diff --git a/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/README.md b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/README.md new file mode 100644 index 0000000000..81d40b4715 --- /dev/null +++ b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/README.md @@ -0,0 +1,109 @@ +# Depth-Recurrent UT + Rank-1 LoRA Per-Iteration Adaptation — val_bpb 1.3342 + +**val_bpb = 1.3342** (1 seed, additional seeds pending H100 access) | **11.39 MB** | 8xH100 SXM + +## Results (8xH100 80GB SXM, PyTorch 2.7.1) + +| Seed | step_avg | steps | pre_quant_bpb | roundtrip_bpb | sliding_bpb | Artifact | +|------|----------|-------|---------------|---------------|-------------|----------| +| 1337 | 125ms | 4,769 | 1.343 | 1.359 | **1.334** | 11,385,022 | +| 42 | — | — | — | — | — | pending | +| 2025 | — | — | — | — | — | pending | + +> Additional seeds pending H100 access. + +## Key Innovation: Rank-1 LoRA for Stable Per-Iteration Adaptation + +This submission introduces **rank-1 LoRA** as the first stable mechanism for per-iteration weight adaptation in depth-recurrent transformers. Each shared block's Q, V, MLP-up, and MLP-down matrices get a unique rank-1 modification at each loop iteration: + +```python +# Rank-1 delta = outer product of two learned vectors +# b: (out_dim,), a: (in_dim,) — both on AdamW, NOT Muon +delta_W = b.unsqueeze(1) * a.unsqueeze(0) # rank-1 matrix +W_effective = W_shared + delta_W # unique per iteration +``` + +Total rank-1 params: ~9K (negligible — 0.04% of model). The vectors are stored on **AdamW** (not Muon), which is critical for stability. + +### Why Rank-8 LoRA Diverges (and Rank-1 Doesn't) + +We conducted 8 training runs with rank-8 LoRA, systematically varying optimizer (Muon vs AdamW), learning rate (0.005-0.010), warmup strategy (0-2000 steps), and gradient scaling (1/num_loops). **All 8 diverged** between steps 1500-4000. + +**Root cause: Muon's Newton-Schulz scale factor asymmetry.** + +Muon applies `scale = sqrt(rows/cols)` to each parameter update. For rank-8 LoRA: + +| Matrix | Shape | Muon scale | +|--------|-------|:---:| +| LoRA B (out x rank) | (576, 8) | sqrt(72) = **8.49x** | +| LoRA A (rank x in) | (8, 576) | max(1, 8/576)^0.5 = **1.0x** | + +The B matrices receive updates **8.5x larger** than A matrices. This creates a positive feedback loop: B grows fast, which increases dL/dA (since dA = B^T @ dL/dW), which grows A, which makes B@A larger, accelerating divergence. + +**Rank-1 fix**: With rank=1, the LoRA params are 1D **vectors** (not 2D matrices), so they go to AdamW instead of Muon. AdamW has no aspect-ratio scaling — problem eliminated. + +| Attempt | Optimizer | LR | Fix | Result | +|---------|----------|:---:|-----|--------| +| v1 | Muon | 0.010 | None | Diverged step 1500 | +| v2 | Muon | 0.005 | Lower LR | Diverged step 1500 | +| v3 | AdamW (LoRA only) | 0.025 | Separate optimizer | Slow convergence, diverged step 3000 | +| v4 | Muon | 0.010 | Grad zero warmup + 1/3 scale | Grad clip bug: LoRA inflated global norm | +| v5 | Muon | 0.010 | Fixed clip ordering + warmup | Diverged step 3500 (1500 after unfreeze) | +| v6 | Muon (scale=1.0 override) | 0.010 | Override Muon scale for LoRA | Diverged step 4000 | +| v7 | AdamW | 0.010 | LoRA warmup 2000 steps | Partial — survived to end but noisy | +| **v8 (this)** | **AdamW, rank-1** | **0.010** | **Vectors, not matrices** | **Stable! 1.334 BPB** | + +## Architecture + +640d model that **cannot fit as a flat transformer** in 16 MB (would be 18.2 MB at INT6). Depth recurrence enables this width. + +| Parameter | Value | +|-----------|-------| +| Structure | 1 prelude + 4 shared x 3 loops + 1 coda | +| Effective layers | 14 (from 6 unique blocks) | +| Model dim | 640 | +| Heads / KV heads | 10 / 5 (head_dim=64) | +| MLP multiplier | 3.0 (hidden=1920) | +| Activation | LeakyReLU(0.5) squared | +| Rank-1 LoRA | On Q, V, MLP-up, MLP-down per shared effective layer | +| Total rank-1 params | ~9K | +| Vocab | 1024 BPE, tied embeddings | + +### Stability Techniques + +- **Output-LN (Peri-LN)**: RMSNorm on attn/MLP output (not input) for shared blocks. Prevents magnitude information loss across loop iterations. (arXiv:2502.02732) +- **Birkhoff-constrained mixing**: `alpha = sigmoid(logit)` for residual mixing, guaranteeing spectral norm <= 1. Prevents signal blowup. (PR #855) +- **Capped timestep scaling**: Per-effective-layer scale vectors clamped to [-4, 4], stored as FP16 passthrough. Reduces quantization gap by 26-30%. +- **Noisy QAT**: INT6-calibrated noise on shared block weights during training. + +## Training + +| Parameter | Value | +|-----------|-------| +| Optimizer (banks) | Muon (NS5, momentum 0.99) | +| Optimizer (rank-1 LoRA, scalars) | AdamW | +| Matrix LR | 0.010 | +| Grad clip norm | 0.3 | +| Weight decay | 0.04 | +| Batch tokens | 524,288 | +| EMA decay | 0.997 | + +## Artifact + +Only **11.39 MB** — leaves **4.61 MB free** for potential n-gram cache integration. + +``` +Shared block weights (INT6 GPTQ): ~10.5 MB +Rank-1 LoRA vectors (FP16): ~0.02 MB +Embedding + controls: ~0.8 MB +Code: ~0.1 MB +Total: 11.39 MB +``` + +## Credits + +- PR #855 (@aazizyan) — Output-LN, Birkhoff mixing, timestep scaling (first viable 3-loop recurrence) +- PR #895 (@iverbovoy) — Progressive depth, loop embedding concept +- PR #363 (@evangelinehelsinki) — Noisy QAT for recurrence, negative results documentation +- arXiv:2502.02732 — Peri-LN normalization +- arXiv:2502.13181 — RingFormer level signals (inspiration for per-iteration adaptation) diff --git a/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/submission.json b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/submission.json new file mode 100644 index 0000000000..3e4d55e329 --- /dev/null +++ b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/submission.json @@ -0,0 +1,9 @@ +{ + "name": "Depth-Recurrent UT + Rank-1 LoRA Per-Iteration Adaptation", + "val_bpb": 1.3342, + "bytes_total": 11385022, + "blurb": "Universal Transformer (1 prelude + 4 shared x 3 loops + 1 coda = 14 effective layers from 6 unique blocks) at 640d with rank-1 LoRA per-iteration adaptation. Each loop iteration gets a unique weight modification via an outer product of two learned vectors stored on AdamW. This is the first stable per-iteration adaptation for recurrent transformers — rank-8 LoRA diverges due to Muon NS5 scale factor asymmetry on rectangular matrices. Output-LN, Birkhoff mixing, timestep scaling, noisy QAT. Artifact 11.3 MB.", + "author": "Vilhelm Toivonen", + "github_id": "vimeto", + "date": "2026-03-30" +} diff --git a/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_gpt.py b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_gpt.py new file mode 100644 index 0000000000..b65dfe1d36 --- /dev/null +++ b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_gpt.py @@ -0,0 +1,2363 @@ +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 lzma +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "lzma" +except ImportError: + _COMPRESSOR = "lzma" +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 + HAS_FLASH3 = True +except ImportError: + HAS_FLASH3 = False +try: + import kernels + HAS_KERNELS = True +except ImportError: + HAS_KERNELS = False +IS_ROCM = hasattr(torch.version, 'hip') and torch.version.hip is not None +FULL_GPTQ = bool(int(os.environ.get("FULL_GPTQ", "1"))) +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 300.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 6)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 5)) + model_dim = int(os.environ.get("MODEL_DIM", 640)) + num_heads = int(os.environ.get("NUM_HEADS", 10)) + 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)) + 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", 0.010)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + 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", 0)) + 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", 50)) + 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"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.5)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 2)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + # Universal Transformer parameters + num_prelude = int(os.environ.get("NUM_PRELUDE", 1)) + num_shared = int(os.environ.get("NUM_SHARED", 4)) + num_coda = int(os.environ.get("NUM_CODA", 1)) + num_loops = int(os.environ.get("NUM_LOOPS", 3)) + noisy_qat = bool(int(os.environ.get("NOISY_QAT", "1"))) + shared_muon_wd = float(os.environ.get("SHARED_MUON_WD", 0.02)) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +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), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + if not self._built: + self._build() + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + prev_ag_handle = None + prev_m = None + sharded = self._distributed and hasattr(self, '_rs_futures') + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + 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: + update = g.add(buf, alpha=momentum) + else: + update = buf + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + if hasattr(self, '_rs_futures'): + del self._rs_futures + return loss + +# --- Tokenizer evaluation helpers --- + +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("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + 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) + +# --- Quantization helpers --- + +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,dtg_gate,ve_layer_scales,ve_shared.scale," + "ts_attn,ts_mlp,resid_mix_logit,attn_out_norm,mlp_out_norm,r1_", + ).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: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + 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 + 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) + 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(): + 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: + 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 + _is_shared: 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: + 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): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + 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, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, 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, + train_seq_len: int, + ): + 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") + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=train_seq_len) + self.use_xsa = False + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + 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, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if HAS_FLASH3 and not IS_ROCM: + y = flash_attn_3_func(q, k, v, causal=True) + else: + q_sdpa = q.transpose(1, 2).contiguous() + k_sdpa = k.transpose(1, 2).contiguous() + v_sdpa = v.transpose(1, 2).contiguous() + if self.num_kv_heads < self.num_heads: + rep = self.num_heads // self.num_kv_heads + k_sdpa = k_sdpa.repeat_interleave(rep, dim=1) + v_sdpa = v_sdpa.repeat_interleave(rep, dim=1) + y = F.scaled_dot_product_attention(q_sdpa, k_sdpa, v_sdpa, is_causal=True) + y = y.transpose(1, 2).contiguous() + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)) + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + +class Block(nn.Module): + """Transformer block supporting both Pre-LN (prelude/coda) and Output-LN (shared) modes. + + block_mode: + "preln" -- standard pre-norm: norm before attn/mlp + "outputln" -- output-norm (Peri-LN): norm after attn/mlp output + """ + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + train_seq_len: int, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + block_mode: str = "preln", + ): + super().__init__() + self.block_mode = block_mode + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len) + self.mlp = MLP(dim, mlp_mult) + + if block_mode == "preln": + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + 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()) + elif block_mode == "outputln": + # Output-LN: normalize the output of attn/mlp instead of the input + self.attn_out_norm = RMSNorm() + self.mlp_out_norm = RMSNorm() + # Birkhoff-constrained residual mixing via sigmoid + self.resid_mix_logit = nn.Parameter(torch.tensor(0.0, dtype=torch.float32)) + else: + raise ValueError(f"Unknown block_mode: {block_mode}") + + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + + def forward( + self, x: Tensor, x0: Tensor, + q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, + up_w: Tensor, down_w: Tensor, + v_embed: Tensor | None = None, + ts_attn: Tensor | None = None, + ts_mlp: Tensor | None = None, + ) -> Tensor: + if self.block_mode == "preln": + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed) + if ts_attn is not None: + attn_out = attn_out * ts_attn[None, None, :] + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + mlp_out = self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if ts_mlp is not None: + mlp_out = mlp_out * ts_mlp[None, None, :] + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * mlp_out + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out + else: + # Output-LN mode for shared blocks + alpha = torch.sigmoid(self.resid_mix_logit.to(dtype=x.dtype)) + x_in = alpha * x + (1.0 - alpha) * x0 + attn_out = self.attn_out_norm(self.attn(x_in, q_w, k_w, v_w, out_w, v_embed=v_embed)) + if ts_attn is not None: + attn_out = attn_out * ts_attn[None, None, :] + x_out = x_in + attn_out + mlp_out = self.mlp_out_norm(self.mlp(x_out, up_w, down_w)) + if ts_mlp is not None: + mlp_out = mlp_out * ts_mlp[None, None, :] + x_out = x_out + mlp_out + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out + + +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, + train_seq_len: int = 2048, + 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, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + num_prelude: int = 1, + num_shared: int = 4, + num_coda: int = 1, + num_loops: int = 3, + noisy_qat: bool = True, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) + 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.noisy_qat = noisy_qat + + # Universal Transformer structure + self.num_prelude = num_prelude + self.num_shared = num_shared + self.num_coda = num_coda + self.num_loops = num_loops + self.num_unique = num_prelude + num_shared + num_coda + self.num_effective = num_prelude + num_shared * num_loops + num_coda + # num_layers is used for bank sizing = num_unique (unique physical blocks) + self.num_layers = self.num_unique + + 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) + + # Skip connection: prelude output -> coda input (if both exist) + self.has_skip = num_prelude > 0 and num_coda > 0 + if self.has_skip: + self.skip_weights = nn.Parameter(torch.ones(min(num_prelude, num_coda), model_dim, dtype=torch.float32)) + else: + self.skip_weights = nn.Parameter(torch.zeros(0, dtype=torch.float32)) + + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + + # Parameter banks sized for unique blocks only + n = self.num_unique + self.qo_bank = nn.Parameter(torch.empty(2 * n, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * n, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(n, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(n, model_dim, mlp_dim)) + + # Timestep scaling: sized for fixed num_effective layers + self.ts_attn = nn.Parameter(torch.ones(self.num_effective, model_dim, dtype=torch.float32)) + self.ts_mlp = nn.Parameter(torch.ones(self.num_effective, model_dim, dtype=torch.float32)) + + # Rank-1 LoRA: per-iteration vector pairs for Q, V, MLP_up, MLP_down + # delta_W = b ⊗ a (outer product), applied as: out += b * (a · x) + # All vectors → AdamW (no Muon scale issues) + num_eff_shared = num_shared * num_loops + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.r1_enabled = bool(int(os.environ.get("R1_LORA", "1"))) and num_eff_shared > 0 + if self.r1_enabled: + s = 1.0 / math.sqrt(model_dim) + # a vectors (input-side): initialized small random + self.r1_q_a = nn.Parameter(torch.randn(num_eff_shared, model_dim) * s) + self.r1_v_a = nn.Parameter(torch.randn(num_eff_shared, model_dim) * s) + self.r1_up_a = nn.Parameter(torch.randn(num_eff_shared, model_dim) * s) + self.r1_down_a = nn.Parameter(torch.randn(num_eff_shared, mlp_dim) * (1.0 / math.sqrt(mlp_dim))) + # b vectors (output-side): initialized to ZERO (delta starts at zero) + self.r1_q_b = nn.Parameter(torch.zeros(num_eff_shared, model_dim)) + self.r1_v_b = nn.Parameter(torch.zeros(num_eff_shared, kv_dim)) + self.r1_up_b = nn.Parameter(torch.zeros(num_eff_shared, mlp_dim)) + self.r1_down_b = nn.Parameter(torch.zeros(num_eff_shared, model_dim)) + + # Build blocks: prelude (preln), shared (outputln), coda (preln) + block_kwargs = dict( + dim=model_dim, num_heads=num_heads, num_kv_heads=num_kv_heads, + mlp_mult=mlp_mult, rope_base=rope_base, qk_gain_init=qk_gain_init, + train_seq_len=train_seq_len, ln_scale=ln_scale, dtg=dtg, + ) + blocks: list[Block] = [] + for i in range(num_prelude): + blocks.append(Block(**block_kwargs, layer_idx=i, block_mode="preln")) + for i in range(num_shared): + blocks.append(Block(**block_kwargs, layer_idx=num_prelude + i, block_mode="outputln")) + for i in range(num_coda): + blocks.append(Block(**block_kwargs, layer_idx=num_prelude + num_shared + i, block_mode="preln")) + self.blocks = nn.ModuleList(blocks) + + if rope_dims > 0: + hd = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(hd, base=rope_base, train_seq_len=train_seq_len, rope_dims=rope_dims) + + # VE layers: map effective layer indices + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + + # XSA: apply to last N effective layers + if xsa_last_n > 0: + # For shared blocks, XSA is on the block itself (all loops) + # Map effective layer indices back to unique block indices + xsa_eff_start = max(0, self.num_effective - xsa_last_n) + xsa_unique = set() + for eff_i in range(xsa_eff_start, self.num_effective): + uid = self._eff_to_unique(eff_i) + xsa_unique.add(uid) + for uid in xsa_unique: + self.blocks[uid].attn.use_xsa = True + + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + self._init_weights() + + def _eff_to_unique(self, eff_idx: int) -> int: + """Map effective layer index to unique block index.""" + if eff_idx < self.num_prelude: + return eff_idx + eff_idx -= self.num_prelude + if eff_idx < self.num_shared * self.num_loops: + return self.num_prelude + (eff_idx % self.num_shared) + eff_idx -= self.num_shared * self.num_loops + return self.num_prelude + self.num_shared + eff_idx + + 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) + n = self.num_unique + proj_scale = 1.0 / math.sqrt(2 * self.num_effective) + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) + nn.init.zeros_(self.qo_bank.data[n + i]) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) + nn.init.zeros_(self.mlp_down_bank.data[i]) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + 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) + + def _get_ve(self, eff_layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or eff_layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(eff_layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def _noisy_qat_weight(self, w: Tensor) -> Tensor: + """Add INT6 quantization noise to weight during training (STE-style).""" + if not self.training or not self.noisy_qat: + return w + with torch.no_grad(): + amax = w.float().abs().amax(dim=1, keepdim=True).clamp_min(1e-12) + step_size = amax / 31.0 + noise = (torch.rand_like(w) - 0.5) * step_size.to(w.dtype) + return w + noise + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_unique + 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 + ve_cache: dict = {} + + prelude_out: Tensor | None = None + eff_idx = 0 + + # Prelude blocks (standard Pre-LN) + for i in range(self.num_prelude): + uid = i + ve = self._get_ve(eff_idx, input_ids, ve_cache) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + self.qo_bank[uid], self.kv_bank[uid], self.kv_bank[n + uid], + self.qo_bank[n + uid], self.mlp_up_bank[uid], self.mlp_down_bank[uid], + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + if self.has_skip and self.num_prelude > 0: + prelude_out = x + + # Shared blocks (Output-LN) x num_loops + for loop in range(self.num_loops): + for si in range(self.num_shared): + uid = self.num_prelude + si + esi = loop * self.num_shared + si # effective shared index + ve = self._get_ve(eff_idx, input_ids, ve_cache) + # Apply noisy QAT to shared block weights during training + q_w = self._noisy_qat_weight(self.qo_bank[uid]) + o_w = self._noisy_qat_weight(self.qo_bank[n + uid]) + k_w = self._noisy_qat_weight(self.kv_bank[uid]) + v_w = self._noisy_qat_weight(self.kv_bank[n + uid]) + up_w = self._noisy_qat_weight(self.mlp_up_bank[uid]) + down_w = self._noisy_qat_weight(self.mlp_down_bank[uid]) + # Add rank-1 LoRA deltas: delta = b ⊗ a (outer product) + if self.r1_enabled: + q_w = q_w + self.r1_q_b[esi].unsqueeze(1) * self.r1_q_a[esi].unsqueeze(0) + v_w = v_w + self.r1_v_b[esi].unsqueeze(1) * self.r1_v_a[esi].unsqueeze(0) + up_w = up_w + self.r1_up_b[esi].unsqueeze(1) * self.r1_up_a[esi].unsqueeze(0) + down_w = down_w + self.r1_down_b[esi].unsqueeze(1) * self.r1_down_a[esi].unsqueeze(0) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + q_w, k_w, v_w, o_w, up_w, down_w, + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + + # Skip connection: prelude -> coda + if self.has_skip and prelude_out is not None: + x = x + self.skip_weights[0].to(dtype=x.dtype)[None, None, :] * prelude_out + + # Coda blocks (standard Pre-LN) + for i in range(self.num_coda): + uid = self.num_prelude + self.num_shared + i + ve = self._get_ve(eff_idx, input_ids, ve_cache) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + self.qo_bank[uid], self.kv_bank[uid], self.kv_bank[n + uid], + self.qo_bank[n + uid], self.mlp_up_bank[uid], self.mlp_down_bank[uid], + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + + 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: + n = self.num_unique + 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 + ve_cache: dict = {} + prelude_out: Tensor | None = None + eff_idx = 0 + for i in range(self.num_prelude): + uid = i + ve = self._get_ve(eff_idx, input_ids, ve_cache) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + self.qo_bank[uid], self.kv_bank[uid], self.kv_bank[n + uid], + self.qo_bank[n + uid], self.mlp_up_bank[uid], self.mlp_down_bank[uid], + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + if self.has_skip and self.num_prelude > 0: + prelude_out = x + for loop in range(self.num_loops): + for si in range(self.num_shared): + uid = self.num_prelude + si + esi = loop * self.num_shared + si + ve = self._get_ve(eff_idx, input_ids, ve_cache) + q_w = self.qo_bank[uid] + o_w = self.qo_bank[n + uid] + k_w = self.kv_bank[uid] + v_w = self.kv_bank[n + uid] + up_w = self.mlp_up_bank[uid] + down_w = self.mlp_down_bank[uid] + if self.r1_enabled: + q_w = q_w + self.r1_q_b[esi].unsqueeze(1) * self.r1_q_a[esi].unsqueeze(0) + v_w = v_w + self.r1_v_b[esi].unsqueeze(1) * self.r1_v_a[esi].unsqueeze(0) + up_w = up_w + self.r1_up_b[esi].unsqueeze(1) * self.r1_up_a[esi].unsqueeze(0) + down_w = down_w + self.r1_down_b[esi].unsqueeze(1) * self.r1_down_a[esi].unsqueeze(0) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + q_w, k_w, v_w, o_w, up_w, down_w, + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + if self.has_skip and prelude_out is not None: + x = x + self.skip_weights[0].to(dtype=x.dtype)[None, None, :] * prelude_out + for i in range(self.num_coda): + uid = self.num_prelude + self.num_shared + i + ve = self._get_ve(eff_idx, input_ids, ve_cache) + ts_a = self.ts_attn[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + ts_m = self.ts_mlp[eff_idx].clamp(-4, 4).to(dtype=x.dtype) + x = self.blocks[uid]( + x, x0, + self.qo_bank[uid], self.kv_bank[uid], self.kv_bank[n + uid], + self.qo_bank[n + uid], self.mlp_up_bank[uid], self.mlp_down_bank[uid], + v_embed=ve, ts_attn=ts_a, ts_mlp=ts_m, + ) + eff_idx += 1 + 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]: + 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() + if IS_ROCM: + compiled_logits = torch.compile(base_model.forward_logits, mode="default", fullgraph=False) + else: + 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 + +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, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + 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, num_chunks - 1) + chunk_windows[ci].append(ws) + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_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) + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + 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, prev = y_batch[i, s:wlen], 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() + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.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 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_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): + 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, args.ttt_grad_clip) + optimizer.step() + if rank == 0 and (ci % 10 == 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 + log0(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() + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +# --- GPTQ-lite int6 quantization --- + +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, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def gptq_quantize_layer(weight: Tensor, hessian: Tensor = None, clip_range: int = 31, + block_size: int = 128) -> tuple[Tensor, Tensor]: + t32 = weight.float() + if t32.ndim != 2 or hessian is None: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone().to(weight.device) + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q = None; best_scale = None; best_err = float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale + + +def collect_hessians(model: nn.Module, train_loader, args, device, grad_accum_steps, + num_batches: int = 256, log_fn=None) -> dict[str, Tensor]: + """Collect H = X^T X for each linear layer and for bank forward passes. + Adapted for Universal Transformer: banks are indexed by unique block id.""" + hessians = {} + hooks = [] + n = model.num_unique + + for block_idx, block in enumerate(model.blocks): + q_key = f"blocks.{block_idx}.attn.c_q.weight" + k_key = f"blocks.{block_idx}.attn.c_k.weight" + v_key = f"blocks.{block_idx}.attn.c_v.weight" + o_key = f"blocks.{block_idx}.attn.proj.weight" + up_key = f"blocks.{block_idx}.mlp.fc.weight" + down_key = f"blocks.{block_idx}.mlp.proj.weight" + + model_dim = args.model_dim + kv_dim = args.num_kv_heads * (model_dim // args.num_heads) + mlp_dim = int(args.mlp_mult * model_dim) + + for key, cols in [(q_key, model_dim), (k_key, model_dim), (v_key, model_dim), + (o_key, model_dim), (up_key, model_dim), (down_key, mlp_dim)]: + hessians[key] = torch.zeros(cols, cols, dtype=torch.float32, device='cpu') + + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + param_name = name + ".weight" + cols = module.weight.shape[1] + hessians[param_name] = torch.zeros(cols, cols, dtype=torch.float32, device='cpu') + def make_hook(pname): + def hook_fn(module, input, output): + x = input[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pname] += (x.T @ x).cpu() + return hook_fn + h = module.register_forward_hook(make_hook(param_name)) + hooks.append(h) + + _orig_block_forwards = [] + for block_idx, block in enumerate(model.blocks): + orig_forward = block.forward + _orig_block_forwards.append(orig_forward) + + q_key = f"blocks.{block_idx}.attn.c_q.weight" + k_key = f"blocks.{block_idx}.attn.c_k.weight" + v_key = f"blocks.{block_idx}.attn.c_v.weight" + o_key = f"blocks.{block_idx}.attn.proj.weight" + up_key = f"blocks.{block_idx}.mlp.fc.weight" + down_key = f"blocks.{block_idx}.mlp.proj.weight" + + def make_wrapped_forward(blk, orig_fwd, qk, kk, vk, ok, upk, dnk): + def wrapped_forward(x, x0, q_w, k_w, v_w, out_w, up_w, down_w, v_embed=None, ts_attn=None, ts_mlp=None): + # Capture inputs for Hessian collection + if blk.block_mode == "preln": + mix = blk.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_input = (blk.attn_norm(x_in) * blk.ln_scale_factor).detach().float() + else: + alpha = torch.sigmoid(blk.resid_mix_logit.to(dtype=x.dtype)) + x_in = alpha * x + (1.0 - alpha) * x0 + attn_input = x_in.detach().float() + + if attn_input.ndim == 3: + attn_flat = attn_input.reshape(-1, attn_input.shape[-1]) + else: + attn_flat = attn_input + hessians[qk] += (attn_flat.T @ attn_flat).cpu() + hessians[kk] += (attn_flat.T @ attn_flat).cpu() + hessians[vk] += (attn_flat.T @ attn_flat).cpu() + + bsz, seqlen, dim = x_in.shape + attn = blk.attn + qv = F.linear(attn_input.to(x.dtype), q_w.to(x.dtype)).reshape(bsz, seqlen, attn.num_heads, attn.head_dim) + kv = F.linear(attn_input.to(x.dtype), k_w.to(x.dtype)).reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + vv = F.linear(attn_input.to(x.dtype), v_w.to(x.dtype)) + if v_embed is not None: + vv = vv + v_embed + vv = vv.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + qv = F.rms_norm(qv, (qv.size(-1),)) + kv = F.rms_norm(kv, (kv.size(-1),)) + cos, sin = attn.rotary(seqlen, x.device, qv.dtype) + qv = apply_rotary_emb(qv, cos, sin, attn.rope_dims) + kv = apply_rotary_emb(kv, cos, sin, attn.rope_dims) + qv = qv * attn.q_gain.to(dtype=qv.dtype)[None, None, :, None] + if HAS_FLASH3 and not IS_ROCM: + y_attn = flash_attn_3_func(qv, kv, vv, causal=True) + else: + q_sdpa = qv.transpose(1, 2).contiguous() + k_sdpa = kv.transpose(1, 2).contiguous() + v_sdpa = vv.transpose(1, 2).contiguous() + if attn.num_kv_heads < attn.num_heads: + rep = attn.num_heads // attn.num_kv_heads + k_sdpa = k_sdpa.repeat_interleave(rep, dim=1) + v_sdpa = v_sdpa.repeat_interleave(rep, dim=1) + y_attn = F.scaled_dot_product_attention(q_sdpa, k_sdpa, v_sdpa, is_causal=True) + y_attn = y_attn.transpose(1, 2).contiguous() + if attn.use_xsa: + y_attn = attn._xsa_efficient(y_attn, vv) + y_flat = y_attn.reshape(bsz, seqlen, dim).detach().float() + if y_flat.ndim == 3: + y_2d = y_flat.reshape(-1, y_flat.shape[-1]) + else: + y_2d = y_flat + hessians[ok] += (y_2d.T @ y_2d).cpu() + + result = orig_fwd(x, x0, q_w, k_w, v_w, out_w, up_w, down_w, v_embed=v_embed, ts_attn=ts_attn, ts_mlp=ts_mlp) + + attn_out_val = F.linear(y_attn.reshape(bsz, seqlen, dim).to(x.dtype), out_w.to(x.dtype)) + if blk.block_mode == "preln": + x_out = x_in + blk.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out_val + mlp_input = (blk.mlp_norm(x_out) * blk.ln_scale_factor).detach().float() + else: + x_out = x_in + blk.attn_out_norm(attn_out_val) + mlp_input = x_out.detach().float() + + if mlp_input.ndim == 3: + mlp_flat = mlp_input.reshape(-1, mlp_input.shape[-1]) + else: + mlp_flat = mlp_input + hessians[upk] += (mlp_flat.T @ mlp_flat).cpu() + + up_out = F.leaky_relu(F.linear(mlp_input.to(x.dtype), up_w.to(x.dtype)), negative_slope=0.5) + down_input = up_out.square().detach().float() + if down_input.ndim == 3: + down_flat = down_input.reshape(-1, down_input.shape[-1]) + else: + down_flat = down_input + hessians[dnk] += (down_flat.T @ down_flat).cpu() + + return result + return wrapped_forward + + block.forward = make_wrapped_forward(block, orig_forward, q_key, k_key, v_key, o_key, up_key, down_key) + + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + + for block_idx, block in enumerate(model.blocks): + block.forward = _orig_block_forwards[block_idx] + + for h in hooks: + h.remove() + for name in hessians: + H = hessians[name] + H /= num_batches + model.train() + if log_fn is not None: + log_fn(f"gptq:collected hessians for {len(hessians)} layers, {num_batches} calibration batches") + return hessians + + +def _state_dict_key_to_module_name(key: str) -> str: + if key.endswith(".weight"): + return key[:-7] + if key.endswith(".bias"): + return key[:-5] + return key + +# --- Bank <-> individual weight conversion --- + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names. + num_layers here is num_unique (the number of unique physical blocks).""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + log_fn=None): + 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] = {} + gptq_count = 0 + fallback_count = 0 + 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 + if cat in int6_cats and t.ndim >= 1: + H = hessians.get(name) if hessians is not None else None + if FULL_GPTQ and H is not None and t.ndim == 2 and H.shape[0] == t.shape[1]: + try: + q, s = gptq_quantize_layer(t, H) + gptq_count += 1 + meta[name] = {"type": "int6", "method": "gptq"} + except Exception as e: + if log_fn: + log_fn(f"gptq:cholesky_failed layer={name} error={e}") + q, s = quantize_int6_per_row(t) + fallback_count += 1 + meta[name] = {"type": "int6", "method": "gptq_fallback"} + else: + q, s = quantize_int6_per_row(t) + fallback_count += 1 + if FULL_GPTQ and log_fn and H is not None: + log_fn(f"gptq:shape_mismatch layer={name} H={H.shape} W={t.shape}") + meta[name] = {"type": "int6"} + result[name + ".q"] = q + result[name + ".scale"] = s + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + if log_fn is not None: + log_fn(f"gptq:quantized {gptq_count} layers with full GPTQ, {fallback_count} with GPTQ-lite fallback") + 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 = "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 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + if not IS_ROCM: + 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) + if FULL_GPTQ: + log0("gptq:full_hessian block_size:128 percdamp:0.01 calib_batches:256") + else: + log0("gptq:disabled (using GPTQ-lite fallback)") + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["rocm-smi" if IS_ROCM else "nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + 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}") + + 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, + train_seq_len=args.train_seq_len, + 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, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + num_prelude=args.num_prelude, + num_shared=args.num_shared, + num_coda=args.num_coda, + num_loops=args.num_loops, + noisy_qat=args.noisy_qat, + ).to(device).bfloat16() + + # Banks stay FP32 + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + + if IS_ROCM: + _inductor_config = __import__("torch._inductor.config", fromlist=["config"]) + _inductor_config.shape_padding = False + compiled_model = torch.compile(base_model, mode="default", fullgraph=False) + else: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = compiled_model + + num_eff = base_model.num_effective + log0("=" * 60) + log0("UNIVERSAL TRANSFORMER CONFIG:") + log0(f" num_prelude: {args.num_prelude}") + log0(f" num_shared: {args.num_shared}") + log0(f" num_coda: {args.num_coda}") + log0(f" num_loops: {args.num_loops}") + log0(f" num_unique_blocks: {base_model.num_unique}") + log0(f" num_effective_layers: {num_eff}") + log0(f" noisy_qat: {args.noisy_qat}") + log0(f" shared_muon_wd: {args.shared_muon_wd}") + log0("=" * 60) + log0("FEATURE VERIFICATION:") + log0(f" parallel_muon: True (3-phase overlapped, no DDP)") + log0(f" parameter_banking: True (qo_bank, kv_bank, mlp_up_bank, mlp_down_bank)") + log0(f" xsa_last_n: {args.xsa_last_n} (XSA on last N effective layers)") + log0(f" late_qat_threshold: {args.late_qat_threshold} (QAT activation point)") + log0(f" warmdown_iters: {args.warmdown_iters}") + log0(f" bigram_vocab_size: {args.bigram_vocab_size}") + log0(f" train_batch_tokens: {args.train_batch_tokens} (global batch)") + log0(f" compression: {_COMPRESSOR}") + log0(f" full_gptq: {FULL_GPTQ}") + log0(f" leaky_relu_sq: True (LeakyReLU(0.5)^2)") + log0(f" ema_decay: 0.997") + log0(f" swa_enabled: {args.swa_enabled}") + log0(f" timestep_scaling: True ({base_model.num_effective} effective layers)") + log0("=" * 60) + + # Optimizer split: separate shared vs non-shared bank params + # Identify which bank slices are shared vs prelude/coda + n_unique = base_model.num_unique + shared_start = args.num_prelude + shared_end = args.num_prelude + args.num_shared + + # All bank params go to single Muon optimizer (WD split handled at step time) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + + block_named_params = list(base_model.blocks.named_parameters()) + 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) + # Timestep scaling params + scalar_params.append(base_model.ts_attn) + scalar_params.append(base_model.ts_mlp) + # Rank-1 LoRA vectors → AdamW (not Muon — they're 1D vectors) + r1_params = [p for n, p in base_model.named_parameters() if 'r1_' in n and p.numel() > 0] + scalar_params.extend(r1_params) + 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) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + _adam_fused = not IS_ROCM + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=_adam_fused, + ) + 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=_adam_fused, + ) + optimizer_head = None + 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=_adam_fused, + ) + optimizers.insert(1, optimizer_head) + replicated_params: list[nn.Parameter] = [] + for pg in optimizer_tok.param_groups: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + if optimizer_head is not None: + replicated_params.append(base_model.lm_head.weight) + 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} (unique blocks={base_model.num_unique}, effective_layers={num_eff})") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_blocks = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_blocks:{xsa_blocks}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.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): + 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() + optimizer_muon.launch_reduce_scatters() + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + optimizer_muon.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() + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + 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): + 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_mom = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac_mom) * args.muon_momentum_warmup_start + frac_mom * 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) + optimizer_muon.launch_reduce_scatters() + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + optimizer_muon.step() + zero_grad_all() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms" + ) + 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" + ) + + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + # Need to recompile since we changed the model state + if IS_ROCM: + compiled_model = torch.compile(base_model, mode="default", fullgraph=False) + else: + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, base_model.num_unique) + + hessians = None + if FULL_GPTQ: + log0("gptq:collecting Hessians from calibration data...") + torch.cuda.synchronize() + t_hess = time.perf_counter() + hessians = collect_hessians( + base_model, train_loader, args, device, grad_accum_steps, + num_batches=256, + log_fn=log0, + ) + torch.cuda.synchronize() + log0(f"gptq:hessian_collection_time:{1000.0 * (time.perf_counter() - t_hess):.0f}ms") + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}, + hessians=hessians, log_fn=log0) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + if _COMPRESSOR == "lzma": + quant_blob = lzma.compress(quant_raw, preset=6) + elif _COMPRESSOR == "zstd": + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) + else: + quant_blob = 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") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk) if _COMPRESSOR == "lzma" else zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + deq_state = _rebank_state_dict(deq_unbanked, base_model.num_unique, 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, + train_seq_len=args.train_seq_len, + 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, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + num_prelude=args.num_prelude, num_shared=args.num_shared, + num_coda=args.num_coda, num_loops=args.num_loops, + noisy_qat=False, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + 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) + if IS_ROCM: + compiled_eval = torch.compile(eval_model, mode="default", fullgraph=False) + else: + 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}") + 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}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + 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}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_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 if args.eval_stride > 0 else 64, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") + if distributed: + dist.destroy_process_group() + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_seed1337.log b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_seed1337.log new file mode 100644 index 0000000000..b370e52bed --- /dev/null +++ b/records/track_10min_16mb/2026-03-30_UT_Rank1LoRA_OutputLN_Birkhoff/train_seed1337.log @@ -0,0 +1,128 @@ +Running Python 3.12.11 | packaged by Anaconda, Inc. | (main, Jun 5 2025, 13:09:17) [GCC 11.2.0] +Running PyTorch 2.7.1+cu124 + + + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/tmp/fineweb_1024_bpe.model +train_loader:dataset:pgolf_data train_shards:80 +val_loader:shards pattern=/tmp/pgolf_data/fineweb_val_*.bin tokens:62021632 +============================================================ +UNIVERSAL TRANSFORMER CONFIG: + num_prelude: 1 + num_shared: 4 + num_coda: 1 + num_loops: 3 + num_unique_blocks: 6 + num_effective_layers: 14 + noisy_qat: True + shared_muon_wd: 0.02 +============================================================ +FEATURE VERIFICATION: + parallel_muon: True (3-phase overlapped, no DDP) + parameter_banking: True (qo_bank, kv_bank, mlp_up_bank, mlp_down_bank) + xsa_last_n: 11 (XSA on last N effective layers) + late_qat_threshold: 0.5 (QAT activation point) + warmdown_iters: 3000 + bigram_vocab_size: 3072 + train_batch_tokens: 524288 (global batch) + compression: lzma + full_gptq: True + leaky_relu_sq: True (LeakyReLU(0.5)^2) + ema_decay: 0.997 + swa_enabled: True + timestep_scaling: True (14 effective layers) +============================================================ +model_params:23533636 (unique blocks=6, effective_layers=14) +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_blocks:[1, 2, 3, 4, 5] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:10 num_kv_heads:5 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.01 scalar_lr:0.025 +train_batch_tokens:524288 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9345 val_bpb:4.1070 train_time:0ms step_avg:0.00ms +step:1/20000 train_loss:6.9344 train_time:24ms step_avg:24.60ms +step:2/20000 train_loss:7.7972 train_time:134ms step_avg:67.45ms +step:3/20000 train_loss:8.9017 train_time:263ms step_avg:87.69ms +step:4/20000 train_loss:9.0880 train_time:384ms step_avg:96.08ms +step:5/20000 train_loss:9.4610 train_time:509ms step_avg:101.88ms +step:6/20000 train_loss:8.9306 train_time:632ms step_avg:105.39ms +step:7/20000 train_loss:6.9689 train_time:755ms step_avg:107.99ms +step:8/20000 train_loss:6.1356 train_time:884ms step_avg:110.56ms +step:9/20000 train_loss:5.8424 train_time:1004ms step_avg:111.65ms +step:10/20000 train_loss:5.8372 train_time:1132ms step_avg:113.29ms +step:500/20000 train_loss:3.0781 train_time:62465ms step_avg:124.93ms +step:1000/20000 train_loss:2.7492 train_time:125142ms step_avg:125.14ms +step:1500/20000 train_loss:2.5205 train_time:187794ms step_avg:125.20ms +step:2000/20000 train_loss:2.3970 train_time:250507ms step_avg:125.25ms +step:2500/20000 train_loss:2.2234 train_time:313222ms step_avg:125.29ms +step:3000/20000 train_loss:2.2610 train_time:375884ms step_avg:125.30ms +late_qat:enabled step:3270 scale:0.4998 +step:3500/20000 train_loss:2.2639 train_time:438593ms step_avg:125.31ms +step:4000/20000 train_loss:2.3042 train_time:501272ms step_avg:125.32ms +step:4000/20000 val_loss:2.2974 val_bpb:1.3606 train_time:501385ms step_avg:125.35ms +swa:start step:4200 +step:4500/20000 train_loss:2.3893 train_time:564052ms step_avg:125.34ms +step:4769/20000 val_loss:2.2828 val_bpb:1.3520 train_time:598000ms step_avg:125.39ms +stopping_early: wallclock_cap train_time:598000ms step:4769/20000 +peak memory allocated: 23841 MiB reserved: 24648 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:2.2684 val_bpb:1.3434 eval_time:4391ms +Serialized model: 91788132 bytes +Code size: 110742 bytes +gptq:collecting Hessians from calibration data... +gptq:collected hessians for 38 layers, 256 calibration batches +gptq:hessian_collection_time:218732ms +gptq:cholesky_failed layer=blocks.0.attn.proj.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 130 is not positive-definite). +gptq:cholesky_failed layer=blocks.1.attn.c_q.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 9 is not positive-definite). +gptq:cholesky_failed layer=blocks.2.attn.c_q.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 45 is not positive-definite). +gptq:cholesky_failed layer=blocks.3.attn.c_q.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 40 is not positive-definite). +gptq:cholesky_failed layer=blocks.4.attn.c_q.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 511 is not positive-definite). +gptq:cholesky_failed layer=blocks.1.attn.c_k.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 9 is not positive-definite). +gptq:cholesky_failed layer=blocks.1.attn.c_v.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 9 is not positive-definite). +gptq:cholesky_failed layer=blocks.2.attn.c_k.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 45 is not positive-definite). +gptq:cholesky_failed layer=blocks.2.attn.c_v.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 45 is not positive-definite). +gptq:cholesky_failed layer=blocks.3.attn.c_k.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 40 is not positive-definite). +gptq:cholesky_failed layer=blocks.3.attn.c_v.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 40 is not positive-definite). +gptq:cholesky_failed layer=blocks.4.attn.c_k.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 511 is not positive-definite). +gptq:cholesky_failed layer=blocks.4.attn.c_v.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 511 is not positive-definite). +gptq:cholesky_failed layer=blocks.0.mlp.fc.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 4 is not positive-definite). +gptq:cholesky_failed layer=blocks.1.mlp.fc.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 9 is not positive-definite). +gptq:cholesky_failed layer=blocks.2.mlp.fc.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 34 is not positive-definite). +gptq:cholesky_failed layer=blocks.3.mlp.fc.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 166 is not positive-definite). +gptq:cholesky_failed layer=blocks.4.mlp.fc.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 560 is not positive-definite). +gptq:cholesky_failed layer=blocks.0.mlp.proj.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 3 is not positive-definite). +gptq:cholesky_failed layer=blocks.1.mlp.proj.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 2 is not positive-definite). +gptq:cholesky_failed layer=blocks.2.mlp.proj.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 2 is not positive-definite). +gptq:cholesky_failed layer=blocks.4.mlp.proj.weight error=linalg.cholesky: The factorization could not be completed because the input is not positive-definite (the leading minor of order 2 is not positive-definite). +gptq:quantized 15 layers with full GPTQ, 22 with GPTQ-lite fallback +Serialized model int6+lzma: 11274280 bytes +Total submission size int6+lzma: 11385022 bytes +Total submission size int8+zlib: 11385022 bytes +final_int6_roundtrip val_loss:2.2950 val_bpb:1.3592 eval_time:20691ms +final_int6_roundtrip_exact val_loss:2.29500863 val_bpb:1.35923338 +final_int6_sliding_window_s64 val_loss:2.2527 val_bpb:1.3342 stride:64 eval_time:153033ms +final_int6_sliding_window_s64_exact val_loss:2.25266760 val_bpb:1.33416017 +final_int8_zlib_roundtrip_exact val_loss:2.25266760 val_bpb:1.33416017