diff --git a/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/README.md b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/README.md new file mode 100644 index 000000000..1aeca14fb --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/README.md @@ -0,0 +1,59 @@ +**val_bpb: 1.4525** (sliding window, stride=128, GA+VR combined) | **13.2 MB** (control model) | 1xRTX3090, 1000 steps, 131K batch + +## Non-record: Value Residual (-0.015 BPB) + Gated Attention (-0.003 BPB) with ablations + +Two novel architecture modifications and one negative result. The goal is to share validated techniques with controlled ablation data. + +### Contributions + +1. **Value Residual (ResFormer)** -- -0.015 BPB standalone. Caches raw V vectors from layer 0 and mixes them into all subsequent layers: `V_n = lambda1 * V_0 + lambda2 * V_current`. Both lambdas are learnable (init 0.5), 18 scalars total for 9L. Preserves token identity through depth, especially beneficial for deep-narrow architectures (512d). Based on arXiv:2410.17897 (ACL 2025). Enable: `VALUE_RESIDUAL=1`. + +2. **Gated Attention** -- -0.003 BPB standalone. Per-head sigmoid gate after SDPA: `Y' = Y * sigmoid(X @ W_gate + b_gate)`. Bias init 4.0 (sigmoid ~0.98, near no-op at start). Eliminates attention sinks where softmax forces distribution over irrelevant keys. ~37K params for 9L/8H. Based on arXiv:2505.06708 (NeurIPS 2025 Best Paper). Enable: `GATED_ATTENTION=1`. + +3. **PPM-C Context Mixer** -- +0.0018 BPB (negative result). Classical Prediction by Partial Matching (order 2) blended with neural softmax at eval time (`mixed = 0.95*neural + 0.05*ppm`). On SmearGate+BigramHash models, the neural predictions already capture bigram patterns; PPM just dilutes them. Zero artifact cost but no benefit. + +The two positive techniques **stack additively** for -0.017 BPB combined (no interference). + +### Ablation Results + +Controlled A/B test: v1024 9L 2xMLP, SmearGate + BigramHash(4096) + OrthoInit + WD 0.04, 131K batch, 1000 steps, RTX 3090. + +| Config | Sliding BPB | Delta vs Control | +|--------|-------------|------------------| +| Control (no novel techniques) | 1.4697 | -- | +| Gated Attention only | 1.4665 | -0.0032 | +| **Value Residual only** | **1.4546** | **-0.0151** | +| **GA + VR combined** | **1.4525** | **-0.0172** | +| PPM-C (alpha=0.95, order=2) | 1.2900* | +0.0018 (worse) | + +*PPM-C tested on stronger pre-trained model (524K batch, 2000 steps, baseline 1.2882 BPB). + +### Relationship to community techniques + +**Value Residual vs VE128 (Shared Value Embeddings):** Both preserve value information across depth, but through different mechanisms. VE128 shares a learned embedding matrix; Value Residual skip-connects layer 0's computed V vectors. Whether they are complementary or redundant has not been tested. + +**Gated Attention vs XSA:** Both address attention quality through different failure modes. XSA removes self-value bias; Gated Attention allows heads to suppress output entirely. They may be complementary. + +### Reproducibility + +Ablation cost: ~$0.70 (3x RTX 3090 pods, ~20 min each). + +```bash +TORCHDYNAMO_DISABLE=1 VOCAB_SIZE=1024 NUM_LAYERS=9 MLP_MULT=2 \ +TRAIN_SEQ_LEN=1024 TRAIN_BATCH_TOKENS=131072 ITERATIONS=1000 \ +SMEAR_GATE=1 BIGRAM_HASH=1 BIGRAM_BUCKETS=4096 ORTHO_INIT=1 \ +WEIGHT_DECAY_MUON=0.04 WEIGHT_DECAY_ADAM=0.04 \ +GATED_ATTENTION=1 VALUE_RESIDUAL=1 \ +EVAL_SEQ_LEN=1024 EVAL_STRIDE=128 QUANT_BITS=6 \ +torchrun --standalone --nproc_per_node=1 train_gpt.py +``` + +### What we'd do with more compute + +A production run (11L MLP3x + full community stack + VR + GA, 9500 steps, 524K batch) is in progress on 1xA6000. If Value Residual's -0.015 BPB holds at the frontier, it could push SOTA from ~1.125 to ~1.110 BPB. Results will be submitted in a follow-up record PR if competitive. + +### Files + +- `README.md` -- This writeup +- `submission.json` -- Submission metadata +- `train_gpt.py` -- Training script with Value Residual, Gated Attention, XSA, EMA, Partial RoPE, LN Scale implementations diff --git a/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/submission.json b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/submission.json new file mode 100644 index 000000000..953108b11 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/submission.json @@ -0,0 +1,18 @@ +{ + "author": "Anant Goel", + "github_id": "anantdgoel", + "name": "Value Residual + Gated Attention ablation", + "blurb": "Two novel architecture modifications: Value Residual (-0.015 BPB) caches layer-0 V vectors as skip connections, Gated Attention (-0.003 BPB) adds per-head sigmoid gates. They stack additively for -0.017 BPB. PPM-C context mixing is a negative result (+0.002 BPB) on SmearGate+BigramHash models.", + "date": "2026-03-22", + "track": "non-record-16mb", + "val_bpb": 1.4525, + "pre_quant_val_bpb": 1.4525, + "step_stop": 1000, + "wallclock_seconds": 5506, + "hardware": "1xRTX3090 (ablations), 1xA6000 (validation)", + "novel_contributions": [ + "Value Residual: layer-0 V shortcut to all layers, 18 learnable scalars, -0.015 BPB (arXiv:2410.17897)", + "Gated Attention: per-head sigmoid gate, ~37K params, -0.003 BPB (arXiv:2505.06708)", + "PPM-C context mixer: negative result, +0.002 BPB on SmearGate+BigramHash models" + ] +} diff --git a/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/train_gpt.py b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/train_gpt.py new file mode 100644 index 000000000..d6d4bf420 --- /dev/null +++ b/records/track_non_record_16mb/2026-03-22_ValueResidual_GatedAttention_anantdgoel/train_gpt.py @@ -0,0 +1,1645 @@ +""" +The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. + +Hard stop: `train_gpt.py` and `train_gpt_mlx.py` must never be longer than 1500 lines. +""" + +from __future__ import annotations + +import copy +import glob +import io +import json +import math +import os +import random +import subprocess +import sys +import time +import uuid +from pathlib import Path + +try: + import zstandard +except ImportError: + zstandard = None + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# HYPERPARAMETERS +# 9 transformer blocks at width 512, 8 heads with 4 KV heads (GQA), vocab 1024, tied embeddings + +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + + + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 2)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + weight_decay_muon = float(os.environ.get("WEIGHT_DECAY_MUON", 0.02)) + weight_decay_adam = float(os.environ.get("WEIGHT_DECAY_ADAM", 0.01)) + + + mtp_k = int(os.environ.get("MTP_K", 0)) + mtp_alpha = float(os.environ.get("MTP_ALPHA", 0.3)) + + + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 0)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 0)) + + + # Full-model SGD TTT at eval (adapts all weights via SGD over val data). + sgd_ttt = bool(int(os.environ.get("SGD_TTT", "0"))) + sgd_ttt_lr = float(os.environ.get("SGD_TTT_LR", 0.002)) + sgd_ttt_momentum = float(os.environ.get("SGD_TTT_MOMENTUM", 0.9)) + sgd_ttt_epochs = int(os.environ.get("SGD_TTT_EPOCHS", 2)) + sgd_ttt_seq_len = int(os.environ.get("SGD_TTT_SEQ_LEN", 1024)) + sgd_ttt_param_set = os.environ.get("SGD_TTT_PARAM_SET", "control") + # Attention pointer/copy: use last-layer attention as copy distribution. + pointer_copy = bool(int(os.environ.get("POINTER_COPY", "0"))) + pointer_copy_lambda = float(os.environ.get("POINTER_COPY_LAMBDA", 0.05)) + meta_ttt = bool(int(os.environ.get("META_TTT", "0"))) + meta_ttt_param_set = os.environ.get("META_TTT_PARAM_SET", "control") + meta_ttt_start_frac = float(os.environ.get("META_TTT_START_FRAC", 0.5)) + meta_ttt_every = int(os.environ.get("META_TTT_EVERY", 4)) + meta_ttt_seq_len = int(os.environ.get("META_TTT_SEQ_LEN", 512)) + meta_ttt_inner_lr = float(os.environ.get("META_TTT_INNER_LR", 0.03)) + meta_ttt_inner_steps = int(os.environ.get("META_TTT_INNER_STEPS", 1)) + meta_ttt_weight = float(os.environ.get("META_TTT_WEIGHT", 0.5)) + # Eval-time adaptation: unigram cache mixture and online gradient descent on vocab bias. + cache_mixture = bool(int(os.environ.get("CACHE_MIXTURE", "0"))) + cache_lambda = float(os.environ.get("CACHE_LAMBDA", 0.02)) + cache_decay = float(os.environ.get("CACHE_DECAY", 0.995)) + ogd_bias = bool(int(os.environ.get("OGD_BIAS", "0"))) + ogd_lr = float(os.environ.get("OGD_LR", 0.1)) + smear_gate = bool(int(os.environ.get("SMEAR_GATE", "0"))) + bigram_hash = bool(int(os.environ.get("BIGRAM_HASH", "0"))) + bigram_buckets = int(os.environ.get("BIGRAM_BUCKETS", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + ortho_init = bool(int(os.environ.get("ORTHO_INIT", "0"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + xsa_layers = int(os.environ.get("XSA_LAYERS", 0)) + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + partial_rope_dims = int(os.environ.get("PARTIAL_ROPE_DIMS", 0)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "0"))) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "0"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.2)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + train_only = bool(int(os.environ.get("TRAIN_ONLY", "0"))) + artifact_filename = os.environ.get("ARTIFACT_FILENAME", "") + load_artifact = os.environ.get("LOAD_ARTIFACT", "") + +def get_artifact_compressor() -> str: + compressor = os.environ.get("ARTIFACT_COMPRESSOR", "zstd" if zstandard is not None else "zlib") + if compressor not in ("zlib", "zstd"): + raise ValueError(f"ARTIFACT_COMPRESSOR must be zlib or zstd, got {compressor}") + if compressor == "zstd" and zstandard is None: + raise RuntimeError("ARTIFACT_COMPRESSOR=zstd requires zstandard to be installed") + return compressor + +# MUON OPTIMIZER (from modded-nanogpt) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + # Per-row normalization (NorMuon). + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + row_norms = g.norm(dim=-1, keepdim=True).clamp_min(1e-7) + g = g / row_norms + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + if wd > 0: + p.data.mul_(1.0 - lr * wd) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + +# TOKENIZER-AGNOSTIC EVALUATION +# BPB (bits-per-byte) instead of val loss, so we need per-token byte counts. + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +def _count_token_bytes( + prev_ids: Tensor, tgt_ids: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, +) -> float: + tb = base_bytes_lut[tgt_ids].to(torch.int16) + ( + has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids] + ).to(torch.int16) + return tb.to(torch.float64).sum().item() + +BOS_ID = 1 + +def _build_sliding_windows(val_tokens: Tensor, seq_len: int, stride: int, doc_isolated: bool): + """Build sliding window start positions. If doc_isolated, reset at BOS boundaries.""" + N = val_tokens.numel() - 1 + if not doc_isolated: + return list(range(0, N - seq_len + 1, stride)) + # Find document boundaries and build windows per document. + bos_positions = (val_tokens[:N] == BOS_ID).nonzero(as_tuple=True)[0].tolist() + if not bos_positions or bos_positions[0] != 0: + bos_positions = [0] + bos_positions + bos_positions.append(N) + starts: list[int] = [] + for di in range(len(bos_positions) - 1): + doc_start, doc_end = bos_positions[di], bos_positions[di + 1] + doc_len = doc_end - doc_start + if doc_len < 2: + continue + pos = doc_start + while pos + 1 <= doc_end: + starts.append(pos) + pos += stride + if pos + seq_len > doc_end: + # Last window aligned to document end. + if doc_end - seq_len > starts[-1]: + starts.append(max(doc_start, doc_end - seq_len)) + break + return starts + +def eval_val_sliding( + args: Hyperparameters, model: nn.Module, device: torch.device, + val_tokens: Tensor, base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float, int, float]: + """Sliding-window evaluation with optional document isolation.""" + seq_len = args.eval_seq_len or args.train_seq_len + stride = args.eval_stride or seq_len + N = val_tokens.numel() - 1 + batch_seqs = max(1, min(args.val_batch_size // seq_len, 64)) + starts = _build_sliding_windows(val_tokens, seq_len, stride, False) + loss_sum, tok_count, byte_count = 0.0, 0, 0.0 + model.eval() + with torch.inference_mode(): + for bi in range(0, len(starts), batch_seqs): + bs = starts[bi : bi + batch_seqs] + bsz = len(bs) + x = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(bs): + end = min(ws + seq_len, N) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x[i, :wlen] = chunk[:-1] + y[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = model.get_logits(x) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), y.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(bs): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum().item() + tok_count += wlen - s + byte_count += _count_token_bytes( + x[i, s:wlen], y[i, s:wlen], + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + model.train() + val_loss = loss_sum / tok_count + return val_loss, val_loss / math.log(2.0) * tok_count / byte_count, tok_count, byte_count + +# FULL-MODEL SGD TTT + ATTENTION POINTER/COPY + +def eval_val_sgd_ttt( + args: Hyperparameters, model: nn.Module, device: torch.device, + val_tokens: Tensor, base_bytes_lut: Tensor, has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float, int, float]: + """Two-phase SGD TTT: (1) adapt weights via SGD over val data, (2) score with sliding window. + Optionally mix with recency-weighted pointer/copy distribution during scoring.""" + seq_len = args.sgd_ttt_seq_len or args.train_seq_len + stride = args.eval_stride or seq_len + N = val_tokens.numel() - 1 + use_pointer = args.pointer_copy + ptr_lambda = args.pointer_copy_lambda + vocab_size = model.tok_emb.num_embeddings if hasattr(model, "tok_emb") else args.vocab_size + named_ttt_params = _named_ttt_params(model, args.sgd_ttt_param_set) + if not named_ttt_params: + raise ValueError(f"No parameters matched SGD_TTT_PARAM_SET={args.sgd_ttt_param_set}") + ttt_params = [param for _, param in named_ttt_params] + saved = {name: param.data.clone() for name, param in named_ttt_params} + for param in model.parameters(): + param.requires_grad_(False) + for param in ttt_params: + param.requires_grad_(True) + opt = torch.optim.SGD(ttt_params, lr=args.sgd_ttt_lr, momentum=args.sgd_ttt_momentum) + _reset_rotary_caches(model) + # Phase 1: Adapt — SGD over val data with non-overlapping windows (fast) + model.eval() + adapt_starts = list(range(0, N - seq_len + 1, seq_len)) + for epoch in range(args.sgd_ttt_epochs): + for s in adapt_starts: + chunk = val_tokens[s : s + seq_len + 1].to(device=device, dtype=torch.int64) + x, y = chunk[:-1].unsqueeze(0), chunk[1:].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = model.get_logits(x) + loss = F.cross_entropy(logits.reshape(-1, logits.size(-1)).float(), y.reshape(-1), reduction="mean") + opt.zero_grad() + loss.backward() + torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0) + opt.step() + # Phase 2: Score — sliding window eval on adapted model (no gradients) + # Optional eval-time adaptation: cache_mixture (decayed unigram counts) and ogd_bias (online vocab bias). + for p in model.parameters(): + p.requires_grad_(False) + use_cache = args.cache_mixture + use_ogd = args.ogd_bias + ogd_b = torch.zeros(vocab_size, device=device) if use_ogd else None + cache_counts = torch.zeros(vocab_size, device=device, dtype=torch.float64) if use_cache else None + cache_sum = 0.0 + batch_seqs = max(1, min(args.val_batch_size // seq_len, 64)) + starts = list(range(0, N - seq_len + 1, stride)) + loss_sum, tok_count, byte_count = 0.0, 0, 0.0 + with torch.inference_mode(): + for bi in range(0, len(starts), batch_seqs): + bs = starts[bi : bi + batch_seqs] + bsz = len(bs) + x = torch.stack([val_tokens[s : s + seq_len] for s in bs]).to(device=device, dtype=torch.int64) + y = torch.stack([val_tokens[s + 1 : s + seq_len + 1] for s in bs]).to(device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = model.get_logits(x) + for i, s in enumerate(bs): + sc = seq_len - stride if s > 0 else 0 + sl, st = logits[i, sc:, :].float(), y[i, sc:] + if use_ogd: + sl = sl + ogd_b + if use_cache and cache_sum > 0: + # Mix model softmax with decayed unigram cache distribution. + log_pm = F.log_softmax(sl, dim=-1) + lam = args.cache_lambda + target_log_pm = log_pm.gather(1, st.unsqueeze(1)).squeeze(1).to(torch.float64) + la = target_log_pm + math.log(1 - lam) + lb = torch.log(lam * cache_counts[st] / cache_sum + 1e-30) + loss_sum += (-torch.logaddexp(la, lb)).sum().item() + elif use_pointer and s > 0: + log_pm = F.log_softmax(sl, dim=-1).gather(1, st.unsqueeze(1)).squeeze(1) + ctx_ids = x[i, :sc] + recency = torch.exp(0.01 * (torch.arange(sc, device=device, dtype=torch.float32) - sc)) + cdist = torch.zeros(vocab_size, device=device) + cdist.scatter_add_(0, ctx_ids.long(), recency) + cdist = cdist / cdist.sum().clamp(min=1) + pc = cdist[st].to(torch.float64) + la = log_pm.to(torch.float64) + math.log(1 - ptr_lambda) + lb = torch.log(ptr_lambda * pc + 1e-30) + loss_sum += (-torch.logaddexp(la, lb)).sum().item() + else: + loss_sum += F.cross_entropy(sl, st, reduction="sum").item() + tok_count += st.numel() + byte_count += _count_token_bytes(x[i, sc:], st, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + # Update cache counts with scored tokens (decayed). + if use_cache: + decay_factor = args.cache_decay ** st.numel() + cache_counts *= decay_factor + cache_sum = cache_sum * decay_factor + st.numel() + cache_counts.scatter_add_(0, st.long(), torch.ones_like(st, dtype=torch.float64)) + # Update OGD bias: gradient of CE w.r.t. bias = softmax - one_hot. + if use_ogd: + probs = F.softmax(sl.detach(), dim=-1).mean(0) + one_hot_mean = torch.zeros(vocab_size, device=device) + one_hot_mean.scatter_add_(0, st.long(), torch.ones(st.numel(), device=device) / st.numel()) + ogd_b.sub_(args.ogd_lr * (probs - one_hot_mean)) + # Restore original weights + with torch.no_grad(): + for name, param in named_ttt_params: + param.data.copy_(saved[name]) + model.train() + val_loss = loss_sum / tok_count + return val_loss, val_loss / math.log(2.0) * tok_count / byte_count, tok_count, byte_count + +# POST-TRAINING QUANTIZATION (int8 + zlib) + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,bigram.scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +FP16_KEEP_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get("FP16_KEEP_NAME_PATTERNS", "tok_emb").split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + """Quantize to int6 [-32, 31] with per-row scaling.""" + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1e-12).to(torch.float16) + scale = scale.clamp_min(torch.finfo(torch.float16).tiny) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(max(amax / 31.0, 1e-12), dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + return q, scale + +def quantize_int8_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + """Quantize to int8 [-127, 127] with per-row scaling.""" + t32 = t.float() + if t32.ndim == 2: + clip_abs = torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) if t32.numel() else torch.empty((t32.shape[0],), dtype=torch.float32) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def mixed_quantize(state_dict: dict[str, Tensor], num_layers: int, quant_bits: int = 6): + """Mixed-precision quantization: fp16 passthrough for embeddings and last-2-layer c_k, + int6/int8 for MLP+attention weights, fp16 for small/control tensors.""" + # Build fp16 keep patterns: tok_emb + last 2 layers' c_k + fp16_patterns = list(FP16_KEEP_NAME_PATTERNS) + for li in range(max(0, num_layers - 2), num_layers): + fp16_patterns.append(f"blocks.{li}.attn.c_k") + quantize_fn = quantize_int6_per_row if quant_bits == 6 else quantize_int8_per_row + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if any(p in name for p in fp16_patterns): + result[name] = t.to(dtype=torch.float16).contiguous() + meta[name] = "passthrough_fp16" + continue + cat = _classify_param(name) + if cat in ("mlp", "attn") and t.ndim >= 1: + q, s = quantize_fn(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{quant_bits}"} + else: + q, s = quantize_int8_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return {"w": result, "m": meta} + +def dequantize_mixed(obj: dict[str, object], template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + if obj.get("__quant_format__") == "int8_clean_per_row_v1": # legacy format + out, qm, po = {}, obj.get("qmeta", {}), obj.get("passthrough_orig_dtypes", {}) + for n, q in obj["quantized"].items(): + dt, s = getattr(torch, obj["dtypes"][n]), obj["scales"][n] + out[n] = (q.float() * s.float().view(q.shape[0], *([1]*(q.ndim-1)))).to(dt) if (qm.get(n,{}).get("scheme")=="per_row" or s.ndim>0) else (q.float()*float(s.item())).to(dt) + for n, t in obj["passthrough"].items(): + out[n] = t.to(dtype=getattr(torch, po[n])) if po.get(n) else t.detach().cpu() + return out + result, meta, out = obj["w"], obj["m"], {} + for name, orig in template_sd.items(): + info = meta[name] + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + out[name] = t.to(orig.dtype) if (t.dtype == torch.float16 and orig.dtype in (torch.float32, torch.bfloat16)) else t + continue + q, s = result[name + ".q"], result[name + ".scale"] + out[name] = (q.float() * s.float().view(q.shape[0], *([1]*(q.ndim-1)))).to(orig.dtype) if s.ndim > 0 else (q.float()*float(s.item())).to(orig.dtype) + return out + +# DATA LOADING + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +# TRANSFORMER MODULES + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + +class CastedLinear(nn.Linear): + def forward(self, x: Tensor) -> Tensor: + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + +def _reset_rotary_caches(module: nn.Module) -> None: + for submodule in module.modules(): + if hasattr(submodule, "_seq_len_cached"): + submodule._seq_len_cached = 0 + +def _named_ttt_params(model: nn.Module, param_set: str) -> list[tuple[str, nn.Parameter]]: + named_params = list(model.named_parameters()) + if param_set == "all": + return named_params + if param_set not in ("control", "control_qk"): + raise ValueError(f"Unsupported TTT param set: {param_set}") + qk_names = { + f"blocks.{li}.attn.c_k.weight" + for li in range(max(0, len(model.blocks) - 2), len(model.blocks)) + } + selected: list[tuple[str, nn.Parameter]] = [] + for name, param in named_params: + is_control = any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + if is_control or (param_set == "control_qk" and name in qk_names): + selected.append((name, param)) + return selected + +def _meta_ttt_step( + args: Hyperparameters, + train_model: nn.Module, + base_model: nn.Module, + x_support: Tensor, + y_support: Tensor, + x_query: Tensor, + y_query: Tensor, + grad_scale: float, +) -> Tensor | None: + named_fast_params = _named_ttt_params(base_model, args.meta_ttt_param_set) + if not named_fast_params: + return None + fast_params = [param for _, param in named_fast_params] + saved = [param.detach().clone() for param in fast_params] + try: + for _ in range(args.meta_ttt_inner_steps): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + support_loss = train_model(x_support, y_support) + grads = torch.autograd.grad(support_loss, fast_params, allow_unused=True) + with torch.no_grad(): + for param, grad in zip(fast_params, grads, strict=True): + if grad is not None: + param.add_(grad, alpha=-args.meta_ttt_inner_lr) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + query_loss = train_model(x_query, y_query) + (args.meta_ttt_weight * query_loss * grad_scale).backward() + return query_loss.detach() + finally: + with torch.no_grad(): + for param, orig in zip(fast_params, saved, strict=True): + param.copy_(orig) + _reset_rotary_caches(base_model) + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, rope_dims: int = 0): + super().__init__() + self.rope_dims = rope_dims if rope_dims > 0 else dim + rd = self.rope_dims + inv_freq = 1.0 / (base ** (torch.arange(0, rd, 2, dtype=torch.float32) / rd)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + rd = cos.size(-1) * 2 + if rd < x.size(-1): + x_rope, x_pass = x[..., :rd], x[..., rd:] + half = rd // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rot = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rot, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + partial_rope_dims: int = 0, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base, rope_dims=partial_rope_dims) + self.use_xsa = False + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Remove self-value projection via GQA-aware reshape.""" + B, H, T, D = y.shape + Hkv = v.size(1) + group = H // Hkv + y_g = y.reshape(B, Hkv, group, T, D) + vn = F.normalize(v, dim=-1).unsqueeze(2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, H, T, D) + + def forward(self, x: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, + k, + v, + attn_mask=None, + is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + gate = torch.sigmoid(self.attn_gate(x)) + y = y * gate.unsqueeze(-1).transpose(1, 2) + y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y), raw_v + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + +class SmearGate(nn.Module): + """Blend each token's embedding with the previous token's embedding.""" + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + """Hash consecutive token pairs into a learned embedding table.""" + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: float, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + partial_rope_dims: int = 0, + layer_idx: int = 0, + ln_scale: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual, + partial_rope_dims=partial_rope_dims) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x: Tensor, x0: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self.ln_scale_factor + attn_out, raw_v = self.attn(self.attn_norm(x) * s, v0=v0) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x) * s) + return x, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: float, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_k: int = 0, + mtp_alpha: float = 0.3, + smear_gate: bool = False, + bigram_buckets: int = 0, + bigram_dim: int = 128, + ortho_init: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + xsa_layers: int = 0, + partial_rope_dims: int = 0, + ln_scale: bool = False, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.smear = SmearGate(model_dim) if smear_gate else None + self.bigram = BigramHashEmbedding(bigram_buckets, bigram_dim, model_dim) if bigram_buckets > 0 else None + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual, + partial_rope_dims=partial_rope_dims, layer_idx=i, ln_scale=ln_scale) + for i in range(num_layers) + ]) + if xsa_layers > 0: + for i in range(max(0, num_layers - xsa_layers), num_layers): + self.blocks[i].attn.use_xsa = True + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_k = mtp_k + self.mtp_alpha = mtp_alpha + self.mtp_head = CastedLinear(model_dim, vocab_size, bias=False) if mtp_k > 0 else None + if self.mtp_head is not None: + self.mtp_head._zero_init = True + self._ortho_init = ortho_init + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif self._ortho_init and module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + # muP output scaling for projection layers. + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _encode(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + if self.smear is not None: + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + for i, block in enumerate(self.blocks): + if i < self.num_encoder_layers: + x, raw_v = block(x, x0, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + else: + dec_idx = i - self.num_encoder_layers + if dec_idx < self.num_skip_weights and skips: + x = x + self.skip_weights[dec_idx].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = block(x, x0, v0=v0) + return self.final_norm(x) + + def _project_to_logits(self, hidden: Tensor) -> Tensor: + logits_proj = F.linear(hidden, self.tok_emb.weight) if self.tie_embeddings else self.lm_head(hidden) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + def get_logits(self, input_ids: Tensor) -> Tensor: + return self._project_to_logits(self._encode(input_ids)) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + hidden = self._encode(input_ids) + logits = self._project_to_logits(hidden.reshape(-1, hidden.size(-1))) + loss = F.cross_entropy(logits.float(), target_ids.reshape(-1), reduction="mean") + if self.mtp_head is not None and self.training: + k = self.mtp_k + T = hidden.size(1) + mtp_logits = self.logit_softcap * torch.tanh( + self.mtp_head(hidden[:, : T - k + 1, :]) / self.logit_softcap + ) + mtp_loss = F.cross_entropy( + mtp_logits.reshape(-1, mtp_logits.size(-1)).float(), + target_ids[:, k - 1 :].reshape(-1), + reduction="mean", + ) + loss = loss + self.mtp_alpha * mtp_loss + return loss + +# TRAINING + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + artifact_compressor = get_artifact_compressor() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # DISTRIBUTED + CUDA SETUP + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # TOKENIZER + VALIDATION METRIC SETUP + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + # MODEL + OPTIMIZER SETUP + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_k=args.mtp_k, + mtp_alpha=args.mtp_alpha, + smear_gate=args.smear_gate, + bigram_buckets=args.bigram_buckets if args.bigram_hash else 0, + bigram_dim=args.bigram_dim, + ortho_init=args.ortho_init, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + xsa_layers=args.xsa_layers, + partial_rope_dims=args.partial_rope_dims, + ln_scale=args.ln_scale, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False, find_unused_parameters=(args.gated_attention or args.value_residual)) if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + if base_model.smear is not None: + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + optimizer_tok = torch.optim.AdamW( + tok_params, betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.weight_decay_adam, fused=True, + ) + optimizer_muon = Muon( + matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay_muon, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.weight_decay_adam, fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.AdamW( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.weight_decay_adam, fused=True, + ) + optimizers.insert(1, optimizer_head) + if base_model.mtp_head is not None: + optimizer_mtp = torch.optim.AdamW( + [{"params": [base_model.mtp_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.weight_decay_adam, fused=True, + ) + optimizers.append(optimizer_mtp) + + # MTP params are excluded from the export artifact — don't count them. + export_param_names = {n for n, _ in base_model.named_parameters() if 'mtp_head' not in n} + n_params = sum(p.numel() for n, p in base_model.named_parameters() if n in export_param_names) + log0(f"model_params:{n_params}") + log0(f"layers:{len(base_model.blocks)}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0( + f"sgd_ttt:param_set:{args.sgd_ttt_param_set} meta_ttt:{args.meta_ttt} " + f"meta_param_set:{args.meta_ttt_param_set} artifact_compressor:{artifact_compressor}" + ) + log0(f"seed:{args.seed}") + + # DATA LOADER & MODEL WARMUP + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + if args.load_artifact: + log0(f"load_artifact: loading pre-trained model from {args.load_artifact}") + with open(args.load_artifact, "rb") as f: + la_blob = f.read() + if artifact_compressor == "zstd": + la_dctx = zstandard.ZstdDecompressor() + la_raw = la_dctx.decompress(la_blob) + else: + import zlib + la_raw = zlib.decompress(la_blob) + la_state = torch.load(io.BytesIO(la_raw), map_location="cpu") + la_template = {k: v.detach().cpu() for k, v in base_model.state_dict().items() if 'mtp_head' not in k} + la_sd = dequantize_mixed(la_state, la_template) + base_model.load_state_dict(la_sd, strict=(args.mtp_k == 0)) + log0("load_artifact: model loaded, skipping training, proceeding to eval") + quant_filename = args.load_artifact + sd_cpu = la_template + if args.eval_stride > 0: + lut_args = (base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + sw_loss, sw_bpb, sw_tokens, sw_bytes = eval_val_sliding(args, base_model, device, val_tokens, *lut_args) + log0( + f"sliding_window eval_stride:{args.eval_stride} val_loss:{sw_loss:.4f} val_bpb:{sw_bpb:.4f} " + f"scored_tokens:{sw_tokens} scored_bytes:{sw_bytes:.0f}" + ) + log0(f"sliding_window_exact val_loss:{sw_loss:.8f} val_bpb:{sw_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # MAIN TRAINING LOOP + + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + training_time_ms = 0.0 + stop_after_step: int | None = None + meta_global_tokens = max(args.meta_ttt_seq_len * world_size * grad_accum_steps, args.meta_ttt_seq_len) + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + run_meta_ttt = ( + args.meta_ttt + and args.meta_ttt_every > 0 + and step >= int(args.iterations * args.meta_ttt_start_frac) + and step % args.meta_ttt_every == 0 + ) + zero_grad_all() + train_loss = torch.zeros((), device=device) + meta_query_loss = torch.zeros((), device=device) if run_meta_ttt else None + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if run_meta_ttt: + x_support, y_support = train_loader.next_batch(meta_global_tokens, args.meta_ttt_seq_len, grad_accum_steps) + x_query, y_query = train_loader.next_batch(meta_global_tokens, args.meta_ttt_seq_len, grad_accum_steps) + q_loss = _meta_ttt_step(args, model, base_model, x_support, y_support, x_query, y_query, grad_scale) + if q_loss is not None and meta_query_loss is not None: + meta_query_loss += q_loss.detach() + train_loss /= grad_accum_steps + if meta_query_loss is not None: + meta_query_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + + # EMA: update shadow weights every step. + if ema_state is not None: + d = args.ema_decay + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(d).add_(t.detach().float(), alpha=1.0 - d) + + # SWA: collect checkpoints during warmdown phase (disabled when EMA is active). + if args.swa_enabled and not args.ema_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().float() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu().float() + swa_count += 1 + + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + msg = ( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + if meta_query_loss is not None: + msg += f" meta_query_loss:{meta_query_loss.item():.4f}" + log0(msg) + + # Needed to sync whether we've reached the wallclock cap. + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + # Apply EMA if enabled. + if ema_state is not None: + log0("ema:applying EMA weights") + avg_sd = {name: t.to(dtype=base_model.state_dict()[name].dtype) for name, t in ema_state.items()} + del ema_state + base_model.load_state_dict(avg_sd, strict=True) + + # Apply SWA if collected (only when EMA is not used). + if args.swa_enabled and not args.ema_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_sd = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + base_model.load_state_dict(avg_sd, strict=True) + + # SERIALIZATION + ROUNDTRIP VALIDATION + # Exclude MTP head from export (training-only auxiliary head). + export_sd = {k: v for k, v in base_model.state_dict().items() if 'mtp_head' not in k} + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_bits = int(os.environ.get("QUANT_BITS", "6")) + quant_label = f"int{quant_bits}+{artifact_compressor}" + quant_obj = mixed_quantize(sd_cpu, args.num_layers, quant_bits=quant_bits) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + if artifact_compressor == "zstd": + cctx = zstandard.ZstdCompressor(level=22) + quant_blob = cctx.compress(quant_raw) + else: + import zlib + quant_blob = zlib.compress(quant_raw, level=9) + quant_filename = args.artifact_filename or f"final_model.{quant_label}.ptz" + if master_process: + with open(quant_filename, "wb") as f: + f.write(quant_blob) + with open("artifact_info.json", "w", encoding="utf-8") as f: + json.dump({ + "artifact_filename": quant_filename, + "artifact_compressor": artifact_compressor, + "quant_bits": quant_bits, + "tokenizer_path": args.tokenizer_path, + "data_path": args.data_path, + }, f, indent=2) + quant_file_bytes = os.path.getsize(quant_filename) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model {quant_label}: {quant_file_bytes} bytes") + log0(f"Total submission size {quant_label}: {quant_file_bytes + code_bytes} bytes") + log0(f"artifact_filename:{quant_filename} artifact_compressor:{artifact_compressor}") + + if args.train_only: + log0("train_only: skipping eval, model saved.") + if distributed: + dist.destroy_process_group() + return + + if distributed: + dist.barrier() + with open(quant_filename, "rb") as f: + quant_blob_disk = f.read() + if artifact_compressor == "zstd": + dctx = zstandard.ZstdDecompressor() + quant_raw_disk = dctx.decompress(quant_blob_disk) + else: + import zlib + quant_raw_disk = zlib.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_raw_disk), map_location="cpu") + base_model.load_state_dict(dequantize_mixed(quant_state, sd_cpu), strict=(args.mtp_k == 0)) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_{quant_label}_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_quant_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + # Special eval modes (sliding window, SGD TTT) + lut_args = (base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + if args.eval_stride > 0: + sw_loss, sw_bpb, sw_tokens, sw_bytes = eval_val_sliding(args, base_model, device, val_tokens, *lut_args) + log0( + f"sliding_window eval_stride:{args.eval_stride} val_loss:{sw_loss:.4f} val_bpb:{sw_bpb:.4f} " + f"scored_tokens:{sw_tokens} scored_bytes:{sw_bytes:.0f}" + ) + log0(f"sliding_window_exact val_loss:{sw_loss:.8f} val_bpb:{sw_bpb:.8f}") + if args.sgd_ttt: + original_param_set = args.sgd_ttt_param_set + eval_modes = [("sgd_ttt", original_param_set)] + if original_param_set != "all": + eval_modes.append(("sgd_ttt_full", "all")) + for label, param_set in eval_modes: + args.sgd_ttt_param_set = param_set + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb, ttt_tokens, ttt_bytes = eval_val_sgd_ttt( + args, base_model, device, val_tokens, *lut_args) + torch.cuda.synchronize() + log0( + f"{label} param_set:{param_set} lr:{args.sgd_ttt_lr} momentum:{args.sgd_ttt_momentum} " + f"epochs:{args.sgd_ttt_epochs} pointer:{args.pointer_copy} cache:{args.cache_mixture} " + f"ogd:{args.ogd_bias} val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"scored_tokens:{ttt_tokens} scored_bytes:{ttt_bytes:.0f} " + f"time:{1000*(time.perf_counter()-t_ttt):.0f}ms" + ) + log0(f"{label}_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + args.sgd_ttt_param_set = original_param_set + + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main()