diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/README.md b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/README.md new file mode 100644 index 0000000000..df5c0f857d --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/README.md @@ -0,0 +1,86 @@ +# 12L + Catalytic Residuals + BigramHash(10240) + SWA + Late QAT + +**val_bpb: 1.14662** (mean of 3 seeds, sliding window stride=64, post int6+zstd quantization roundtrip) + +## Run Command + +```bash +# Setup (once) +pip install sentencepiece zstandard +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 80 + +# Train + evaluate (default seed=1337) +torchrun --standalone --nproc_per_node=8 train_gpt.py + +# With specific seed +SEED=42 torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +All parameters are set as defaults in `train_gpt.py`. No env vars needed. + +## 3-Seed Results + +| Seed | val_bpb | artifact_bytes | valid | +|------|---------|---------------|-------| +| 1337 | 1.14749 | 14,014,540 | yes | +| 42 | 1.14575 | 14,104,510 | yes | +| 7 | 1.14662 | 14,385,363 | yes | +| **Mean** | **1.14662** | | | +| **Std** | **0.00071** | | | + +## Key Techniques + +### Catalytic Residual Connections (Novel) +- Replace `x + f(x)` with `x + c * f(x)`, where `c` is a learned per-dimension vector +- Initialized to ones (starts as standard residual) +- Provides dimension-wise gain control on residual connections +- Consistent -0.024 bpb improvement at zero computational overhead (~11K extra params) + +### 12 Layers (Depth Scaling) +- Standard stack uses 10-11 layers, leaving significant budget headroom +- 12 layers validated as the depth sweet spot (-0.023 bpb vs 11L) +- 13L+ shows diminishing returns due to throughput cost + +### BigramHash(10240) +- Hash consecutive token pairs into 10240-bucket embedding table (dim=128) +- Projected to model_dim=512 via learned linear +- -0.070 bpb improvement over BigramHash(2048) + +### Late QAT (Quantization-Aware Training) +- STE (straight-through estimator) int6 quantization in the final 4% of training +- Forward uses quantized weights, backward gets full-precision gradients +- Closes quantization gap to ~0.015 bpb + +### SWA (Stochastic Weight Averaging) +- Collect checkpoints from last 20% of warmdown +- Average ~16 checkpoints for smoother weight landscape + +## Architecture +- 12 layers, 512 dim, 8 heads, 4 KV heads (GQA) +- MLP 3x expansion (hidden=1536), relu^2 activation +- SmearGate + BigramHash(10240, dim=128) +- Orthogonal init with muP-scaled output projections +- U-Net skip connections, tied embeddings (FP16) +- XSA (cross-sequence attention) on last 4 layers +- Value Embeddings (dim=128) on layers 10, 11 +- Partial RoPE (16/64 dims) +- LN Scale (1/sqrt(layer_idx+1)) + +## Training Hyperparameters +- Muon optimizer: matrix_lr=0.04, WD=0.042, momentum=0.95 +- AdamW for embeddings/scalars: WD=0.042 +- warmdown=4000 iters, warmup=20 steps +- seq_len=2048, batch=786K tokens +- SWA: start_scale=0.2, every 50 steps +- Late QAT: threshold=0.25 + +## Evaluation +- Int6+zstd quantization roundtrip +- Sliding window eval: stride=64 +- No TTT (test-time training) + +## Training Metrics +- ~5,370 steps in 600s (~112 ms/step) on 8xH100 SXM +- Peak memory: ~25 GB per GPU + +Built on the standard stack from PR #180 by @thwu1. diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/submission.json b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/submission.json new file mode 100644 index 0000000000..034782530f --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/submission.json @@ -0,0 +1,9 @@ +{ + "name": "12L + Catalytic Residuals + BigramHash(10240) + SWA + Late QAT", + "val_loss": 1.93601, + "bytes_total": 14104510, + "blurb": "12 layers with Catalytic Residual Connections (learned per-dimension residual gain, -0.024 bpb at zero overhead). BigramHash 10240 buckets. SWA start_frac=0.2. Late QAT (STE int6 in final 4%). Int6+zstd quantization. Sliding window eval stride=64. Mean of 3 seeds: 1.14662 (std 0.00071). Built on PR #180 standard stack.", + "author": "zacharygoldfine", + "github_id": "zacharygoldfine", + "date": "2026-03-22" +} diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_gpt.py b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_gpt.py new file mode 100644 index 0000000000..53b2804f68 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_gpt.py @@ -0,0 +1,1851 @@ +""" +final_competition_recipe.py — Parameter Golf Competition Entry + +Best recipe: 12L + W6 Catalytic Residuals + BigramHash(10240) + Late QAT + SWA +Based on PR #180 standard stack with validated improvements from Phase 3 screening. + +COMPETITION DEFAULTS (ready for 8×H100, validated with 3 seeds): + NUM_LAYERS=12 # 12L validated as depth sweet spot (-0.023 bpb vs 11L) + CATALYTIC=1 # W6 Catalytic Residuals (-0.024 bpb, zero overhead) + BIGRAM_VOCAB_SIZE=10240 # Big bigram (-0.070 bpb) + XSA_LAST_N=4 # Cross-sequence attention on last 4 layers + LATE_QAT=1 # STE quantization in final 4% + QAT_THRESHOLD=0.25 # QAT activates when scale < 0.25 + EMA_ENABLED=0 # Disabled (SWA works better at this step count) + SWA_ENABLED=1 # SWA averaging of checkpoints + TTT_ENABLED=0 # Disabled + EVAL_STRIDE=64 # Sliding window eval (stride=64) + SLEEP_CONSOLIDATION=0 # Disabled + USE_CUDNN_SDPA=0 # Flash SDP (cuDNN has stride issues) + +Mean val_bpb: 1.14662 (3 seeds, sliding window stride=64) +""" + +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 + +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" + +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 + +try: + from flash_attn.flash_attn_interface import flash_attn_func as _fa3_func + _HAS_FA = True +except ImportError: + _HAS_FA = False + +# ----------------------------- +# 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", 4000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_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)) + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) # COMPETITION: 12L validated best + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + catalytic = int(os.environ.get("CATALYTIC", "1")) + 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", 0.05)) + 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)) + 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)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) # COMPETITION: SWA enabled + swa_every = int(os.environ.get("SWA_EVERY", 50)) + muon_wd = float(os.environ.get("MUON_WD", 0.042)) + adam_wd = float(os.environ.get("ADAM_WD", 0.042)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 10240)) # COMPETITION: big bigram validated + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + + # TTT (Test-Time Training) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) # COMPETITION: disabled + ttt_lr = float(os.environ.get("TTT_LR", 0.008)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 25)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + + # PR315 techniques (all opt-in, disabled by default) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) # 0 means full dims; 16 = partial (default) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + late_qat = bool(int(os.environ.get("LATE_QAT", "1"))) # COMPETITION: proven at 20K steps + qat_threshold = float(os.environ.get("QAT_THRESHOLD", "0.25")) # COMPETITION: validated threshold + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # COMPETITION: XSA on last 4 layers + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "0"))) # COMPETITION: disabled, using SWA + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + swa_start_scale = float(os.environ.get("SWA_START_SCALE", 0.2)) + + # Sleep Consolidation: replay hard batches in final 20% of training + sleep_consolidation = bool(int(os.environ.get("SLEEP_CONSOLIDATION", "0"))) # COMPETITION: disabled + sleep_buffer_size = int(os.environ.get("SLEEP_BUFFER_SIZE", "32")) + sleep_start_frac = float(os.environ.get("SLEEP_START_FRAC", "0.8")) # start replaying at 80% done + sleep_mix_ratio = float(os.environ.get("SLEEP_MIX_RATIO", "0.3")) # 30% replays, 70% fresh + +# ----------------------------- +# 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: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + 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) + + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + 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, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < 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}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // 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 * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, 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) + +# ----------------------------- +# 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,smear,ve_layer_scales,ve_shared.scale", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +def 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]: + 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() + 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() + 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): + _qat_enabled: bool = False + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, 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): + # NTK-aware RoPE: auto-scales base frequency when seq_len exceeds train_seq_len. + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.rope_dims = rope_dims if rope_dims > 0 else dim + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + rd = self.rope_dims + inv_freq = 1.0 / (base ** (torch.arange(0, rd, 2, dtype=torch.float32) / rd)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + rd = cos.size(-1) * 2 + if rd < x.size(-1): + x_rope, x_pass = x[..., :rd], x[..., rd:] + half = rd // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rot = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rot, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + rope_dims: int = 0, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + self.use_sdpa = bool(int(os.environ.get("USE_CUDNN_SDPA", "0"))) + self.use_xsa = False + 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, train_seq_len=1024, rope_dims=rope_dims) + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Subtract self-value projection via GQA-aware reshape (no repeat_interleave).""" + 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) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + 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] + if self.use_sdpa: + q2, k2, v2 = q.transpose(1,2), k.transpose(1,2), v.transpose(1,2) + if self.num_kv_heads < self.num_heads: + rep = self.num_heads // self.num_kv_heads + k2 = k2.repeat_interleave(rep, dim=1) + v2 = v2.repeat_interleave(rep, dim=1) + y = F.scaled_dot_product_attention(q2, k2, v2, is_causal=True).transpose(1,2) + elif _HAS_FA: + y = _fa3_func(q, k, v, causal=True) + else: + q2, k2, v2 = q.transpose(1,2), k.transpose(1,2), v.transpose(1,2) + if self.num_kv_heads < self.num_heads: + rep = self.num_heads // self.num_kv_heads + k2 = k2.repeat_interleave(rep, dim=1) + v2 = v2.repeat_interleave(rep, dim=1) + y = F.scaled_dot_product_attention(q2, k2, v2, is_causal=True).transpose(1,2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = torch.relu(self.fc(x)) + return self.proj(x.square()) + + +class 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, + rope_dims: int = 0, + layer_idx: int = 0, + ln_scale: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, rope_dims=rope_dims) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + if Hyperparameters.catalytic: + self.attn_gate = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_gate = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self.ln_scale_factor + attn_out = self.attn(self.attn_norm(x) * s, v_embed=v_embed) + ag = self.attn_gate.to(dtype=x.dtype)[None, None, :] if hasattr(self, 'attn_gate') else 1.0 + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * ag * attn_out + mg = self.mlp_gate.to(dtype=x.dtype)[None, None, :] if hasattr(self, 'mlp_gate') else 1.0 + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * mg * self.mlp(self.mlp_norm(x) * s) + return x + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + rope_dims: int = 0, + ln_scale: bool = False, + xsa_last_n: int = 0, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + ): + 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.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(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, + rope_dims=rope_dims, + layer_idx=i, + ln_scale=ln_scale, + ) + for i in range(num_layers) + ] + ) + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = num_kv_heads * (model_dim // num_heads) + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + 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) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + 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() + layer_idx = self.num_encoder_layers + i + ve = self._get_ve(layer_idx, input_ids, ve_cache) + x = self.blocks[layer_idx](x, x0, v_embed=ve) + + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, 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_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + + return main_loss + + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + 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() + layer_idx = self.num_encoder_layers + i + ve = self._get_ve(layer_idx, input_ids, ve_cache) + x = self.blocks[layer_idx](x, x0, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + + +# ----------------------------- +# SLIDING WINDOW EVALUATION +# ----------------------------- + +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, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# ----------------------------- +# INT6 MIXED QUANTIZATION (transplanted from working diagnostic scripts) +# ----------------------------- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + +def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + row_max = t32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0).to(torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8) + return q, scale + amax = t32.abs().max().item() + scale = torch.tensor(amax / 31.0 if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8) + return q, scale + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + # tok_emb.weight falls through to int8 via "embed" category + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +# ----------------------------- +# TTT (TEST-TIME TRAINING) +# ----------------------------- + +def ttt_adapt(args, base_model, device, val_tokens, rank=0, world_size=1, log_fn=None): + """Full-weight TTT: SGD adaptation on val data with DDP across all GPUs.""" + seq_len = args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + batch_seqs = args.ttt_batch_seqs + + frozen_params = set() + if args.ttt_freeze_blocks > 0: + for i, block in enumerate(base_model.blocks): + if i < args.ttt_freeze_blocks: + for p in block.parameters(): + p.requires_grad_(False) + frozen_params.add(id(p)) + + ttt_params = [p for p in base_model.parameters() if p.requires_grad] + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + + my_start = (total_seqs * rank) // world_size + my_end = (total_seqs * (rank + 1)) // world_size + + base_model.train() + t0 = time.perf_counter() + + for epoch in range(args.ttt_epochs): + epoch_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + epoch_tokens = torch.zeros((), device=device, dtype=torch.float64) + + for batch_start in range(my_start, my_end, batch_seqs): + batch_end = min(batch_start + batch_seqs, my_end) + raw_start = batch_start * seq_len + raw_end = batch_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + + epoch_loss_sum += loss.detach().to(torch.float64) * y.numel() + epoch_tokens += float(y.numel()) + + if world_size > 1: + dist.all_reduce(epoch_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(epoch_tokens, op=dist.ReduceOp.SUM) + + elapsed = time.perf_counter() - t0 + if log_fn: + log_fn(f"ttt_epoch:{epoch+1}/{args.ttt_epochs} loss:{epoch_loss_sum.item()/max(epoch_tokens.item(),1):.4f} time:{elapsed:.1f}s") + + for p in base_model.parameters(): + p.requires_grad_(True) + + if log_fn: + log_fn(f"ttt:done elapsed={time.perf_counter()-t0:.1f}s") + + +# ----------------------------- +# 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}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + # Fast math knobs + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + + use_cudnn_sdpa = bool(int(os.environ.get("USE_CUDNN_SDPA", "0"))) + if use_cudnn_sdpa: + enable_cudnn_sdp(True) + enable_flash_sdp(False) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + else: + enable_cudnn_sdp(False) + enable_flash_sdp(True) + if not _HAS_FA: + enable_mem_efficient_sdp(True) + enable_math_sdp(True) + else: + enable_mem_efficient_sdp(False) + enable_math_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"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_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 + # ----------------------------- + + CastedLinear._qat_enabled = args.qat_enabled + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + xsa_last_n=args.xsa_last_n, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + ).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) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + if base_model.ve_shared is not None: + scalar_params.append(base_model.ve_shared.scale) + for p in base_model.ve_layer_scales.parameters(): + scalar_params.append(p) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + 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()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"sdp_backends:cudnn={use_cudnn_sdpa} flash={not use_cudnn_sdpa and _HAS_FA} mem_efficient={not _HAS_FA and not use_cudnn_sdpa} math={not _HAS_FA and not use_cudnn_sdpa}") + 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 + # ----------------------------- + + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + + # Sleep Consolidation: buffer of hardest batches for replay in final phase + import heapq + sleep_buffer: list[tuple[float, int, tuple[Tensor, Tensor]]] = [] # (neg_loss, uid, (x, y)) + sleep_uid = 0 # unique id for heap tie-breaking + sleep_active = False + + 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) + if args.late_qat and scale < args.qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + 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) + # Sleep Consolidation: mix in hard batch replays in final phase + use_replay = False + if args.sleep_consolidation and sleep_active and len(sleep_buffer) > 0: + if random.random() < args.sleep_mix_ratio: + _, _, (rx, ry) = random.choice(sleep_buffer) + x, y = rx.to(x.device), ry.to(y.device) + use_replay = True + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + # Sleep Consolidation: store hard batches (before sleep phase starts) + if args.sleep_consolidation and not sleep_active and not use_replay: + loss_val = loss.item() + if len(sleep_buffer) < args.sleep_buffer_size: + heapq.heappush(sleep_buffer, (loss_val, sleep_uid, (x.detach().cpu(), y.detach().cpu()))) + sleep_uid += 1 + elif loss_val > sleep_buffer[0][0]: + heapq.heapreplace(sleep_buffer, (loss_val, sleep_uid, (x.detach().cpu(), y.detach().cpu()))) + sleep_uid += 1 + train_loss /= grad_accum_steps + # Sleep Consolidation: activate replay phase + if args.sleep_consolidation and not sleep_active: + progress = step / args.iterations if args.iterations > 0 else 1.0 + if progress >= args.sleep_start_frac: + sleep_active = True + log0(f"sleep_consolidation:active step:{step} buffer_size:{len(sleep_buffer)}") + + 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() + + if ema_state is not None: + d = args.ema_decay + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(d).add_(t.detach().float(), alpha=1.0 - d) + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + if args.swa_enabled and scale < args.swa_start_scale and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + + 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" + ) + + if ema_state is not None: + log0("ema:applying EMA weights") + avg_state = {name: t.to(dtype=base_model.state_dict()[name].dtype) for name, t in ema_state.items()} + del ema_state + base_model.load_state_dict(avg_state, strict=True) + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + avg_state = {name: (t / swa_count).to(dtype=base_model.state_dict()[name].dtype) + for name, t in swa_state.items()} + base_model.load_state_dict(avg_state, strict=True) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"}) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + + # Roundtrip: decompress + dequantize into fresh model + eval + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, xsa_last_n=args.xsa_last_n, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + + # TTT: adapt model on validation data before eval + if args.ttt_enabled: + if master_process: + log0(f"ttt:start lr={args.ttt_lr} momentum={args.ttt_momentum} epochs={args.ttt_epochs}") + t_ttt = time.perf_counter() + ttt_adapt(args, eval_model, device, val_tokens, rank=rank, world_size=world_size, log_fn=log0) + if master_process: + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + + # Standard non-overlapping eval (sanity check) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_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_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + + # Sliding window eval (submission score) + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + + # Second sliding window eval at stride=64 for submission comparison + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + + # === COMPETITION RESULTS SUMMARY === + if master_process: + total_time = training_time_ms / 1000.0 + artifact_mb = (quant_file_bytes + code_bytes) / (1024 * 1024) + log0("") + log0("=" * 50) + log0("=== COMPETITION RESULTS ===") + log0("=" * 50) + log0(f" Pre-quant val_bpb: {train_loss.item():.4f} (last train)") + log0(f" Post-quant val_bpb: {q_val_bpb:.4f}") + if args.eval_stride > 0: + log0(f" Sliding window val_bpb: {sw_val_bpb:.4f} (stride={args.eval_stride})") + log0(f" Artifact size: {artifact_mb:.2f} MB / 16.00 MB") + log0(f" Training time: {total_time:.1f}s ({total_time/60:.1f} min)") + log0(f" Steps completed: {step}") + log0(f" ms/step: {training_time_ms/step:.1f}") + log0(f" Layers: {args.num_layers} Catalytic: {args.catalytic} Bigram: {args.bigram_vocab_size}") + log0(f" Late QAT: {args.late_qat} EMA: {args.ema_enabled} SWA: {args.swa_enabled}") + log0(f" Sleep Consolidation: {args.sleep_consolidation}") + log0(f" TTT: {args.ttt_enabled} Eval stride: {args.eval_stride}") + log0("=" * 50) + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed1337.log b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed1337.log new file mode 100644 index 0000000000..b8b7b2c744 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed1337.log @@ -0,0 +1,108 @@ +W0322 17:26:54.376000 98204 torch/distributed/run.py:792] +W0322 17:26:54.376000 98204 torch/distributed/run.py:792] ***************************************** +W0322 17:26:54.376000 98204 torch/distributed/run.py:792] 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. +W0322 17:26:54.376000 98204 torch/distributed/run.py:792] ***************************************** +logs/5fe46e66-4773-4ea9-b021-7425aeb939a9.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:30416484 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +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:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:786432 train_seq_len:2048 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:6.9293 val_bpb:4.1039 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9320 train_time:185ms step_avg:185.02ms +step:2/20000 train_loss:9.6216 train_time:295ms step_avg:147.46ms +step:3/20000 train_loss:8.5067 train_time:410ms step_avg:136.62ms +step:4/20000 train_loss:7.7020 train_time:525ms step_avg:131.19ms +step:5/20000 train_loss:7.2805 train_time:639ms step_avg:127.86ms +step:6/20000 train_loss:7.2110 train_time:755ms step_avg:125.76ms +step:7/20000 train_loss:7.0927 train_time:869ms step_avg:124.19ms +step:8/20000 train_loss:7.1304 train_time:983ms step_avg:122.93ms +step:9/20000 train_loss:6.7573 train_time:1099ms step_avg:122.07ms +step:10/20000 train_loss:6.1886 train_time:1215ms step_avg:121.47ms +step:200/20000 train_loss:2.3817 train_time:22063ms step_avg:110.32ms +step:400/20000 train_loss:2.4212 train_time:44258ms step_avg:110.65ms +step:600/20000 train_loss:2.3404 train_time:66387ms step_avg:110.65ms +step:800/20000 train_loss:2.2414 train_time:88609ms step_avg:110.76ms +step:1000/20000 train_loss:2.2765 train_time:110720ms step_avg:110.72ms +step:1000/20000 val_loss:2.2225 val_bpb:1.3163 train_time:110734ms step_avg:110.73ms +step:1200/20000 train_loss:2.3565 train_time:132941ms step_avg:110.78ms +step:1400/20000 train_loss:2.1787 train_time:155168ms step_avg:110.83ms +step:1600/20000 train_loss:2.0689 train_time:177291ms step_avg:110.81ms +step:1800/20000 train_loss:2.1468 train_time:199499ms step_avg:110.83ms +step:2000/20000 train_loss:2.0692 train_time:221652ms step_avg:110.83ms +step:2000/20000 val_loss:2.1346 val_bpb:1.2643 train_time:221667ms step_avg:110.83ms +step:2200/20000 train_loss:2.1357 train_time:244879ms step_avg:111.31ms +step:2400/20000 train_loss:2.0694 train_time:267528ms step_avg:111.47ms +step:2600/20000 train_loss:2.1037 train_time:290768ms step_avg:111.83ms +step:2800/20000 train_loss:2.1473 train_time:314038ms step_avg:112.16ms +step:3000/20000 train_loss:2.1465 train_time:336771ms step_avg:112.26ms +step:3000/20000 val_loss:2.0771 val_bpb:1.2302 train_time:336784ms step_avg:112.26ms +step:3200/20000 train_loss:2.1545 train_time:358956ms step_avg:112.17ms +step:3400/20000 train_loss:1.9961 train_time:381029ms step_avg:112.07ms +step:3600/20000 train_loss:2.0665 train_time:403207ms step_avg:112.00ms +step:3800/20000 train_loss:2.0358 train_time:425304ms step_avg:111.92ms +step:4000/20000 train_loss:1.9338 train_time:447487ms step_avg:111.87ms +step:4000/20000 val_loss:2.0250 val_bpb:1.1993 train_time:447501ms step_avg:111.88ms +step:4200/20000 train_loss:2.1053 train_time:469660ms step_avg:111.82ms +late_qat:enabled step:4369 scale:0.2498 +step:4400/20000 train_loss:1.9830 train_time:491738ms step_avg:111.76ms +swa:start step:4600 +step:4600/20000 train_loss:1.7860 train_time:513918ms step_avg:111.72ms +step:4800/20000 train_loss:2.3662 train_time:536215ms step_avg:111.71ms +step:5000/20000 train_loss:2.0339 train_time:558550ms step_avg:111.71ms +step:5000/20000 val_loss:1.9551 val_bpb:1.1579 train_time:558598ms step_avg:111.72ms +step:5200/20000 train_loss:1.9679 train_time:580815ms step_avg:111.70ms +step:5371/20000 val_loss:1.9316 val_bpb:1.1440 train_time:599966ms step_avg:111.70ms +stopping_early: wallclock_cap train_time:599966ms step:5371/20000 +peak memory allocated: 22248 MiB reserved: 22452 MiB +swa:applying averaged 16 checkpoints +Serialized model: 117782520 bytes +Code size: 80877 bytes +Serialized model int6+zstd: 13933663 bytes +Total submission size int6+zstd: 14014540 bytes +final_int6_roundtrip val_loss:1.9778 val_bpb:1.1714 eval_time:20401ms +final_int6_roundtrip_exact val_loss:1.97780547 val_bpb:1.17136780 +final_int6_sliding_window val_loss:1.9375 val_bpb:1.1475 stride:64 eval_time:98371ms +final_int6_sliding_window_exact val_loss:1.93747980 val_bpb:1.14748771 + +================================================== +=== COMPETITION RESULTS === +================================================== + Pre-quant val_bpb: 2.1385 (last train) + Post-quant val_bpb: 1.1714 + Sliding window val_bpb: 1.1475 (stride=64) + Artifact size: 13.37 MB / 16.00 MB + Training time: 600.0s (10.0 min) + Steps completed: 5371 + ms/step: 111.7 + Layers: 12 Catalytic: 1 Bigram: 10240 + Late QAT: True EMA: False SWA: True + Sleep Consolidation: False + TTT: False Eval stride: 64 +================================================== diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed42.log b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed42.log new file mode 100644 index 0000000000..1056430081 --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed42.log @@ -0,0 +1,108 @@ +W0322 17:41:18.014000 99625 torch/distributed/run.py:792] +W0322 17:41:18.014000 99625 torch/distributed/run.py:792] ***************************************** +W0322 17:41:18.014000 99625 torch/distributed/run.py:792] 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. +W0322 17:41:18.014000 99625 torch/distributed/run.py:792] ***************************************** +logs/dee1a284-8ee9-4c17-a823-a974a8b68e1e.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:30416484 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +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:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:42 +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:6.9316 val_bpb:4.1053 train_time:0ms step_avg:0.03ms +step:1/20000 train_loss:6.9327 train_time:162ms step_avg:161.85ms +step:2/20000 train_loss:9.9783 train_time:272ms step_avg:135.85ms +step:3/20000 train_loss:8.6554 train_time:388ms step_avg:129.33ms +step:4/20000 train_loss:7.7880 train_time:503ms step_avg:125.66ms +step:5/20000 train_loss:7.3174 train_time:619ms step_avg:123.70ms +step:6/20000 train_loss:7.2076 train_time:734ms step_avg:122.26ms +step:7/20000 train_loss:7.0720 train_time:849ms step_avg:121.22ms +step:8/20000 train_loss:6.9911 train_time:963ms step_avg:120.38ms +step:9/20000 train_loss:6.6814 train_time:1079ms step_avg:119.89ms +step:10/20000 train_loss:6.1561 train_time:1194ms step_avg:119.41ms +step:200/20000 train_loss:2.3488 train_time:22113ms step_avg:110.57ms +step:400/20000 train_loss:2.4068 train_time:44327ms step_avg:110.82ms +step:600/20000 train_loss:2.3276 train_time:66488ms step_avg:110.81ms +step:800/20000 train_loss:2.2273 train_time:88743ms step_avg:110.93ms +step:1000/20000 train_loss:2.2663 train_time:110883ms step_avg:110.88ms +step:1000/20000 val_loss:2.2154 val_bpb:1.3121 train_time:110897ms step_avg:110.90ms +step:1200/20000 train_loss:2.3428 train_time:133103ms step_avg:110.92ms +step:1400/20000 train_loss:2.1770 train_time:155349ms step_avg:110.96ms +step:1600/20000 train_loss:2.0632 train_time:177469ms step_avg:110.92ms +step:1800/20000 train_loss:2.1455 train_time:199737ms step_avg:110.96ms +step:2000/20000 train_loss:2.0631 train_time:221900ms step_avg:110.95ms +step:2000/20000 val_loss:2.1316 val_bpb:1.2625 train_time:221914ms step_avg:110.96ms +step:2200/20000 train_loss:2.1276 train_time:244616ms step_avg:111.19ms +step:2400/20000 train_loss:2.0593 train_time:267239ms step_avg:111.35ms +step:2600/20000 train_loss:2.1036 train_time:290515ms step_avg:111.74ms +step:2800/20000 train_loss:2.1450 train_time:314086ms step_avg:112.17ms +step:3000/20000 train_loss:2.1455 train_time:336907ms step_avg:112.30ms +step:3000/20000 val_loss:2.0743 val_bpb:1.2285 train_time:336922ms step_avg:112.31ms +step:3200/20000 train_loss:2.1512 train_time:359105ms step_avg:112.22ms +step:3400/20000 train_loss:1.9953 train_time:381199ms step_avg:112.12ms +step:3600/20000 train_loss:2.0614 train_time:403422ms step_avg:112.06ms +step:3800/20000 train_loss:2.0347 train_time:425539ms step_avg:111.98ms +step:4000/20000 train_loss:1.9324 train_time:447730ms step_avg:111.93ms +step:4000/20000 val_loss:2.0212 val_bpb:1.1971 train_time:447744ms step_avg:111.94ms +step:4200/20000 train_loss:2.1008 train_time:469916ms step_avg:111.88ms +late_qat:enabled step:4366 scale:0.2498 +step:4400/20000 train_loss:1.9759 train_time:492010ms step_avg:111.82ms +swa:start step:4600 +step:4600/20000 train_loss:1.7878 train_time:514202ms step_avg:111.78ms +step:4800/20000 train_loss:2.3621 train_time:536558ms step_avg:111.78ms +step:5000/20000 train_loss:2.0319 train_time:558907ms step_avg:111.78ms +step:5000/20000 val_loss:1.9515 val_bpb:1.1558 train_time:558955ms step_avg:111.79ms +step:5200/20000 train_loss:1.9637 train_time:581170ms step_avg:111.76ms +step:5368/20000 val_loss:1.9282 val_bpb:1.1420 train_time:600072ms step_avg:111.79ms +stopping_early: wallclock_cap train_time:600072ms step:5368/20000 +peak memory allocated: 22248 MiB reserved: 22350 MiB +swa:applying averaged 16 checkpoints +Serialized model: 117782520 bytes +Code size: 80877 bytes +Serialized model int6+zstd: 14023633 bytes +Total submission size int6+zstd: 14104510 bytes +final_int6_roundtrip val_loss:1.9748 val_bpb:1.1696 eval_time:21758ms +final_int6_roundtrip_exact val_loss:1.97480102 val_bpb:1.16958840 +final_int6_sliding_window val_loss:1.9345 val_bpb:1.1457 stride:64 eval_time:98236ms +final_int6_sliding_window_exact val_loss:1.93454046 val_bpb:1.14574686 + +================================================== +=== COMPETITION RESULTS === +================================================== + Pre-quant val_bpb: 1.9400 (last train) + Post-quant val_bpb: 1.1696 + Sliding window val_bpb: 1.1457 (stride=64) + Artifact size: 13.45 MB / 16.00 MB + Training time: 600.1s (10.0 min) + Steps completed: 5368 + ms/step: 111.8 + Layers: 12 Catalytic: 1 Bigram: 10240 + Late QAT: True EMA: False SWA: True + Sleep Consolidation: False + TTT: False Eval stride: 64 +================================================== diff --git a/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed7.log b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed7.log new file mode 100644 index 0000000000..6c1d70e0eb --- /dev/null +++ b/records/track_10min_16mb/2026-03-22_12L_CatalyticResiduals_BigBigram/train_seed7.log @@ -0,0 +1,108 @@ +W0322 17:55:46.158000 101024 torch/distributed/run.py:792] +W0322 17:55:46.158000 101024 torch/distributed/run.py:792] ***************************************** +W0322 17:55:46.158000 101024 torch/distributed/run.py:792] 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. +W0322 17:55:46.158000 101024 torch/distributed/run.py:792] ***************************************** +logs/4142ca15-55eb-46bd-9100-f3238113b93f.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:30416484 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +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:0.05 head_lr:0.0 matrix_lr:0.04 scalar_lr:0.04 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:7 +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:6.9288 val_bpb:4.1036 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9305 train_time:159ms step_avg:159.21ms +step:2/20000 train_loss:9.8817 train_time:269ms step_avg:134.62ms +step:3/20000 train_loss:8.6670 train_time:385ms step_avg:128.27ms +step:4/20000 train_loss:7.7515 train_time:501ms step_avg:125.25ms +step:5/20000 train_loss:7.2045 train_time:616ms step_avg:123.16ms +step:6/20000 train_loss:7.0933 train_time:732ms step_avg:121.94ms +step:7/20000 train_loss:6.9881 train_time:847ms step_avg:120.94ms +step:8/20000 train_loss:7.2235 train_time:961ms step_avg:120.15ms +step:9/20000 train_loss:6.8761 train_time:1076ms step_avg:119.58ms +step:10/20000 train_loss:6.2638 train_time:1192ms step_avg:119.23ms +step:200/20000 train_loss:2.3548 train_time:22108ms step_avg:110.54ms +step:400/20000 train_loss:2.4018 train_time:44306ms step_avg:110.76ms +step:600/20000 train_loss:2.3286 train_time:66446ms step_avg:110.74ms +step:800/20000 train_loss:2.2300 train_time:88700ms step_avg:110.88ms +step:1000/20000 train_loss:2.2691 train_time:110849ms step_avg:110.85ms +step:1000/20000 val_loss:2.2179 val_bpb:1.3135 train_time:110864ms step_avg:110.86ms +step:1200/20000 train_loss:2.3483 train_time:133128ms step_avg:110.94ms +step:1400/20000 train_loss:2.1757 train_time:155439ms step_avg:111.03ms +step:1600/20000 train_loss:2.0633 train_time:177682ms step_avg:111.05ms +step:1800/20000 train_loss:2.1425 train_time:199979ms step_avg:111.10ms +step:2000/20000 train_loss:2.0675 train_time:222191ms step_avg:111.10ms +step:2000/20000 val_loss:2.1308 val_bpb:1.2620 train_time:222207ms step_avg:111.10ms +step:2200/20000 train_loss:2.1304 train_time:245479ms step_avg:111.58ms +step:2400/20000 train_loss:2.0626 train_time:268130ms step_avg:111.72ms +step:2600/20000 train_loss:2.1011 train_time:291389ms step_avg:112.07ms +step:2800/20000 train_loss:2.1445 train_time:314619ms step_avg:112.36ms +step:3000/20000 train_loss:2.1408 train_time:337346ms step_avg:112.45ms +step:3000/20000 val_loss:2.0742 val_bpb:1.2284 train_time:337360ms step_avg:112.45ms +step:3200/20000 train_loss:2.1515 train_time:359536ms step_avg:112.35ms +step:3400/20000 train_loss:1.9968 train_time:381640ms step_avg:112.25ms +step:3600/20000 train_loss:2.0630 train_time:403812ms step_avg:112.17ms +step:3800/20000 train_loss:2.0332 train_time:425925ms step_avg:112.09ms +step:4000/20000 train_loss:1.9342 train_time:448128ms step_avg:112.03ms +step:4000/20000 val_loss:2.0235 val_bpb:1.1984 train_time:448143ms step_avg:112.04ms +step:4200/20000 train_loss:2.1056 train_time:470310ms step_avg:111.98ms +late_qat:enabled step:4361 scale:0.2499 +step:4400/20000 train_loss:1.9829 train_time:492394ms step_avg:111.91ms +swa:start step:4600 +step:4600/20000 train_loss:1.7873 train_time:514577ms step_avg:111.86ms +step:4800/20000 train_loss:2.3662 train_time:536914ms step_avg:111.86ms +step:5000/20000 train_loss:2.0342 train_time:559234ms step_avg:111.85ms +step:5000/20000 val_loss:1.9539 val_bpb:1.1572 train_time:559288ms step_avg:111.86ms +step:5200/20000 train_loss:1.9649 train_time:581487ms step_avg:111.82ms +step:5365/20000 val_loss:1.9310 val_bpb:1.1437 train_time:600032ms step_avg:111.84ms +stopping_early: wallclock_cap train_time:600032ms step:5365/20000 +peak memory allocated: 22249 MiB reserved: 22446 MiB +swa:applying averaged 16 checkpoints +Serialized model: 117782520 bytes +Code size: 80877 bytes +Serialized model int6+zstd: 14304486 bytes +Total submission size int6+zstd: 14385363 bytes +final_int6_roundtrip val_loss:1.9764 val_bpb:1.1705 eval_time:21291ms +final_int6_roundtrip_exact val_loss:1.97640853 val_bpb:1.17054046 +final_int6_sliding_window val_loss:1.9360 val_bpb:1.1466 stride:64 eval_time:98311ms +final_int6_sliding_window_exact val_loss:1.93602036 val_bpb:1.14662334 + +================================================== +=== COMPETITION RESULTS === +================================================== + Pre-quant val_bpb: 1.9276 (last train) + Post-quant val_bpb: 1.1705 + Sliding window val_bpb: 1.1466 (stride=64) + Artifact size: 13.72 MB / 16.00 MB + Training time: 600.0s (10.0 min) + Steps completed: 5365 + ms/step: 111.8 + Layers: 12 Catalytic: 1 Bigram: 10240 + Late QAT: True EMA: False SWA: True + Sleep Consolidation: False + TTT: False Eval stride: 64 +==================================================