import os # Setup env vars for robust Kaggle training os.environ["NCCL_P2P_DISABLE"] = "1" os.environ["NCCL_IB_DISABLE"] = "1" os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1" os.environ["HF_HUB_DISABLE_PROGRESS_BARS"] = "1" import math import time import argparse import logging from pathlib import Path import torch import torch.distributed as dist from torch.nn.parallel import DistributedDataParallel as DDP from torch.cuda.amp import GradScaler import numpy as np from transformers import AutoTokenizer from huggingface_hub import hf_hub_download, HfApi from huggingface_hub.utils import EntryNotFoundError, HfHubHTTPError try: from kaggle_secrets import UserSecretsClient HAS_KAGGLE_SECRETS = True except ImportError: HAS_KAGGLE_SECRETS = False logging.basicConfig(level=logging.INFO, format="%(asctime)s [%(levelname)s] %(message)s") logger = logging.getLogger(__name__) def get_hf_token(): token = os.environ.get("HF_TOKEN") if token: logger.info("Loaded HF_TOKEN from environment.") return token if HAS_KAGGLE_SECRETS: try: token = UserSecretsClient().get_secret("HF_TOKEN") if token: os.environ["HF_TOKEN"] = token logger.info("Loaded HF_TOKEN from kaggle_secrets.") return token except Exception as e: logger.warning(f"Failed to get HF_TOKEN from kaggle_secrets: {e}") raise ValueError("HF_TOKEN not found in environment or kaggle_secrets.") def setup_ddp(): if "RANK" not in os.environ or "WORLD_SIZE" not in os.environ: return 0, 0, 1, torch.device("cuda" if torch.cuda.is_available() else "cpu"), False dist.init_process_group("nccl") local_rank = int(os.environ["LOCAL_RANK"]) device = torch.device(f"cuda:{local_rank}") torch.cuda.set_device(device) return int(os.environ["RANK"]), local_rank, int(os.environ["WORLD_SIZE"]), device, True def load_tokenizer_robust(hf_token: str, model_repo: str, subfolder: str, is_main: bool): if is_main: logger.info("Loading tokenizer...") local_candidates = [ "/kaggle/working/ViuRec/tokenizer", "/kaggle/input/viuai-500m-tokenizer", "./tokenizer", "./viuai-500m-tokenizer" ] for fast in [True, False]: for p_str in local_candidates: p = Path(p_str) if (p / "tokenizer.json").exists() or (p / "vocab.json").exists(): try: return AutoTokenizer.from_pretrained(str(p), local_files_only=True, use_fast=fast, trust_remote_code=True) except Exception as e: if "sentencepiece" in str(e).lower() or "tiktoken" in str(e).lower(): continue try: return AutoTokenizer.from_pretrained(model_repo, subfolder=subfolder, token=hf_token, use_fast=fast, trust_remote_code=True) except Exception as e: pass try: return AutoTokenizer.from_pretrained(model_repo, token=hf_token, use_fast=fast, trust_remote_code=True) except Exception as e: pass if is_main: logger.warning("Tokenizer failed, using dummy vocab 64000") return None def load_checkpoint(model, optimizer, scaler, api, token, repo_id, filename="checkpoints/latest_chk.pt", strict=False): local_path = "latest_chk.pt" try: logger.info(f"Checking for existing checkpoint on HF: {filename}") hf_hub_download(repo_id=repo_id, filename=filename, local_dir=".", token=token) # Security Note: weights_only=False is used here to allow loading the optimizer state safely inside our trusted environment. ckpt = torch.load(local_path, map_location="cpu", weights_only=False) missing, unexpected = model.load_state_dict(ckpt["model"], strict=strict) if missing or unexpected: logger.warning(f"Checkpoint loaded with Missing keys: {missing} | Unexpected keys: {unexpected}") else: logger.info("Model weights loaded successfully.") optimizer.load_state_dict(ckpt["optimizer"]) if "scaler" in ckpt and scaler is not None: scaler.load_state_dict(ckpt["scaler"]) step = ckpt.get("step", 0) logger.info(f"Resuming from step {step}") Path(local_path).unlink(missing_ok=True) return step except (EntryNotFoundError, HfHubHTTPError): logger.info("No checkpoint found on HF. Starting from scratch.") return 0 except Exception as e: logger.error(f"Error loading checkpoint: {e}") return 0 class ShardPool: def __init__(self, api, repo_id, work_dir, token, vocab_size, max_pool=3): self.api = api self.repo_id = repo_id self.work_dir = Path(work_dir) self.work_dir.mkdir(parents=True, exist_ok=True) self.token = token self.max_pool = max_pool self.pool = [] self.available_files = [] self.vocab_size = vocab_size self._refresh_file_list() def _refresh_file_list(self): files = self.api.list_repo_files(repo_id=self.repo_id, repo_type="dataset", token=self.token) self.available_files = [f for f in files if f.startswith("shards/hi_indiccorp_") and f.endswith(".bin")] np.random.shuffle(self.available_files) def fill(self, ctx_len): max_attempts = self.max_pool * 5 attempts = 0 while len(self.pool) < self.max_pool: if attempts >= max_attempts: raise RuntimeError("Failed to fill ShardPool: too many invalid shards.") if not self.available_files: self._refresh_file_list() f = self.available_files.pop() local_p = hf_hub_download(repo_id=self.repo_id, repo_type="dataset", filename=f, local_dir=self.work_dir, token=self.token) # Safe dtype selection dtype = np.uint16 if self.vocab_size <= 65535 else np.uint32 mmap_arr = np.memmap(local_p, dtype=dtype, mode="r") if len(mmap_arr) < ctx_len + 1: logger.warning(f"Shard {f} is too small ({len(mmap_arr)} tokens). Discarding.") del mmap_arr Path(local_p).unlink(missing_ok=True) attempts += 1 continue self.pool.append((local_p, mmap_arr)) logger.info(f"Loaded shard {f}") def sample_batch(self, batch_size, ctx_len, rng): xs, ys = [], [] while len(xs) < batch_size: idx = rng.integers(0, len(self.pool)) _, tokens = self.pool[idx] max_start = len(tokens) - ctx_len - 1 if max_start <= 0: continue i = rng.integers(0, max_start) xs.append(np.array(tokens[i:i+ctx_len], dtype=np.int64)) ys.append(np.array(tokens[i+1:i+1+ctx_len], dtype=np.int64)) return np.stack(xs), np.stack(ys) def rotate_one(self): if not self.pool: return old_path, old_mmap = self.pool.pop(0) del old_mmap try: Path(old_path).unlink(missing_ok=True) logger.info(f"Rotated out shard {Path(old_path).name}") except Exception as e: logger.warning(f"Failed to delete {old_path}: {e}") def main(): parser = argparse.ArgumentParser() parser.add_argument("--max-steps", type=int, default=100) parser.add_argument("--micro-batch", type=int, default=2) parser.add_argument("--grad-accum", type=int, default=16) parser.add_argument("--eval-every", type=int, default=500) parser.add_argument("--save-every", type=int, default=500) parser.add_argument("--rotate-every", type=int, default=500) parser.add_argument("--lr", type=float, default=3e-4) parser.add_argument("--weight-decay", type=float, default=0.1) parser.add_argument("--ctx-len", type=int, default=2048) parser.add_argument("--vocab-size", type=int, default=None) args = parser.parse_args() token = get_hf_token() rank, local_rank, world_size, device, is_ddp = setup_ddp() is_main = (rank == 0) # Random seed seed = 42 + rank torch.manual_seed(seed) np.random.seed(seed) rng = np.random.default_rng(seed) # Load resources tokenizer = load_tokenizer_robust(token, "ViuAI/ViuRec", "tokenizer", is_main) vocab_size = args.vocab_size or (len(tokenizer) if tokenizer is not None else 64000) from model import ViuResonance100M, ModelConfig config = ModelConfig(vocab_size=vocab_size) model = ViuResonance100M(config).to(device) if is_ddp: model = DDP(model, device_ids=[local_rank], output_device=local_rank, broadcast_buffers=False) unwrapped_model = model.module else: unwrapped_model = model # Optional torch.compile fallback try: model = torch.compile(model) if is_main: logger.info("Compiled model successfully.") except Exception as e: if is_main: logger.warning(f"Failed to compile model: {e}. Falling back to eager mode.") # Optimizer try: optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, betas=(0.9, 0.95), fused=True) except: if is_main: logger.warning("fused=True failed for AdamW. Falling back to non-fused.") optimizer = torch.optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, betas=(0.9, 0.95)) # GradScaler for safe fp16/bf16 scaler = GradScaler(enabled=not torch.cuda.is_bf16_supported()) api = HfApi() start_step = load_checkpoint(unwrapped_model, optimizer, scaler, api, token, "ViuAI/ViuRec", strict=False) work_dir = Path("/kaggle/working" if Path("/kaggle/working").exists() else ".") shard_pool = ShardPool(api, "ViuAI/ViuCorp", work_dir / "shards", token, vocab_size) shard_pool.fill(args.ctx_len) # Val Data Safely (cap at 20M tokens) val_tokens = None if is_main: logger.info("Loading validation data...") try: val_files = [f for f in api.list_repo_files(repo_id="ViuAI/ViuCorp", repo_type="dataset", token=token) if f.startswith("val_shards/")] if val_files: val_p = hf_hub_download(repo_id="ViuAI/ViuCorp", repo_type="dataset", filename=val_files[0], local_dir=work_dir, token=token) dtype = np.uint16 if vocab_size <= 65535 else np.uint32 val_arr = np.memmap(val_p, dtype=dtype, mode="r") max_toks = 20_000_000 val_tokens = np.array(val_arr[:max_toks]) # Copy to RAM, safely capped del val_arr Path(val_p).unlink(missing_ok=True) except Exception as e: logger.warning(f"Failed to load val data: {e}") @torch.no_grad() def evaluate(): if val_tokens is None or len(val_tokens) < args.ctx_len + 1: return 0.0 unwrapped_model.eval() val_loss = 0.0 val_iters = 20 for _ in range(val_iters): i = rng.integers(0, len(val_tokens) - args.ctx_len - 1) x = torch.from_numpy(val_tokens[i:i+args.ctx_len].astype(np.int64)).unsqueeze(0).to(device, non_blocking=True) y = torch.from_numpy(val_tokens[i+1:i+1+args.ctx_len].astype(np.int64)).unsqueeze(0).to(device, non_blocking=True) with torch.autocast(device_type="cuda", dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16): # Pass use_checkpoint=False for evaluation _, loss = unwrapped_model(x, y, use_checkpoint=False) val_loss += loss.item() unwrapped_model.train() return val_loss / val_iters # Training Loop model.train() if is_main: logger.info(f"Starting training for {args.max_steps} steps...") for step in range(start_step, args.max_steps): # LR Schedule warmup = 1000 progress = min(1.0, step / max(1, warmup)) lr = args.lr * progress if step < warmup else args.lr * (0.5 * (1.0 + math.cos(math.pi * (step - warmup) / max(1, args.max_steps - warmup)))) for param_group in optimizer.param_groups: param_group['lr'] = lr optimizer.zero_grad(set_to_none=True) accum_loss = 0.0 t0 = time.time() for micro_i in range(args.grad_accum): x_np, y_np = shard_pool.sample_batch(args.micro_batch, args.ctx_len, rng) x = torch.from_numpy(x_np).to(device, non_blocking=True) y = torch.from_numpy(y_np).to(device, non_blocking=True) with torch.autocast(device_type="cuda", dtype=torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16): _, loss = model(x, y) loss = loss / args.grad_accum scaler.scale(loss).backward() accum_loss += loss.item() del x, y, loss # Safe Gradient Clipping with Scaler scaler.unscale_(optimizer) torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) # Check for NaNs has_nan = not math.isfinite(accum_loss) for p in model.parameters(): if p.grad is not None and not torch.isfinite(p.grad).all(): has_nan = True break if has_nan: if is_main: logger.warning(f"NaN/Inf detected at step {step}! Skipping optimizer step.") optimizer.zero_grad(set_to_none=True) continue scaler.step(optimizer) scaler.update() dt = time.time() - t0 if is_main and step % 10 == 0: tok_s = (args.micro_batch * args.grad_accum * args.ctx_len * world_size) / dt if dt > 0 else 0 mem_alloc = torch.cuda.memory_allocated(device) / 1e9 mem_res = torch.cuda.memory_reserved(device) / 1e9 logger.info(f"step {step:5d}/{args.max_steps} | loss {accum_loss:.4f} | lr {lr:.2e} | {tok_s:.0f} tok/s | mem alloc={mem_alloc:.2f}GB res={mem_res:.2f}GB") if step > start_step and step % args.rotate_every == 0: shard_pool.rotate_one() shard_pool.fill(args.ctx_len) if step > start_step and step % args.eval_every == 0 and is_main: val_loss = evaluate() logger.info(f"--- Eval step {step}: val_loss {val_loss:.4f} ---") if step > start_step and step % args.save_every == 0: if is_ddp: dist.barrier() if is_main: logger.info(f"Saving checkpoint at step {step}...") chk = { "model": unwrapped_model.state_dict(), "optimizer": optimizer.state_dict(), "scaler": scaler.state_dict(), "step": step, } torch.save(chk, "latest_chk.pt") try: api.upload_file( path_or_fileobj="latest_chk.pt", path_in_repo="checkpoints/latest_chk.pt", repo_id="ViuAI/ViuRec", token=token ) logger.info("Upload successful.") except Exception as e: logger.error(f"Upload failed: {e}") if is_ddp: dist.barrier() if is_ddp: dist.destroy_process_group() if __name__ == "__main__": main()