diff --git a/records/track_10min_16mb/2026-03-20_NovelSOTA/README.md b/records/track_10min_16mb/2026-03-20_NovelSOTA/README.md new file mode 100644 index 0000000000..b3bf205a3e --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_NovelSOTA/README.md @@ -0,0 +1,36 @@ +# Non-record: 11L int5/int6 + XSA + online TTT w/ decay prior (single-run val_bpb=1.1520) + +Built on the stack from PRs #198, #180, #162, #164, #265, #254. + +## What's new here + +- **Pre-Q/K RMSNorm**: extra `rms_norm` on attention input before Q and K projections only (V gets raw input). Motivated by Steinmetz et al. 2025; stabilizes the RoPE-facing path under int5/int6. +- **Online causal TTT with decay prior**: full-weight SGD adaptation during eval, but with a Krause-style decay (`p += λ(p₀ − p)` after each step) to prevent drift. Adapts MLP weights in the last 3 blocks only, following TTT-E2E's finding that attention is unstable to adapt. +- **Reptile meta-learning (last 10%)**: K=1 inner SGD step + Reptile interpolation in the final 10% of training. Teaches the model to be adaptable for eval-time TTT. + +## Stack (from prior work) + +11L 512d 8h/4kv, MLP 3×, relu², tied fp16 embed, vocab 1024, seq 2048, U-Net skips, SmearGate, BigramHash(10240), OrthoInit + muP, Muon WD=0.04, SWA/200, int5-MLP/int6-attn + zstd-22, XSA in last 3 layers (#265), sliding window stride=64. + +## Results + +| Seed | val_bpb (TTT+sliding) | val_bpb (roundtrip, non-sliding) | Artifact | +|------|-----------------------|----------------------------------|----------| +| 1337 | 1.1520 | see train.log | 15.1 MB | + +Single seed, not a record submission. Posting as a non-record to share the TTT+decay approach. + +## Reproduce + +```bash +python3 data/cached_challenge_fineweb.py --variant sp1024 +SEED=1337 torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## References + +- Krause et al. 2017 (dynamic evaluation / decay prior): arXiv:1709.07432 +- Steinmetz et al. 2025 (extra RMSNorm): arXiv:2505.08823 +- Sun et al. 2025 (TTT-E2E): arXiv:2512.23675 +- Zhai 2026 (XSA): arXiv:2603.09078 +- Nichol & Schulman 2018 (Reptile): arXiv:1803.02999 diff --git a/records/track_10min_16mb/2026-03-20_NovelSOTA/submission.json b/records/track_10min_16mb/2026-03-20_NovelSOTA/submission.json new file mode 100644 index 0000000000..b4bb0cb248 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_NovelSOTA/submission.json @@ -0,0 +1,11 @@ +{ + "author": "Jack Young", + "github_id": "JWLBOYCE", + "name": "11L int5/int6 + XSA + online TTT w/ decay prior", + "blurb": "Standard SOTA stack plus pre-Q/K RMSNorm, online causal TTT with Krause decay prior, and Reptile meta-learning", + "date": "2026-03-21", + "val_loss": 1.9452, + "val_bpb": 1.1520, + "bytes_total": 15164971, + "bytes_code": 61581 +} diff --git a/records/track_10min_16mb/2026-03-20_NovelSOTA/train.log b/records/track_10min_16mb/2026-03-20_NovelSOTA/train.log new file mode 100644 index 0000000000..63a1974819 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_NovelSOTA/train.log @@ -0,0 +1,116 @@ +logs/seed1337.txt +val_tokens:62021633 (raw:62021846) +model_params:27878489 layers:11 dim:512 mlp_mult:3 +matrix_lr:0.025 muon_wd:0.04 adam_wd:0.04 grad_clip:0.3 +seq_len:2048 warmdown:3000 swa:True/200 +seed:1337 world_size:8 xsa_last_n:3 +meta_ttt:True start_frac:0.9 inner_steps:1 +extra_linear_rmsnorm:False meta_ttt_log_every:50 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9308 val_bpb:4.1048 train_time:0ms step_avg:0.11ms +step:1/20000 train_loss:6.9315 train_time:126ms step_avg:126.20ms +step:2/20000 train_loss:8.6088 train_time:184ms step_avg:91.87ms +step:3/20000 train_loss:8.5768 train_time:254ms step_avg:84.74ms +step:4/20000 train_loss:8.5364 train_time:326ms step_avg:81.48ms +step:5/20000 train_loss:6.8914 train_time:385ms step_avg:76.98ms +step:6/20000 train_loss:7.4635 train_time:454ms step_avg:75.59ms +step:7/20000 train_loss:6.6596 train_time:522ms step_avg:74.60ms +step:8/20000 train_loss:6.5727 train_time:591ms step_avg:73.87ms +step:9/20000 train_loss:6.3492 train_time:660ms step_avg:73.30ms +step:10/20000 train_loss:6.0478 train_time:729ms step_avg:72.86ms +step:200/20000 train_loss:2.7013 train_time:13926ms step_avg:69.63ms +step:400/20000 train_loss:2.2225 train_time:27855ms step_avg:69.64ms +step:600/20000 train_loss:2.4434 train_time:41820ms step_avg:69.70ms +step:800/20000 train_loss:2.1991 train_time:55817ms step_avg:69.77ms +step:1000/20000 train_loss:2.3044 train_time:69821ms step_avg:69.82ms +step:1000/20000 val_loss:2.2540 val_bpb:1.3349 train_time:69836ms step_avg:69.84ms +step:1200/20000 train_loss:2.3298 train_time:83836ms step_avg:69.86ms +step:1400/20000 train_loss:2.3668 train_time:97845ms step_avg:69.89ms +step:1600/20000 train_loss:2.0337 train_time:111848ms step_avg:69.91ms +step:1800/20000 train_loss:2.1397 train_time:125835ms step_avg:69.91ms +step:2000/20000 train_loss:2.1815 train_time:139806ms step_avg:69.90ms +step:2000/20000 val_loss:2.1645 val_bpb:1.2819 train_time:139821ms step_avg:69.91ms +checkpoint saved: ckpt_step2000.pt +step:2200/20000 train_loss:1.9967 train_time:153839ms step_avg:69.93ms +step:2400/20000 train_loss:2.1533 train_time:167785ms step_avg:69.91ms +step:2600/20000 train_loss:2.3611 train_time:181742ms step_avg:69.90ms +step:2800/20000 train_loss:2.1717 train_time:195681ms step_avg:69.89ms +step:3000/20000 train_loss:2.1580 train_time:209617ms step_avg:69.87ms +step:3000/20000 val_loss:2.1228 val_bpb:1.2573 train_time:209633ms step_avg:69.88ms +step:3200/20000 train_loss:2.1186 train_time:223554ms step_avg:69.86ms +step:3400/20000 train_loss:2.0955 train_time:237476ms step_avg:69.85ms +step:3600/20000 train_loss:2.0377 train_time:251406ms step_avg:69.83ms +step:3800/20000 train_loss:2.1456 train_time:265334ms step_avg:69.82ms +step:4000/20000 train_loss:2.1133 train_time:279257ms step_avg:69.81ms +step:4000/20000 val_loss:2.1045 val_bpb:1.2464 train_time:279272ms step_avg:69.82ms +checkpoint saved: ckpt_step4000.pt +step:4200/20000 train_loss:2.1081 train_time:293315ms step_avg:69.84ms +step:4400/20000 train_loss:2.0458 train_time:307231ms step_avg:69.83ms +step:4600/20000 train_loss:1.9071 train_time:321149ms step_avg:69.81ms +step:4800/20000 train_loss:2.2009 train_time:335089ms step_avg:69.81ms +step:5000/20000 train_loss:1.9578 train_time:349049ms step_avg:69.81ms +step:5000/20000 val_loss:2.0944 val_bpb:1.2404 train_time:349065ms step_avg:69.81ms +step:5200/20000 train_loss:2.1162 train_time:362949ms step_avg:69.80ms +step:5400/20000 train_loss:2.1372 train_time:376847ms step_avg:69.79ms +step:5600/20000 train_loss:2.1297 train_time:390763ms step_avg:69.78ms +step:5800/20000 train_loss:2.0792 train_time:404674ms step_avg:69.77ms +step:6000/20000 train_loss:2.1539 train_time:418581ms step_avg:69.76ms +step:6000/20000 val_loss:2.0788 val_bpb:1.2312 train_time:418597ms step_avg:69.77ms +checkpoint saved: ckpt_step6000.pt +step:6200/20000 train_loss:2.0245 train_time:432497ms step_avg:69.76ms +step:6400/20000 train_loss:2.1455 train_time:446397ms step_avg:69.75ms +step:6600/20000 train_loss:2.0532 train_time:460297ms step_avg:69.74ms +step:6800/20000 train_loss:2.0991 train_time:474197ms step_avg:69.74ms +step:7000/20000 train_loss:2.1310 train_time:488097ms step_avg:69.73ms +step:7000/20000 val_loss:2.0337 val_bpb:1.2045 train_time:488113ms step_avg:69.73ms +step:7200/20000 train_loss:2.0245 train_time:501997ms step_avg:69.72ms +step:7400/20000 train_loss:2.0812 train_time:515897ms step_avg:69.72ms +step:7600/20000 train_loss:2.0116 train_time:529797ms step_avg:69.71ms +step:7800/20000 train_loss:1.9853 train_time:543697ms step_avg:69.71ms +meta_ttt:inner step:7847/20000 inner:1/1 loss:1.9632 +meta_ttt:inner step:7850/20000 inner:1/1 loss:1.9291 +meta_ttt:inner step:7855/20000 inner:1/1 loss:2.0563 +meta_ttt:inner step:7860/20000 inner:1/1 loss:2.0297 +meta_ttt:inner step:7865/20000 inner:1/1 loss:1.9586 +meta_ttt:inner step:7870/20000 inner:1/1 loss:1.9981 +meta_ttt:inner step:7875/20000 inner:1/1 loss:2.0206 +meta_ttt:inner step:7880/20000 inner:1/1 loss:2.0113 +meta_ttt:inner step:7882/20000 inner:1/1 loss:1.9841 +stopping_early: wallclock_cap train_time:600000ms step:7883 +peak_mem: 14844 MiB +post_train: skipping SWA averaging +post_train: saving raw checkpoint +Raw checkpoint saved: 107886527 bytes +post_train: quantizing artifact +Artifact: 15103390 bytes, code: 61581 bytes, total: 15164971 bytes +post_train: loading quantized roundtrip +post_train: starting roundtrip eval +final_roundtrip val_loss:1.9688 val_bpb:1.1660 eval_time:8542ms +final_roundtrip_exact val_loss:1.96884521 val_bpb:1.16601372 +post_train: starting online TTT eval +ttt_eval:progress windows:3550/15143 rank:0 partial_bpb:1.1582 +ttt_eval:progress windows:7000/15143 rank:0 partial_bpb:1.1555 +ttt_eval:progress windows:10000/15143 rank:0 partial_bpb:1.1545 +ttt_eval:progress windows:13000/15143 rank:0 partial_bpb:1.1551 +ttt_eval:progress windows:15100/15143 rank:0 partial_bpb:1.1539 +final_ttt val_loss:1.9452 val_bpb:1.1520 eval_time:311712ms +final_ttt_exact val_loss:1.94518395 val_bpb:1.15204743 diff --git a/records/track_10min_16mb/2026-03-20_NovelSOTA/train_gpt.py b/records/track_10min_16mb/2026-03-20_NovelSOTA/train_gpt.py new file mode 100644 index 0000000000..024f0077b8 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_NovelSOTA/train_gpt.py @@ -0,0 +1,1327 @@ +""" +11L int5/int6 + XSA + online TTT w/ decay prior. +Built on #198, #180, #162, #164, #265, #254. +torchrun --standalone --nproc_per_node=8 train_gpt.py +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import time +import uuid +from pathlib import Path + +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 + +# zstandard is required - fail fast if missing rather than silently falling back +# to zlib (which produces a different, larger artifact that may exceed 16MB). +import zstandard as _zstd + +_ZSTD_COMPRESSOR = _zstd.ZstdCompressor(level=22) +_ZSTD_DECOMPRESSOR = _zstd.ZstdDecompressor() + +def compress_blob(data: bytes) -> bytes: + return _ZSTD_COMPRESSOR.compress(data) + +def decompress_blob(data: bytes) -> bytes: + return _ZSTD_DECOMPRESSOR.decompress(data) + +COMPRESS_EXT = "int6.zst" + +# ───────────────────────────────────────────── +# HYPERPARAMETERS +# ───────────────────────────────────────────── + +class Hyp: + 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", 11)) + 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 = int(os.environ.get("MLP_MULT", 3)) + 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)) + + # Optimizer + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + 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)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + 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)) + + # SWA + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 200)) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.5)) + + # SmearGate + BigramHash + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 10240)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + # XSA (Exclusive Self Attention) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 3)) + + # Extra RMSNorm before every linear projection is experimental and off by default. + extra_linear_rmsnorm = bool(int(os.environ.get("EXTRA_LINEAR_RMSNORM", "0"))) + + # Sliding window eval + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + + # Safety: checkpoint interval + ckpt_every = int(os.environ.get("CKPT_EVERY", 2000)) + + # Reptile Meta-TTT + meta_ttt_enabled = bool(int(os.environ.get("META_TTT_ENABLED", "1"))) + meta_ttt_start_frac = float(os.environ.get("META_TTT_START_FRAC", 0.90)) + meta_ttt_inner_steps = int(os.environ.get("META_TTT_INNER_STEPS", 1)) + meta_ttt_inner_lr = float(os.environ.get("META_TTT_INNER_LR", 2e-3)) + meta_ttt_epsilon = float(os.environ.get("META_TTT_EPSILON", 0.3)) + meta_ttt_log_every = int(os.environ.get("META_TTT_LOG_EVERY", 50)) + + # Online Causal TTT Eval + ttt_eval_lr = float(os.environ.get("TTT_EVAL_LR", 2e-3)) + ttt_eval_momentum = float(os.environ.get("TTT_EVAL_MOMENTUM", 0.9)) + ttt_eval_grad_clip = float(os.environ.get("TTT_EVAL_GRAD_CLIP", 1.0)) + ttt_eval_decay = float(os.environ.get("TTT_EVAL_DECAY", 0.02)) + ttt_eval_adapt_last_n = int(os.environ.get("TTT_EVAL_ADAPT_LAST_N", 3)) + + +CONTROL_PATTERNS = ( + "attn_scale", "mlp_scale", "resid_mix", "q_gain", + "skip_weight", "skip_weights", "smear", "bigram_scale", +) + +# ───────────────────────────────────────────── +# MUON OPTIMIZER WITH WEIGHT DECAY +# ───────────────────────────────────────────── + +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, + weight_decay: float = 0.0, nesterov: bool = True): + super().__init__(params, dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + weight_decay=weight_decay, nesterov=nesterov)) + + @torch.no_grad() + def step(self, closure=None): + 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"] + wd = group["weight_decay"] + + 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) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + 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() + + +# ───────────────────────────────────────────── +# TOKENIZER + BPB UTILS +# ───────────────────────────────────────────── + +def build_sentencepiece_luts(sp, vocab_size, device): + 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("\u2581"): + 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 _reduce_bpb(val_loss_sum, val_token_count, val_byte_count): + """All-reduce across ranks and compute (val_loss, bits_per_byte).""" + 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).item() + bpb = (val_loss / math.log(2.0)) * (val_token_count.item() / val_byte_count.item()) + return val_loss, bpb + + +# ───────────────────────────────────────────── +# SLIDING WINDOW EVAL (stride=64) +# ───────────────────────────────────────────── + +def eval_val_sliding( + base_model, val_tokens, eval_seq_len, eval_stride, device, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + rank, world_size, batch_size=32, +): + """Sliding window eval with batching for GPU efficiency. + + Windows are padded to eval_seq_len and processed in batches of `batch_size`, + giving ~30x speedup over single-window iteration. With ~10M val tokens, + stride=64, 8 GPUs: ~19.5K windows/rank / 32 per batch = ~610 forward + passes, well within the 10-minute eval budget. + """ + total_tokens = val_tokens.numel() + window_starts = list(range(0, total_tokens - 1, eval_stride)) + my_starts = [ws for i, ws in enumerate(window_starts) if i % world_size == rank] + + 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) + + sc = base_model.logit_softcap + base_model.eval() + with torch.inference_mode(): + for batch_idx in range(0, len(my_starts), batch_size): + batch_ws_list = my_starts[batch_idx : batch_idx + batch_size] + + # Pre-filter valid windows and collect metadata + x_list = [] + y_list = [] + score_ranges = [] + for ws in batch_ws_list: + wend = min(ws + eval_seq_len + 1, total_tokens) + wlen = wend - ws - 1 + if wlen < 1: + continue + score_start = 0 if ws == 0 else max(eval_seq_len - eval_stride, 0) + if score_start >= wlen: + continue + chunk = val_tokens[ws : wend].to(dtype=torch.int64) + x_list.append(chunk[:-1]) + y_list.append(chunk[1:]) + score_ranges.append((score_start, wlen)) + + if not x_list: + continue + + # Pad to uniform length and stack into a batch tensor + max_len = max(t.numel() for t in x_list) + B = len(x_list) + x_batch = torch.zeros(B, max_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(B, max_len, dtype=torch.int64, device=device) + for i, (xi, yi) in enumerate(zip(x_list, y_list)): + L = xi.numel() + x_batch[i, :L] = xi + y_batch[i, :L] = yi + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + raw_logits = base_model.forward_logits(x_batch) + + logits = sc * torch.tanh(raw_logits / sc) + + for i, (score_start, wlen) in enumerate(score_ranges): + window_logits = logits[i, :wlen].float() + window_y = y_batch[i, :wlen] + nll = F.cross_entropy(window_logits, window_y, reduction="none") + + scored_nll = nll[score_start:wlen] + scored_x = x_batch[i, score_start:wlen] + scored_y = y_batch[i, score_start:wlen] + + val_loss_sum += scored_nll.to(torch.float64).sum() + val_token_count += float(scored_nll.numel()) + + token_bytes = base_bytes_lut[scored_y].to(torch.int16) + token_bytes += (has_leading_space_lut[scored_y] & ~is_boundary_token_lut[scored_x]).to(torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + val_loss, bpb = _reduce_bpb(val_loss_sum, val_token_count, val_byte_count) + base_model.train() + return val_loss, bpb + + +def eval_val_simple(args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut): + """Non-sliding eval for mid-training validation (faster).""" + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + 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() + + val_loss, bpb = _reduce_bpb(val_loss_sum, val_token_count, val_byte_count) + model.train() + return val_loss, bpb + + +# ───────────────────────────────────────────── +# ONLINE CAUSAL TTT EVAL WITH DECAY PRIOR +# ───────────────────────────────────────────── + +def eval_val_ttt( + base_model, val_tokens, eval_seq_len, eval_stride, device, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + rank, world_size, args, log_fn=None, +): + """Online Causal TTT Eval: interleave scoring and adaptation with decay prior. + + TTT requires sequential window processing (each backward adapts the model + for subsequent windows), so batching is not possible. Instead, we use a + coarser stride for the adaptation+scoring loop to keep wall-clock time + within budget: + + ttt_stride = max(eval_stride, eval_seq_len // 4) + + With eval_seq_len=2048 this gives ttt_stride=512, producing ~19.5K total + windows (~2.4K per rank with 8 GPUs). At ~5ms per forward+backward, + that's ~12 seconds -- well within the 10-minute eval cap. + + For each sliding window: + 1. Forward pass -> score tokens + 2. Backward pass -> SGD update on MLP params in last N blocks + 3. Apply decay prior: p.data += decay * (theta_original - p.data) + """ + total_tokens = val_tokens.numel() + # Use coarser stride for TTT to stay within wall-clock budget. + # Standard sliding-window with stride=64 would create ~156K windows, + # each needing forward+backward -- far too slow for a 10-minute cap. + ttt_stride = max(eval_stride, eval_seq_len // 4) + window_starts = list(range(0, total_tokens - 1, ttt_stride)) + my_starts = [ws for i, ws in enumerate(window_starts) if i % world_size == rank] + + 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) + + # Identify MLP params in last N blocks to adapt + adapt_last_n = args.ttt_eval_adapt_last_n + num_blocks = len(base_model.blocks) + adapt_block_indices = list(range(max(0, num_blocks - adapt_last_n), num_blocks)) + adapt_params = [] + for bi in adapt_block_indices: + block = base_model.blocks[bi] + for name, p in block.mlp.named_parameters(): + if "weight" in name and ("fc" in name or "proj" in name): + adapt_params.append(p) + + # Save original params for decay prior + theta_original = {id(p): p.data.clone() for p in adapt_params} + + # Set up SGD optimizer for adaptation + ttt_optimizer = torch.optim.SGD( + adapt_params, lr=args.ttt_eval_lr, momentum=args.ttt_eval_momentum, + ) + + decay = args.ttt_eval_decay + sc = base_model.logit_softcap + + base_model.eval() + # We need gradients for TTT adaptation, so no inference_mode + for p in base_model.parameters(): + p.requires_grad_(False) + for p in adapt_params: + p.requires_grad_(True) + + total_my_windows = len(my_starts) + for window_idx, ws in enumerate(my_starts): + wend = min(ws + eval_seq_len + 1, total_tokens) + wlen = wend - ws - 1 + if wlen < 1: + continue + chunk = val_tokens[ws : wend].to(device=device, dtype=torch.int64) + x = chunk[:-1].unsqueeze(0) + y = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + raw_logits = base_model.forward_logits(x) + + logits = sc * torch.tanh(raw_logits.squeeze(0) / sc) + nll = F.cross_entropy(logits.float(), y, reduction="none") + + score_start = 0 if ws == 0 else max(eval_seq_len - ttt_stride, 0) + if score_start >= wlen: + continue + scored_nll = nll[score_start:wlen] + scored_x = x.squeeze(0)[score_start:wlen] + scored_y = y[score_start:wlen] + + val_loss_sum += scored_nll.detach().to(torch.float64).sum() + val_token_count += float(scored_nll.numel()) + + token_bytes = base_bytes_lut[scored_y].to(torch.int16) + token_bytes += (has_leading_space_lut[scored_y] & ~is_boundary_token_lut[scored_x]).to(torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + # Backward pass for adaptation (use mean NLL over full window) + adapt_loss = nll.mean() + ttt_optimizer.zero_grad() + adapt_loss.backward() + + if args.ttt_eval_grad_clip > 0: + torch.nn.utils.clip_grad_norm_(adapt_params, args.ttt_eval_grad_clip) + + ttt_optimizer.step() + + if decay > 0: + with torch.no_grad(): + for p in adapt_params: + p.data.add_(theta_original[id(p)] - p.data, alpha=decay) + + if log_fn is not None and total_my_windows > 0 and ( + (window_idx + 1) % 50 == 0 or (window_idx + 1) == total_my_windows + ): + log_fn( + f"ttt_eval:progress windows:{window_idx + 1}/{total_my_windows} " + f"rank:{rank} partial_bpb:{(val_loss_sum.item() / max(val_token_count.item(), 1.0)) / math.log(2.0) * (val_token_count.item() / max(val_byte_count.item(), 1.0)):.4f}" + ) + + # Restore original params after TTT eval + with torch.no_grad(): + for p in adapt_params: + p.data.copy_(theta_original[id(p)]) + + # Re-enable gradients for all params + for p in base_model.parameters(): + p.requires_grad_(True) + + val_loss, bpb = _reduce_bpb(val_loss_sum, val_token_count, val_byte_count) + base_model.train() + return val_loss, bpb + + +# ───────────────────────────────────────────── +# MIXED INT5/INT6 QUANTIZATION + ZSTD +# ───────────────────────────────────────────── +# Int5 for MLP weights ([-16,15], 3 zero high bits → zstd compresses at ~1.88x) +# Int6 for attention weights ([-32,31], 2 zero high bits → zstd compresses at ~1.51x) +# This saves ~1.8MB vs all-int6, funding BigramHash(10240). + +INT6_CLIP = 31 # [-32, 31] +INT5_CLIP = 15 # [-16, 15] +FP16_KEEP_PATTERNS = ("tok_emb",) +MLP_PATTERNS = (".mlp.",) # tensors matching these get int5 + +def _quantize_intN_per_row(t: Tensor, clip: int) -> tuple[Tensor, Tensor]: + t32 = t.float() + row_max = t32.abs().amax(dim=1) + scale = (row_max / clip).clamp_min(1.0 / clip).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale[:, None].float()), -(clip + 1), clip).to(torch.int8) + return q.contiguous(), scale.contiguous() + +def quantize_state_dict_int6(state_dict): + quantized, scales, dtypes = {}, {}, {} + passthrough, passthrough_orig_dtypes = {}, {} + stats = {"param_count": 0, "payload_bytes": 0} + + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + stats["param_count"] += t.numel() + + # FP16 passthrough for embeddings + if any(p in name for p in FP16_KEEP_PATTERNS): + pt = t.to(torch.float16).contiguous() + passthrough[name] = pt + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["payload_bytes"] += pt.numel() * pt.element_size() + continue + + # Small tensors: keep as fp16 + if t.numel() <= 65536: + if any(p in name for p in CONTROL_PATTERNS): + passthrough[name] = t.float().contiguous() + stats["payload_bytes"] += t.numel() * 4 + elif t.is_floating_point(): + pt = t.to(torch.float16).contiguous() + passthrough[name] = pt + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["payload_bytes"] += pt.numel() * 2 + else: + passthrough[name] = t + stats["payload_bytes"] += t.numel() * t.element_size() + continue + + # Large 2D float tensors: int5 for MLP, int6 for attention/other + if t.is_floating_point() and t.ndim == 2: + clip = INT5_CLIP if any(p in name for p in MLP_PATTERNS) else INT6_CLIP + q, s = _quantize_intN_per_row(t, clip) + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["payload_bytes"] += q.numel() * 1 + s.numel() * 2 + elif t.is_floating_point(): + # 1D float: per-tensor int8 + clip_abs = float(t.float().abs().max().item()) if t.numel() else 0.0 + sc = max(clip_abs / 127.0, 1.0 / 127.0) + q = torch.clamp(torch.round(t.float() / sc), -127, 127).to(torch.int8).contiguous() + quantized[name] = q + scales[name] = torch.tensor(sc, dtype=torch.float32) + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["payload_bytes"] += q.numel() * 1 + 4 + else: + passthrough[name] = t + stats["payload_bytes"] += t.numel() * t.element_size() + + obj = { + "__quant_format__": "int6_per_row_v1", + "quantized": quantized, "scales": scales, "dtypes": dtypes, + "passthrough": passthrough, "passthrough_orig_dtypes": passthrough_orig_dtypes, + } + return obj, stats + +def dequantize_state_dict_int6(obj): + out = {} + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype).contiguous() + else: + out[name] = (q.float() * float(s.item())).to(dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().cpu().contiguous() + orig = passthrough_orig_dtypes.get(name) + if isinstance(orig, str): + out_t = out_t.to(getattr(torch, orig)).contiguous() + out[name] = out_t + return out + + +# ───────────────────────────────────────────── +# DATA LOADING +# ───────────────────────────────────────────── + +def load_data_shard(file): + header_bytes = 256 * np.dtype(" 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance() + 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, rank, world_size, device): + self.rank, self.world_size, self.device = rank, world_size, device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens, seq_len, grad_accum_steps): + 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=None): + super().__init__() + self.eps = eps + def forward(self, x): + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x): + return F.linear(x, self.weight.to(x.dtype), self.bias.to(x.dtype) if self.bias is not None else None) + + +class SmearGate(nn.Module): + """Learned per-dim gate blending each token with previous token embedding.""" + def __init__(self, dim): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x): + 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 BigramHash(nn.Module): + """Hash-based bigram embedding for additive token-pair context.""" + def __init__(self, bigram_vocab, bigram_dim, model_dim): + super().__init__() + self.bigram_vocab = bigram_vocab + self.embed = nn.Embedding(bigram_vocab, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) + nn.init.zeros_(self.proj.weight) + self.bigram_scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def forward(self, token_ids): + t = token_ids.to(torch.int32) + mod = self.bigram_vocab - 1 + h = torch.empty_like(t) + h[..., 0] = mod # sentinel for position 0 + h[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + emb = self.embed(h.long()) + return self.proj(emb) * self.bigram_scale.to(dtype=emb.dtype) + + +class Rotary(nn.Module): + def __init__(self, dim, base=10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._cache = (0, None, None) + + def forward(self, seq_len, device, dtype): + if self._cache[0] != seq_len or self._cache[1] is None or self._cache[1].device != device: + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cache = (seq_len, freqs.cos()[None, None], freqs.sin()[None, None]) + return self._cache[1].to(dtype), self._cache[2].to(dtype) + + +def apply_rotary_emb(x, cos, sin): + 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, num_heads, num_kv_heads, rope_base, qk_gain_init, use_xsa=False): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + self.use_xsa = use_xsa + kv_dim = 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) + + def _xsa_efficient(self, y, v): + """Exclusive Self Attention: remove value-direction component from output. + y: [B, T, H, D], v: [B, T, Hkv, D]""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x): + bsz, seqlen, dim = x.shape + # Extra RMSNorm before Q/K projections (Steinmetz 2025): stabilizes activation + # scales entering the fragile RoPE path, improving post-quantization quality. + x_qk = F.rms_norm(x, (x.size(-1),)) + q = self.c_q(x_qk).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x_qk).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) + 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)) + # y: [B, H, T, D], v: [B, Hkv, T, D] + if self.use_xsa: + # Transpose to [B, T, H, D] and [B, T, Hkv, D] + y_bt = y.transpose(1, 2) # [B, T, H, D] + v_bt = v.transpose(1, 2) # [B, T, Hkv, D] + y_bt = self._xsa_efficient(y_bt, v_bt) + return self.proj(y_bt.contiguous().reshape(bsz, seqlen, dim)) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim, mlp_mult): + super().__init__() + hidden = 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): + return self.proj(torch.relu(self.fc(x)).square()) + + +class Block(nn.Module): + def __init__(self, dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init, use_xsa=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, use_xsa=use_xsa) + 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()) + + def forward(self, x, x0): + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + 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)) + return x + + +class GPT(nn.Module): + def __init__(self, vocab_size, num_layers, model_dim, num_heads, num_kv_heads, + mlp_mult, tie_embeddings, tied_embed_init_std, logit_softcap, + rope_base, qk_gain_init, bigram_vocab, bigram_dim, xsa_last_n=3): + super().__init__() + self.tie_embeddings = tie_embeddings + self.logit_softcap = logit_softcap + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.smear_gate = SmearGate(model_dim) + self.bigram_hash = BigramHash(bigram_vocab, bigram_dim, model_dim) + 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, + use_xsa=(i >= num_layers - xsa_last_n)) + for i in range(num_layers) + ]) + 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._init_weights(num_layers, tied_embed_init_std) + + def _init_weights(self, num_layers, std): + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=std) + out_scale = 1.0 / math.sqrt(2 * num_layers) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + else: + nn.init.orthogonal_(module.weight, gain=1.0) + # muP: scale output projections in transformer blocks only + if "blocks." in name and ".proj." in name and "c_" not in name: + module.weight.data.mul_(out_scale) + + def forward(self, input_ids, target_ids): + logits = self.forward_logits(input_ids) + targets = target_ids.reshape(-1) + logits_flat = logits.reshape(-1, logits.size(-1)) + logits_capped = self.logit_softcap * torch.tanh(logits_flat / self.logit_softcap) + return F.cross_entropy(logits_capped.float(), targets, reduction="mean") + + def forward_logits(self, input_ids): + x = self.tok_emb(input_ids) + x = self.smear_gate(x) + x = x + self.bigram_hash(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + skips = [] + + for i in range(self.num_encoder_layers): + x = self.blocks[i](x, x0) + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x = self.blocks[self.num_encoder_layers + i](x, x0) + + x = self.final_norm(x) + if self.tie_embeddings: + return F.linear(x, self.tok_emb.weight) + return self.lm_head(x) + + +def restore_low_dim_params_to_fp32(module): + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(p in name for p in CONTROL_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +# ───────────────────────────────────────────── +# MAIN +# ───────────────────────────────────────────── + +def main(): + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyp() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + # Distributed 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 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 + if grad_accum_steps <= 0: + raise ValueError(f"WORLD_SIZE={world_size} too large: grad_accum_steps would be 0") + grad_scale = 1.0 / grad_accum_steps + 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 = rank == 0 + + 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: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg, console=True): + if not master: return + if console: print(msg, flush=True) + if logfile: + with open(logfile, "a") as f: print(msg, file=f) + + log0(code, console=False) + log0("=" * 80, console=False) + + # Seed + random.seed(args.seed); np.random.seed(args.seed) + torch.manual_seed(args.seed); torch.cuda.manual_seed_all(args.seed) + + # Tokenizer + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError(f"VOCAB_SIZE={args.vocab_size} != tokenizer vocab_size={int(sp.vocab_size())}") + + # Validation data - load and align to seq_len for simple eval + val_files_list = [Path(p) for p in sorted(glob.glob(args.val_files))] + val_tokens_raw = torch.cat([load_data_shard(f) for f in val_files_list]).contiguous() + usable = ((val_tokens_raw.numel() - 1) // args.train_seq_len) * args.train_seq_len + val_tokens = val_tokens_raw[:usable + 1] + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(sp, args.vocab_size, device) + log0(f"val_tokens:{val_tokens.numel()} (raw:{val_tokens_raw.numel()})") + + # Model + 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, + bigram_vocab=args.bigram_vocab_size, bigram_dim=args.bigram_dim, xsa_last_n=args.xsa_last_n, + ).to(device).bfloat16() + for m in base_model.modules(): + if isinstance(m, CastedLinear): m.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizers + block_named = list(base_model.blocks.named_parameters()) + matrix_params = [p for n, p in block_named if p.ndim == 2 and not any(c in n for c in CONTROL_PATTERNS)] + scalar_params = [p for n, p in block_named if p.ndim < 2 or any(c in n for c in CONTROL_PATTERNS)] + + # BigramHash: proj weight goes to Muon (it's a real weight matrix), + # but the embedding table goes to Adam (it's an embedding, not a weight matrix) + for n, p in base_model.bigram_hash.named_parameters(): + if n == "proj.weight": + matrix_params.append(p) + else: + scalar_params.append(p) + # SmearGate params -> Adam + for p in base_model.smear_gate.parameters(): + scalar_params.append(p) + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + opt_tok = torch.optim.AdamW( + [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, weight_decay=args.adam_wd, fused=True, + ) + opt_muon = Muon(matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.muon_wd) + for g in opt_muon.param_groups: g["base_lr"] = args.matrix_lr + opt_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.adam_wd, fused=True, + ) + optimizers = [opt_tok, opt_muon, opt_scalar] + if base_model.lm_head is not None: + opt_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.adam_wd, fused=True, + ) + optimizers.insert(1, opt_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params} layers:{args.num_layers} dim:{args.model_dim} mlp_mult:{args.mlp_mult}") + log0(f"matrix_lr:{args.matrix_lr} muon_wd:{args.muon_wd} adam_wd:{args.adam_wd} grad_clip:{args.grad_clip_norm}") + log0(f"seq_len:{args.train_seq_len} warmdown:{args.warmdown_iters} swa:{args.swa_enabled}/{args.swa_every}") + log0(f"seed:{args.seed} world_size:{world_size} xsa_last_n:{args.xsa_last_n}") + log0(f"meta_ttt:{args.meta_ttt_enabled} start_frac:{args.meta_ttt_start_frac} inner_steps:{args.meta_ttt_inner_steps}") + log0(f"extra_linear_rmsnorm:{args.extra_linear_rmsnorm} meta_ttt_log_every:{args.meta_ttt_log_every}") + + # Data loader + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all(): + for opt in optimizers: opt.zero_grad(set_to_none=True) + + def apply_optimizers(step, scale): + """Muon momentum warmup, LR schedule, grad clip, optimizer step, zero grad.""" + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_mom = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for g in opt_muon.param_groups: g["momentum"] = muon_mom + for opt in optimizers: + for g in opt.param_groups: + g["lr"] = g["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() + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step, elapsed_ms): + if args.warmdown_iters <= 0: return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + if warmdown_start <= step < args.iterations: + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) + return 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 + + def sync_reached_cap(approx_ms): + reached_cap = max_wallclock_ms is not None and approx_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + t_cap = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(t_cap, op=dist.ReduceOp.MAX) + reached_cap = bool(t_cap.item()) + return reached_cap + + # Warmup (torch.compile priming) + if args.warmup_steps > 0: + init_state = {n: t.detach().cpu().clone() for n, t in base_model.state_dict().items()} + init_opt_states = [copy.deepcopy(o.state_dict()) for o in optimizers] + model.train() + for ws in range(args.warmup_steps): + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: model.require_backward_grad_sync = ms == 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): + wl = model(x, y) + (wl * grad_scale).backward() + for o in optimizers: o.step() + zero_grad_all() + log0(f"warmup_step:{ws+1}/{args.warmup_steps}") + base_model.load_state_dict(init_state, strict=True) + for o, s in zip(optimizers, init_opt_states): o.load_state_dict(s) + zero_grad_all() + if distributed: model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # SWA state + swa_state = None + swa_count = 0 + + # Training loop + training_time_ms = 0.0 + stop_after_step = None + torch.cuda.synchronize() + t0 = time.perf_counter() + + # Pre-compute Reptile meta-TTT param set (constant across steps) + _num_blocks = len(base_model.blocks) + _adapt_indices = range(max(0, _num_blocks - args.ttt_eval_adapt_last_n), _num_blocks) + meta_adapt_ids = set() + for _bi in _adapt_indices: + for _n, _p in base_model.blocks[_bi].mlp.named_parameters(): + if "weight" in _n and ("fc" in _n or "proj" in _n): + meta_adapt_ids.add(id(_p)) + + 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_simple( + 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 step:{step}") + break + + # Safety checkpoint + if args.ckpt_every > 0 and step > 0 and step % args.ckpt_every == 0 and master: + torch.save(base_model.state_dict(), f"ckpt_step{step}.pt") + log0(f"checkpoint saved: ckpt_step{step}.pt") + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + + # Compute elapsed fraction for meta-TTT check + if max_wallclock_ms is not None and max_wallclock_ms > 0: + elapsed_frac = elapsed_ms / max_wallclock_ms + else: + elapsed_frac = step / max(args.iterations, 1) + + # SWA collection + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {k: v.detach().cpu().clone() for k, v in base_model.state_dict().items()} + swa_count = 1 + else: + for k, v in base_model.state_dict().items(): + swa_state[k] += v.detach().cpu() + swa_count += 1 + + # ─── Reptile Meta-TTT (last 15% of training) ─── + if args.meta_ttt_enabled and elapsed_frac >= args.meta_ttt_start_frac: + if args.meta_ttt_log_every > 0 and ( + step == 0 + or step == args.iterations - 1 + or step % args.meta_ttt_log_every == 0 + ): + log0( + f"meta_ttt:start step:{step}/{args.iterations} " + f"elapsed_frac:{elapsed_frac:.4f} inner_steps:{args.meta_ttt_inner_steps}" + ) + # Save current params theta_0 (only adapted MLP params need cloning) + theta_0 = {n: p.data.clone() for n, p in base_model.named_parameters() + if id(p) in meta_adapt_ids} + + # Inner loop: K SGD steps on consecutive batches (simulating TTT) + for _inner in range(args.meta_ttt_inner_steps): + approx_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if sync_reached_cap(approx_ms): + if stop_after_step is None: + stop_after_step = step + log0( + f"meta_ttt:aborting_for_wallclock step:{step}/{args.iterations} " + f"inner_step:{_inner}/{args.meta_ttt_inner_steps} train_time:{approx_ms:.0f}ms" + ) + break + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: model.require_backward_grad_sync = ms == 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): + inner_loss = model(x, y) + (inner_loss * grad_scale).backward() + # SGD step only on MLP params in last N blocks (matching eval TTT) + with torch.no_grad(): + for p in base_model.parameters(): + if p.grad is not None and id(p) in meta_adapt_ids: + p.data.sub_(args.meta_ttt_inner_lr * p.grad) + zero_grad_all() + if args.meta_ttt_log_every > 0 and ( + (_inner + 1) == args.meta_ttt_inner_steps + or step % args.meta_ttt_log_every == 0 + ): + log0( + f"meta_ttt:inner step:{step}/{args.iterations} " + f"inner:{_inner + 1}/{args.meta_ttt_inner_steps} loss:{inner_loss.item():.4f}" + ) + + if stop_after_step is not None and step >= stop_after_step: + continue + + # Reptile interpolate only adapted params + with torch.no_grad(): + for n, p in base_model.named_parameters(): + if id(p) in meta_adapt_ids: + p.data.copy_(theta_0[n] + args.meta_ttt_epsilon * (p.data - theta_0[n])) + + # Get eval batch and compute eval_loss for the normal optimizer step + zero_grad_all() + for ms in range(grad_accum_steps): + if distributed: model.require_backward_grad_sync = ms == 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): + eval_loss = model(x, y) + (eval_loss * grad_scale).backward() + + apply_optimizers(step, scale) + + step += 1 + approx_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0): + log0(f"step:{step}/{args.iterations} [meta-ttt] train_loss:{eval_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + else: + # ─── Normal training step ─── + zero_grad_all() + train_loss = torch.zeros((), device=device) + for ms in range(grad_accum_steps): + if distributed: model.require_backward_grad_sync = ms == 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): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + apply_optimizers(step, scale) + + step += 1 + approx_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0): + log0(f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_ms:.0f}ms step_avg:{approx_ms / step:.2f}ms") + + approx_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + reached_cap = sync_reached_cap(approx_ms) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0(f"peak_mem: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB") + + # ─── POST-TRAINING PIPELINE ─── + + # 1. Apply SWA + if args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"SWA: averaging {swa_count} checkpoints") + avg = {k: (v / swa_count).to(dtype=base_model.state_dict()[k].dtype) for k, v in swa_state.items()} + base_model.load_state_dict(avg, strict=True) + else: + log0("post_train: skipping SWA averaging") + + # 2. Save raw checkpoint + log0("post_train: saving raw checkpoint") + if master: + torch.save(base_model.state_dict(), "final_model.pt") + log0(f"Raw checkpoint saved: {os.path.getsize('final_model.pt')} bytes") + + # 3. Int6 + zstd quantization + log0("post_train: quantizing artifact") + quant_obj, quant_stats = quantize_state_dict_int6(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_blob = compress_blob(quant_buf.getvalue()) + artifact_name = f"final_model.{COMPRESS_EXT}" + if master: + with open(artifact_name, "wb") as f: + f.write(quant_blob) + artifact_bytes = os.path.getsize(artifact_name) + code_bytes = len(code.encode("utf-8")) + total_bytes = artifact_bytes + code_bytes + log0(f"Artifact: {artifact_bytes} bytes, code: {code_bytes} bytes, total: {total_bytes} bytes") + if total_bytes > 16_000_000: + raise RuntimeError(f"artifact too large: total {total_bytes} exceeds 16MB limit") + + # 4. Verify artifact < 16MB (already done in step 3 logging) + + # 5. Load quantized model (roundtrip) + log0("post_train: loading quantized roundtrip") + if distributed: dist.barrier() + with open(artifact_name, "rb") as f: + rt_blob = f.read() + rt_state = torch.load(io.BytesIO(decompress_blob(rt_blob)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int6(rt_state), strict=True) + + # 6. Run simple eval for baseline number + log0("post_train: starting roundtrip eval") + torch.cuda.synchronize() + t_eval = time.perf_counter() + q_loss, q_bpb = eval_val_simple( + 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_roundtrip val_loss:{q_loss:.4f} val_bpb:{q_bpb:.4f} eval_time:{1000*(time.perf_counter()-t_eval):.0f}ms") + log0(f"final_roundtrip_exact val_loss:{q_loss:.8f} val_bpb:{q_bpb:.8f}") + + # 7. Run online TTT eval with decay prior (the real score) + log0("post_train: starting online TTT eval") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_ttt( + base_model, val_tokens_raw, args.eval_seq_len, args.eval_stride, device, + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, rank, world_size, args, log_fn=log0, + ) + torch.cuda.synchronize() + log0(f"final_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} eval_time:{1000*(time.perf_counter()-t_ttt):.0f}ms") + log0(f"final_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + + if distributed: dist.destroy_process_group() + + +if __name__ == "__main__": + main()