From 515f6e8749a34099008230aaad6ea7eadeeda49c Mon Sep 17 00:00:00 2001 From: Amaljith Kuttamath Date: Sun, 22 Mar 2026 21:48:32 -0400 Subject: [PATCH 1/4] Record: 11L VR+GA + EMA + AdamW TTT (val_bpb=1.0891) Value Residual + Gated Attention on PR #442 stack. Single seed (1337), 8xH100 SXM, 14.2 MB artifact. --- .../README.md | 67 + .../submission.json | 43 + .../train_gpt.py | 1844 +++++++++++++++++ .../train_seed1337.log | 111 + 4 files changed, 2065 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md create mode 100644 records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/submission.json create mode 100644 records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_seed1337.log diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md new file mode 100644 index 0000000000..53b308b39b --- /dev/null +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md @@ -0,0 +1,67 @@ +# Record: 11L EMA + Value Residual + Gated Attention + AdamW TTT (val_bpb=1.0891) + +**val_bpb = 1.0891** (sliding window stride=64, seed 1337) | **14.2 MB** artifact | 8xH100 SXM, 600s + +## Approach + +Two architecture changes on top of the PR #442 recipe (11L EMA + AdamW TTT): + +**Value Residual** (ResFormer, arXiv:2410.17897): Each attention block receives the raw V from the first block. A learned 2-element lambda blends first-block V with current V before attention. Block 0 passes V through unchanged (no lambda parameter). Adds 2 params per layer (layers 1-10 only). + +**Gated Attention** (arXiv:2505.06708): Per-head sigmoid gate on attention output. Learned weight matrix (dim x num_heads) + bias initialized to 4.0 (near-open gate at init). Adds 4104 params per layer. + +Both techniques were ablated individually in PR #413 (-0.015 and -0.003 bpb respectively, -0.017 combined). This is the first validation on the full competitive stack with AdamW TTT. + +## Results (seed 1337, 8xH100 SXM) + +| Metric | Value | +|--------|-------| +| Training steps | 6,021 (wallclock capped at 600s) | +| Step time | 99.66 ms/step | +| Pre-quant val_bpb | 1.1545 | +| Post-quant roundtrip val_bpb | 1.0964 | +| **Sliding window val_bpb (s=64)** | **1.0891** | +| Artifact size | 14,195,825 bytes | +| Peak GPU memory | 21,374 MiB | +| TTT time | 171.8s | + +## Comparison to prior SOTA + +| Submission | Best BPB | Steps | Step time | +|-----------|----------|-------|-----------| +| **Ours** | **1.0891** | 6,021 | 99.7 ms | +| PR #442 (sjp611) | 1.0992 | 4,612 | ~137 ms | +| PR #481 (mrdavtan) | 1.0959 | 7,101 | ~84 ms | + +## Key findings + +1. VR+GA adds ~300K params (27.1M vs 26.8M) with negligible throughput cost +2. Faster step time (99.7ms vs PR #442's 137ms) yields 38% more training steps +3. AdamW TTT recovers 0.065 bpb from quantized model (1.1545 -> 1.0891 with sliding window) + +## Config + +All hyperparameters set as defaults in train_gpt.py. Key settings: + +``` +NUM_LAYERS=11 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4 +MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 +ITERATIONS=9000 WARMDOWN_ITERS=1200 +EMA_ENABLED=1 EMA_DECAY=0.997 +VALUE_RESIDUAL=1 GATED_ATTENTION=1 +TTT_ENABLED=1 TTT_LR=0.0005 TTT_EPOCHS=10 +EVAL_STRIDE=64 +``` + +## Run command + +```bash +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` + +## Credits + +- **PR #442** (sjp611): AdamW TTT, base recipe +- **PR #398** (felipe-parodi): EMA, aggressive TTT findings +- **PR #413**: Value Residual + Gated Attention ablation +- **PR #315** (jfprincz): Foundation architecture (U-Net skips, SmearGate, orthogonal init) diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/submission.json b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/submission.json new file mode 100644 index 0000000000..28375bf24c --- /dev/null +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/submission.json @@ -0,0 +1,43 @@ +{ + "track": "10min_16mb", + "val_bpb": 1.0891, + "val_bpb_exact": 1.08909943, + "seeds": { + "1337": 1.08909943 + }, + "artifact_bytes": 14195825, + "code_bytes": 78596, + "hardware": "8xH100 SXM", + "training_seconds": 600, + "training_steps": 6021, + "step_avg_ms": 99.66, + "ttt_epochs": 10, + "ttt_optimizer": "adamw", + "ttt_lr": 0.0005, + "ttt_seconds": 171.8, + "eval_stride": 64, + "eval_seq_len": 2048, + "model": { + "num_layers": 11, + "model_dim": 512, + "num_heads": 8, + "num_kv_heads": 4, + "mlp_mult": 3, + "vocab_size": 1024, + "params": 27137221 + }, + "techniques": [ + "Value Residual (ResFormer)", + "Gated Attention", + "EMA (decay=0.997)", + "AdamW TTT (10 epochs)", + "SmearGate", + "BigramHash (4096 buckets)", + "Orthogonal init", + "U-Net skip connections", + "GPTQ-lite quantization", + "2% magnitude pruning", + "Int6 + zlib compression", + "Sliding window eval (stride=64)" + ] +} diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py new file mode 100644 index 0000000000..91dcd24190 --- /dev/null +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py @@ -0,0 +1,1844 @@ +""" +train_gpt_submit.py — Submission v2: wider MLP + STE int6 QAT + MTP + seq2048 + NTK RoPE + +fp16 embed + late-K passthrough + sliding window eval. +""" + +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_interface import flash_attn_func as flash_attn_3_func +except ImportError: + from flash_attn import flash_attn_func as flash_attn_3_func + +# ----------------------------- +# 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", 9000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + 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", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + 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", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + 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"))) + swa_every = int(os.environ.get("SWA_EVERY", 200)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 0)) + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "1"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + rope_dims = int(os.environ.get("ROPE_DIMS", 0)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "0"))) + late_qat = bool(int(os.environ.get("LATE_QAT", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "1"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "1"))) + + # TTT (Test-Time Training) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.0005)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 10)) + 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)) + +# ----------------------------- +# 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", + ).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, + value_residual: bool = False, + layer_idx: int = 0, + gated_attention: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = rope_dims + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.use_xsa = False + self._value_residual = value_residual + if value_residual and layer_idx > 0: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + self._gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + + 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, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self._value_residual else None + if self._value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + fa_dtype = torch.bfloat16 + y = flash_attn_3_func(q.to(fa_dtype), k.to(fa_dtype), v.to(fa_dtype), causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self._gated_attention: + gate = torch.sigmoid(self.attn_gate(x)) + y = y * gate.unsqueeze(-1) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v + + +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 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, + value_residual: bool = False, + gated_attention: 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, value_residual=value_residual, + layer_idx=layer_idx, gated_attention=gated_attention) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x: Tensor, x0: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self.ln_scale_factor + attn_out, raw_v = self.attn(self.attn_norm(x) * s, v0=v0) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x) * s) + return x, raw_v + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: 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, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + value_residual: bool = False, + gated_attention: bool = False, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.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, + value_residual=value_residual, + gated_attention=gated_attention, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.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 + 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._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 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 + v0 = None + skips: list[Tensor] = [] + + for i in range(self.num_encoder_layers): + x, raw_v = self.blocks[i](x, x0, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = self.blocks[self.num_encoder_layers + i](x, x0, v0=v0) + + 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 + v0 = None + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x, raw_v = self.blocks[i](x, x0, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = self.blocks[self.num_encoder_layers + i](x, x0, v0=v0) + 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() + base_model.bfloat16() + 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 + + +# ----------------------------- +# TEST-TIME TRAINING (TTT) +# ----------------------------- + +def ttt_adapt(args: Hyperparameters, base_model: nn.Module, device: torch.device, + val_tokens: Tensor, rank: int = 0, world_size: int = 1, + log_fn=None) -> None: + """Single-phase AdamW TTT on validation data (PR #442 approach). + All blocks unfrozen, AdamW lr=0.0005, 10 epochs.""" + seq_len = args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + batch_seqs = args.ttt_batch_seqs + + ttt_params = [p for p in base_model.parameters() if p.requires_grad] + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0) + + 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} " + f"loss:{epoch_loss_sum.item()/max(epoch_tokens.item(),1):.4f} time:{elapsed:.1f}s") + + if log_fn: + log_fn(f"ttt:done elapsed={time.perf_counter()-t0:.1f}s") + + +# ----------------------------- +# FULL GPTQ QUANTIZATION +# ----------------------------- + +def gptq_quantize_int6(W: Tensor, H: Tensor, blocksize: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ: column-wise int6 quantization with Hessian error compensation. + W: (out_features, in_features) weight matrix + H: (in_features, in_features) Hessian = X^T @ X from calibration data + Returns: (quantized_int8, per_row_scale_fp16) + """ + rows, cols = W.shape + W = W.float().clone() + Q = torch.zeros_like(W) + # Damping for numerical stability + damp = percdamp * torch.diag(H).mean() + H = H.float() + damp * torch.eye(cols, device=H.device, dtype=torch.float32) + # Compute H_inv via Cholesky + try: + H_inv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + except torch.linalg.LinAlgError: + H = H + 0.1 * torch.eye(cols, device=H.device, dtype=torch.float32) + H_inv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + # Compute per-row scales from original weights + row_max = W.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + # Block-wise GPTQ + for col_start in range(0, cols, blocksize): + col_end = min(col_start + blocksize, cols) + bs = col_end - col_start + W_block = W[:, col_start:col_end].clone() + Q_block = torch.zeros_like(W_block) + Err_block = torch.zeros_like(W_block) + H_inv_block = H_inv[col_start:col_end, col_start:col_end] + for j in range(bs): + w = W_block[:, j] + d = H_inv_block[j, j] + q = torch.clamp(torch.round(w / scale), -32, 31) + Q_block[:, j] = q + err = (w - q * scale) / d + Err_block[:, j] = err + # Update remaining columns in this block + W_block[:, j + 1:] -= err.unsqueeze(1) * H_inv_block[j, j + 1:].unsqueeze(0) + Q[:, col_start:col_end] = Q_block + # Propagate block error to all remaining columns + if col_end < cols: + W[:, col_end:] -= Err_block @ H_inv[col_start:col_end, col_end:] + return Q.to(torch.int8), scale.to(torch.float16) + + +# Calibration data collection hook for GPTQ +_gptq_hessians: dict[str, Tensor] = {} +_gptq_nsamples: dict[str, int] = {} + +class GPTQHook: + """Forward hook to accumulate H = X^T @ X for a linear layer.""" + def __init__(self, name: str): + self.name = name + def __call__(self, module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + n = x.shape[0] + H = x.T @ x + if self.name not in _gptq_hessians: + _gptq_hessians[self.name] = H + _gptq_nsamples[self.name] = n + else: + _gptq_hessians[self.name] += H + _gptq_nsamples[self.name] += n + + +# ----------------------------- +# 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: + # GPTQ-lite: search 5 clip percentiles per row, pick min MSE + row_max = t32.abs().amax(dim=1) + best_q = None + best_scale = None + best_mse = None + for clip_frac in [0.95, 0.975, 0.99, 0.999, 1.0]: + clip_val = row_max * clip_frac + s = (clip_val / 31.0).clamp_min(1.0 / 31.0) + clipped = torch.clamp(t32, -clip_val[:, None], clip_val[:, None]) + q = torch.clamp(torch.round(clipped / s[:, None]), -32, 31) + recon = q * s[:, None] + mse = ((t32 - recon) ** 2).mean(dim=1) + if best_mse is None: + best_mse = mse + best_q = q + best_scale = s + else: + improved = mse < best_mse + best_mse = torch.where(improved, mse, best_mse) + best_q = torch.where(improved[:, None], q, best_q) + best_scale = torch.where(improved, s, best_scale) + return best_q.to(torch.int8), best_scale.to(torch.float16) + 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: + # Use Full GPTQ if Hessian available, else GPTQ-lite + if name in _gptq_hessians and t.ndim == 2: + H = _gptq_hessians[name] / _gptq_nsamples[name] + q, s = gptq_quantize_int6(t, H) + else: + 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 + + +# ----------------------------- +# 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 + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + value_residual=args.value_residual, + gated_attention=args.gated_attention, + ).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) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.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("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + + training_time_ms = 0.0 + stop_after_step: int | None = None + 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) + qat_threshold = float(os.environ.get("QAT_THRESHOLD", "0.1")) + if args.late_qat and scale < 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) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + 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) + + if args.swa_enabled and not args.ema_enabled and scale < 0.5 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().float().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].add_(t.detach().float()) + 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()} + del swa_state + 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") + + # GPTQ calibration: collect Hessians from a few training batches + log0("gptq_calibration:start") + _gptq_hessians.clear() + _gptq_nsamples.clear() + hooks = [] + for name, module in base_model.named_modules(): + if isinstance(module, CastedLinear) and module.weight.numel() > 65536: + param_name = name + ".weight" + hooks.append(module.register_forward_hook(GPTQHook(param_name))) + base_model.eval() + base_model.bfloat16() + calib_batches = 8 + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + with torch.no_grad(): + for i in range(calib_batches): + x_calib, y_calib = calib_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + base_model(x_calib, y_calib) + for h in hooks: + h.remove() + # Move Hessians to CPU + for k in _gptq_hessians: + _gptq_hessians[k] = _gptq_hessians[k].cpu() + log0(f"gptq_calibration:done layers={len(_gptq_hessians)}") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # Magnitude pruning: zero out smallest 2% of weights in large 2D tensors + prune_pct = 0.02 + for name, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536 and v.is_floating_point(): + thresh = torch.quantile(v.abs().float(), prune_pct) + v[v.abs() < thresh] = 0.0 + log0(f"magnitude_pruning:{prune_pct*100:.0f}%") + 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, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + value_residual=args.value_residual, + gated_attention=args.gated_attention, + ).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 distributed: + dist.barrier() + for block in eval_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} optimizer=adamw") + t_ttt = time.perf_counter() + ttt_adapt(args, eval_model, device, val_tokens, + rank=rank, world_size=world_size, log_fn=log0) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + if distributed: + dist.barrier() + + 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}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_seed1337.log b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_seed1337.log new file mode 100644 index 0000000000..e58a8d050c --- /dev/null +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_seed1337.log @@ -0,0 +1,111 @@ +W0323 01:20:18.806000 790 torch/distributed/run.py:803] +W0323 01:20:18.806000 790 torch/distributed/run.py:803] ***************************************** +W0323 01:20:18.806000 790 torch/distributed/run.py:803] 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. +W0323 01:20:18.806000 790 torch/distributed/run.py:803] ***************************************** +logs/submission_vr_ga_8xh100.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:27137221 +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:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:9000 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:0/9000 val_loss:6.9286 val_bpb:4.1035 train_time:0ms step_avg:0.01ms +step:1/9000 train_loss:6.9308 train_time:157ms step_avg:157.33ms +step:2/9000 train_loss:8.5186 train_time:253ms step_avg:126.59ms +step:3/9000 train_loss:7.7960 train_time:351ms step_avg:116.85ms +step:4/9000 train_loss:7.2352 train_time:448ms step_avg:112.04ms +step:5/9000 train_loss:6.9703 train_time:546ms step_avg:109.10ms +step:6/9000 train_loss:6.8758 train_time:643ms step_avg:107.15ms +step:7/9000 train_loss:6.7993 train_time:741ms step_avg:105.87ms +step:8/9000 train_loss:6.7274 train_time:838ms step_avg:104.76ms +step:9/9000 train_loss:6.4122 train_time:936ms step_avg:103.96ms +step:10/9000 train_loss:6.0712 train_time:1033ms step_avg:103.31ms +step:200/9000 train_loss:2.3410 train_time:19876ms step_avg:99.38ms +step:400/9000 train_loss:2.3965 train_time:39809ms step_avg:99.52ms +step:600/9000 train_loss:2.3173 train_time:59664ms step_avg:99.44ms +step:800/9000 train_loss:2.2112 train_time:79587ms step_avg:99.48ms +step:1000/9000 train_loss:2.2458 train_time:99454ms step_avg:99.45ms +step:1000/9000 val_loss:2.1982 val_bpb:1.3019 train_time:99460ms step_avg:99.46ms +step:1200/9000 train_loss:2.3253 train_time:119389ms step_avg:99.49ms +step:1400/9000 train_loss:2.1541 train_time:139354ms step_avg:99.54ms +step:1600/9000 train_loss:2.0472 train_time:159237ms step_avg:99.52ms +step:1800/9000 train_loss:2.1392 train_time:179195ms step_avg:99.55ms +step:2000/9000 train_loss:2.0610 train_time:199106ms step_avg:99.55ms +step:2000/9000 val_loss:2.1259 val_bpb:1.2591 train_time:199112ms step_avg:99.56ms +step:2200/9000 train_loss:2.1526 train_time:219067ms step_avg:99.58ms +step:2400/9000 train_loss:2.0683 train_time:238976ms step_avg:99.57ms +step:2600/9000 train_loss:2.1166 train_time:258952ms step_avg:99.60ms +step:2800/9000 train_loss:2.1646 train_time:278902ms step_avg:99.61ms +step:3000/9000 train_loss:2.1725 train_time:298804ms step_avg:99.60ms +step:3000/9000 val_loss:2.1060 val_bpb:1.2473 train_time:298810ms step_avg:99.60ms +step:3200/9000 train_loss:2.1883 train_time:318775ms step_avg:99.62ms +step:3400/9000 train_loss:2.0357 train_time:338664ms step_avg:99.61ms +step:3600/9000 train_loss:2.1123 train_time:358622ms step_avg:99.62ms +step:3800/9000 train_loss:2.0893 train_time:378510ms step_avg:99.61ms +step:4000/9000 train_loss:1.9989 train_time:398482ms step_avg:99.62ms +step:4000/9000 val_loss:2.0906 val_bpb:1.2382 train_time:398490ms step_avg:99.62ms +step:4200/9000 train_loss:2.1813 train_time:418465ms step_avg:99.63ms +step:4400/9000 train_loss:2.0676 train_time:438370ms step_avg:99.63ms +step:4600/9000 train_loss:1.8856 train_time:458399ms step_avg:99.65ms +step:4800/9000 train_loss:2.4675 train_time:478299ms step_avg:99.65ms +step:5000/9000 train_loss:2.1431 train_time:498271ms step_avg:99.65ms +step:5000/9000 val_loss:2.0600 val_bpb:1.2201 train_time:498277ms step_avg:99.66ms +step:5200/9000 train_loss:2.0637 train_time:518172ms step_avg:99.65ms +step:5400/9000 train_loss:2.0583 train_time:538134ms step_avg:99.65ms +step:5600/9000 train_loss:1.9514 train_time:558099ms step_avg:99.66ms +step:5800/9000 train_loss:1.9783 train_time:578002ms step_avg:99.66ms +step:6000/9000 train_loss:1.9101 train_time:597951ms step_avg:99.66ms +step:6000/9000 val_loss:1.9501 val_bpb:1.1550 train_time:597958ms step_avg:99.66ms +step:6021/9000 val_loss:1.9493 val_bpb:1.1545 train_time:600040ms step_avg:99.66ms +stopping_early: wallclock_cap train_time:600040ms step:6021/9000 +peak memory allocated: 21374 MiB reserved: 21544 MiB +ema:applying EMA weights +Serialized model: 106408383 bytes +Code size: 78596 bytes +gptq_calibration:start +gptq_calibration:done layers=66 +magnitude_pruning:2% +Serialized model int6+zlib: 14117229 bytes +Total submission size int6+zlib: 14195825 bytes +ttt:start lr=0.0005 epochs=10 optimizer=adamw +ttt_epoch:1/10 loss:1.9762 time:17.3s +ttt_epoch:2/10 loss:1.9449 time:34.5s +ttt_epoch:3/10 loss:1.9258 time:51.6s +ttt_epoch:4/10 loss:1.9090 time:68.8s +ttt_epoch:5/10 loss:1.8938 time:86.0s +ttt_epoch:6/10 loss:1.8810 time:103.1s +ttt_epoch:7/10 loss:1.8717 time:120.3s +ttt_epoch:8/10 loss:1.8611 time:137.4s +ttt_epoch:9/10 loss:1.8504 time:154.6s +ttt_epoch:10/10 loss:1.8419 time:171.7s +ttt:done elapsed=171.8s +ttt:elapsed=171.8s +final_int6_roundtrip val_loss:1.8512 val_bpb:1.0964 eval_time:2166ms +final_int6_roundtrip_exact val_loss:1.85118995 val_bpb:1.09637896 +final_int6_sliding_window val_loss:1.8389 val_bpb:1.0891 stride:64 eval_time:94817ms +final_int6_sliding_window_exact val_loss:1.83889390 val_bpb:1.08909943 From 05075423dddbc7d5055b77c8dec4435d772e35f3 Mon Sep 17 00:00:00 2001 From: Amaljith Kuttamath Date: Mon, 23 Mar 2026 18:15:36 -0400 Subject: [PATCH 2/4] =?UTF-8?q?Switch=20to=20legal=20score-first=20TTT=20+?= =?UTF-8?q?=20add=20LeakyReLU=C2=B2?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Pre-eval TTT was non-compliant per issue #402. Now uses score-first TTT: score each chunk before training on it. Added LeakyReLU(0.5)² replacing relu² (proven by #569, #535). Score pending rerun with compute credits. --- .../README.md | 41 +-- .../train_gpt.py | 268 +++++++++++++----- 2 files changed, 203 insertions(+), 106 deletions(-) diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md index 53b308b39b..5ab4db093a 100644 --- a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/README.md @@ -1,43 +1,24 @@ -# Record: 11L EMA + Value Residual + Gated Attention + AdamW TTT (val_bpb=1.0891) +# Record: 11L VR + GA + LeakyReLU² + Legal Score-First TTT (val_bpb=pending) -**val_bpb = 1.0891** (sliding window stride=64, seed 1337) | **14.2 MB** artifact | 8xH100 SXM, 600s +**val_bpb = pending rerun** | 8xH100 SXM, 600s training + legal TTT eval ## Approach -Two architecture changes on top of the PR #442 recipe (11L EMA + AdamW TTT): +Architecture improvements on the standard 11L competitive stack: **Value Residual** (ResFormer, arXiv:2410.17897): Each attention block receives the raw V from the first block. A learned 2-element lambda blends first-block V with current V before attention. Block 0 passes V through unchanged (no lambda parameter). Adds 2 params per layer (layers 1-10 only). **Gated Attention** (arXiv:2505.06708): Per-head sigmoid gate on attention output. Learned weight matrix (dim x num_heads) + bias initialized to 4.0 (near-open gate at init). Adds 4104 params per layer. -Both techniques were ablated individually in PR #413 (-0.015 and -0.003 bpb respectively, -0.017 combined). This is the first validation on the full competitive stack with AdamW TTT. +**LeakyReLU(0.5)²**: Replaces relu² in MLP. Preserves negative gradient flow. Proven by PR #569 and #535. -## Results (seed 1337, 8xH100 SXM) +**Legal score-first TTT**: Score each validation chunk before training on it. Every token evaluated BEFORE the model has seen it. AdamW optimizer, cosine LR across chunks, last 2 blocks + norms unfrozen. -| Metric | Value | -|--------|-------| -| Training steps | 6,021 (wallclock capped at 600s) | -| Step time | 99.66 ms/step | -| Pre-quant val_bpb | 1.1545 | -| Post-quant roundtrip val_bpb | 1.0964 | -| **Sliding window val_bpb (s=64)** | **1.0891** | -| Artifact size | 14,195,825 bytes | -| Peak GPU memory | 21,374 MiB | -| TTT time | 171.8s | +Both VR and GA were ablated individually in PR #413 (-0.015 and -0.003 bpb respectively, -0.017 combined). This is the first validation with legal TTT + LeakyReLU². -## Comparison to prior SOTA +## Previous result (pre-eval TTT, non-compliant) -| Submission | Best BPB | Steps | Step time | -|-----------|----------|-------|-----------| -| **Ours** | **1.0891** | 6,021 | 99.7 ms | -| PR #442 (sjp611) | 1.0992 | 4,612 | ~137 ms | -| PR #481 (mrdavtan) | 1.0959 | 7,101 | ~84 ms | - -## Key findings - -1. VR+GA adds ~300K params (27.1M vs 26.8M) with negligible throughput cost -2. Faster step time (99.7ms vs PR #442's 137ms) yields 38% more training steps -3. AdamW TTT recovers 0.065 bpb from quantized model (1.1545 -> 1.0891 with sliding window) +The initial submission used pre-eval TTT (training on all val data before scoring), which is not competition-legal per issue #402. That result (1.0891) is invalid. This update switches to legal score-first TTT. Score pending rerun. ## Config @@ -49,7 +30,7 @@ MATRIX_LR=0.025 SCALAR_LR=0.025 TIED_EMBED_LR=0.035 ITERATIONS=9000 WARMDOWN_ITERS=1200 EMA_ENABLED=1 EMA_DECAY=0.997 VALUE_RESIDUAL=1 GATED_ATTENTION=1 -TTT_ENABLED=1 TTT_LR=0.0005 TTT_EPOCHS=10 +TTT_ENABLED=1 TTT_LR=0.0001 TTT_EPOCHS=3 TTT_UNFREEZE_BLOCKS=2 EVAL_STRIDE=64 ``` @@ -61,7 +42,7 @@ torchrun --standalone --nproc_per_node=8 train_gpt.py ## Credits -- **PR #442** (sjp611): AdamW TTT, base recipe -- **PR #398** (felipe-parodi): EMA, aggressive TTT findings +- **PR #576** (cmcdnd): Legal score-first TTT implementation, temperature calibration +- **PR #569** (gowtham0992): VRL + LeakyReLU² + Full GPTQ (best non-TTT) - **PR #413**: Value Residual + Gated Attention ablation - **PR #315** (jfprincz): Foundation architecture (U-Net skips, SmearGate, orthogonal init) diff --git a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py index 91dcd24190..fcdedf47ca 100644 --- a/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py +++ b/records/track_10min_16mb/2026-03-23_11L_VR_GA_AdamWTTT_1.0891/train_gpt.py @@ -743,7 +743,7 @@ def __init__(self, dim: int, mlp_mult: int): self.proj._zero_init = True def forward(self, x: Tensor) -> Tensor: - x = torch.relu(self.fc(x)) + x = F.leaky_relu(self.fc(x), negative_slope=0.5) return self.proj(x.square()) @@ -1036,66 +1036,177 @@ def eval_val_sliding( # ----------------------------- -# TEST-TIME TRAINING (TTT) +# LEGAL SCORE-FIRST TTT # ----------------------------- -def ttt_adapt(args: Hyperparameters, base_model: nn.Module, device: torch.device, - val_tokens: Tensor, rank: int = 0, world_size: int = 1, - log_fn=None) -> None: - """Single-phase AdamW TTT on validation data (PR #442 approach). - All blocks unfrozen, AdamW lr=0.0005, 10 epochs.""" - seq_len = args.train_seq_len - total_seqs = (val_tokens.numel() - 1) // seq_len - batch_seqs = args.ttt_batch_seqs +def eval_val_sliding_ttt( + 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, ttt_epochs: int = 3, ttt_lr: float = 0.0001, + ttt_unfreeze_blocks: int = 2, batch_seqs: int = 32, + eval_seq_len: int | None = None, ttt_chunk_tokens: int = 131072, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token is scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 - ttt_params = [p for p in base_model.parameters() if p.requires_grad] - optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.0) + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0 = (lambda msg: print(msg)) if rank == 0 else (lambda msg: None) + log0(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} unfreeze_blocks={ttt_unfreeze_blocks}") - my_start = (total_seqs * rank) // world_size - my_end = (total_seqs * (rank + 1)) // world_size + 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.train() + # Freeze everything, then unfreeze last N blocks + norms + scales + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids: set[int] = set() + for i in range(max(0, num_blocks - ttt_unfreeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # For tied embeddings, unfreeze tok_emb + if hasattr(base_model, 'tie_embeddings') and base_model.tie_embeddings: + base_model.tok_emb.weight.requires_grad_(True) + if id(base_model.tok_emb.weight) not in ttt_param_ids: + ttt_params.append(base_model.tok_emb.weight) + ttt_param_ids.add(id(base_model.tok_emb.weight)) + + n_unfrozen = sum(p.numel() for p in ttt_params) + log0(f"ttt:params unfrozen={n_unfrozen}") + + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) 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) + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue - torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) - optimizer.step() + # --- Phase 1: SCORE this chunk (no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + 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_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_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() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + base_model.train() + for _ep in range(ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + 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): + ttt_loss = base_model(x, y) + ttt_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() + + if rank == 0 and (ci % 100 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") - epoch_loss_sum += loss.detach().to(torch.float64) * y.numel() - epoch_tokens += float(y.numel()) + 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) - if world_size > 1: - dist.all_reduce(epoch_loss_sum, op=dist.ReduceOp.SUM) - dist.all_reduce(epoch_tokens, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) - elapsed = time.perf_counter() - t0 - if log_fn: - log_fn(f"ttt_epoch:{epoch+1}/{args.ttt_epochs} " - f"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) + base_model.eval() - if log_fn: - log_fn(f"ttt:done elapsed={time.perf_counter()-t0:.1f}s") + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb # ----------------------------- @@ -1768,25 +1879,8 @@ def lr_mul(step: int, elapsed_ms: float) -> 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 distributed: - dist.barrier() - for block in eval_model.blocks: - block.attn.rotary._cos_cached = None - block.attn.rotary._sin_cached = None - block.attn.rotary._seq_len_cached = 0 - log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} optimizer=adamw") - t_ttt = time.perf_counter() - ttt_adapt(args, eval_model, device, val_tokens, - rank=rank, world_size=world_size, log_fn=log0) - log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") - if distributed: - dist.barrier() - + # Standard non-overlapping eval (sanity check, no TTT) 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( @@ -1801,7 +1895,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - # Sliding window eval (submission score) + # Sliding window eval WITHOUT TTT (baseline score) sw_seq_len = effective_eval_seq_len if args.eval_stride > 0 and args.eval_stride < sw_seq_len: torch.cuda.synchronize() @@ -1819,22 +1913,44 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) 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: + # Legal score-first TTT + sliding window eval (submission score) + if args.ttt_enabled: + if distributed: + dist.barrier() + for block in eval_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + eval_model.bfloat16() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0001")) + ttt_unfreeze = int(os.environ.get("TTT_UNFREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "131072")) torch.cuda.synchronize() - t_slide64 = time.perf_counter() - sw64_val_loss, sw64_val_bpb = eval_val_sliding( + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=64, + stride=args.eval_stride, + ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_unfreeze_blocks=ttt_unfreeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ) 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" + f"final_int6_ttt_sliding val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" ) - log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int6_ttt_sliding_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + + # Save model (the model after TTT, for inference) + if master_process: + ttt_model_path = f"ttt_model_{args.run_id}.pt" + torch.save(eval_model.state_dict(), ttt_model_path) + log0(f"saved_ttt_model:{ttt_model_path} bytes:{os.path.getsize(ttt_model_path)}") if distributed: dist.destroy_process_group() From a56d7f9f3c2ca7e2052be5b70057875a456ef5a4 Mon Sep 17 00:00:00 2001 From: Amaljith Kuttamath Date: Mon, 23 Mar 2026 22:40:08 -0400 Subject: [PATCH 3/4] Add full technique stack submission script XSA + Partial RoPE + LN Scale + Late QAT + GPTQ + score-first TTT with temp calibration. Untested, needs 1xH100 validation. --- generate.py | 118 +++ train_gpt_submission.py | 1870 +++++++++++++++++++++++++++++++++++++++ 2 files changed, 1988 insertions(+) create mode 100644 generate.py create mode 100644 train_gpt_submission.py diff --git a/generate.py b/generate.py new file mode 100644 index 0000000000..01a20dacd9 --- /dev/null +++ b/generate.py @@ -0,0 +1,118 @@ +""" +Generate text from a trained parameter-golf model checkpoint. + +Usage: + python generate.py --checkpoint final_model.pt --prompt "The" --max_tokens 200 + +Requires the train_gpt.py (or train_gpt_submission.py) in the same directory +for model class definitions. +""" +import argparse +import sys +import os + +# Mock flash_attn before importing train script +import types +mock_fa = types.ModuleType("flash_attn") +def _mock_flash(q, k, v, causal=False): + import torch + import torch.nn.functional as F + B, T, H, D = q.shape + Hkv = k.shape[2] + group = H // Hkv + if group > 1: + k = k.unsqueeze(3).expand(B, T, Hkv, group, D).reshape(B, T, H, D) + v = v.unsqueeze(3).expand(B, T, Hkv, group, D).reshape(B, T, H, D) + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + scale = 1.0 / (D ** 0.5) + attn = torch.matmul(q * scale, k.transpose(-2, -1)) + mask = torch.triu(torch.ones(T, T, device=q.device, dtype=torch.bool), diagonal=1) + attn = attn.masked_fill(mask, float("-inf")) + attn = torch.softmax(attn.float(), dim=-1).to(q.dtype) + out = torch.matmul(attn, v) + return out.transpose(1, 2) + +mock_fa.flash_attn_func = _mock_flash +sys.modules["flash_attn"] = mock_fa +sys.modules["flash_attn_interface"] = mock_fa + +import torch +import sentencepiece as spm + + +def load_model_and_tokenizer(checkpoint_path, script_path="train_gpt_submission.py", + tokenizer_path=None): + sys.path.insert(0, os.path.dirname(os.path.abspath(script_path))) + spec = __import__(os.path.splitext(os.path.basename(script_path))[0]) + + args = spec.Hyperparameters + if tokenizer_path is None: + tokenizer_path = args.tokenizer_path + + model = spec.GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + value_residual=args.value_residual, gated_attention=args.gated_attention, + ) + + state_dict = torch.load(checkpoint_path, map_location="cpu", weights_only=True) + model.load_state_dict(state_dict, strict=True) + model.eval() + + sp = spm.SentencePieceProcessor(model_file=tokenizer_path) + + return model, sp, args + + +@torch.no_grad() +def generate(model, sp, prompt, max_tokens=200, temperature=0.8, top_k=50, device="cpu"): + model = model.to(device).float() + + token_ids = sp.encode(prompt) + tokens = torch.tensor([token_ids], dtype=torch.long, device=device) + + print(f"\n--- Prompt: \"{prompt}\" ---\n") + print(prompt, end="", flush=True) + + for _ in range(max_tokens): + x = tokens[:, -2048:] # context window + logits = model.forward_logits(x) + logits = logits[:, -1, :] / temperature + + if top_k > 0: + v, _ = torch.topk(logits, top_k) + logits[logits < v[:, [-1]]] = float("-inf") + + probs = torch.softmax(logits.float(), dim=-1) + next_token = torch.multinomial(probs, num_samples=1) + tokens = torch.cat([tokens, next_token], dim=1) + + decoded = sp.decode([next_token.item()]) + print(decoded, end="", flush=True) + + print("\n\n--- Done ---") + + +def main(): + parser = argparse.ArgumentParser(description="Generate text from parameter-golf model") + parser.add_argument("--checkpoint", required=True, help="Path to final_model.pt") + parser.add_argument("--script", default="train_gpt_submission.py", help="Training script for model defs") + parser.add_argument("--tokenizer", default=None, help="Path to tokenizer .model file") + parser.add_argument("--prompt", default="The", help="Text prompt") + parser.add_argument("--max_tokens", type=int, default=200) + parser.add_argument("--temperature", type=float, default=0.8) + parser.add_argument("--top_k", type=int, default=50) + parser.add_argument("--device", default="cpu") + args = parser.parse_args() + + model, sp, hparams = load_model_and_tokenizer(args.checkpoint, args.script, args.tokenizer) + generate(model, sp, args.prompt, args.max_tokens, args.temperature, args.top_k, args.device) + + +if __name__ == "__main__": + main() diff --git a/train_gpt_submission.py b/train_gpt_submission.py new file mode 100644 index 0000000000..67cd5e0f0f --- /dev/null +++ b/train_gpt_submission.py @@ -0,0 +1,1870 @@ +""" +train_gpt_submit.py — 11L/512d GPT with VR+GA+XSA+LeakyReLU², Partial RoPE, LN Scale, +EMA, Muon, Late QAT, GPTQ int6+zstd, score-first TTT with temp calibration. +""" + +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_interface import flash_attn_func as flash_attn_3_func +except ImportError: + from flash_attn import flash_attn_func as flash_attn_3_func + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- +# 11L/512d GPT with 3x MLP, GQA (8/4), seq 2048, 9000 iters, 786K batch tokens. +# VR+GA+XSA4+LeakyReLU², Partial RoPE (16/64), LN Scale, EMA, Late QAT, GPTQ int6+zstd. +# Score-first TTT with temp calibration (T=0.98). + +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", 9000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + 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", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + 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", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + 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)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 4096)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "1"))) + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "1"))) + xsa_start_layer = int(os.environ.get("XSA_START_LAYER", -4)) # -4 = last 4 layers + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) # partial RoPE: 16 of 64 head dims + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) + + # TTT (Test-Time Training) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.0001)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_unfreeze_blocks = int(os.environ.get("TTT_UNFREEZE_BLOCKS", 2)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 131072)) + +# ----------------------------- +# 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 to int6 + GPTQ + zstd 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", + ).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() # STE: gradient flows through as if no quantization + 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): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + rd = dim + 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.dim + 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, + value_residual: bool = False, + layer_idx: int = 0, + gated_attention: bool = False, + xsa: bool = False, + rope_dims: int = 0, + train_seq_len: int = 2048, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + rd = rope_dims if rope_dims > 0 else self.head_dim + self.rotary = Rotary(rd, base=rope_base, train_seq_len=train_seq_len) + self._value_residual = value_residual + if value_residual and layer_idx > 0: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + self._gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self._xsa = xsa + + def forward(self, x: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if (self._value_residual or self._xsa) else None + if self._value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + fa_dtype = torch.bfloat16 + y = flash_attn_3_func(q.to(fa_dtype), k.to(fa_dtype), v.to(fa_dtype), causal=True) + if self._xsa and raw_v is not None: + # Gram-Schmidt: remove self-value projection, GQA-aware + y_g = y.reshape(bsz, seqlen, self.num_kv_heads, self.num_heads // self.num_kv_heads, self.head_dim) + vn = F.normalize(raw_v, dim=-1, eps=1e-6).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + y = (y_g - proj).reshape(bsz, seqlen, self.num_heads, self.head_dim) + if self._gated_attention: + gate = torch.sigmoid(self.attn_gate(x)) + y = y * gate.unsqueeze(-1) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v + + +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 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 = F.leaky_relu(self.fc(x), negative_slope=0.5) + 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, + layer_idx: int = 0, + value_residual: bool = False, + gated_attention: bool = False, + xsa: bool = False, + ln_scale: bool = False, + rope_dims: int = 0, + train_seq_len: int = 2048, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + value_residual=value_residual, + layer_idx=layer_idx, gated_attention=gated_attention, + xsa=xsa, rope_dims=rope_dims, + train_seq_len=train_seq_len) + self._ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + s = self._ln_scale_factor + attn_out, raw_v = self.attn(self.attn_norm(x) * s, v0=v0) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x) * s) + return x, raw_v + + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + value_residual: bool = False, + gated_attention: bool = False, + xsa_start_layer: int = -4, + rope_dims: int = 16, + ln_scale: bool = False, + train_seq_len: int = 2048, + ): + 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.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)) + xsa_abs = xsa_start_layer if xsa_start_layer >= 0 else num_layers + xsa_start_layer + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + value_residual=value_residual, + gated_attention=gated_attention, + xsa=(i >= xsa_abs), + ln_scale=ln_scale, + rope_dims=rope_dims, + train_seq_len=train_seq_len, + ) + for i in range(num_layers) + ] + ) + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.register_buffer("inference_temp", torch.tensor(1.0)) + 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 _forward_body(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + for i in range(self.num_encoder_layers): + x, raw_v = self.blocks[i](x, x0, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + x, _ = self.blocks[self.num_encoder_layers + i](x, x0, v0=v0) + return self.final_norm(x) + + def _logits(self, x: Tensor) -> Tensor: + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return logits / self.inference_temp + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self._forward_body(input_ids) + logits = self._logits(x.reshape(-1, x.size(-1))) + return F.cross_entropy(logits.float(), target_ids.reshape(-1), reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + return self._logits(self._forward_body(input_ids)) + + +# ----------------------------- +# 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() + base_model.bfloat16() + 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 + + +# ----------------------------- +# LEGAL SCORE-FIRST TTT +# ----------------------------- + +def eval_val_sliding_ttt( + 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, ttt_epochs: int, ttt_lr: float, + ttt_unfreeze_blocks: int = 2, batch_seqs: int = 32, + eval_seq_len: int | None = None, ttt_chunk_tokens: int = 131072, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token is scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0 = (lambda msg: print(msg)) if rank == 0 else (lambda msg: None) + log0(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} unfreeze_blocks={ttt_unfreeze_blocks}") + + 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) + + # Freeze everything, then unfreeze last N blocks + norms + scales + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids: set[int] = set() + for i in range(max(0, num_blocks - ttt_unfreeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # For tied embeddings, unfreeze tok_emb + if hasattr(base_model, 'tie_embeddings') and base_model.tie_embeddings: + base_model.tok_emb.weight.requires_grad_(True) + if id(base_model.tok_emb.weight) not in ttt_param_ids: + ttt_params.append(base_model.tok_emb.weight) + ttt_param_ids.add(id(base_model.tok_emb.weight)) + + n_unfrozen = sum(p.numel() for p in ttt_params) + log0(f"ttt:params unfrozen={n_unfrozen}") + + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + + # --- Phase 1: SCORE this chunk (no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + 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_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[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() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cosine decay over trainable chunks only (last chunk is skipped), + # with floor at 10% to avoid wasting late-chunk adaptation budget. + trainable_chunks = max(num_chunks - 1, 1) + cos_lr = ttt_lr * max(0.5 * (1.0 + math.cos(math.pi * ci / trainable_chunks)), 0.1) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + base_model.train() + for _ep in range(ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + 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): + ttt_loss = base_model(x, y) + ttt_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() + + if rank == 0 and (ci % 100 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + 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() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# ----------------------------- +# FULL GPTQ QUANTIZATION +# ----------------------------- + +def gptq_quantize_int6(W: Tensor, H: Tensor, blocksize: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ: column-wise int6 quantization with Hessian error compensation. + W: (out_features, in_features) weight matrix + H: (in_features, in_features) Hessian = X^T @ X from calibration data + Returns: (quantized_int8, per_row_scale_fp16) + """ + rows, cols = W.shape + W = W.float().clone() + Q = torch.zeros_like(W) + # Damping for numerical stability + damp = percdamp * torch.diag(H).mean() + H = H.float() + damp * torch.eye(cols, device=H.device, dtype=torch.float32) + # Compute H_inv via Cholesky + try: + H_inv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + except torch.linalg.LinAlgError: + H = H + 0.1 * torch.eye(cols, device=H.device, dtype=torch.float32) + H_inv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + # Compute per-row scales from original weights + row_max = W.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + # Block-wise GPTQ + for col_start in range(0, cols, blocksize): + col_end = min(col_start + blocksize, cols) + bs = col_end - col_start + W_block = W[:, col_start:col_end].clone() + Q_block = torch.zeros_like(W_block) + Err_block = torch.zeros_like(W_block) + H_inv_block = H_inv[col_start:col_end, col_start:col_end] + for j in range(bs): + w = W_block[:, j] + d = H_inv_block[j, j] + q = torch.clamp(torch.round(w / scale), -32, 31) + Q_block[:, j] = q + err = (w - q * scale) / d + Err_block[:, j] = err + # Update remaining columns in this block + W_block[:, j + 1:] -= err.unsqueeze(1) * H_inv_block[j, j + 1:].unsqueeze(0) + Q[:, col_start:col_end] = Q_block + # Propagate block error to all remaining columns + if col_end < cols: + W[:, col_end:] -= Err_block @ H_inv[col_start:col_end, col_end:] + return Q.to(torch.int8), scale.to(torch.float16) + + +# Calibration data collection hook for GPTQ +_gptq_hessians: dict[str, Tensor] = {} +_gptq_nsamples: dict[str, int] = {} + +class GPTQHook: + """Forward hook to accumulate H = X^T @ X for a linear layer.""" + def __init__(self, name: str): + self.name = name + def __call__(self, module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + n = x.shape[0] + H = x.T @ x + if self.name not in _gptq_hessians: + _gptq_hessians[self.name] = H + _gptq_nsamples[self.name] = n + else: + _gptq_hessians[self.name] += H + _gptq_nsamples[self.name] += n + + +# ----------------------------- +# 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: + # GPTQ-lite: search 5 clip percentiles per row, pick min MSE + row_max = t32.abs().amax(dim=1) + best_q = None + best_scale = None + best_mse = None + for clip_frac in [0.95, 0.975, 0.99, 0.999, 1.0]: + clip_val = row_max * clip_frac + s = (clip_val / 31.0).clamp_min(1.0 / 31.0) + clipped = torch.clamp(t32, -clip_val[:, None], clip_val[:, None]) + q = torch.clamp(torch.round(clipped / s[:, None]), -32, 31) + recon = q * s[:, None] + mse = ((t32 - recon) ** 2).mean(dim=1) + if best_mse is None: + best_mse = mse + best_q = q + best_scale = s + else: + improved = mse < best_mse + best_mse = torch.where(improved, mse, best_mse) + best_q = torch.where(improved[:, None], q, best_q) + best_scale = torch.where(improved, s, best_scale) + return best_q.to(torch.int8), best_scale.to(torch.float16) + 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: + # Use Full GPTQ if Hessian available, else GPTQ-lite + if name in _gptq_hessians and t.ndim == 2: + H = _gptq_hessians[name] / _gptq_nsamples[name] + q, s = gptq_quantize_int6(t, H) + else: + 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 + + +# ----------------------------- +# 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 + + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + # ----------------------------- + # TOKENIZER + VALIDATION METRIC SETUP + # ----------------------------- + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + 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 + # ----------------------------- + + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + value_residual=args.value_residual, + gated_attention=args.gated_attention, + xsa_start_layer=args.xsa_start_layer, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + train_seq_len=args.train_seq_len, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + + # Optimizer split: + # - token embedding (Adam) uses EMBED_LR + # - untied lm_head (Adam) uses HEAD_LR + # - matrix params in transformer blocks use MATRIX_LR via Muon + # - vectors/scalars use SCALAR_LR via Adam + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + 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) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.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()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + # ----------------------------- + # DATA LOADER & MODEL WARMUP + # ----------------------------- + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + + # Warmup primes the compiled forward/backward/optimizer paths, then we restore the + # initial weights/optimizer state so measured training starts from the true init. + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + # ----------------------------- + # MAIN TRAINING LOOP + # ----------------------------- + + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + + training_time_ms = 0.0 + stop_after_step: int | None = None + 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_threshold > 0 and scale < args.late_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) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + 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) + + 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" + ) + + CastedLinear._qat_enabled = False + 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) + + # ----------------------------- + # SERIALIZATION + ROUNDTRIP VALIDATION + # ----------------------------- + + export_sd = base_model.state_dict() + + 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") + + # GPTQ calibration: collect Hessians from a few training batches + log0("gptq_calibration:start") + _gptq_hessians.clear() + _gptq_nsamples.clear() + hooks = [] + for name, module in base_model.named_modules(): + if isinstance(module, CastedLinear) and module.weight.numel() > 65536: + param_name = name + ".weight" + hooks.append(module.register_forward_hook(GPTQHook(param_name))) + base_model.eval() + base_model.bfloat16() + calib_batches = 8 + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + with torch.no_grad(): + for i in range(calib_batches): + x_calib, y_calib = calib_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + base_model(x_calib, y_calib) + for h in hooks: + h.remove() + # Move Hessians to CPU + for k in _gptq_hessians: + _gptq_hessians[k] = _gptq_hessians[k].cpu() + log0(f"gptq_calibration:done layers={len(_gptq_hessians)}") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # Magnitude pruning: zero out smallest 2% of float weights BEFORE quantization + # so GPTQ can compensate for the pruned zeros rather than having its compensation undone. + prune_pct = 0.02 + for name, t in sd_cpu.items(): + if t.is_floating_point() and t.ndim == 2 and t.numel() > 65536: + cat = _classify_param(name) + if cat in ("mlp", "attn"): + thresh = torch.quantile(t.abs().float(), prune_pct) + t[t.abs() <= thresh] = 0.0 + log0(f"magnitude_pruning:{prune_pct*100:.0f}% (pre-quantization)") + 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, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + value_residual=args.value_residual, gated_attention=args.gated_attention, + xsa_start_layer=args.xsa_start_layer, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + train_seq_len=args.train_seq_len, + ).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) + + # Standard non-overlapping eval (sanity check, no TTT) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + 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 WITHOUT TTT (baseline 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}") + + # Legal score-first TTT + sliding window eval (submission score) + if args.ttt_enabled: + if distributed: + dist.barrier() + for block in eval_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + eval_model.inference_temp.fill_(args.ttt_temperature) + eval_model.bfloat16() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + ttt_epochs=args.ttt_epochs, ttt_lr=args.ttt_lr, + ttt_unfreeze_blocks=args.ttt_unfreeze_blocks, + eval_seq_len=sw_seq_len, + ttt_chunk_tokens=args.ttt_chunk_tokens, + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt_sliding val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_sliding_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + + # Save model (the model after TTT, for inference) + if master_process: + ttt_model_path = f"ttt_model_{args.run_id}.pt" + torch.save(eval_model.state_dict(), ttt_model_path) + log0(f"saved_ttt_model:{ttt_model_path} bytes:{os.path.getsize(ttt_model_path)}") + + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() From 3c4c4bd3dfec3718e00b929ec25191e53d55598d Mon Sep 17 00:00:00 2001 From: Amaljith Kuttamath Date: Mon, 23 Mar 2026 23:31:36 -0400 Subject: [PATCH 4/4] Fix catastrophic quant gap + TTT eval crash - Deep-clone state dict before bf16 calibration cast (was silently corrupting fp32 weights to bf16 before GPTQ, causing 0.328 bpb gap) - Keep tok_emb.weight as fp16 passthrough instead of int8 quantization - Fix TTT eval: keep CastedLinear weights in fp32 for stable AdamW - Remove torch.compile from TTT chunk loop (re-compilation + weight mutation = crash) - Add quant diagnostics: GPTQ key matching + per-layer error stats --- train_gpt_submission.py | 51 ++++++++++++++++++++++++++++++++++++----- 1 file changed, 45 insertions(+), 6 deletions(-) diff --git a/train_gpt_submission.py b/train_gpt_submission.py index 67cd5e0f0f..9253854026 100644 --- a/train_gpt_submission.py +++ b/train_gpt_submission.py @@ -1060,7 +1060,8 @@ def eval_val_sliding_ttt( my_windows = windows[my_s:my_e] base_model.eval() - compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + # Don't torch.compile inside the loop — weights change via TTT, + # and repeated recompilation causes errors/OOM. with torch.inference_mode(): for bi in range(0, len(my_windows), batch_seqs): batch_ws = my_windows[bi:bi + batch_seqs] @@ -1076,7 +1077,7 @@ def eval_val_sliding_ttt( x_batch[i, :wlen] = chunk_tok[:-1] y_batch[i, :wlen] = chunk_tok[1:] with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - logits = compiled_logits(x_batch) + logits = base_model.forward_logits(x_batch) nll = F.cross_entropy( logits.reshape(-1, logits.size(-1)).float(), y_batch.reshape(-1), reduction="none", @@ -1290,7 +1291,11 @@ def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): result[name] = t.float() meta[name] = "passthrough_ctrl" continue - # tok_emb.weight falls through to int8 via "embed" category + # Keep tok_emb (tied embedding/lm_head) in fp16 — highest-value precision decision + if "tok_emb" in name or "lm_head" in name: + result[name] = t.to(torch.float16) + meta[name] = "passthrough_fp16" + continue if cat in int6_cats and t.ndim >= 1: # Use Full GPTQ if Hessian available, else GPTQ-lite if name in _gptq_hessians and t.ndim == 2: @@ -1708,7 +1713,9 @@ def lr_mul(step: int, elapsed_ms: float) -> float: # SERIALIZATION + ROUNDTRIP VALIDATION # ----------------------------- - export_sd = base_model.state_dict() + # Deep copy state dict BEFORE any bf16 cast for calibration — + # state_dict() returns tensor references, not copies. + export_sd = {k: v.detach().clone().cpu() for k, v in base_model.state_dict().items()} if master_process: torch.save(export_sd, "final_model.pt") @@ -1741,7 +1748,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float: _gptq_hessians[k] = _gptq_hessians[k].cpu() log0(f"gptq_calibration:done layers={len(_gptq_hessians)}") - sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + sd_cpu = export_sd # already cloned to CPU above # Magnitude pruning: zero out smallest 2% of float weights BEFORE quantization # so GPTQ can compensate for the pruned zeros rather than having its compensation undone. prune_pct = 0.02 @@ -1752,6 +1759,14 @@ def lr_mul(step: int, elapsed_ms: float) -> float: thresh = torch.quantile(t.abs().float(), prune_pct) t[t.abs() <= thresh] = 0.0 log0(f"magnitude_pruning:{prune_pct*100:.0f}% (pre-quantization)") + # Diagnostic: check how many layers have GPTQ Hessians + gptq_keys = set(_gptq_hessians.keys()) + sd_keys = set(sd_cpu.keys()) + matched = gptq_keys & sd_keys + unmatched = gptq_keys - sd_keys + log0(f"gptq_diag: hessian_keys={len(gptq_keys)} sd_keys={len(sd_keys)} matched={len(matched)} unmatched={len(unmatched)}") + if unmatched: + log0(f"gptq_diag: UNMATCHED hessian keys (first 5): {list(unmatched)[:5]}") 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) @@ -1776,6 +1791,26 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + # Diagnostic: measure per-layer quant error + if master_process: + total_mse = 0.0 + total_params = 0 + worst_layers = [] + for name in sd_cpu: + if name in deq_state and sd_cpu[name].is_floating_point() and sd_cpu[name].ndim == 2: + orig = sd_cpu[name].float() + recon = deq_state[name].float() + mse = ((orig - recon) ** 2).mean().item() + rmse = mse ** 0.5 + rel_err = rmse / (orig.abs().mean().item() + 1e-8) + total_mse += mse * orig.numel() + total_params += orig.numel() + worst_layers.append((rel_err, name, mse)) + worst_layers.sort(reverse=True) + log0(f"quant_diag: avg_mse={total_mse/max(total_params,1):.8f}") + for rel, name, mse in worst_layers[:5]: + log0(f"quant_diag: worst {name} rel_err={rel:.6f} mse={mse:.8f}") + 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, @@ -1835,7 +1870,11 @@ def lr_mul(step: int, elapsed_ms: float) -> float: block.attn.rotary._sin_cached = None block.attn.rotary._seq_len_cached = 0 eval_model.inference_temp.fill_(args.ttt_temperature) - eval_model.bfloat16() + # Keep CastedLinear weights in fp32 for stable AdamW TTT updates; + # only non-CastedLinear modules go to bf16 for inference speed. + for m in eval_model.modules(): + if not isinstance(m, CastedLinear): + m.bfloat16() torch.cuda.synchronize() t_ttt = time.perf_counter() ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt(