training_code/train_dpo.py

"""
DPO (Direct Preference Optimization) training for the 1B Transformer.

Takes the SFT model and aligns it with human preferences using
UltraFeedback preference pairs.

DPO Loss:
  L = -log sigma(beta * (log pi(yw|x)/pi_ref(yw|x) - log pi(yl|x)/pi_ref(yl|x)))

Launch: torchrun --nproc_per_node=8 train_dpo.py
"""

import os
import sys
import math
import time
import json
import datetime

import torch
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data.distributed import DistributedSampler

sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from model.config import ModelConfig
from model.transformer import Transformer
from model.data import get_tokenizer
from model.dpo_data import DPODataset, dpo_collate_fn


# === Config ===
SFT_CHECKPOINT = "/jfs/deepak-kumar/checkpoints_sft/sft_final.pt"
DPO_CHECKPOINT_DIR = "/jfs/deepak-kumar/checkpoints_dpo"
LOG_DIR = "/home/jovyan/training/logs"
DATA_CACHE = "/jfs/deepak-kumar/data"

NUM_EPOCHS = 1
BATCH_SIZE_PER_GPU = 2
GRADIENT_ACCUMULATION = 4       # effective batch = 2 * 8 * 4 = 64
MAX_SEQ_LEN = 1024
LEARNING_RATE = 5e-7            # very low LR for DPO
MIN_LR = 1e-7
WARMUP_STEPS = 100
WEIGHT_DECAY = 0.01
GRAD_CLIP = 1.0
BETA = 0.1                      # DPO temperature
LOG_INTERVAL = 10
SAVE_INTERVAL = 200


def get_cosine_lr(step, warmup_steps, total_steps, max_lr, min_lr):
    if step < warmup_steps:
        return max_lr * step / max(warmup_steps, 1)
    progress = (step - warmup_steps) / max(total_steps - warmup_steps, 1)
    return min_lr + 0.5 * (max_lr - min_lr) * (1 + math.cos(math.pi * progress))


def get_per_token_logps(model, input_ids, prompt_lens):
    """
    Compute sum of log probabilities for response tokens only.
    input_ids: [B, S] full sequence (prompt + response)
    prompt_lens: [B] where response starts
    Returns: [B] sum of log probs over response tokens
    """
    # Clone input to avoid inplace issues with shared RoPE buffers
    inp = input_ids[:, :-1].contiguous()
    with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
        logits, _ = model(inp)

    labels = input_ids[:, 1:].contiguous()
    log_probs = F.log_softmax(logits.float(), dim=-1)
    token_logps = log_probs.gather(2, labels.unsqueeze(2)).squeeze(2)

    B, S = token_logps.shape
    mask = torch.zeros_like(token_logps)
    for b in range(B):
        pl = prompt_lens[b].item()
        response_start = max(0, pl - 1)
        seq_len = (labels[b] != 0).sum().item()
        mask[b, response_start:seq_len] = 1.0

    return (token_logps * mask).sum(dim=1)


def dpo_loss(policy_chosen_logps, policy_rejected_logps,
             ref_chosen_logps, ref_rejected_logps, beta=0.1):
    """Compute DPO loss and metrics."""
    chosen_rewards = beta * (policy_chosen_logps - ref_chosen_logps)
    rejected_rewards = beta * (policy_rejected_logps - ref_rejected_logps)

    logits = chosen_rewards - rejected_rewards
    loss = -F.logsigmoid(logits).mean()

    with torch.no_grad():
        chosen_better = (chosen_rewards > rejected_rewards).float().mean()
        reward_margin = (chosen_rewards - rejected_rewards).mean()

    return loss, chosen_better.item(), reward_margin.item()


def main():
    dist.init_process_group("nccl", timeout=datetime.timedelta(minutes=30))
    rank = int(os.environ.get("RANK", 0))
    local_rank = int(os.environ.get("LOCAL_RANK", 0))
    world_size = int(os.environ.get("WORLD_SIZE", 1))
    torch.cuda.set_device(local_rank)
    device = torch.device(f"cuda:{local_rank}")

    if rank == 0:
        os.makedirs(DPO_CHECKPOINT_DIR, exist_ok=True)
        os.makedirs(LOG_DIR, exist_ok=True)
        print("=" * 70)
        print("  DPO: PREFERENCE ALIGNMENT FOR 1B TRANSFORMER")
        print("=" * 70)

    tokenizer = get_tokenizer()
    special_tokens = ["<|user|>", "<|assistant|>", "<|end|>"]
    vocab = tokenizer.get_vocab()
    new_tokens = [t for t in special_tokens if t not in vocab]
    if new_tokens:
        tokenizer.add_tokens(new_tokens, special_tokens=True)

    model_config = ModelConfig()
    model_config.vocab_size = len(tokenizer)

    if rank == 0:
        print(f"[Init] Loading SFT model from {SFT_CHECKPOINT}")

    # Policy model (trainable)
    policy = Transformer(model_config)
    ckpt = torch.load(SFT_CHECKPOINT, map_location="cpu", weights_only=False)
    policy.load_state_dict(ckpt["model"])
    sft_step = ckpt.get("step", 0)
    if rank == 0:
        print(f"[Init] SFT model loaded (step {sft_step})")

    # Reference model (frozen copy)
    ref_model = Transformer(model_config)
    ref_model.load_state_dict(ckpt["model"])
    del ckpt

    policy = policy.to(device)
    ref_model = ref_model.to(device).bfloat16()
    ref_model.eval()
    for p in ref_model.parameters():
        p.requires_grad = False

    policy = DDP(policy, device_ids=[local_rank])

    if rank == 0:
        n = sum(p.numel() for p in policy.parameters())
        print(f"[Init] Params: {n:,} | GPUs: {world_size}x H100")
        print(f"[Init] Beta: {BETA} | LR: {LEARNING_RATE}")

    # Dataset
    dataset = DPODataset(
        tokenizer=tokenizer,
        max_seq_len=MAX_SEQ_LEN,
        split="train",
        cache_dir=DATA_CACHE,
    )

    sampler = DistributedSampler(dataset, num_replicas=world_size, rank=rank, shuffle=True)
    dataloader = torch.utils.data.DataLoader(
        dataset,
        batch_size=BATCH_SIZE_PER_GPU,
        sampler=sampler,
        num_workers=4,
        pin_memory=True,
        collate_fn=lambda b: dpo_collate_fn(b, pad_id=tokenizer.pad_token_id),
    )

    steps_per_epoch = len(dataloader) // GRADIENT_ACCUMULATION
    total_steps = steps_per_epoch * NUM_EPOCHS

    if rank == 0:
        eff_batch = BATCH_SIZE_PER_GPU * world_size * GRADIENT_ACCUMULATION
        print(f"[Init] Dataset: {len(dataset):,} preference pairs")
        print(f"[Init] Effective batch: {eff_batch} | Steps/epoch: {steps_per_epoch}")
        print(f"[Init] Total steps: {total_steps}")
        print("-" * 70)

    decay_params = [p for n, p in policy.named_parameters() if p.dim() >= 2 and p.requires_grad]
    nodecay_params = [p for n, p in policy.named_parameters() if p.dim() < 2 and p.requires_grad]
    optimizer = torch.optim.AdamW([
        {"params": decay_params, "weight_decay": WEIGHT_DECAY},
        {"params": nodecay_params, "weight_decay": 0.0},
    ], lr=LEARNING_RATE, betas=(0.9, 0.95), fused=True)

    policy.train()
    global_step = 0
    running_loss = 0.0
    running_acc = 0.0
    running_margin = 0.0
    t0 = time.time()

    log_file = open(os.path.join(LOG_DIR, "dpo_log.jsonl"), "w") if rank == 0 else None

    for epoch in range(NUM_EPOCHS):
        sampler.set_epoch(epoch)
        data_iter = iter(dataloader)

        if rank == 0:
            print(f"\n[Epoch {epoch + 1}/{NUM_EPOCHS}]")

        while True:
            optimizer.zero_grad(set_to_none=True)
            batch_loss = 0.0
            batch_acc = 0.0
            batch_margin = 0.0
            valid_micros = 0

            for _ in range(GRADIENT_ACCUMULATION):
                try:
                    batch = next(data_iter)
                except StopIteration:
                    break

                chosen_ids = batch["chosen_ids"].to(device, non_blocking=True)
                rejected_ids = batch["rejected_ids"].to(device, non_blocking=True)
                prompt_lens = batch["prompt_lens"].to(device, non_blocking=True)

                policy_chosen_logps = get_per_token_logps(policy, chosen_ids, prompt_lens)
                policy_rejected_logps = get_per_token_logps(policy, rejected_ids, prompt_lens)

                with torch.no_grad():
                    ref_chosen_logps = get_per_token_logps(ref_model, chosen_ids, prompt_lens)
                    ref_rejected_logps = get_per_token_logps(ref_model, rejected_ids, prompt_lens)

                loss, acc, margin = dpo_loss(
                    policy_chosen_logps, policy_rejected_logps,
                    ref_chosen_logps, ref_rejected_logps,
                    beta=BETA,
                )
                loss = loss / GRADIENT_ACCUMULATION
                loss.backward()

                batch_loss += loss.item()
                batch_acc += acc
                batch_margin += margin
                valid_micros += 1

            if valid_micros == 0:
                break

            torch.nn.utils.clip_grad_norm_(policy.parameters(), GRAD_CLIP)

            lr = get_cosine_lr(global_step, WARMUP_STEPS, total_steps, LEARNING_RATE, MIN_LR)
            for pg in optimizer.param_groups:
                pg["lr"] = lr

            optimizer.step()
            global_step += 1
            running_loss += batch_loss
            running_acc += batch_acc / valid_micros
            running_margin += batch_margin / valid_micros

            if global_step % LOG_INTERVAL == 0:
                avg_loss = running_loss / LOG_INTERVAL
                avg_acc = running_acc / LOG_INTERVAL
                avg_margin = running_margin / LOG_INTERVAL
                elapsed = time.time() - t0
                pct = 100.0 * global_step / total_steps
                eta = (elapsed / max(global_step, 1)) * (total_steps - global_step)

                if rank == 0:
                    gpu_mem = torch.cuda.max_memory_allocated(device) / 1e9
                    print(
                        f"  [Step {global_step:>5d}/{total_steps}] "
                        f"loss={avg_loss:.4f} | acc={avg_acc:.1%} | "
                        f"margin={avg_margin:.3f} | lr={lr:.2e} | "
                        f"GPU={gpu_mem:.1f}GB | {pct:.1f}% | ETA={eta/60:.0f}m",
                        flush=True,
                    )
                    if log_file:
                        log_file.write(json.dumps({
                            "step": global_step, "loss": round(avg_loss, 4),
                            "accuracy": round(avg_acc, 4),
                            "reward_margin": round(avg_margin, 4),
                            "lr": lr, "elapsed_s": round(elapsed, 1),
                        }) + "\n")
                        log_file.flush()

                running_loss = 0.0
                running_acc = 0.0
                running_margin = 0.0

            if global_step % SAVE_INTERVAL == 0:
                dist.barrier()
                if rank == 0:
                    path = os.path.join(DPO_CHECKPOINT_DIR, f"dpo_step_{global_step}.pt")
                    torch.save({
                        "step": global_step,
                        "model": policy.module.state_dict(),
                        "config": model_config.__dict__,
                        "vocab_size": model_config.vocab_size,
                    }, path)
                    print(f"  >> Checkpoint: {path}", flush=True)
                dist.barrier()

    # Final save
    dist.barrier()
    if rank == 0:
        final_path = os.path.join(DPO_CHECKPOINT_DIR, "dpo_final.pt")
        torch.save({
            "step": global_step,
            "model": policy.module.state_dict(),
            "config": model_config.__dict__,
            "vocab_size": model_config.vocab_size,
        }, final_path)
        total_time = time.time() - t0
        print("=" * 70)
        print(f"  DPO COMPLETE")
        print(f"  Steps: {global_step:,} | Epochs: {NUM_EPOCHS}")
        print(f"  Time: {total_time/60:.1f} minutes")
        print(f"  Final model: {final_path}")
        print("=" * 70)
        if log_file:
            log_file.close()

    dist.destroy_process_group()


if __name__ == "__main__":
    main()