cfg["micro_batch"],
        shuffle=True,
        num_workers=cfg["num_workers"],
        pin_memory=cfg["pin_memory"],
        drop_last=True,
        prefetch_factor=4 if cfg["num_workers"] > 0 else None,
        persistent_workers=cfg["num_workers"] > 0,
    )
    val_loader = None
    if val_dataset is not None:
        val_loader = torch.utils.data.DataLoader(
            val_dataset,
            batch_size=cfg["micro_batch"],
            shuffle=False,
            num_workers=min(2, cfg["num_workers"]),
            pin_memory=cfg["pin_memory"],
            drop_last=False,
        )

    optimizer = torch.optim.AdamW(
        trainable_parameters,
        lr=cfg["lr"],
        weight_decay=cfg["weight_decay"],
        betas=cfg["betas"],
        eps=cfg["eps"],
    )
    scaler = GradScaler(enabled=(device.type == "cuda" and dtype == torch.float16))

    steps_per_epoch = max(1, len(train_loader) // cfg["grad_accum"])
    total_steps = steps_per_epoch * cfg["epochs"]
    warmup_steps = min(cfg["lr_warmup_steps"], max(1, total_steps // 5))
    out_dir = Path(cfg["out_dir"])
    out_dir.mkdir(parents=True, exist_ok=True)
    best_val_loss = float("inf")
    step = 0

    latest_path = out_dir / "latest.pt"
    if latest_path.exists():
        checkpoint = torch.load(latest_path, map_location=device, weights_only=True)
        model.load_state_dict(checkpoint["adapter"], strict=False)
        optimizer.load_state_dict(checkpoint["optimizer"])
        step = checkpoint["step"]

    print(SEP)
    print("  LUNA 100M - LoRA SFT")
    print(SEP)
    print(f"  Base checkpoint : {ckpt_path}")
    print(f"  Train dataset   : {cfg['train_json']}")
    print(f"  Val dataset     : {cfg['val_json']}")
    print(f"  Output dir      : {out_dir}")
    print(f"  Device          : {hw['gpu_name']} ({hw['vram_gb']:.1f} GB)")
    print(f"  Precision       : {cfg['precision']} dtype={dtype}")
    print(f"  LoRA modules    : {', '.join(replaced)}")
    print(f"  Trainable params: {trainable_params:,} / {total_params:,}")
    print(f"  micro_batch     : {cfg['micro_batch']}")
    print(f"  grad_accum      : {cfg['grad_accum']}")
    print(f"  effective_batch : {effective_batch}")
    print(f"  Train samples   : {len(train_dataset):,}")
    print(f"  Val samples     : {len(val_dataset):,}" if val_dataset is not None else "  Val samples     : 0")
    print(SEP)

    if cfg["eval_prompts"] and step == 0:
        run_eval_prompts(model, tokenizer, cfg["eval_prompts"], device, 0, out_dir)

    model.train()
    run_t0 = time.perf_counter()

    for epoch in range(cfg["epochs"]):
        micro_step = 0
        for input_ids, loss_mask in train_loader:
            current_global_step = epoch * steps_per_epoch + (micro_step // cfg["grad_accum"])
            if current_global_step < step and (micro_step % cfg["grad_accum"] == cfg["grad_accum"] - 1):
                micro_step += 1
                continue
            if current_global_step >= total_steps:
                break

            input_ids = input_ids.to(device, non_blocking=True)
            loss_mask = loss_mask.to(device, non_blocking=True)
            step_start = time.perf_counter()

            with autocast(device_type=device.type, dtype=dtype, enabled=(device.type == "cuda")):
                _, loss = model(input_ids, targets=input_ids, loss_mask=loss_mask, return_logits=False)
                loss = loss / cfg["grad_accum"]

            scaler.scale(loss).backward()
            micro_step += 1

            if micro_step % cfg["grad_accum"] != 0:
                continue

            scaler.unscale_(optimizer)
            torch.nn.utils.clip_grad_norm_(trainable_parameters, cfg["max_norm"])
            lr_now = cosine_lr(step, warmup_steps, total_steps, cfg["lr"], cfg["min_lr"])
            for param_group in optimizer.param_groups:
                param_group["lr"] = lr_now

            scaler.step(optimizer)
            scaler.update()
            optimizer.zero_grad(set_to_none=True)

            if device.type == "cuda":
                torch.cuda.synchronize()

            dt = time.perf_counter() - step_start
            step += 1

            if step % cfg["log_interval"] == 0 or step <= 3:
                tokens_step = effective_batch * cfg["seq_len"]
                tps = tokens_step / max(dt, 1e-6)
                vram = torch.cuda.max_memory_allocated() / 1024**3 if device.type == "cuda" else 0
                eta_h = (total_steps - step) * dt / 3600
                print(
                    f"  step {step:6d}/{total_steps} | epoch {epoch + 1}/{cfg['epochs']} | "
                    f"loss {loss.item() * cfg['grad_accum']:.4f} | lr {lr_now:.2e} | "
                    f"{tps:,.0f} tok/s | VRAM {vram:.1f}GB | ETA {eta_h:.1f}h"
                )

            if step % cfg["save_interval"] == 0 or step == total_steps:
                step_dir = out_dir / f"step-{step:06d}"
                step_dir.mkdir(parents=True, exist_ok=True)
                adapter_state = get_lora_state_dict(model)
                torch.save(adapter_state, step_dir / "adapter_model.pt")
                torch.save(
                    {
                        "step": step,
                        "adapter": adapter_state,
                        "optimizer": optimizer.state_dict(),
                        "epoch": epoch,
                        "loss": loss.item() * cfg["grad_accum"],
                    },
                    latest_path,
                )
                print(f"  Saved -> {step_dir}")

            if step % cfg["eval_interval"] == 0 or step == total_steps:
                if val_loader is not None:
                    model.eval()
                    val_loss_sum = 0.0
                    val_count = 0
                    with torch.no_grad():
                        for val_ids, val_mask in val_loader:
                            val_ids = val_ids.to(device, non_blocking=True)
                            val_mask = val_mask.to(device, non_blocking=True)
                            with autocast(device_type=device.type, dtype=dtype, enabled=(device.type == "cuda")):
                                _, val_loss = model(val_ids, targets=val_ids, loss_mask=val_mask, return_logits=False)
                            val_loss_sum += val_loss.item()
                            val_count += 1
                            if val_count >= 50:
                                break
                    avg_val = val_loss_sum / max(val_count, 1)
                    print(f"  Val loss: {avg_val:.4f}")
                    if avg_val < best_val_loss:
                        best_val_loss = avg_val
                        torch.save(get_lora_state_dict(model), out_dir / "best_adapter_model.pt")
                        print("  New best! Saved best_adapter_model.pt")
                    model.train()

                if cfg["eval_prompts"]:
                    run_eval_prompts(model, tokenizer, cfg["eval_prompts"], device, step, out_dir)

    final_dir = out_dir / "final"
    final_dir.mkdir(parents=True, exist_ok=True)
    torch.save(get_lora_state_dict(model), final_dir / "adapter_model.pt")
    torch.save(
        {
            "step": step,
            "adapter": get_lora_state_dict(model),
            "lora_rank": cfg["lora_rank"],
            "lora_alpha": cfg["lora_alpha"],
            "lora_dropout": cfg["lora_dropout"],
            "target_modules": cfg["target_modules"],
            "base_checkpoint": str(ckpt_path),
        },
        final_dir / "adapter_bundle.pt",
    )

    total_h = (time.perf_counter() - run_t0) / 3600
    print(SEP)
    print(f"  LoRA SFT complete in {total_h:.2f}h -> {final_dir}")
    print(f"  Best val loss: {best_val_loss:.4f}")
    print(SEP)


def parse_args():
    parser = argparse.ArgumentParser(description="LUNA 100M - LoRA SFT")
    parser.add_argument("--config", default="rag_mcp_lora_config.yaml")
    parser.add_argument("--pretrained_ckpt", default=None)
    parser.add_argument("--train_json", default=None)
    parser.add_argument("--val_json", default=None)
    parser.add_argument("--out_dir", default=None)
    parser.add_argument("--epochs", type=int, default=None)
    return parser.parse_args()


def main():
    args = parse_args()
    cfg = load_config(args.config)
    for key in ("pretrained_ckpt", "train_json", "val_json", "out_dir"):
        value = getattr(args, key)
        if value:
            cfg[key] = value
    if args.epochs is not None:
        cfg["epochs"] = args.epochs
    train(cfg)


if __name__ == "__main__":
    main()