diff --git a/.gitignore b/.gitignore index 3423c416a7..9260888ec4 100644 --- a/.gitignore +++ b/.gitignore @@ -8,4 +8,6 @@ data/manifest.json data/docs_selected.jsonl .mypy_cache/ .venv -logs/ \ No newline at end of file +logs/ +final_model.* +sweep.sh \ No newline at end of file diff --git a/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/README.md b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/README.md new file mode 100644 index 0000000000..7fdbb74758 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/README.md @@ -0,0 +1,58 @@ +## Depth Recurrence + Cross-Repeat Skip + Value Embeddings + +Beats naive baseline (1.2244) by 0.005 bpb using 3.1x fewer training steps through stateful depth recurrence. + +val_bpb = 1.2196 (sliding window eval on int8+zlib roundtrip model, stride=256) +val_bpb = 1.2533 (standard int8+zlib roundtrip) + +### Architecture + +Replaced the baseline's 9 unique transformer blocks with 3 shared blocks repeated 4 times (12 effective layers). Trades unique parameters for effective depth. + +Changes from baseline: +- Depth recurrence: 3 blocks x 4 repeats = 12 effective layers (vs 9 in baseline) +- Cross-Repeat Skip (original): each block gets a weighted residual of its own output from the previous repeat, turning stateless recurrence into stateful. Per-repeat learned scales, ~7.5K params total. +- Value Embeddings: 2 extra embedding tables mixed into the residual stream at each effective layer with learned scales. From snimu's modded-nanogpt record. +- Loop Embedding: learned per-layer vector added before each block as depth-wise positional encoding. +- Model dim 832 (vs 512), 8 heads, 4 KV heads, MLP 2x +- Removed U-Net skip connections (Cross-Repeat Skip covers this role) +- 17.14M params, 12.83MB artifact + +### Training + +LR x0.3 from baseline — recurrence amplifies gradients through 4 passes, so optimal LR is much lower. Found via sweep of 10 configs on RTX 3060. + +MATRIX_LR=0.012, SCALAR_LR=0.012, TIED_EMBED_LR=0.015, GRAD_CLIP_NORM=0.3, WARMDOWN_ITERS=3000, TRAIN_SEQ_LEN=1024. + +Tested train@2048 but 1024 gives more steps (133ms vs 253ms/step) which matters more for this architecture. Standard Muon + Adam. + +### Evaluation + +Sliding window eval: window=1024, stride=256 on the int8+zlib roundtrip model. Eval time 209s on 8xH100. + +### Results (8xH100, 600s wallclock) + +4494 steps, 133ms/step avg. Pre-quant 1.2487, roundtrip 1.2533, sliding window 1.2196. Artifact 12.83MB, quant degradation 0.005 bpb, peak memory ~29GB/GPU. + +### Ablations (RTX 3060, 2000 steps each) + +- Cross-Repeat Skip: -0.041 bpb +- Value Embeddings (2 tables): -0.079 bpb +- LR x0.3: -0.052 bpb +- Sliding window eval: -0.034 bpb +- WARMDOWN_ITERS=3000: -0.027 bpb + +### Development + +All experiments, ablations, and hyperparameter sweeps done on a single RTX 3060 12GB. Cloud GPUs (1xH200, 6xH100) used only for validation. Final run on 8xH100. + +### Command + +``` +RUN_ID=submission_8xh100 \ +QUANT_LEVELS=127 \ +TTT_STEPS=0 \ +EVAL_STRIDE=256 \ +EVAL_SEQ_LEN=1024 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` diff --git a/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/submission.json b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/submission.json new file mode 100644 index 0000000000..f04f129d16 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/submission.json @@ -0,0 +1,16 @@ +{ + "author": "Ivan Verbovoy", + "github_id": "iverbovoy", + "name": "Depth Recurrence + Cross-Repeat Skip + Value Embeddings + Sliding Window", + "blurb": "3 unique blocks x 4 repeats (12 effective layers), dim=832, with Cross-Repeat Skip (stateful recurrence), 2 Value Embedding tables, LR x0.3, sliding window eval (stride=256). 4494 steps in 600s on 8xH100.", + "date": "2026-03-20T02:00:00Z", + "val_loss": 2.05921204, + "val_bpb": 1.21958209, + "roundtrip_val_loss": 2.11612232, + "roundtrip_val_bpb": 1.25328684, + "step_stop": 4494, + "wallclock_seconds": 600.133, + "bytes_total": 12829176, + "bytes_model_int8_zlib": 12771121, + "bytes_code": 58055 +} diff --git a/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train.log b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train.log new file mode 100644 index 0000000000..d9a0c15299 --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train.log @@ -0,0 +1,84 @@ +W0320 00:54:42.000000 1050 torch/distributed/run.py:852] +W0320 00:54:42.000000 1050 torch/distributed/run.py:852] ***************************************** +W0320 00:54:42.000000 1050 torch/distributed/run.py:852] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0320 00:54:42.000000 1050 torch/distributed/run.py:852] ***************************************** +logs/submission_8xh100.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:17140056 +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.015 head_lr:0.0 matrix_lr:0.012 scalar_lr:0.012 +train_batch_tokens:524288 train_seq_len:1024 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +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:6.9766 val_bpb:4.1319 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9765 train_time:162ms step_avg:161.95ms +step:2/20000 train_loss:9.0581 train_time:218ms step_avg:109.04ms +step:3/20000 train_loss:7.8439 train_time:342ms step_avg:114.12ms +step:4/20000 train_loss:6.5913 train_time:466ms step_avg:116.40ms +step:5/20000 train_loss:6.1067 train_time:589ms step_avg:117.72ms +step:6/20000 train_loss:6.3514 train_time:712ms step_avg:118.70ms +step:7/20000 train_loss:5.9725 train_time:836ms step_avg:119.39ms +step:8/20000 train_loss:5.8139 train_time:958ms step_avg:119.78ms +step:9/20000 train_loss:5.5629 train_time:1081ms step_avg:120.13ms +step:10/20000 train_loss:5.3728 train_time:1206ms step_avg:120.64ms +step:200/20000 train_loss:2.7739 train_time:26609ms step_avg:133.05ms +step:400/20000 train_loss:2.3107 train_time:53543ms step_avg:133.86ms +step:600/20000 train_loss:2.5249 train_time:80122ms step_avg:133.54ms +step:800/20000 train_loss:2.2710 train_time:106824ms step_avg:133.53ms +step:1000/20000 train_loss:2.3610 train_time:133649ms step_avg:133.65ms +step:1000/20000 val_loss:2.3206 val_bpb:1.3744 train_time:133722ms step_avg:133.72ms +step:1200/20000 train_loss:2.3700 train_time:160457ms step_avg:133.71ms +step:1400/20000 train_loss:2.4196 train_time:187085ms step_avg:133.63ms +step:1600/20000 train_loss:2.0826 train_time:213643ms step_avg:133.53ms +step:1800/20000 train_loss:2.1817 train_time:240257ms step_avg:133.48ms +step:2000/20000 train_loss:2.2342 train_time:266823ms step_avg:133.41ms +step:2000/20000 val_loss:2.2137 val_bpb:1.3111 train_time:266903ms step_avg:133.45ms +step:2200/20000 train_loss:2.0469 train_time:293423ms step_avg:133.37ms +step:2400/20000 train_loss:2.1757 train_time:320078ms step_avg:133.37ms +step:2600/20000 train_loss:2.3756 train_time:346626ms step_avg:133.32ms +step:2800/20000 train_loss:2.2012 train_time:373394ms step_avg:133.35ms +step:3000/20000 train_loss:2.1910 train_time:400062ms step_avg:133.35ms +step:3000/20000 val_loss:2.1585 val_bpb:1.2784 train_time:400147ms step_avg:133.38ms +step:3200/20000 train_loss:2.1485 train_time:426762ms step_avg:133.36ms +step:3400/20000 train_loss:2.1171 train_time:453425ms step_avg:133.36ms +step:3600/20000 train_loss:2.0703 train_time:480073ms step_avg:133.35ms +step:3800/20000 train_loss:2.1774 train_time:506627ms step_avg:133.32ms +step:4000/20000 train_loss:2.1156 train_time:532930ms step_avg:133.23ms +step:4000/20000 val_loss:2.1201 val_bpb:1.2556 train_time:533004ms step_avg:133.25ms +step:4200/20000 train_loss:2.1277 train_time:561906ms step_avg:133.79ms +step:4400/20000 train_loss:2.0541 train_time:588700ms step_avg:133.80ms +step:4494/20000 val_loss:2.1084 val_bpb:1.2487 train_time:600133ms step_avg:133.54ms +stopping_early: wallclock_cap train_time:600133ms step:4494/20000 +peak memory allocated: 21771 MiB reserved: 21818 MiB +Serialized model: 63387167 bytes +Code size: 58055 bytes +Total submission size: 63445222 bytes +Serialized model int8+zlib: 12771121 bytes (payload:17243616 raw_torch:17261176 payload_ratio:3.68x) +Total submission size int8+zlib: 12829176 bytes +final_int8_zlib_roundtrip val_loss:2.1161 val_bpb:1.2533 eval_time:3709ms +final_int8_zlib_roundtrip_exact val_loss:2.11612232 val_bpb:1.25328684 +final_sliding_window val_loss:2.0592 val_bpb:1.2196 window:1024 stride:256 eval_time:209349ms +final_sliding_window_exact val_loss:2.05921204 val_bpb:1.21958209 diff --git a/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train_gpt.py b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train_gpt.py new file mode 100644 index 0000000000..aa83a930be --- /dev/null +++ b/records/track_10min_16mb/2026-03-20_DepthRecurrence_CrossRepeatSkip/train_gpt.py @@ -0,0 +1,1365 @@ +""" +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 math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +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 + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# Default Simple Baseline run: +# - 9 transformer blocks at width 512 +# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion +# - vocab size 1024, sequence length 1024, tied embeddings +# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + 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)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + 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)) + + # Training length. + 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", 1024)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + ttt_steps = int(os.environ.get("TTT_STEPS", 0)) + ttt_lr = float(os.environ.get("TTT_LR", 1e-4)) + + # Sliding window eval. + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 256)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 3)) + num_repeats = int(os.environ.get("NUM_REPEATS", 4)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 832)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = int(os.environ.get("MLP_MULT", 2)) + num_value_embeds = int(os.environ.get("NUM_VALUE_EMBEDS", 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)) + + # Optimizer hyperparameters. + 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", 0.015)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.012)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.012)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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)) + +# ----------------------------- +# MUON OPTIMIZER +# ----------------------------- +# +# As borrowed from modded-nanogpt +# Background on Muon: https://kellerjordan.github.io/posts/muon/ + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. + # Muon uses this to normalize matrix-shaped gradients before applying them. + 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): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + ) + + @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) + # Scale correction from Muon reference implementations. + 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) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# ----------------------------- +# TOKENIZER-AGNOSTIC EVALUATION SETUP +# ----------------------------- +# +# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. +# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. +# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. +# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. + +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}") + # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. + 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]: + # Validation computes two metrics: + # - val_loss: token cross-entropy (natural log) + # - val_bpb: tokenizer-agnostic compression metric used by the challenge + 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 eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + """Sliding window eval: each window is eval_seq_len tokens, advancing by eval_stride. + Loss is scored only on the last eval_stride tokens per window.""" + seq_len = args.eval_seq_len + stride = args.eval_stride + total_tokens = val_tokens.numel() + + starts: list[int] = [] + pos = 0 + while pos + seq_len < total_tokens: + starts.append(pos) + pos += stride + total_windows = len(starts) + win_start = (total_windows * rank) // world_size + win_end = (total_windows * (rank + 1)) // world_size + score_offset = seq_len - stride + + 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) + + base_model.eval() + with torch.no_grad(): + for wi in range(win_start, win_end): + s = starts[wi] + window = val_tokens[s : s + seq_len + 1].to(device=device, dtype=torch.int64) + x = window[:-1].unsqueeze(0) + y = window[1:].unsqueeze(0) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = base_model.forward_logits(x) + + tail_logits = logits[0, score_offset:, :].float() + tail_targets = y[0, score_offset:] + per_token_loss = F.cross_entropy(tail_logits, tail_targets, reduction="none") + val_loss_sum += per_token_loss.to(torch.float64).sum() + val_token_count += float(stride) + + tail_prev = x[0, score_offset:] + tail_tgt = y[0, score_offset:] + token_bytes = base_bytes_lut[tail_tgt].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tail_tgt] & ~is_boundary_token_lut[tail_prev]).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() + base_model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + + +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + 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]: + # Test-Time Training: adapt the model on each validation batch before evaluating. + # For each batch: save weights → K gradient steps → evaluate → restore weights. + if args.ttt_steps <= 0: + return eval_val(args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + + 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) + + # Save original weights once + saved_state = {k: v.detach().clone() for k, v in base_model.state_dict().items()} + + 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) + + # TTT: adapt on this batch + model.train() + for _ttt_step in range(args.ttt_steps): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(x, y) + ttt_loss.backward() + with torch.no_grad(): + for p in base_model.parameters(): + if p.grad is not None: + p -= args.ttt_lr * p.grad + p.grad = None + + # Evaluate with adapted model + model.eval() + with torch.no_grad(): + 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() + + # Restore original weights + base_model.load_state_dict(saved_state, strict=True) + + 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) + + +# ----------------------------- +# POST-TRAINING QUANTIZATION +# ----------------------------- +# +# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. +# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. +# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. + +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", + ).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 +) +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 +# Int6 quantization: ±31 instead of ±127. Stored as int8 but zlib compresses better. +QUANT_LEVELS = int(os.environ.get("QUANT_LEVELS", 127)) # 127 = int8, 31 = int6 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + ql = QUANT_LEVELS # 31 for int6, 127 for int8 + 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 / ql).clamp_min(1.0 / ql) + q = torch.clamp(torch.round(clipped / scale[:, None]), -ql, ql).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 / ql if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -ql, ql).to(torch.int8).contiguous() + return q, scale + +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + # Single supported clean-script export format: + # - per-row int8 for 2D float tensors + # - per-tensor int8 for other float tensors + # - exact passthrough for non-floats + # - passthrough for small float tensors, stored as fp16 to save bytes + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + # Small float tensors are cheap enough to keep directly. We still downcast + # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + 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 qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + # Broadcast the saved row scale back across trailing dimensions. + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + # Restore small tensors, undoing the temporary fp16 storage cast if needed. + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + 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: + # Each call consumes a contiguous chunk from the shared token stream, then slices out + # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. + 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): + # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + def forward(self, x: Tensor) -> Tensor: + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, self.weight.to(x.dtype), bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + # Keep small/control parameters in fp32 even when the model body runs in bf16. + 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() + + +class Rotary(nn.Module): + # Caches cos/sin tables per sequence length on the current device. + def __init__(self, dim: int, base: float = 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._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: + 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, + ): + 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) + + def forward(self, x: Tensor) -> Tensor: + 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) + 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 = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) + return self.proj(y) + + +class MLP(nn.Module): + # relu^2 MLP from the original modded-nanogpt setup + def __init__(self, dim: int, mlp_mult: int): + 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: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + 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: Tensor, x0: Tensor) -> Tensor: + 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: int, + num_layers: int, + num_repeats: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + num_value_embeds: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + ): + 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.num_repeats = num_repeats + effective_depth = num_layers * num_repeats + self.tok_emb = nn.Embedding(vocab_size, model_dim) + # Value embeddings: extra embedding tables mixed into each effective layer + self.num_value_embeds = num_value_embeds + if num_value_embeds > 0: + self.value_embeds = nn.ModuleList([nn.Embedding(vocab_size, model_dim) for _ in range(num_value_embeds)]) + self.value_scales = nn.Parameter(torch.zeros(effective_depth, num_value_embeds, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + ) + for i in range(num_layers) + ] + ) + # Loop embedding: tells the model which effective layer it's at + self.loop_embed = nn.Parameter(torch.zeros(effective_depth, model_dim, dtype=torch.float32)) + # Cross-repeat skip: each block remembers its output from previous repeat + # Per-repeat scales (repeat 0 has no prev, so num_repeats-1 scales per block) + self.cross_repeat_scales = nn.Parameter(torch.zeros(num_layers, num_repeats - 1, model_dim, dtype=torch.float32)) + 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() + + 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) + for module in self.modules(): + if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x0 = x + + # Pre-compute value embeddings once + ve_list: list[Tensor] = [] + if self.num_value_embeds > 0: + for ve in self.value_embeds: + ve_list.append(ve(input_ids)) # (bsz, seq, dim) + + num_blocks = len(self.blocks) + prev_block_outputs: list[Tensor | None] = [None] * num_blocks + layer_idx = 0 + for repeat in range(self.num_repeats): + for block_idx, block in enumerate(self.blocks): + x = x + self.loop_embed[layer_idx].to(dtype=x.dtype) + # Value embeddings: add weighted extra embeddings at each layer + for ve_idx, ve_out in enumerate(ve_list): + vs = self.value_scales[layer_idx, ve_idx].to(dtype=x.dtype) + x = x + vs[None, None, :] * ve_out + # Cross-repeat skip: mix in this block's output from previous repeat + if repeat > 0 and prev_block_outputs[block_idx] is not None: + scale = self.cross_repeat_scales[block_idx, repeat - 1].to(dtype=x.dtype) + x = x + scale[None, None, :] * prev_block_outputs[block_idx] + x = block(x, x0) + prev_block_outputs[block_idx] = x.detach() if not self.training else x + layer_idx += 1 + + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return logits + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + logits = self.forward_logits(input_ids) + logits = logits.reshape(-1, logits.size(-1)) + targets = target_ids.reshape(-1) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + 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}") + grad_accum_steps = max(1, 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, + num_repeats=args.num_repeats, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + num_value_embeds=args.num_value_embeds, + 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, + ).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) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + 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) + ] + scalar_params.append(base_model.loop_embed) + scalar_params.append(base_model.cross_repeat_scales) + if base_model.num_value_embeds > 0: + scalar_params.append(base_model.value_scales) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + embed_params = [base_model.tok_emb.weight] + if base_model.num_value_embeds > 0: + embed_params.extend(ve.weight for ve in base_model.value_embeds) + optimizer_tok = torch.optim.Adam( + [{"params": embed_params, "lr": token_lr, "base_lr": token_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + 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.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + 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"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 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + 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 + # ----------------------------- + + training_time_ms = 0.0 + stop_after_step: int | None = None + 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) + zero_grad_all() + train_loss = torch.zeros((), device=device) + 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() + train_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 + 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: + log0( + 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" + ) + + # 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" + ) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce + # the compressed int8+zlib artifact and validate the round-tripped weights. + + if master_process: + torch.save(base_model.state_dict(), "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"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zlib.compress(quant_raw, level=9) + quant_raw_bytes = len(quant_raw) + if master_process: + with open("final_model.int8.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize("final_model.int8.ptz") + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int8+zlib: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open("final_model.int8.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + 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_int8_zlib_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_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + # Sliding window eval + if args.eval_stride > 0: + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_sw = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"window:{args.eval_seq_len} stride:{args.eval_stride} " + f"eval_time:{1000.0 * (time.perf_counter() - t_sw):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + + # TTT eval: adapt model on each batch before evaluating + if args.ttt_steps > 0: + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, + base_model, + 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_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"ttt_steps:{args.ttt_steps} ttt_lr:{args.ttt_lr} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/sweep.sh b/sweep.sh new file mode 100755 index 0000000000..5e6f9f8aa0 --- /dev/null +++ b/sweep.sh @@ -0,0 +1,46 @@ +#!/bin/bash +# Hyperparameter sweep — run overnight on 3060 +# Each run: 2000 steps, batch 8K, no TTT + +export ITERATIONS=2000 +export TRAIN_BATCH_TOKENS=8192 +export VAL_LOSS_EVERY=0 +export VAL_BATCH_SIZE=8192 +export MAX_WALLCLOCK_SECONDS=0 +export TTT_STEPS=0 + +echo "=== Starting sweep at $(date) ===" + +# 1. Baseline (current defaults: matrix_lr=0.04, embed_lr=0.05, scalar_lr=0.04) +echo "--- Run 1: baseline ---" +RUN_ID=sweep_baseline torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(model_params|step:2000|final_int8_zlib_roundtrip_exact)" + +# 2. All lr x1.5 +echo "--- Run 2: lr x1.5 ---" +RUN_ID=sweep_lr15 MATRIX_LR=0.06 TIED_EMBED_LR=0.075 SCALAR_LR=0.06 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 3. All lr x2.0 +echo "--- Run 3: lr x2.0 ---" +RUN_ID=sweep_lr20 MATRIX_LR=0.08 TIED_EMBED_LR=0.1 SCALAR_LR=0.08 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 4. All lr x0.5 +echo "--- Run 4: lr x0.5 ---" +RUN_ID=sweep_lr05 MATRIX_LR=0.02 TIED_EMBED_LR=0.025 SCALAR_LR=0.02 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 5. Lower embed_lr ratio (embed_lr = 0.3x matrix_lr) +echo "--- Run 5: low embed_lr ---" +RUN_ID=sweep_lowemb TIED_EMBED_LR=0.012 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 6. Longer warmdown (2400 iters) +echo "--- Run 6: warmdown_iters=2400 ---" +RUN_ID=sweep_wd2400 WARMDOWN_ITERS=2400 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 7. Higher muon momentum +echo "--- Run 7: muon_momentum=0.98 ---" +RUN_ID=sweep_mom98 MUON_MOMENTUM=0.98 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +# 8. Matrix lr x1.5 + lower embed +echo "--- Run 8: matrix_lr=0.06 + embed_lr=0.02 ---" +RUN_ID=sweep_combo MATRIX_LR=0.06 TIED_EMBED_LR=0.02 torchrun --standalone --nproc_per_node=1 train_gpt.py 2>&1 | grep -E "(step:2000|final_int8_zlib_roundtrip_exact)" + +echo "=== Sweep done at $(date) ===" diff --git a/train_gpt.py b/train_gpt.py index 0deb0565f5..6261f87b7d 100644 --- a/train_gpt.py +++ b/train_gpt.py @@ -52,20 +52,29 @@ class Hyperparameters: # Training length. iterations = int(os.environ.get("ITERATIONS", 20000)) - warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + 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", 1024)) max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + ttt_steps = int(os.environ.get("TTT_STEPS", 0)) + ttt_lr = float(os.environ.get("TTT_LR", 1e-4)) + + # Sliding window eval. + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 256)) # Model shape. vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) - num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_layers = int(os.environ.get("NUM_LAYERS", 3)) + num_repeats = int(os.environ.get("NUM_REPEATS", 4)) num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) - model_dim = int(os.environ.get("MODEL_DIM", 512)) + model_dim = int(os.environ.get("MODEL_DIM", 832)) num_heads = int(os.environ.get("NUM_HEADS", 8)) mlp_mult = int(os.environ.get("MLP_MULT", 2)) + num_value_embeds = int(os.environ.get("NUM_VALUE_EMBEDS", 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)) @@ -73,10 +82,10 @@ class Hyperparameters: # Optimizer hyperparameters. 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_lr = float(os.environ.get("TIED_EMBED_LR", 0.015)) tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) - matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) - scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.012)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.012)) muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) @@ -84,7 +93,6 @@ class Hyperparameters: 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", 0.0)) # ----------------------------- # MUON OPTIMIZER @@ -277,6 +285,153 @@ def eval_val( model.train() return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + """Sliding window eval: each window is eval_seq_len tokens, advancing by eval_stride. + Loss is scored only on the last eval_stride tokens per window.""" + seq_len = args.eval_seq_len + stride = args.eval_stride + total_tokens = val_tokens.numel() + + starts: list[int] = [] + pos = 0 + while pos + seq_len < total_tokens: + starts.append(pos) + pos += stride + total_windows = len(starts) + win_start = (total_windows * rank) // world_size + win_end = (total_windows * (rank + 1)) // world_size + score_offset = seq_len - stride + + 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) + + base_model.eval() + with torch.no_grad(): + for wi in range(win_start, win_end): + s = starts[wi] + window = val_tokens[s : s + seq_len + 1].to(device=device, dtype=torch.int64) + x = window[:-1].unsqueeze(0) + y = window[1:].unsqueeze(0) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = base_model.forward_logits(x) + + tail_logits = logits[0, score_offset:, :].float() + tail_targets = y[0, score_offset:] + per_token_loss = F.cross_entropy(tail_logits, tail_targets, reduction="none") + val_loss_sum += per_token_loss.to(torch.float64).sum() + val_token_count += float(stride) + + tail_prev = x[0, score_offset:] + tail_tgt = y[0, score_offset:] + token_bytes = base_bytes_lut[tail_tgt].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tail_tgt] & ~is_boundary_token_lut[tail_prev]).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() + base_model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + + +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + 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]: + # Test-Time Training: adapt the model on each validation batch before evaluating. + # For each batch: save weights → K gradient steps → evaluate → restore weights. + if args.ttt_steps <= 0: + return eval_val(args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut) + + 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) + + # Save original weights once + saved_state = {k: v.detach().clone() for k, v in base_model.state_dict().items()} + + 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) + + # TTT: adapt on this batch + model.train() + for _ttt_step in range(args.ttt_steps): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(x, y) + ttt_loss.backward() + with torch.no_grad(): + for p in base_model.parameters(): + if p.grad is not None: + p -= args.ttt_lr * p.grad + p.grad = None + + # Evaluate with adapted model + model.eval() + with torch.no_grad(): + 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() + + # Restore original weights + base_model.load_state_dict(saved_state, strict=True) + + 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) + + # ----------------------------- # POST-TRAINING QUANTIZATION # ----------------------------- @@ -306,6 +461,8 @@ def eval_val( INT8_PER_ROW_SCALE_DTYPE = torch.float16 INT8_CLIP_PERCENTILE = 99.99984 INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +# Int6 quantization: ±31 instead of ±127. Stored as int8 but zlib compresses better. +QUANT_LEVELS = int(os.environ.get("QUANT_LEVELS", 31)) # 31 = int6, 127 = int8 def tensor_nbytes(t: Tensor) -> int: return int(t.numel()) * int(t.element_size()) @@ -319,24 +476,22 @@ def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, s return t def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + ql = QUANT_LEVELS # 31 for int6, 127 for int8 t32 = t.float() if t32.ndim == 2: - # Matrices get one scale per row, which usually tracks output-channel - # ranges much better than a single tensor-wide scale. 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() + scale = (clip_abs / ql).clamp_min(1.0 / ql) + q = torch.clamp(torch.round(clipped / scale[:, None]), -ql, ql).to(torch.int8).contiguous() return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() - # Vectors / scalars use a simpler per-tensor scale. 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() + scale = torch.tensor(clip_abs / ql if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -ql, ql).to(torch.int8).contiguous() return q, scale def quantize_state_dict_int8(state_dict: dict[str, Tensor]): @@ -650,10 +805,12 @@ def __init__( self, vocab_size: int, num_layers: int, + num_repeats: int, model_dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, + num_value_embeds: int, tie_embeddings: bool, tied_embed_init_std: float, logit_softcap: float, @@ -666,11 +823,14 @@ def __init__( self.tie_embeddings = tie_embeddings self.tied_embed_init_std = tied_embed_init_std self.logit_softcap = logit_softcap + self.num_repeats = num_repeats + effective_depth = num_layers * num_repeats self.tok_emb = nn.Embedding(vocab_size, 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)) + # Value embeddings: extra embedding tables mixed into each effective layer + self.num_value_embeds = num_value_embeds + if num_value_embeds > 0: + self.value_embeds = nn.ModuleList([nn.Embedding(vocab_size, model_dim) for _ in range(num_value_embeds)]) + self.value_scales = nn.Parameter(torch.zeros(effective_depth, num_value_embeds, model_dim, dtype=torch.float32)) self.blocks = nn.ModuleList( [ Block( @@ -684,6 +844,11 @@ def __init__( for i in range(num_layers) ] ) + # Loop embedding: tells the model which effective layer it's at + self.loop_embed = nn.Parameter(torch.zeros(effective_depth, model_dim, dtype=torch.float32)) + # Cross-repeat skip: each block remembers its output from previous repeat + # Per-repeat scales (repeat 0 has no prev, so num_repeats-1 scales per block) + self.cross_repeat_scales = nn.Parameter(torch.zeros(num_layers, num_repeats - 1, model_dim, dtype=torch.float32)) 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: @@ -697,23 +862,36 @@ def _init_weights(self) -> None: if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): nn.init.zeros_(module.weight) - def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + def forward_logits(self, input_ids: Tensor) -> Tensor: x = self.tok_emb(input_ids) x = F.rms_norm(x, (x.size(-1),)) x0 = x - skips: list[Tensor] = [] - - # First half stores skips; second half reuses them in reverse order. - 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).reshape(-1, x.size(-1)) - targets = target_ids.reshape(-1) + + # Pre-compute value embeddings once + ve_list: list[Tensor] = [] + if self.num_value_embeds > 0: + for ve in self.value_embeds: + ve_list.append(ve(input_ids)) # (bsz, seq, dim) + + num_blocks = len(self.blocks) + prev_block_outputs: list[Tensor | None] = [None] * num_blocks + layer_idx = 0 + for repeat in range(self.num_repeats): + for block_idx, block in enumerate(self.blocks): + x = x + self.loop_embed[layer_idx].to(dtype=x.dtype) + # Value embeddings: add weighted extra embeddings at each layer + for ve_idx, ve_out in enumerate(ve_list): + vs = self.value_scales[layer_idx, ve_idx].to(dtype=x.dtype) + x = x + vs[None, None, :] * ve_out + # Cross-repeat skip: mix in this block's output from previous repeat + if repeat > 0 and prev_block_outputs[block_idx] is not None: + scale = self.cross_repeat_scales[block_idx, repeat - 1].to(dtype=x.dtype) + x = x + scale[None, None, :] * prev_block_outputs[block_idx] + x = block(x, x0) + prev_block_outputs[block_idx] = x.detach() if not self.training else x + layer_idx += 1 + + x = self.final_norm(x) if self.tie_embeddings: logits_proj = F.linear(x, self.tok_emb.weight) else: @@ -721,6 +899,12 @@ def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: raise RuntimeError("lm_head is required when tie_embeddings=False") logits_proj = self.lm_head(x) logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return logits + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + logits = self.forward_logits(input_ids) + logits = logits.reshape(-1, logits.size(-1)) + targets = target_ids.reshape(-1) return F.cross_entropy(logits.float(), targets, reduction="mean") @@ -745,9 +929,7 @@ def main() -> None: 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_accum_steps = max(1, 8 // world_size) grad_scale = 1.0 / grad_accum_steps if not torch.cuda.is_available(): raise RuntimeError("CUDA is required") @@ -826,10 +1008,12 @@ def log0(msg: str, console: bool = True) -> None: base_model = GPT( vocab_size=args.vocab_size, num_layers=args.num_layers, + num_repeats=args.num_repeats, model_dim=args.model_dim, num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + num_value_embeds=args.num_value_embeds, tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, logit_softcap=args.logit_softcap, @@ -859,11 +1043,16 @@ def log0(msg: str, console: bool = True) -> None: 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) + scalar_params.append(base_model.loop_embed) + scalar_params.append(base_model.cross_repeat_scales) + if base_model.num_value_embeds > 0: + scalar_params.append(base_model.value_scales) token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + embed_params = [base_model.tok_emb.weight] + if base_model.num_value_embeds > 0: + embed_params.extend(ve.weight for ve in base_model.value_embeds) optimizer_tok = torch.optim.Adam( - [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + [{"params": embed_params, "lr": token_lr, "base_lr": token_lr}], betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, @@ -1118,6 +1307,56 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + # Sliding window eval + if args.eval_stride > 0: + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_sw = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"window:{args.eval_seq_len} stride:{args.eval_stride} " + f"eval_time:{1000.0 * (time.perf_counter() - t_sw):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + + # TTT eval: adapt model on each batch before evaluating + if args.ttt_steps > 0: + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, + base_model, + 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_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"ttt_steps:{args.ttt_steps} ttt_lr:{args.ttt_lr} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: dist.destroy_process_group()