From 17836df54df77a51e47120c775cd7e722862c3e0 Mon Sep 17 00:00:00 2001 From: ethan Date: Tue, 24 Mar 2026 15:49:11 +0800 Subject: [PATCH 1/5] Record: int5 GPTQ + Soft-Round QAT + 33.6M model (3-seed mean val_bpb=1.1162) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit int5 GPTQ quantization with Hessian-aware error compensation enables 33.6M params in 16MB. Soft-Round QAT (differentiable tanh rounding, alpha 1→16) replaces STE for better training quality at zero cost. 3-seed results: - Seed 1337: val_bpb=1.1155, artifact=15.82MB - Seed 42: val_bpb=1.1163, artifact=15.42MB - Seed 7: val_bpb=1.1167, artifact=15.37MB - Mean: 1.1162 (std 0.0006) --- .../README.md | 75 + .../submission.json | 22 + .../train_gpt.py | 2273 +++++++++++++++++ .../train_seed1337.log | 172 ++ .../train_seed42.log | 172 ++ .../train_seed7.log | 172 ++ 6 files changed, 2886 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log create mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md new file mode 100644 index 000000000..fc738e0a9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md @@ -0,0 +1,75 @@ +# Record: int5 GPTQ + 33.6M model + Soft-Round QAT + Legal Score-First TTT + +## Summary + +**3-seed mean val_bpb = 1.1162 (std 0.0006)** + +int5 GPTQ quantization (values in [-15, 15], 31 unique levels) with Hessian-aware error compensation enables a 33.6M parameter model to fit under 16MB. Soft-Round QAT replaces STE hard rounding with differentiable tanh-based rounding (alpha annealing 1→16) for better training quality at zero cost. Combined with early QAT at threshold 0.5, EMA 0.997, and legal score-first AdamW TTT with cosine LR decay across chunks. + +## Key Innovations + +1. **int5 quantization** — 31 unique values ([-15,15]) stored as int8, ~0.46 bytes/param after zstd. Lower entropy = better compression ratio than int6. +2. **GPTQ error compensation** — Hessian-aware column reordering + Cholesky error redistribution. 256-sample calibration on training data. +3. **33.6M param model** — MHA 8/8 (full attention), BigramHash 8192, MLP 3.5x (1792), enabled by int5 compression. +4. **Soft-Round QAT** — Differentiable rounding `s_α(y) = floor(y) + 0.5 * tanh(α·r) / tanh(α/2) + 0.5` replaces STE. Alpha anneals from 1→16 during QAT steps. Better gradient flow = better training quality at zero computational cost. +5. **Early QAT 0.5** — QAT clipping matched to int5 range (0.9995 percentile / 15.0), ~1750 QAT steps. +6. **EMA 0.997** — Exponential moving average of weights, tuned from 0.9985. +7. **Legal score-first TTT** — every token scored BEFORE any gradient update using it. Cosine LR decay across chunks. + +## Architecture + +- 11 layers, model_dim=512, 8 heads / 8 KV heads (MHA), MLP 3.5x relu² +- XSA on all 11 layers +- Partial RoPE 16/64, LN Scale (1/√(layer+1)) +- SmearGate + OrthoInit +- BigramHash 8192, Shared VE128 (layers 9,10) +- Tight SWA (every 50) + EMA 0.997 +- Muon lr=0.025, WD=0.04 +- FA3 Hopper, ~98ms/step → ~6120 steps in 600s +- **33.6M params**, int5 GPTQ + zstd-22, 2% magnitude pruning + +## Quantization Pipeline + +1. **Early QAT** (threshold 0.5): QAT-aware training with int5 clipping (scale = row_clip / 15.0, clamp [-16, 15]) +2. **GPTQ** (post-training): 256-sample Hessian calibration, per-row optimal scales (5-percentile search), column reordering by Hessian diagonal, block-128 Cholesky error compensation +3. **int5 quantization** (range [-15, 15], 31 levels) stored as int8 +4. **zstd-22** compression +5. **2% magnitude pruning** + +## Legal Score-First TTT + +- Val data split into 131072-token chunks (474 chunks) +- For each chunk: **score first** (sliding window stride=32, inference_mode), **then** adapt +- AdamW (lr=0.0001, wd=0.0), 3 epochs per chunk, cosine LR across chunks +- Last 2 blocks + norms + lm_head unfrozen (~5.8M / 33.6M params) +- Last chunk never trained on +- Every token scored BEFORE any gradient update using it +- Manual grad all_reduce (no DDP wrapper) + +## Results + +| Seed | TTT BPB | Artifact | +|------|---------|----------| +| 1337 | **1.1155** | 15,822,078 bytes | +| 42 | **1.1163** | 15,415,405 bytes | +| 7 | **1.1167** | 15,368,627 bytes | +| **Mean** | **1.1162** | | +| **Std** | **0.0006** | | + +## Reproduction + +```bash +# On 8xH100 SXM: +pip install --break-system-packages zstandard +# Build FA3 Hopper (see repo README for instructions) +python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 80 + +SEED=1337 SKIP_SLIDING=1 PRUNE_PCT=0.02 \ +SOFT_ROUND_QAT=1 \ +TTT_EPOCHS=3 TTT_LR=0.0001 TTT_OPTIMIZER=adamw \ +TTT_FREEZE_BLOCKS=2 TTT_CHUNK_TOKENS=131072 \ +TTT_TEMPERATURE=0.98 INT6_LAST_N=0 \ +PPM_ALPHA=1.0 BYTE_WEIGHTED_TTT=0 USE_CACHE=0 \ +ADAPTIVE_LR=0 USE_MIXER=0 \ +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json new file mode 100644 index 000000000..8edbcae58 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json @@ -0,0 +1,22 @@ +{ + "author": "Ethan Yang", + "github_id": "EthanYangTW", + "name": "Record: int5 GPTQ + 33.6M model + Soft-Round QAT + Legal Score-First TTT", + "blurb": "int5 GPTQ quantization ([-15,15], 31 levels) with Hessian-aware error compensation enables 33.6M params in 16MB. Soft-Round QAT (differentiable tanh rounding, alpha 1→16) replaces STE for better training quality. MHA 8/8, BigramHash 8192, MLP 3.5x (1792), XSA all 11 layers, Early QAT 0.5, EMA 0.997, legal score-first AdamW TTT with cosine LR decay.", + "date": "2026-03-24T00:00:00Z", + "val_bpb": 1.1162, + "val_bpb_std": 0.0006, + "val_loss_seed1337": 1.88347869, + "val_bpb_seed1337": 1.11550587, + "val_loss_seed42": 1.88480123, + "val_bpb_seed42": 1.11628915, + "val_loss_seed7": 1.88543543, + "val_bpb_seed7": 1.11666477, + "bytes_seed1337": 15822078, + "bytes_seed42": 15415405, + "bytes_seed7": 15368627, + "model_params": 33580124, + "quantization": "int5 GPTQ ([-15,15], 31 levels) + Soft-Round QAT", + "compression": "zstd-22", + "ttt": "legal score-first AdamW, 3 epochs, cosine LR across chunks" +} diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py new file mode 100644 index 000000000..757c51358 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py @@ -0,0 +1,2273 @@ +"""V25: LeakyReLU^2 + TempCal + Mixed int5/int6 + 33.6M model.""" +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import zlib +from pathlib import Path +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + _HAS_FA3 = True + except ImportError: + _HAS_FA3 = False + flash_attn_3_func = None + +# ===================== PPM N-gram Model ===================== +import collections + +class PPMModel: + """Prediction by Partial Matching — online n-gram model. + + Builds token-level n-gram statistics from already-scored validation tokens. + At prediction time, backs off from order K down to order 0 (unigram), + blending probabilities via escape mechanism. + + This is 100% legal: it only uses tokens that have already been scored. + """ + + def __init__(self, max_order: int = 6, vocab_size: int = 1024): + self.max_order = max_order + self.vocab_size = vocab_size + # counts[order][context_tuple] = Counter of next tokens + self.counts: list[dict[tuple, collections.Counter]] = [ + {} for _ in range(max_order + 1) + ] + self.total_tokens = 0 + + def update(self, tokens): + """Add observed tokens to the model. tokens = 1D list/array of token IDs.""" + tokens = list(tokens) + self.total_tokens += len(tokens) + for i, tok in enumerate(tokens): + for order in range(min(i, self.max_order) + 1): + ctx = tuple(tokens[i - order:i]) if order > 0 else () + if ctx not in self.counts[order]: + self.counts[order][ctx] = collections.Counter() + self.counts[order][ctx][tok] += 1 + + def predict_probs(self, context_tokens, device=None): + """Return probability distribution over vocab given context. + + Uses PPM Method C escape mechanism: + At each order, allocate escape probability = num_unique / (num_unique + total_count) + and distribute remaining probability proportional to counts. + """ + import torch + context = list(context_tokens) + + # Start with uniform distribution (order -1) + probs = torch.ones(self.vocab_size, dtype=torch.float32) / self.vocab_size + if device is not None: + probs = probs.to(device) + + # Blend from lowest order up to highest + for order in range(min(len(context), self.max_order) + 1): + ctx = tuple(context[-order:]) if order > 0 else () + if ctx in self.counts[order]: + counter = self.counts[order][ctx] + total = sum(counter.values()) + unique = len(counter) + # Escape probability (PPM Method C) + escape = unique / (unique + total) + # Build distribution for this order + order_probs = torch.zeros(self.vocab_size, dtype=torch.float32) + if device is not None: + order_probs = order_probs.to(device) + for tok, cnt in counter.items(): + if tok < self.vocab_size: + order_probs[tok] = cnt / total + # Blend: (1-escape) * order_probs + escape * lower_order_probs + probs = (1.0 - escape) * order_probs + escape * probs + + return probs + + def predict_batch(self, context_batch, targets, device=None): + """Compute log-probs for a batch of (context, target) pairs. + + context_batch: list of token lists (each is the context for one prediction) + targets: tensor of target token IDs + + Returns tensor of log-probabilities (same shape as targets). + """ + import torch + log_probs = torch.zeros(len(targets), dtype=torch.float32) + if device is not None: + log_probs = log_probs.to(device) + + for i, (ctx, tgt) in enumerate(zip(context_batch, targets)): + probs = self.predict_probs(ctx, device=device) + log_probs[i] = torch.log(probs[tgt] + 1e-10) + + return log_probs + + +class FastPPMModel: + """Faster PPM using batch numpy operations and hash-based lookup. + + Trades memory for speed. Maintains only orders 0-4 for tractability. + Uses a single forward pass over scored tokens to update. + """ + + def __init__(self, max_order: int = 4, vocab_size: int = 1024): + self.max_order = max_order + self.vocab_size = vocab_size + # For each order, store: hash(context) -> array of counts + self.counts = [{} for _ in range(max_order + 1)] + self.total_tokens = 0 + self._unigram = None # cached unigram distribution + + def update_chunk(self, tokens): + """Update with a chunk of tokens. tokens = list or 1D numpy/torch array.""" + if hasattr(tokens, 'cpu'): + tokens = tokens.cpu().tolist() + elif hasattr(tokens, 'tolist'): + tokens = tokens.tolist() + + n = len(tokens) + self.total_tokens += n + + for i in range(n): + tok = tokens[i] + for order in range(min(i, self.max_order) + 1): + if order == 0: + ctx_key = 0 # empty context + else: + ctx_key = hash(tuple(tokens[i-order:i])) + + if ctx_key not in self.counts[order]: + self.counts[order][ctx_key] = {} + d = self.counts[order][ctx_key] + d[tok] = d.get(tok, 0) + 1 + + self._unigram = None # invalidate cache + + def score_sequence(self, tokens, start_pos=0): + """Score a sequence, returning NLL for each position from start_pos. + + Returns list of -log2(prob) for each token (bits, not nats). + """ + import math + if hasattr(tokens, 'cpu'): + tokens = tokens.cpu().tolist() + elif hasattr(tokens, 'tolist'): + tokens = tokens.tolist() + + scores = [] + for i in range(start_pos, len(tokens)): + prob = self._predict_one(tokens, i) + scores.append(-math.log2(max(prob, 1e-10))) + return scores + + def get_log_probs_tensor(self, tokens, start_pos, device): + """Get log probabilities as a tensor for interpolation with neural model.""" + import torch, math + if hasattr(tokens, 'cpu'): + tokens = tokens.cpu().tolist() + elif hasattr(tokens, 'tolist'): + tokens = tokens.tolist() + + n = len(tokens) - start_pos + log_probs = torch.zeros(n, dtype=torch.float32, device=device) + + for i in range(start_pos, len(tokens)): + prob = self._predict_one(tokens, i) + log_probs[i - start_pos] = math.log(max(prob, 1e-10)) + + return log_probs + + def _predict_one(self, tokens, pos): + """Predict probability of tokens[pos] given tokens[:pos].""" + target = tokens[pos] + + # Start with uniform + prob = 1.0 / self.vocab_size + + for order in range(min(pos, self.max_order) + 1): + if order == 0: + ctx_key = 0 + else: + ctx_key = hash(tuple(tokens[pos-order:pos])) + + if ctx_key in self.counts[order]: + d = self.counts[order][ctx_key] + total = sum(d.values()) + unique = len(d) + escape = unique / (unique + total) + + count = d.get(target, 0) + if count > 0: + order_prob = count / total + prob = (1.0 - escape) * order_prob + escape * prob + else: + prob = escape * prob + + return prob + + + + +class ExactMatchCache: + """Hash-based exact-match n-gram cache. + + Stores (context_hash → Counter of next tokens) from already-scored tokens. + For repeated patterns in val data, gives near-perfect predictions. + """ + + def __init__(self, orders=(3, 4, 5, 6, 7, 8), vocab_size=1024): + self.orders = orders + self.vocab_size = vocab_size + # For each order: hash(context) -> {token: count} + self.tables = {o: {} for o in orders} + self.total_tokens = 0 + + def update_chunk(self, tokens): + """Add a chunk of tokens to the cache.""" + if hasattr(tokens, 'cpu'): + tokens = tokens.cpu().tolist() + elif hasattr(tokens, 'tolist'): + tokens = tokens.tolist() + + n = len(tokens) + self.total_tokens += n + + for i in range(n): + tok = tokens[i] + for order in self.orders: + if i >= order: + ctx = hash(tuple(tokens[i-order:i])) + if ctx not in self.tables[order]: + self.tables[order][ctx] = {} + d = self.tables[order][ctx] + d[tok] = d.get(tok, 0) + 1 + + def predict_one(self, tokens, pos): + """Get probability of tokens[pos] given exact context matches. + + Returns (prob, matched_order) or (None, -1) if no match. + Uses highest-order match available. + """ + target = tokens[pos] + + # Try highest order first + for order in sorted(self.orders, reverse=True): + if pos >= order: + ctx = hash(tuple(tokens[pos-order:pos])) + if ctx in self.tables[order]: + d = self.tables[order][ctx] + total = sum(d.values()) + if total >= 2: # require at least 2 observations + prob = d.get(target, 0) / total + return prob, order + + return None, -1 + + def get_interpolation_nll(self, tokens, pos, neural_nll, alpha_cache=0.3): + """Interpolate cache prediction with neural model NLL. + + Args: + tokens: full token sequence (list) + pos: position to predict + neural_nll: neural model NLL for this position (float) + alpha_cache: weight for cache (0.3 = 30% cache, 70% neural) + + Returns: interpolated NLL + """ + import math + cache_prob, order = self.predict_one(tokens, pos) + + if cache_prob is not None and order >= 4: + neural_prob = math.exp(-neural_nll) + # Higher weight for longer matches + weight = min(alpha_cache * (order / max(self.orders)), 0.5) + mixed_prob = (1 - weight) * neural_prob + weight * cache_prob + return -math.log(max(mixed_prob, 1e-10)) + + return neural_nll + + +# ===================== Interpolation Helper ===================== + +def interpolate_with_ppm(neural_logits, ppm_model, tokens, window_start, seq_len, + stride, alpha=0.85, device=None): + """Interpolate neural model logits with PPM predictions. + + Args: + neural_logits: (batch, seq_len, vocab) from neural model + ppm_model: FastPPMModel instance + tokens: full val_tokens tensor + window_start: starting position of this window + seq_len: sequence length + stride: stride for scoring + alpha: weight for neural model (1-alpha for PPM) + device: torch device + + Returns: + interpolated NLL values for scored positions + """ + import torch + # For now, just return neural logits — PPM interpolation happens at the NLL level + # We compute PPM log-probs and do log-space interpolation + pass + + +class LogisticContextMixer: + """GPU-vectorized logistic context mixing (inspired by PAQ compression). + + Maintains GPU-resident n-gram count tables and learns online mixing weights + using the Hedge/multiplicative-weights algorithm. All operations are batched + tensor ops — no Python per-token loops. + + Experts: + 0: Neural model (logits passed in) + 1: Unigram frequencies from scored tokens + 2: Bigram frequencies (prev_token → next_token) + """ + + def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): + self.V = vocab_size + self.device = device + self.eta = eta # Hedge learning rate + self.K = 3 # number of experts + + # Expert weights (log-domain for numerical stability) + self.log_weights = torch.zeros(self.K, device=device) + + # N-gram count tables (GPU-resident) + self.uni_counts = torch.zeros(vocab_size, device=device) + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device) + self.total_tokens = 0 + + def update(self, tokens): + """Update n-gram tables with newly scored tokens. Fully vectorized.""" + if hasattr(tokens, 'cpu'): + t = tokens.to(self.device).long() + else: + t = torch.tensor(tokens, device=self.device, dtype=torch.long) + + n = t.numel() + if n == 0: + return + self.total_tokens += n + + # Unigram: bincount + self.uni_counts.scatter_add_(0, t, torch.ones(n, device=self.device)) + + # Bigram: scatter_add into [V, V] table + if n >= 2: + ctx = t[:-1] + nxt = t[1:] + bi_idx = ctx * self.V + nxt + flat = torch.zeros(self.V * self.V, device=self.device) + flat.scatter_add_(0, bi_idx, torch.ones(n - 1, device=self.device)) + self.bi_counts += flat.reshape(self.V, self.V) + + def get_expert_log_probs(self, neural_logits, x_batch, y_batch, wlens): + """Get log-probability of targets from each expert. All GPU-vectorized. + + Args: + neural_logits: [bsz, seq_len, V] neural model logits + x_batch: [bsz, seq_len] input tokens (context) + y_batch: [bsz, seq_len] target tokens + wlens: list of actual lengths per sequence + + Returns: + expert_nll: [bsz, seq_len, K] NLL from each expert + """ + bsz, slen, V = neural_logits.shape + + # Expert 0: Neural model + neural_lp = F.log_softmax(neural_logits, dim=-1) + neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) # [bsz, slen] + + # Expert 1: Unigram + uni_total = self.uni_counts.sum() + if uni_total > 0: + uni_probs = (self.uni_counts + 0.1) / (uni_total + 0.1 * self.V) + uni_lp = uni_probs.log() + uni_nll = -uni_lp[y_batch] # [bsz, slen] + else: + uni_nll = torch.full((bsz, slen), math.log(self.V), device=self.device) + + # Expert 2: Bigram P(next | prev) + bi_total = self.bi_counts.sum(dim=1, keepdim=True) # [V, 1] + if bi_total.sum() > 0: + bi_probs = (self.bi_counts + 0.1) / (bi_total + 0.1 * self.V) # [V, V] + bi_lp = bi_probs.log() + # Lookup: for each position, prev=x_batch, next=y_batch + prev_flat = x_batch.reshape(-1) # [bsz*slen] + next_flat = y_batch.reshape(-1) + bi_nll_flat = -bi_lp[prev_flat, next_flat] + bi_nll = bi_nll_flat.reshape(bsz, slen) + else: + bi_nll = torch.full((bsz, slen), math.log(self.V), device=self.device) + + # Stack: [bsz, slen, K] + return torch.stack([neural_nll, uni_nll, bi_nll], dim=-1) + + def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): + """Compute mixed NLL using current expert weights. Returns [bsz, slen] NLL. + + Uses log-domain mixing: NLL_mixed = -log(sum_k w_k * exp(-NLL_k)) + """ + if self.total_tokens < 10000: + # Not enough data for n-grams — just use neural + return F.cross_entropy( + neural_logits.reshape(-1, neural_logits.size(-1)), + y_batch.reshape(-1), reduction="none" + ).reshape(neural_logits.shape[0], neural_logits.shape[1]) + + expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] + + # Log-domain mixing: log(sum_k w_k * p_k) = logsumexp(log_w_k + log_p_k) + log_w = self.log_weights - self.log_weights.logsumexp(0) # normalize + # expert_lp = -expert_nll [bsz, slen, K] + mixed_lp = (-expert_nll + log_w.unsqueeze(0).unsqueeze(0)).logsumexp(dim=-1) # [bsz, slen] + + return -mixed_lp # mixed NLL + + def update_weights(self, neural_logits, x_batch, y_batch, wlens): + """Update expert weights using Hedge algorithm on this batch's losses.""" + if self.total_tokens < 10000: + return + + with torch.no_grad(): + expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] + + # Mean loss per expert across valid positions + mask = torch.zeros(expert_nll.shape[0], expert_nll.shape[1], device=self.device) + for i, wl in enumerate(wlens): + mask[i, :wl] = 1.0 + + # Masked mean NLL per expert + masked_nll = expert_nll * mask.unsqueeze(-1) + expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) # [K] + + # Hedge update: log_w -= eta * loss + self.log_weights -= self.eta * expert_mean_loss + + +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", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.5)) + 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", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + 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", 32)) + int6_last_n = int(os.environ.get("INT6_LAST_N", 2)) # last N layers use int6, rest use int5 + ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration + 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", 8192)) + 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") + prune_pct = float(os.environ.get("PRUNE_PCT", 0.02)) + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("▁"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + 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) +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", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_Q = 0.9999984 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + +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 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) + +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 + _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) + _use_soft_round: bool = False # enable soft-round QAT instead of STE + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + self._clip_range = 15 # default int5, set to 31 for int6 layers + + @staticmethod + def soft_round(y: Tensor, alpha: float) -> Tensor: + """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). + s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 + where r = y - floor(y) - 0.5 (centered fractional part) + """ + fl = torch.floor(y) + r = y - fl - 0.5 + return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + cr = self._clip_range + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + if CastedLinear._use_soft_round: + # Soft-Round QAT: differentiable rounding with temperature annealing + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_scaled = w32 / scale[:, None] + w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) + w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) + w = w_q # fully differentiable path + else: + # Original STE QAT + with torch.no_grad(): + w32 = self.weight.float() + row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) + scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * 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): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False + + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + B, T, H, D = y.shape + Hkv = v.size(-2) + y_g = y.reshape(B, T, Hkv, H // Hkv, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, 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_FA3: + y = flash_attn_3_func(q, k, v, causal=True).contiguous() + else: + y = F.scaled_dot_product_attention( + q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), + attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y) + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) + +class Block(nn.Module): + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, + rope_base: float, qk_gain_init: float, layer_idx: int = 0, + ln_scale: bool = False, dtg: bool = False): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + 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, v_embed: Tensor | None = None) -> Tensor: + 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, v_embed=v_embed) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + 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, + 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"): + 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.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList([ + Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, + qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) + for i in range(num_layers) + ]) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() + 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 + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + def forward_logits(self, input_ids: Tensor) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x = self.blocks[i](x, x0, v_embed=ve) + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x = self.blocks[bi](x, x0, v_embed=ve) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +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() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + + # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache + if rank == 0: + print(" ttt: pre-compiling forward+backward kernels...", flush=True) + _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + _dummy_logits = base_model.forward_logits(_dummy_x) + _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) + _dummy_loss.backward() + base_model.zero_grad(set_to_none=True) + if rank == 0: + print(" ttt: pre-compile done", flush=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, ttt_epochs: int = 3, ttt_lr: float = 0.001, + ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, + batch_seqs: int = 32, eval_seq_len: int | None = None, + ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", + ttt_temp: float = 1.0, + ppm_alpha: float = 0.85, + byte_weighted_ttt: bool = True, + use_cache: bool = True, + cache_alpha: float = 0.3, + adaptive_lr: bool = True, + adaptive_lr_max_mult: float = 3.0, +) -> tuple[float, float]: + """Legal score-first TTT: score each chunk, then train on it. + Every token scored BEFORE any update that could use it.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Initialize GPU-vectorized logistic context mixer + use_mixer = os.environ.get("USE_MIXER", "1") == "1" + mixer = LogisticContextMixer( + vocab_size=val_tokens.to(torch.int32).max().item() + 1, + device=device, + eta=float(os.environ.get("MIXER_ETA", "0.1")), + ) if use_mixer else None + if use_mixer and rank == 0: + print(f" Logistic context mixer enabled: eta={mixer.eta}") + if adaptive_lr and rank == 0: + print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on scored token position + num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + if rank == 0: + print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " + f"windows={len(window_starts)} stride={stride} " + f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " + f"freeze_first={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) + + # Freeze everything, then selectively unfreeze for TTT + num_blocks = len(base_model.blocks) + for p in base_model.parameters(): + p.requires_grad_(False) + ttt_params = [] + ttt_param_ids = set() + use_qttt = os.environ.get("QTTT", "0") == "1" + if use_qttt: + # qTTT: only unfreeze Q projections in last N blocks + norms + head + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for name, p in base_model.blocks[i].named_parameters(): + if "c_q" in name: + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + else: + # Standard: unfreeze all params in last N blocks + for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(True) + ttt_params.append(p) + ttt_param_ids.add(id(p)) + # Unfreeze norms, scales, lm_head + for name, p in base_model.named_parameters(): + if "norm" in name or "scale" in name or "lm_head" in name: + p.requires_grad_(True) + if id(p) not in ttt_param_ids: + ttt_params.append(p) + ttt_param_ids.add(id(p)) + + if rank == 0: + n_unfrozen = sum(p.numel() for p in ttt_params) + n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) + print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") + + if ttt_optimizer == "adamw": + optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) + else: + optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) + + # Polyak averaging (TTT weight EMA) for smoother scoring + use_polyak = os.environ.get("USE_POLYAK", "1") == "1" + polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) + if use_polyak: + polyak_state = {id(p): p.data.clone() for p in ttt_params} + if rank == 0: + print(f" Polyak averaging enabled: decay={polyak_decay}") + + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + + # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # Swap in Polyak-averaged weights for scoring + if use_polyak and ci > 0: + _saved_weights = {} + for p in ttt_params: + _saved_weights[id(p)] = p.data.clone() + p.data.copy_(polyak_state[id(p)]) + + 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) + logits_scaled = logits.float() / ttt_temp + + # Adaptive temperature: sharpen confident predictions more + if ttt_temp != 1.0: + with torch.no_grad(): + probs_for_entropy = F.softmax(logits.float(), dim=-1) + token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) + max_ent = math.log(logits.size(-1)) + # Confident tokens (low entropy) get more sharpening + adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) + adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) + logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) + + # Logistic context mixing (GPU-vectorized) or plain CE + if mixer is not None: + nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) + mixer.update_weights(logits_scaled, x_batch, y_batch, wlens) + else: + nll = F.cross_entropy( + logits_scaled.reshape(-1, logits_scaled.size(-1)), + 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() + + # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- + chunk_start_tok = ci * ttt_chunk_tokens + chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) + if mixer is not None: + mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) + + # Swap back training weights after scoring + if use_polyak and ci > 0: + for p in ttt_params: + p.data.copy_(_saved_weights[id(p)]) + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and ttt_epochs > 0: + chunk_start = ci * ttt_chunk_tokens + chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) + chunk_seqs = (chunk_end - chunk_start) // seq_len + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) + if chunk_seqs > 0: + # Cosine LR across chunks + adaptive scaling + cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + if adaptive_lr: + # Increase LR as we've seen more data (more confident adaptation) + progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks + lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress + cos_lr = cos_lr * lr_mult + 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(ttt_epochs): + if rank == 0 and ci < 3: + print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + if byte_weighted_ttt: + # Byte-weighted loss: tokens covering more bytes matter more + ttt_logits = base_model.forward_logits(x) + per_token_loss = F.cross_entropy( + ttt_logits.reshape(-1, ttt_logits.size(-1)), + y.reshape(-1), reduction='none' + ).reshape(y.shape) + byte_weights = base_bytes_lut[y].float() + byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() + ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() + else: + ttt_loss = base_model(x, y) + ttt_loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + # Update Polyak EMA after each step + if use_polyak: + for p in ttt_params: + polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) + if rank == 0 and ci < 3: + print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if rank == 0: + print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +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 = 15) -> 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 _find_best_row_scales(W: Tensor, clip_range: int = 15) -> Tensor: + """Find optimal per-row scales by searching percentile clipping thresholds.""" + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), 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) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s + +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 15, + block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation.""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) + +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + return hessians + +def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: + """Return clip_range based on which layer the param belongs to.""" + import re + m = re.search(r'blocks\.(\d+)\.', name) + if m: + layer_idx = int(m.group(1)) + if layer_idx >= num_layers - int6_last_n: + return 31 # int6 + return 15 # int5 + +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + num_layers: int = 11, int6_last_n: int = 2) -> tuple[dict, dict]: + """GPTQ quantization with mixed int5/int6 precision. int6 for last int6_last_n layers, int5 for rest.""" + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count = 0, 0 + int5_params, int6_params = 0, 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 + cr = _get_layer_clip_range(name, num_layers, int6_last_n) + if cr == 31: + int6_params += t.numel() + else: + int5_params += t.numel() + if cat in int6_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=cr) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t, clip_range=cr) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + elif cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t, clip_range=cr) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} + naive_count += 1 + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) + print(f"mixed_precision: {int5_params} int5 params, {int6_params} int6 params", flush=True) + return result, meta + + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta + +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +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 + 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(f"Python {sys.version} PyTorch {torch.__version__}", 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, + 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, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + 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) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + # Set int6 clip_range for last N layers (mixed precision) + int6_start = args.num_layers - args.int6_last_n + for i, block in enumerate(base_model.blocks): + if i >= int6_start: + for m in block.modules(): + if isinstance(m, CastedLinear): + m._clip_range = 31 # int6 + if master_process: + int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) + int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) + log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") + log0(f"model_params:{n_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") + log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s 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 + # TTT_ONLY mode: skip training, load saved model, run TTT eval + if os.environ.get("TTT_ONLY", "0") == "1": + log0("TTT_ONLY mode: skipping training, loading saved model...") + sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, 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, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + log0(f"TTT_ONLY: model loaded, starting TTT eval...") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + 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) + # Anneal soft-round alpha based on QAT progress + if CastedLinear._use_soft_round and CastedLinear._qat_enabled: + qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) + CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" + if CastedLinear._use_soft_round and master_process: + log0(f"soft_round_qat:enabled initial_alpha=1.0") + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + 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" + ) + raw_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} + best_bpb = float('inf') + best_label = "raw" + best_state = raw_state + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(ema_sd, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + ema_val_loss, ema_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:{ema_val_loss:.4f} val_bpb:{ema_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + if ema_val_bpb < best_bpb: + best_bpb = ema_val_bpb + best_label = "ema" + best_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} + if swa_state is not None and swa_count > 0: + log0(f"swa:applying SWA weights (count={swa_count})") + swa_sd = {} + for name in current_state: + swa_avg = (swa_state[name].float() / swa_count).to(dtype=current_state[name].dtype) + swa_sd[name] = swa_avg + base_model.load_state_dict(swa_sd, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + swa_val_loss, swa_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_swa val_loss:{swa_val_loss:.4f} val_bpb:{swa_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + if swa_val_bpb < best_bpb: + best_bpb = swa_val_bpb + best_label = "swa" + best_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} + + log0(f"best_averaging:{best_label} val_bpb:{best_bpb:.4f}") + base_model.load_state_dict(best_state, strict=True) + export_sd = base_model.state_dict() + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + if args.prune_pct > 0: + for k, v in sd_cpu.items(): + if v.ndim == 2 and v.numel() > 65536: + thresh = torch.quantile(v.abs().float(), args.prune_pct) + v[v.abs() < thresh] = 0.0 + if master_process: + log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") + # GPTQ calibration + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn"}, gptq_hessians, num_layers=args.num_layers, int6_last_n=args.int6_last_n) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, 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, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) + if sw_seq_len != effective_eval_seq_len and rank == 0: + log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") + if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): + 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}") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) + ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) + ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) + ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) + ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") + log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") + ttt_temp = args.ttt_temperature + log0(f"TTT temperature: {ttt_temp}") + ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) + bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" + log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, + ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, + ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, + ttt_temp=ttt_temp, + ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), + byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", + use_cache=os.environ.get("USE_CACHE", "1") == "1", + cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), + adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", + adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), + ) + torch.cuda.synchronize() + log0( + f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log new file mode 100644 index 000000000..4596180df --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log @@ -0,0 +1,172 @@ +W0324 06:47:48.570000 94727 torch/distributed/run.py:851] +W0324 06:47:48.570000 94727 torch/distributed/run.py:851] ***************************************** +W0324 06:47:48.570000 94727 torch/distributed/run.py:851] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0324 06:47:48.570000 94727 torch/distributed/run.py:851] ***************************************** +logs/0986e1a2-99db-4b93-ba13-1082fe463b5d.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33580124 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s 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:6.9304 val_bpb:4.1046 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9324 train_time:155ms step_avg:154.62ms +step:2/20000 train_loss:8.6499 train_time:244ms step_avg:122.18ms +step:3/20000 train_loss:7.7400 train_time:339ms step_avg:113.00ms +step:4/20000 train_loss:7.2905 train_time:434ms step_avg:108.40ms +step:5/20000 train_loss:7.0203 train_time:528ms step_avg:105.66ms +step:6/20000 train_loss:6.8351 train_time:623ms step_avg:103.87ms +step:7/20000 train_loss:6.7947 train_time:718ms step_avg:102.57ms +step:8/20000 train_loss:6.7258 train_time:812ms step_avg:101.51ms +step:9/20000 train_loss:6.4110 train_time:907ms step_avg:100.79ms +step:10/20000 train_loss:6.0618 train_time:1002ms step_avg:100.17ms +step:500/20000 train_loss:2.3545 train_time:48339ms step_avg:96.68ms +step:1000/20000 train_loss:2.2365 train_time:96843ms step_avg:96.84ms +step:1500/20000 train_loss:2.1818 train_time:145370ms step_avg:96.91ms +step:2000/20000 train_loss:2.0262 train_time:194003ms step_avg:97.00ms +step:2500/20000 train_loss:2.1279 train_time:242644ms step_avg:97.06ms +step:3000/20000 train_loss:2.1145 train_time:291318ms step_avg:97.11ms +step:3500/20000 train_loss:2.1254 train_time:340011ms step_avg:97.15ms +step:4000/20000 train_loss:1.9115 train_time:388714ms step_avg:97.18ms +step:4000/20000 val_loss:2.0024 val_bpb:1.1860 train_time:388719ms step_avg:97.18ms +soft_round_qat:enabled initial_alpha=1.0 +late_qat:enabled step:4424 scale:0.4997 +step:4500/20000 train_loss:2.0582 train_time:437424ms step_avg:97.21ms +step:5000/20000 train_loss:2.0351 train_time:486102ms step_avg:97.22ms +swa:start step:5500 +step:5500/20000 train_loss:1.9473 train_time:534768ms step_avg:97.23ms +step:6000/20000 train_loss:1.8706 train_time:583974ms step_avg:97.33ms +step:6163/20000 val_loss:1.8983 val_bpb:1.1243 train_time:600024ms step_avg:97.36ms +stopping_early: wallclock_cap train_time:600024ms step:6163/20000 +peak memory allocated: 26201 MiB reserved: 26418 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8966 val_bpb:1.1233 eval_time:2368ms +swa:applying SWA weights (count=14) +DIAGNOSTIC post_swa val_loss:1.8982 val_bpb:1.1242 eval_time:2360ms +best_averaging:ema val_bpb:1.1233 +Serialized model: 130956873 bytes +Code size: 106734 bytes +pruning:2.0% magnitude pruning applied +gptq:calibrating with training data... +gptq:calibrated 68 layers in 3.6s +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +Serialized model int6+zstd: 15715344 bytes +Total submission size int6+zstd: 15822078 bytes +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 1.0, Byte-weighted TTT: False +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27799624 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0755 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0637 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0631 + ttt_chunk [1/474] bpb=1.199717 time=1.3s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.9046 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.9036 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.9017 + ttt_chunk [2/474] bpb=1.153948 time=2.4s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.8512 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.8502 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.8490 + ttt_chunk [3/474] bpb=1.127169 time=3.6s + ttt_chunk [4/474] bpb=1.134789 time=4.7s + ttt_chunk [5/474] bpb=1.133690 time=5.8s + ttt_chunk [11/474] bpb=1.116604 time=12.7s + ttt_chunk [21/474] bpb=1.111760 time=24.1s + ttt_chunk [31/474] bpb=1.109901 time=35.5s + ttt_chunk [41/474] bpb=1.118054 time=46.9s + ttt_chunk [51/474] bpb=1.125284 time=58.3s + ttt_chunk [61/474] bpb=1.123258 time=69.7s + ttt_chunk [71/474] bpb=1.124623 time=81.1s + ttt_chunk [81/474] bpb=1.125042 time=92.5s + ttt_chunk [91/474] bpb=1.127019 time=103.9s + ttt_chunk [101/474] bpb=1.123588 time=115.3s + ttt_chunk [111/474] bpb=1.123831 time=126.8s + ttt_chunk [121/474] bpb=1.127096 time=138.2s + ttt_chunk [131/474] bpb=1.127790 time=149.6s + ttt_chunk [141/474] bpb=1.127537 time=161.0s + ttt_chunk [151/474] bpb=1.125756 time=172.4s + ttt_chunk [161/474] bpb=1.126665 time=183.8s + ttt_chunk [171/474] bpb=1.125481 time=195.2s + ttt_chunk [181/474] bpb=1.126243 time=206.6s + ttt_chunk [191/474] bpb=1.125132 time=218.0s + ttt_chunk [201/474] bpb=1.124308 time=229.4s + ttt_chunk [211/474] bpb=1.122924 time=240.9s + ttt_chunk [221/474] bpb=1.123005 time=252.3s + ttt_chunk [231/474] bpb=1.122391 time=263.7s + ttt_chunk [241/474] bpb=1.121305 time=275.1s + ttt_chunk [251/474] bpb=1.122402 time=286.5s + ttt_chunk [261/474] bpb=1.123029 time=297.9s + ttt_chunk [271/474] bpb=1.121517 time=309.3s + ttt_chunk [281/474] bpb=1.121123 time=320.7s + ttt_chunk [291/474] bpb=1.119558 time=332.1s + ttt_chunk [301/474] bpb=1.119987 time=343.5s + ttt_chunk [311/474] bpb=1.119381 time=354.9s + ttt_chunk [321/474] bpb=1.117767 time=366.4s + ttt_chunk [331/474] bpb=1.116735 time=377.8s + ttt_chunk [341/474] bpb=1.115996 time=389.2s + ttt_chunk [351/474] bpb=1.114351 time=400.6s + ttt_chunk [361/474] bpb=1.114833 time=412.0s + ttt_chunk [371/474] bpb=1.114520 time=423.4s + ttt_chunk [381/474] bpb=1.115285 time=434.8s + ttt_chunk [391/474] bpb=1.116303 time=446.2s + ttt_chunk [401/474] bpb=1.116709 time=457.6s + ttt_chunk [411/474] bpb=1.117120 time=469.0s + ttt_chunk [421/474] bpb=1.118532 time=480.4s + ttt_chunk [431/474] bpb=1.117100 time=491.8s + ttt_chunk [441/474] bpb=1.116713 time=503.2s + ttt_chunk [451/474] bpb=1.116048 time=514.7s + ttt_chunk [461/474] bpb=1.116249 time=526.1s + ttt_chunk [471/474] bpb=1.116338 time=537.5s + ttt_chunk [474/474] bpb=1.116206 time=540.0s +ttt:done val_loss=1.883479 val_bpb=1.115506 elapsed=540.0s +final_int6_ttt val_loss:1.8835 val_bpb:1.1155 stride:32 eval_time:540962ms +final_int6_ttt_exact val_loss:1.88347869 val_bpb:1.11550587 diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log new file mode 100644 index 000000000..6ea144408 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log @@ -0,0 +1,172 @@ +W0324 07:15:10.419000 100096 torch/distributed/run.py:851] +W0324 07:15:10.419000 100096 torch/distributed/run.py:851] ***************************************** +W0324 07:15:10.419000 100096 torch/distributed/run.py:851] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0324 07:15:10.419000 100096 torch/distributed/run.py:851] ***************************************** +logs/026bfb42-da81-45bf-b73b-6ae924fd88fa.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33580124 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9335 train_time:154ms step_avg:153.60ms +step:2/20000 train_loss:8.6987 train_time:244ms step_avg:121.91ms +step:3/20000 train_loss:7.7606 train_time:339ms step_avg:112.85ms +step:4/20000 train_loss:7.2812 train_time:434ms step_avg:108.41ms +step:5/20000 train_loss:7.0422 train_time:529ms step_avg:105.73ms +step:6/20000 train_loss:6.9445 train_time:623ms step_avg:103.84ms +step:7/20000 train_loss:6.8297 train_time:718ms step_avg:102.52ms +step:8/20000 train_loss:6.6897 train_time:812ms step_avg:101.55ms +step:9/20000 train_loss:6.3850 train_time:907ms step_avg:100.77ms +step:10/20000 train_loss:5.9825 train_time:1002ms step_avg:100.21ms +step:500/20000 train_loss:2.3564 train_time:48431ms step_avg:96.86ms +step:1000/20000 train_loss:2.2389 train_time:97006ms step_avg:97.01ms +step:1500/20000 train_loss:2.1831 train_time:145627ms step_avg:97.08ms +step:2000/20000 train_loss:2.0279 train_time:194337ms step_avg:97.17ms +step:2500/20000 train_loss:2.1312 train_time:243086ms step_avg:97.23ms +step:3000/20000 train_loss:2.1151 train_time:291847ms step_avg:97.28ms +step:3500/20000 train_loss:2.1228 train_time:340613ms step_avg:97.32ms +step:4000/20000 train_loss:1.9145 train_time:389395ms step_avg:97.35ms +step:4000/20000 val_loss:2.0040 val_bpb:1.1869 train_time:389401ms step_avg:97.35ms +soft_round_qat:enabled initial_alpha=1.0 +late_qat:enabled step:4413 scale:0.4999 +step:4500/20000 train_loss:2.0597 train_time:438163ms step_avg:97.37ms +step:5000/20000 train_loss:2.0373 train_time:486893ms step_avg:97.38ms +swa:start step:5500 +step:5500/20000 train_loss:1.9452 train_time:535639ms step_avg:97.39ms +step:6000/20000 train_loss:1.8708 train_time:584998ms step_avg:97.50ms +step:6152/20000 val_loss:1.8998 val_bpb:1.1252 train_time:600042ms step_avg:97.54ms +stopping_early: wallclock_cap train_time:600042ms step:6152/20000 +peak memory allocated: 26201 MiB reserved: 26418 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8982 val_bpb:1.1242 eval_time:2370ms +swa:applying SWA weights (count=14) +DIAGNOSTIC post_swa val_loss:1.8997 val_bpb:1.1251 eval_time:2371ms +best_averaging:ema val_bpb:1.1242 +Serialized model: 130956873 bytes +Code size: 106734 bytes +pruning:2.0% magnitude pruning applied +gptq:calibrating with training data... +gptq:calibrated 68 layers in 3.6s +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +Serialized model int6+zstd: 15308671 bytes +Total submission size int6+zstd: 15415405 bytes +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 1.0, Byte-weighted TTT: False +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27799624 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0787 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0670 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0662 + ttt_chunk [1/474] bpb=1.201291 time=1.3s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.9070 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.9062 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.9039 + ttt_chunk [2/474] bpb=1.154634 time=2.5s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.8655 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.8648 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.8629 + ttt_chunk [3/474] bpb=1.130046 time=3.6s + ttt_chunk [4/474] bpb=1.137970 time=4.7s + ttt_chunk [5/474] bpb=1.136011 time=5.9s + ttt_chunk [11/474] bpb=1.117481 time=12.7s + ttt_chunk [21/474] bpb=1.112444 time=24.1s + ttt_chunk [31/474] bpb=1.110774 time=35.5s + ttt_chunk [41/474] bpb=1.118885 time=47.0s + ttt_chunk [51/474] bpb=1.126265 time=58.4s + ttt_chunk [61/474] bpb=1.124173 time=69.8s + ttt_chunk [71/474] bpb=1.125678 time=81.2s + ttt_chunk [81/474] bpb=1.126147 time=92.6s + ttt_chunk [91/474] bpb=1.128060 time=104.0s + ttt_chunk [101/474] bpb=1.124603 time=115.5s + ttt_chunk [111/474] bpb=1.124793 time=126.9s + ttt_chunk [121/474] bpb=1.128136 time=138.3s + ttt_chunk [131/474] bpb=1.128812 time=149.7s + ttt_chunk [141/474] bpb=1.128665 time=161.1s + ttt_chunk [151/474] bpb=1.126959 time=172.5s + ttt_chunk [161/474] bpb=1.127822 time=183.9s + ttt_chunk [171/474] bpb=1.126462 time=195.3s + ttt_chunk [181/474] bpb=1.127216 time=206.8s + ttt_chunk [191/474] bpb=1.126189 time=218.2s + ttt_chunk [201/474] bpb=1.125335 time=229.6s + ttt_chunk [211/474] bpb=1.124007 time=241.0s + ttt_chunk [221/474] bpb=1.124128 time=252.4s + ttt_chunk [231/474] bpb=1.123537 time=263.8s + ttt_chunk [241/474] bpb=1.122433 time=275.3s + ttt_chunk [251/474] bpb=1.123539 time=286.7s + ttt_chunk [261/474] bpb=1.124188 time=298.1s + ttt_chunk [271/474] bpb=1.122695 time=309.5s + ttt_chunk [281/474] bpb=1.122278 time=320.9s + ttt_chunk [291/474] bpb=1.120755 time=332.3s + ttt_chunk [301/474] bpb=1.121114 time=343.7s + ttt_chunk [311/474] bpb=1.120507 time=355.2s + ttt_chunk [321/474] bpb=1.118816 time=366.6s + ttt_chunk [331/474] bpb=1.117759 time=378.0s + ttt_chunk [341/474] bpb=1.117018 time=389.4s + ttt_chunk [351/474] bpb=1.115370 time=400.8s + ttt_chunk [361/474] bpb=1.115858 time=412.2s + ttt_chunk [371/474] bpb=1.115523 time=423.6s + ttt_chunk [381/474] bpb=1.116283 time=435.0s + ttt_chunk [391/474] bpb=1.117326 time=446.5s + ttt_chunk [401/474] bpb=1.117734 time=457.9s + ttt_chunk [411/474] bpb=1.118144 time=469.3s + ttt_chunk [421/474] bpb=1.119563 time=480.7s + ttt_chunk [431/474] bpb=1.118129 time=492.1s + ttt_chunk [441/474] bpb=1.117766 time=503.5s + ttt_chunk [451/474] bpb=1.117119 time=515.0s + ttt_chunk [461/474] bpb=1.117338 time=526.4s + ttt_chunk [471/474] bpb=1.117400 time=537.8s + ttt_chunk [474/474] bpb=1.117261 time=540.3s +ttt:done val_loss=1.884801 val_bpb=1.116289 elapsed=540.3s +final_int6_ttt val_loss:1.8848 val_bpb:1.1163 stride:32 eval_time:541278ms +final_int6_ttt_exact val_loss:1.88480123 val_bpb:1.11628915 diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log new file mode 100644 index 000000000..0fe246d62 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log @@ -0,0 +1,172 @@ +W0324 07:20:32.173000 139509 torch/distributed/run.py:803] +W0324 07:20:32.173000 139509 torch/distributed/run.py:803] ***************************************** +W0324 07:20:32.173000 139509 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0324 07:20:32.173000 139509 torch/distributed/run.py:803] ***************************************** +logs/5d3ebe13-8183-4305-965c-55651fc9638b.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) +model_params:33580124 +XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 +lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9311 val_bpb:4.1050 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9326 train_time:153ms step_avg:152.55ms +step:2/20000 train_loss:8.7576 train_time:246ms step_avg:123.04ms +step:3/20000 train_loss:7.7488 train_time:342ms step_avg:114.08ms +step:4/20000 train_loss:7.2030 train_time:438ms step_avg:109.53ms +step:5/20000 train_loss:7.0022 train_time:534ms step_avg:106.79ms +step:6/20000 train_loss:6.8717 train_time:630ms step_avg:104.99ms +step:7/20000 train_loss:6.7821 train_time:726ms step_avg:103.66ms +step:8/20000 train_loss:6.6344 train_time:822ms step_avg:102.72ms +step:9/20000 train_loss:6.3142 train_time:918ms step_avg:101.98ms +step:10/20000 train_loss:5.9783 train_time:1014ms step_avg:101.40ms +step:500/20000 train_loss:2.3552 train_time:49006ms step_avg:98.01ms +step:1000/20000 train_loss:2.2387 train_time:98370ms step_avg:98.37ms +step:1500/20000 train_loss:2.1831 train_time:147750ms step_avg:98.50ms +step:2000/20000 train_loss:2.0243 train_time:197149ms step_avg:98.57ms +step:2500/20000 train_loss:2.1289 train_time:246526ms step_avg:98.61ms +step:3000/20000 train_loss:2.1142 train_time:295868ms step_avg:98.62ms +step:3500/20000 train_loss:2.1203 train_time:345180ms step_avg:98.62ms +step:4000/20000 train_loss:1.9099 train_time:394458ms step_avg:98.61ms +step:4000/20000 val_loss:2.0013 val_bpb:1.1853 train_time:394463ms step_avg:98.62ms +soft_round_qat:enabled initial_alpha=1.0 +late_qat:enabled step:4334 scale:0.4999 +step:4500/20000 train_loss:2.0596 train_time:443781ms step_avg:98.62ms +step:5000/20000 train_loss:2.0348 train_time:493046ms step_avg:98.61ms +swa:start step:5400 +step:5500/20000 train_loss:1.9439 train_time:542488ms step_avg:98.63ms +step:6000/20000 train_loss:1.8689 train_time:592060ms step_avg:98.68ms +step:6081/20000 val_loss:1.8999 val_bpb:1.1252 train_time:600095ms step_avg:98.68ms +stopping_early: wallclock_cap train_time:600095ms step:6081/20000 +peak memory allocated: 26199 MiB reserved: 26784 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.8983 val_bpb:1.1243 eval_time:2377ms +swa:applying SWA weights (count=14) +DIAGNOSTIC post_swa val_loss:1.9001 val_bpb:1.1253 eval_time:2379ms +best_averaging:ema val_bpb:1.1243 +Serialized model: 130956873 bytes +Code size: 106734 bytes +pruning:2.0% magnitude pruning applied +gptq:calibrating with training data... +gptq:calibrated 68 layers in 3.7s +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +gptq_quantize: 66 GPTQ layers, 0 naive layers +mixed_precision: 33423360 int5 params, 0 int6 params +Serialized model int6+zstd: 15261893 bytes +Total submission size int6+zstd: 15368627 bytes +TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw +TTT temperature: 0.98 +PPM alpha: 1.0, Byte-weighted TTT: False +ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 +ttt:params unfrozen=5780500 frozen=27799624 + ttt_train [1] seqs=64 start_train... + ttt_train [1] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=2.0805 + ttt_train [1] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=2.0689 + ttt_train [1] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=2.0683 + ttt_chunk [1/474] bpb=1.200925 time=1.3s + ttt_train [2] seqs=64 start_train... + ttt_train [2] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.9082 + ttt_train [2] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.9075 + ttt_train [2] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.9048 + ttt_chunk [2/474] bpb=1.155071 time=2.5s + ttt_train [3] seqs=64 start_train... + ttt_train [3] epoch=1/3 batches=8 ... + step done ep=1 bs=0 loss=1.8586 + ttt_train [3] epoch=2/3 batches=8 ... + step done ep=2 bs=0 loss=1.8574 + ttt_train [3] epoch=3/3 batches=8 ... + step done ep=3 bs=0 loss=1.8559 + ttt_chunk [3/474] bpb=1.129053 time=3.6s + ttt_chunk [4/474] bpb=1.137770 time=4.8s + ttt_chunk [5/474] bpb=1.136142 time=5.9s + ttt_chunk [11/474] bpb=1.117081 time=12.8s + ttt_chunk [21/474] bpb=1.112525 time=24.2s + ttt_chunk [31/474] bpb=1.110601 time=35.6s + ttt_chunk [41/474] bpb=1.118836 time=47.0s + ttt_chunk [51/474] bpb=1.126156 time=58.4s + ttt_chunk [61/474] bpb=1.124130 time=69.8s + ttt_chunk [71/474] bpb=1.125735 time=81.2s + ttt_chunk [81/474] bpb=1.126269 time=92.7s + ttt_chunk [91/474] bpb=1.128083 time=104.1s + ttt_chunk [101/474] bpb=1.124677 time=115.5s + ttt_chunk [111/474] bpb=1.125049 time=126.9s + ttt_chunk [121/474] bpb=1.128310 time=138.3s + ttt_chunk [131/474] bpb=1.129031 time=149.7s + ttt_chunk [141/474] bpb=1.129021 time=161.1s + ttt_chunk [151/474] bpb=1.127264 time=172.5s + ttt_chunk [161/474] bpb=1.128178 time=184.0s + ttt_chunk [171/474] bpb=1.126890 time=195.4s + ttt_chunk [181/474] bpb=1.127673 time=206.8s + ttt_chunk [191/474] bpb=1.126603 time=218.2s + ttt_chunk [201/474] bpb=1.125778 time=229.6s + ttt_chunk [211/474] bpb=1.124423 time=241.0s + ttt_chunk [221/474] bpb=1.124509 time=252.4s + ttt_chunk [231/474] bpb=1.123899 time=263.8s + ttt_chunk [241/474] bpb=1.122760 time=275.3s + ttt_chunk [251/474] bpb=1.123798 time=286.7s + ttt_chunk [261/474] bpb=1.124407 time=298.1s + ttt_chunk [271/474] bpb=1.122942 time=309.5s + ttt_chunk [281/474] bpb=1.122514 time=320.9s + ttt_chunk [291/474] bpb=1.120992 time=332.3s + ttt_chunk [301/474] bpb=1.121428 time=343.7s + ttt_chunk [311/474] bpb=1.120836 time=355.1s + ttt_chunk [321/474] bpb=1.119179 time=366.5s + ttt_chunk [331/474] bpb=1.118111 time=377.9s + ttt_chunk [341/474] bpb=1.117318 time=389.4s + ttt_chunk [351/474] bpb=1.115673 time=400.8s + ttt_chunk [361/474] bpb=1.116152 time=412.2s + ttt_chunk [371/474] bpb=1.115831 time=423.6s + ttt_chunk [381/474] bpb=1.116577 time=435.0s + ttt_chunk [391/474] bpb=1.117608 time=446.4s + ttt_chunk [401/474] bpb=1.118021 time=457.9s + ttt_chunk [411/474] bpb=1.118453 time=469.2s + ttt_chunk [421/474] bpb=1.119901 time=480.6s + ttt_chunk [431/474] bpb=1.118463 time=491.9s + ttt_chunk [441/474] bpb=1.118084 time=503.3s + ttt_chunk [451/474] bpb=1.117423 time=514.7s + ttt_chunk [461/474] bpb=1.117641 time=526.0s + ttt_chunk [471/474] bpb=1.117706 time=537.4s + ttt_chunk [474/474] bpb=1.117562 time=539.8s +ttt:done val_loss=1.885435 val_bpb=1.116665 elapsed=539.8s +final_int6_ttt val_loss:1.8854 val_bpb:1.1167 stride:32 eval_time:540891ms +final_int6_ttt_exact val_loss:1.88543543 val_bpb:1.11666477 From 198c91340360ee358fa58498005d4a4c8c4c68ff Mon Sep 17 00:00:00 2001 From: ethan Date: Wed, 25 Mar 2026 15:03:04 +0800 Subject: [PATCH 2/5] Record: CROWN-Q + Full GPTQ + SWA/EMA (3-seed mean val_bpb=1.1186) --- .../README.md | 58 + .../submission.json | 20 + .../train_gpt.py | 1917 +++++++++++++++++ .../train_seed1337.log | 137 ++ .../train_seed42.log | 142 ++ .../train_seed7.log | 135 ++ 6 files changed, 2409 insertions(+) create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log create mode 100644 records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md new file mode 100644 index 000000000..53ebf74dd --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md @@ -0,0 +1,58 @@ +# CROWN-Q + Full GPTQ + SWA/EMA Blend + +## Summary + +- **CROWN-Q**: Curvature-weighted quantization variance penalty applied during warmdown. Encourages weights to settle in flat minima where int6 quantization causes less damage. Penalty: `lambda * sum(h_j * delta_j^2 / 12)` where `h_j = w^2` (curvature proxy) and `delta_j = row_max / 15` (quantization step size). +- **Full Cholesky GPTQ**: Hessian-aware quantization with act-order column permutation, block_size=128, 256-sample calibration from training data. All within 585s training budget. +- **SWA/EMA 50/50 blend**: Stochastic Weight Averaging (every 50 steps during warmdown) blended 50/50 with EMA (decay=0.997). +- **Architecture**: 11L, 512d, GQA 8H/4KV, MLP 3x LeakyReLU(0.5)^2, XSA on all 11 layers, VRL, BigramHash 3072, partial RoPE 16/64. +- **Eval**: Sliding window with stride=64. No test-time training. + +## Configuration + +```bash +# Training (585s wallclock, includes GPTQ calibration) +torchrun --standalone --nproc_per_node=8 train_gpt.py + +# Key env vars (all defaults in code): +# CROWNQ_LAMBDA=0.01 — CROWN-Q penalty weight +# CROWNQ_WARMDOWN_ONLY=1 — only apply during warmdown +# LATE_QAT_THRESHOLD=0.15 — QAT activation point +# MAX_WALLCLOCK_SECONDS=585 — training budget +# WARMDOWN_ITERS=4000 — warmdown length +``` + +## Results + +| Seed | Steps | Post-EMA BPB | Sliding BPB | Artifact | +|------|-------|-------------|-------------|----------| +| 1337 | 6613 | 1.1387 | **1.1189** | 15,945,134 | +| 42 | 6612 | 1.1382 | **1.1189** | 15,947,742 | +| 7 | 6612 | 1.1378 | **1.1179** | 15,938,790 | +| **Mean** | | 1.1382 | **1.1186** | | +| **Std** | | | 0.0006 | | + +- Step speed: 87ms/step (FA3 Hopper) +- Quant gap (roundtrip): ~0.004 BPB +- Sliding window eval time: ~75s +- Training time: 585s (under 600s budget) + +## What is CROWN-Q? + +CROWN-Q (Curvature-Regularized Optimization for Weight Noise Quantization) adds a training-time penalty that makes weights more robust to quantization noise: + +1. For each weight matrix, compute the per-row quantization step size `delta = row_max / 15` (int6 range [-15, 15]) +2. Compute quantization variance `delta^2 / 12` (uniform rounding noise) +3. Weight by curvature proxy `h = w^2` (large weights in high-curvature directions) +4. Penalty: `lambda * sum(h * quant_var)` encourages the optimizer to reduce weights in directions where quantization noise is most damaging + +Applied only during warmdown when QAT is active. Zero eval-time cost. + +## Included Files + +- `train_gpt.py` — self-contained training script +- `submission.json` — submission metadata +- `README.md` — this file +- `train_seed1337.log` — seed 1337 training log +- `train_seed42.log` — seed 42 training log +- `train_seed7.log` — seed 7 training log diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json new file mode 100644 index 000000000..0a4b5540f --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json @@ -0,0 +1,20 @@ +{ + "author": "Ethan Yang", + "github_id": "EthanYangTW", + "name": "CROWN-Q + Full GPTQ + SWA/EMA Blend", + "blurb": "Curvature-weighted quantization variance penalty (CROWN-Q) during warmdown reduces quantization damage. Full Cholesky GPTQ with act-order, SWA/EMA 50/50 blend, VRL, XSA-all 11 layers, LeakyReLU(0.5)^2. Sliding window eval only, no TTT.", + "date": "2026-03-25T06:30:00Z", + "val_bpb": 1.1186, + "val_bpb_std": 0.0006, + "val_bpb_seed1337": 1.1189, + "val_loss_seed1337": 1.8891, + "bytes_seed1337": 15945134, + "val_bpb_seed42": 1.1189, + "val_loss_seed42": 1.8891, + "bytes_seed42": 15947742, + "val_bpb_seed7": 1.1179, + "val_loss_seed7": 1.8876, + "bytes_seed7": 15938790, + "bytes_total": 15947742, + "bytes_code": 95390 +} diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py new file mode 100644 index 000000000..d8fd02865 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py @@ -0,0 +1,1917 @@ +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 +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 +from flash_attn_interface import flash_attn_func as flash_attn_3_func +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", 4000)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 585.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = 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", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # tighter: collect more recent checkpoints + 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", 4)) # XSA on last 4 layers (0 = disabled) + 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", 0.15)) + crownq_lambda = float(os.environ.get("CROWNQ_LAMBDA", 0.01)) # CROWN-Q penalty weight + crownq_warmdown_only = bool(int(os.environ.get("CROWNQ_WARMDOWN_ONLY", "1"))) # only apply during warmdown + 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") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 131072)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + wd = group.get("weight_decay", 0.0) + curr = 0 + for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + return loss +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) +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,vrl_lambda", + ).split(",") + if pattern +) +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +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 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) +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 + _soft_tau: float = 1000.0 # High = hard round; low = soft (annealed during QAT) + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + x_norm = w32 / scale[:, None] + # Hard quantized value (forward) + q_hard = torch.clamp(torch.round(x_norm), -31, 31) + # Soft interpolation (backward) for gradient signal + x_floor = x_norm.detach().floor() + frac = x_norm - x_floor + p = torch.sigmoid((frac - 0.5) / max(CastedLinear._soft_tau, 0.01)) + q_soft = torch.clamp(x_floor.detach() + p, -31, 31) + # STE: hard forward, soft backward + q = q_hard.detach() + (q_soft - q_soft.detach()) + w_q = (q * 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, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + kv_dim = self.num_kv_heads * self.head_dim + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] — broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + bsz, seqlen, dim = x.shape + q = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, 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] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + return self.proj(x.square()) +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + 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, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + 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, raw_v +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + use_vrl: bool = False, + ): + super().__init__() + self.use_vrl = use_vrl + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor, lora=None) -> Tensor: + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + 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, return_hidden: bool = False): + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if return_hidden: + return logits, x + return logits +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 = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + 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) + + if log_fn: + log_fn(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) + + # Freeze first N blocks + optionally embeddings + 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 = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(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)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + 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) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + 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() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + 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: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + 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 + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_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) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(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 + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +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 collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + 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 h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = 0.01 * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + 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 + # Act-order: quantize most-activated columns first + 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] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, 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) + 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 mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, block_size=gptq_block_size) + else: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + +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 + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + 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, + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p + for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + matrix_params.append(base_model.bigram.proj.weight) + 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) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + optimizers.insert(1, optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") + 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.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + CastedLinear._qat_enabled = False + CastedLinear._soft_tau = 1000.0 + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + log0("eval_only:TTT disabled, computing sliding window BPB") + 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=64, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + 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} crownq_lambda:{args.crownq_lambda}") + # Anneal soft-rounding temperature: hard for most of QAT, soft at the end + if CastedLinear._qat_enabled: + CastedLinear._soft_tau = 0.1 if scale < 0.02 else 1000.0 + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + # CROWN-Q: curvature-weighted quantization variance penalty + if CastedLinear._qat_enabled and args.crownq_lambda > 0 and (not args.crownq_warmdown_only or scale < 1.0): + crownq_penalty = torch.zeros((), device=device) + for m in base_model.modules(): + if isinstance(m, CastedLinear) and m.weight.ndim == 2: + w = m.weight.float() + row_max = w.abs().amax(dim=1).clamp(min=1e-10) + delta = row_max / 15.0 # step size for int6 range [-15,15] + quant_var = (delta ** 2) / 12.0 + h_proxy = (w ** 2).mean(dim=1) + crownq_penalty = crownq_penalty + (h_proxy * quant_var).sum() + loss = loss + args.crownq_lambda * crownq_penalty + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + # EMA update + 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" + ) + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + ttt_epoch_loss += ttt_loss.item() + 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) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + log0(f"swa:applying {swa_count} snapshots, blending with EMA (0.5/0.5)") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = 0.5 * ema_w + 0.5 * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + 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() + 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()} + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + return lzma.compress(buf.getvalue(), preset=6) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + # Hessian-weighted cost: how sensitive is output to this weight? + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta) + if len(blob) <= target_model_bytes: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values") + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") + quant_raw = quant_buf.getvalue() + quant_blob = lzma.compress(quant_raw, preset=6) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {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)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, # must match training model + 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, + use_vrl=args.vrl, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + CastedLinear._qat_enabled = False + CastedLinear._soft_tau = 1000.0 + 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}") + # Legal score-first TTT (PR#461/549 recipe) + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log new file mode 100644 index 000000000..ea896f40f --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log @@ -0,0 +1,137 @@ +W0325 05:31:25.719000 324884 torch/distributed/run.py:803] +W0325 05:31:25.719000 324884 torch/distributed/run.py:803] ***************************************** +W0325 05:31:25.719000 324884 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 05:31:25.719000 324884 torch/distributed/run.py:803] ***************************************** +logs/4d3c73cf-10ff-45ac-bd31-745110d174e7.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:27124848 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:True active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:585.000 +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:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9313 train_time:140ms step_avg:140.48ms +step:2/20000 train_loss:8.6873 train_time:232ms step_avg:116.02ms +step:3/20000 train_loss:7.8968 train_time:322ms step_avg:107.41ms +step:4/20000 train_loss:7.2122 train_time:413ms step_avg:103.14ms +step:5/20000 train_loss:6.9640 train_time:502ms step_avg:100.48ms +step:6/20000 train_loss:6.9142 train_time:592ms step_avg:98.73ms +step:7/20000 train_loss:6.8328 train_time:682ms step_avg:97.44ms +step:8/20000 train_loss:6.7688 train_time:773ms step_avg:96.58ms +step:9/20000 train_loss:6.4086 train_time:863ms step_avg:95.93ms +step:10/20000 train_loss:6.1321 train_time:954ms step_avg:95.38ms +step:500/20000 train_loss:2.3720 train_time:43330ms step_avg:86.66ms +step:1000/20000 train_loss:2.2526 train_time:86819ms step_avg:86.82ms +step:1500/20000 train_loss:2.2073 train_time:130376ms step_avg:86.92ms +step:2000/20000 train_loss:2.0509 train_time:174067ms step_avg:87.03ms +step:2500/20000 train_loss:2.1560 train_time:217815ms step_avg:87.13ms +step:3000/20000 train_loss:2.1429 train_time:261560ms step_avg:87.19ms +step:3500/20000 train_loss:2.1521 train_time:305345ms step_avg:87.24ms +step:4000/20000 train_loss:1.9467 train_time:349086ms step_avg:87.27ms +step:4000/20000 val_loss:2.0374 val_bpb:1.2067 train_time:349099ms step_avg:87.27ms +step:4500/20000 train_loss:2.0961 train_time:392809ms step_avg:87.29ms +step:5000/20000 train_loss:2.0786 train_time:436507ms step_avg:87.30ms +step:5500/20000 train_loss:1.9949 train_time:480327ms step_avg:87.33ms +swa:start step:5900 +step:6000/20000 train_loss:1.9168 train_time:524158ms step_avg:87.36ms +late_qat:enabled step:6097 scale:0.1498 crownq_lambda:0.01 +step:6500/20000 train_loss:2.0586 train_time:573554ms step_avg:88.24ms +step:6613/20000 val_loss:1.9254 val_bpb:1.1403 train_time:584992ms step_avg:88.46ms +stopping_early: wallclock_cap train_time:584992ms step:6613/20000 +peak memory allocated: 21415 MiB reserved: 21452 MiB +ttt_burst:start epochs:2 buffer_size:100 lr_factor:0.1 +ttt_burst:epoch:1/2 avg_loss:1.9133 +ttt_burst:epoch:2/2 avg_loss:1.8845 +ttt_burst:done +swa:applying 15 snapshots, blending with EMA (0.5/0.5) +DIAGNOSTIC post_ema val_loss:1.9226 val_bpb:1.1387 eval_time:1997ms +Serialized model: 106443805 bytes +Code size: 95390 bytes +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=37.5s +gptq:pre_prune artifact=15862628 target=15854610 +gptq:pruning candidates=4243702 +gptq:pruned 1077 values +Saved quantized model to final_int6_model.pt +Serialized model int6+lzma: 15849744 bytes +Total submission size int6+lzma: 15945134 bytes +final_int6_roundtrip val_loss:1.9290 val_bpb:1.1425 eval_time:6358ms +final_int6_roundtrip_exact val_loss:1.92899174 val_bpb:1.14245756 +final_int6_sliding_window val_loss:1.8891 val_bpb:1.1189 stride:64 eval_time:74911ms +final_int6_sliding_window_exact val_loss:1.88914884 val_bpb:1.11886332 +final_int8_zlib_roundtrip_exact val_loss:1.88914884 val_bpb:1.11886332 +ttt:start lr=0.002 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=27124848 frozen=0 + ttt_chunk [1/474] bpb=1.198568 time=0.7s + ttt_chunk [11/474] bpb=1.116523 time=6.1s + ttt_chunk [21/474] bpb=1.112338 time=11.4s + ttt_chunk [31/474] bpb=1.111323 time=16.7s + ttt_chunk [41/474] bpb=1.119522 time=22.0s + ttt_chunk [51/474] bpb=1.126882 time=27.4s + ttt_chunk [61/474] bpb=1.124971 time=32.7s + ttt_chunk [71/474] bpb=1.126426 time=38.0s +W0325 05:52:14.355000 324884 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers +W0325 05:52:14.358000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324952 closing signal SIGTERM +W0325 05:52:14.360000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324953 closing signal SIGTERM +W0325 05:52:14.361000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324954 closing signal SIGTERM +W0325 05:52:14.363000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324955 closing signal SIGTERM +W0325 05:52:14.369000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324956 closing signal SIGTERM +W0325 05:52:14.385000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324957 closing signal SIGTERM +W0325 05:52:14.395000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324958 closing signal SIGTERM +W0325 05:52:14.396000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324959 closing signal SIGTERM +Traceback (most recent call last): + File "/usr/local/bin/torchrun", line 7, in + sys.exit(main()) + ^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main + run(args) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent + result = agent.run() + ^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper + result = f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run + result = self._invoke_run(role) + ^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run + time.sleep(monitor_interval) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler + raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) +torch.distributed.elastic.multiprocessing.api.SignalException: Process 324884 got signal: 15 diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log new file mode 100644 index 000000000..81ce4b461 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log @@ -0,0 +1,142 @@ +W0325 06:09:26.608000 329471 torch/distributed/run.py:803] +W0325 06:09:26.608000 329471 torch/distributed/run.py:803] ***************************************** +W0325 06:09:26.608000 329471 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 06:09:26.608000 329471 torch/distributed/run.py:803] ***************************************** +logs/a5331a4e-c975-4ec8-a747-a7c42052088a.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:27124848 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:True active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:585.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9301 val_bpb:4.1044 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9323 train_time:138ms step_avg:138.36ms +step:2/20000 train_loss:8.6766 train_time:225ms step_avg:112.49ms +step:3/20000 train_loss:7.8928 train_time:316ms step_avg:105.38ms +step:4/20000 train_loss:7.1944 train_time:407ms step_avg:101.86ms +step:5/20000 train_loss:6.9292 train_time:499ms step_avg:99.73ms +step:6/20000 train_loss:6.8774 train_time:590ms step_avg:98.34ms +step:7/20000 train_loss:6.8066 train_time:681ms step_avg:97.31ms +step:8/20000 train_loss:6.7352 train_time:773ms step_avg:96.63ms +step:9/20000 train_loss:6.4472 train_time:865ms step_avg:96.08ms +step:10/20000 train_loss:6.1363 train_time:956ms step_avg:95.58ms +step:500/20000 train_loss:2.3712 train_time:43466ms step_avg:86.93ms +step:1000/20000 train_loss:2.2539 train_time:87123ms step_avg:87.12ms +step:1500/20000 train_loss:2.2067 train_time:130874ms step_avg:87.25ms +step:2000/20000 train_loss:2.0510 train_time:174641ms step_avg:87.32ms +step:2500/20000 train_loss:2.1515 train_time:218407ms step_avg:87.36ms +step:3000/20000 train_loss:2.1436 train_time:262173ms step_avg:87.39ms +step:3500/20000 train_loss:2.1521 train_time:305903ms step_avg:87.40ms +step:4000/20000 train_loss:1.9447 train_time:349625ms step_avg:87.41ms +step:4000/20000 val_loss:2.0356 val_bpb:1.2056 train_time:349638ms step_avg:87.41ms +step:4500/20000 train_loss:2.0945 train_time:393335ms step_avg:87.41ms +step:5000/20000 train_loss:2.0769 train_time:436988ms step_avg:87.40ms +step:5500/20000 train_loss:1.9918 train_time:480699ms step_avg:87.40ms +swa:start step:5900 +step:6000/20000 train_loss:1.9175 train_time:524459ms step_avg:87.41ms +late_qat:enabled step:6093 scale:0.1499 crownq_lambda:0.01 +step:6500/20000 train_loss:2.0585 train_time:573787ms step_avg:88.27ms +step:6612/20000 val_loss:1.9247 val_bpb:1.1399 train_time:585006ms step_avg:88.48ms +stopping_early: wallclock_cap train_time:585006ms step:6612/20000 +peak memory allocated: 21415 MiB reserved: 21452 MiB +ttt_burst:start epochs:2 buffer_size:100 lr_factor:0.1 +ttt_burst:epoch:1/2 avg_loss:1.9130 +ttt_burst:epoch:2/2 avg_loss:1.8843 +ttt_burst:done +swa:applying 15 snapshots, blending with EMA (0.5/0.5) +DIAGNOSTIC post_ema val_loss:1.9218 val_bpb:1.1382 eval_time:1979ms +Serialized model: 106443805 bytes +Code size: 95390 bytes +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=37.1s +gptq:pre_prune artifact=15909060 target=15854610 +gptq:pruning candidates=4242193 +gptq:pruned 262615 values +Saved quantized model to final_int6_model.pt +Serialized model int6+lzma: 15852352 bytes +Total submission size int6+lzma: 15947742 bytes +final_int6_roundtrip val_loss:1.9290 val_bpb:1.1424 eval_time:6595ms +final_int6_roundtrip_exact val_loss:1.92896314 val_bpb:1.14244063 +final_int6_sliding_window val_loss:1.8891 val_bpb:1.1189 stride:64 eval_time:75758ms +final_int6_sliding_window_exact val_loss:1.88913665 val_bpb:1.11885610 +final_int8_zlib_roundtrip_exact val_loss:1.88913665 val_bpb:1.11885610 +ttt:start lr=0.002 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=27124848 frozen=0 + ttt_chunk [1/474] bpb=1.196551 time=0.7s + ttt_chunk [11/474] bpb=1.116343 time=6.1s + ttt_chunk [21/474] bpb=1.112385 time=11.4s + ttt_chunk [31/474] bpb=1.111105 time=16.7s + ttt_chunk [41/474] bpb=1.119287 time=22.0s + ttt_chunk [51/474] bpb=1.126759 time=27.3s + ttt_chunk [61/474] bpb=1.124851 time=32.8s + ttt_chunk [71/474] bpb=1.126283 time=38.4s + ttt_chunk [81/474] bpb=1.126907 time=44.0s + ttt_chunk [91/474] bpb=1.128848 time=49.6s + ttt_chunk [101/474] bpb=1.125491 time=55.2s + ttt_chunk [111/474] bpb=1.125899 time=60.6s + ttt_chunk [121/474] bpb=1.129329 time=65.9s +W0325 06:31:28.212000 329471 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers +W0325 06:31:28.216000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329539 closing signal SIGTERM +W0325 06:31:28.219000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329540 closing signal SIGTERM +W0325 06:31:28.221000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329541 closing signal SIGTERM +W0325 06:31:28.225000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329542 closing signal SIGTERM +W0325 06:31:28.227000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329543 closing signal SIGTERM +W0325 06:31:28.228000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329544 closing signal SIGTERM +W0325 06:31:28.230000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329545 closing signal SIGTERM +W0325 06:31:28.231000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329546 closing signal SIGTERM +Traceback (most recent call last): + File "/usr/local/bin/torchrun", line 7, in + sys.exit(main()) + ^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main + run(args) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent + result = agent.run() + ^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper + result = f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run + result = self._invoke_run(role) + ^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run + time.sleep(monitor_interval) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler + raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) +torch.distributed.elastic.multiprocessing.api.SignalException: Process 329471 got signal: 15 diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log new file mode 100644 index 000000000..d6c7b54b6 --- /dev/null +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log @@ -0,0 +1,135 @@ +W0325 06:31:47.254000 330566 torch/distributed/run.py:803] +W0325 06:31:47.254000 330566 torch/distributed/run.py:803] ***************************************** +W0325 06:31:47.254000 330566 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0325 06:31:47.254000 330566 torch/distributed/run.py:803] ***************************************** +logs/2dfea1c3-08cf-4034-8e6a-ec6d90ed9fcc.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model +train_loader:dataset:fineweb10B_sp1024 train_shards:80 +val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 +model_params:27124848 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_4 active_layers:[7, 8, 9, 10] +VRL:True active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:585.000 +seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9304 val_bpb:4.1046 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9318 train_time:138ms step_avg:137.66ms +step:2/20000 train_loss:8.7502 train_time:224ms step_avg:112.22ms +step:3/20000 train_loss:7.8945 train_time:316ms step_avg:105.29ms +step:4/20000 train_loss:7.1935 train_time:407ms step_avg:101.78ms +step:5/20000 train_loss:6.9976 train_time:498ms step_avg:99.69ms +step:6/20000 train_loss:6.8845 train_time:589ms step_avg:98.24ms +step:7/20000 train_loss:6.7647 train_time:681ms step_avg:97.22ms +step:8/20000 train_loss:6.6478 train_time:772ms step_avg:96.46ms +step:9/20000 train_loss:6.3741 train_time:863ms step_avg:95.86ms +step:10/20000 train_loss:6.0492 train_time:955ms step_avg:95.50ms +step:500/20000 train_loss:2.3703 train_time:43438ms step_avg:86.88ms +step:1000/20000 train_loss:2.2552 train_time:86983ms step_avg:86.98ms +step:1500/20000 train_loss:2.2004 train_time:130565ms step_avg:87.04ms +step:2000/20000 train_loss:2.0457 train_time:174270ms step_avg:87.14ms +step:2500/20000 train_loss:2.1562 train_time:217972ms step_avg:87.19ms +step:3000/20000 train_loss:2.1406 train_time:261680ms step_avg:87.23ms +step:3500/20000 train_loss:2.1507 train_time:305392ms step_avg:87.25ms +step:4000/20000 train_loss:1.9441 train_time:349115ms step_avg:87.28ms +step:4000/20000 val_loss:2.0349 val_bpb:1.2052 train_time:349128ms step_avg:87.28ms +step:4500/20000 train_loss:2.0925 train_time:392853ms step_avg:87.30ms +step:5000/20000 train_loss:2.0753 train_time:436582ms step_avg:87.32ms +step:5500/20000 train_loss:1.9910 train_time:480377ms step_avg:87.34ms +swa:start step:5900 +step:6000/20000 train_loss:1.9164 train_time:524234ms step_avg:87.37ms +late_qat:enabled step:6096 scale:0.1498 crownq_lambda:0.01 +step:6500/20000 train_loss:2.0575 train_time:573594ms step_avg:88.25ms +step:6613/20000 val_loss:1.9239 val_bpb:1.1394 train_time:585041ms step_avg:88.47ms +stopping_early: wallclock_cap train_time:585041ms step:6613/20000 +peak memory allocated: 21415 MiB reserved: 21452 MiB +ttt_burst:start epochs:2 buffer_size:100 lr_factor:0.1 +ttt_burst:epoch:1/2 avg_loss:1.9118 +ttt_burst:epoch:2/2 avg_loss:1.8830 +ttt_burst:done +swa:applying 15 snapshots, blending with EMA (0.5/0.5) +DIAGNOSTIC post_ema val_loss:1.9211 val_bpb:1.1378 eval_time:1993ms +Serialized model: 106443805 bytes +Code size: 95390 bytes +gptq:collecting hessians batches=256 +gptq:hessians collected layers=68 time=37.2s +gptq:pre_prune artifact=15843400 target=15854610 +Saved quantized model to final_int6_model.pt +Serialized model int6+lzma: 15843400 bytes +Total submission size int6+lzma: 15938790 bytes +final_int6_roundtrip val_loss:1.9275 val_bpb:1.1416 eval_time:6457ms +final_int6_roundtrip_exact val_loss:1.92750067 val_bpb:1.14157447 +final_int6_sliding_window val_loss:1.8876 val_bpb:1.1179 stride:64 eval_time:74943ms +final_int6_sliding_window_exact val_loss:1.88755846 val_bpb:1.11792140 +final_int8_zlib_roundtrip_exact val_loss:1.88755846 val_bpb:1.11792140 +ttt:start lr=0.002 epochs=3 chunks=131072 +ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 +ttt_sliding:params unfrozen=27124848 frozen=0 + ttt_chunk [1/474] bpb=1.194589 time=0.7s + ttt_chunk [11/474] bpb=1.114842 time=6.1s + ttt_chunk [21/474] bpb=1.111008 time=11.4s + ttt_chunk [31/474] bpb=1.110006 time=16.8s + ttt_chunk [41/474] bpb=1.118383 time=22.1s + ttt_chunk [51/474] bpb=1.126049 time=27.5s + ttt_chunk [61/474] bpb=1.123955 time=33.0s + ttt_chunk [71/474] bpb=1.125390 time=38.5s +W0325 06:53:35.253000 330566 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers +W0325 06:53:35.257000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330634 closing signal SIGTERM +W0325 06:53:35.258000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330635 closing signal SIGTERM +W0325 06:53:35.259000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330636 closing signal SIGTERM +W0325 06:53:35.260000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330637 closing signal SIGTERM +W0325 06:53:35.261000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330638 closing signal SIGTERM +W0325 06:53:35.262000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330639 closing signal SIGTERM +W0325 06:53:35.264000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330640 closing signal SIGTERM +W0325 06:53:35.265000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330641 closing signal SIGTERM +Traceback (most recent call last): + File "/usr/local/bin/torchrun", line 7, in + sys.exit(main()) + ^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper + return f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main + run(args) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run + elastic_launch( + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ + return launch_agent(self._config, self._entrypoint, list(args)) + ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent + result = agent.run() + ^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper + result = f(*args, **kwargs) + ^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run + result = self._invoke_run(role) + ^^^^^^^^^^^^^^^^^^^^^^ + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run + time.sleep(monitor_interval) + File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler + raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) +torch.distributed.elastic.multiprocessing.api.SignalException: Process 330566 got signal: 15 From e62706959e90acda70870248dc84eb5a7fefea8b Mon Sep 17 00:00:00 2001 From: ethan Date: Wed, 25 Mar 2026 15:10:36 +0800 Subject: [PATCH 3/5] Remove old submission folder from PR diff --- .../README.md | 75 - .../submission.json | 22 - .../train_gpt.py | 2273 ----------------- .../train_seed1337.log | 172 -- .../train_seed42.log | 172 -- .../train_seed7.log | 172 -- 6 files changed, 2886 deletions(-) delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log delete mode 100644 records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md deleted file mode 100644 index fc738e0a9..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/README.md +++ /dev/null @@ -1,75 +0,0 @@ -# Record: int5 GPTQ + 33.6M model + Soft-Round QAT + Legal Score-First TTT - -## Summary - -**3-seed mean val_bpb = 1.1162 (std 0.0006)** - -int5 GPTQ quantization (values in [-15, 15], 31 unique levels) with Hessian-aware error compensation enables a 33.6M parameter model to fit under 16MB. Soft-Round QAT replaces STE hard rounding with differentiable tanh-based rounding (alpha annealing 1→16) for better training quality at zero cost. Combined with early QAT at threshold 0.5, EMA 0.997, and legal score-first AdamW TTT with cosine LR decay across chunks. - -## Key Innovations - -1. **int5 quantization** — 31 unique values ([-15,15]) stored as int8, ~0.46 bytes/param after zstd. Lower entropy = better compression ratio than int6. -2. **GPTQ error compensation** — Hessian-aware column reordering + Cholesky error redistribution. 256-sample calibration on training data. -3. **33.6M param model** — MHA 8/8 (full attention), BigramHash 8192, MLP 3.5x (1792), enabled by int5 compression. -4. **Soft-Round QAT** — Differentiable rounding `s_α(y) = floor(y) + 0.5 * tanh(α·r) / tanh(α/2) + 0.5` replaces STE. Alpha anneals from 1→16 during QAT steps. Better gradient flow = better training quality at zero computational cost. -5. **Early QAT 0.5** — QAT clipping matched to int5 range (0.9995 percentile / 15.0), ~1750 QAT steps. -6. **EMA 0.997** — Exponential moving average of weights, tuned from 0.9985. -7. **Legal score-first TTT** — every token scored BEFORE any gradient update using it. Cosine LR decay across chunks. - -## Architecture - -- 11 layers, model_dim=512, 8 heads / 8 KV heads (MHA), MLP 3.5x relu² -- XSA on all 11 layers -- Partial RoPE 16/64, LN Scale (1/√(layer+1)) -- SmearGate + OrthoInit -- BigramHash 8192, Shared VE128 (layers 9,10) -- Tight SWA (every 50) + EMA 0.997 -- Muon lr=0.025, WD=0.04 -- FA3 Hopper, ~98ms/step → ~6120 steps in 600s -- **33.6M params**, int5 GPTQ + zstd-22, 2% magnitude pruning - -## Quantization Pipeline - -1. **Early QAT** (threshold 0.5): QAT-aware training with int5 clipping (scale = row_clip / 15.0, clamp [-16, 15]) -2. **GPTQ** (post-training): 256-sample Hessian calibration, per-row optimal scales (5-percentile search), column reordering by Hessian diagonal, block-128 Cholesky error compensation -3. **int5 quantization** (range [-15, 15], 31 levels) stored as int8 -4. **zstd-22** compression -5. **2% magnitude pruning** - -## Legal Score-First TTT - -- Val data split into 131072-token chunks (474 chunks) -- For each chunk: **score first** (sliding window stride=32, inference_mode), **then** adapt -- AdamW (lr=0.0001, wd=0.0), 3 epochs per chunk, cosine LR across chunks -- Last 2 blocks + norms + lm_head unfrozen (~5.8M / 33.6M params) -- Last chunk never trained on -- Every token scored BEFORE any gradient update using it -- Manual grad all_reduce (no DDP wrapper) - -## Results - -| Seed | TTT BPB | Artifact | -|------|---------|----------| -| 1337 | **1.1155** | 15,822,078 bytes | -| 42 | **1.1163** | 15,415,405 bytes | -| 7 | **1.1167** | 15,368,627 bytes | -| **Mean** | **1.1162** | | -| **Std** | **0.0006** | | - -## Reproduction - -```bash -# On 8xH100 SXM: -pip install --break-system-packages zstandard -# Build FA3 Hopper (see repo README for instructions) -python3 data/cached_challenge_fineweb.py --variant sp1024 --train-shards 80 - -SEED=1337 SKIP_SLIDING=1 PRUNE_PCT=0.02 \ -SOFT_ROUND_QAT=1 \ -TTT_EPOCHS=3 TTT_LR=0.0001 TTT_OPTIMIZER=adamw \ -TTT_FREEZE_BLOCKS=2 TTT_CHUNK_TOKENS=131072 \ -TTT_TEMPERATURE=0.98 INT6_LAST_N=0 \ -PPM_ALPHA=1.0 BYTE_WEIGHTED_TTT=0 USE_CACHE=0 \ -ADAPTIVE_LR=0 USE_MIXER=0 \ -torchrun --standalone --nproc_per_node=8 train_gpt.py -``` diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json deleted file mode 100644 index 8edbcae58..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/submission.json +++ /dev/null @@ -1,22 +0,0 @@ -{ - "author": "Ethan Yang", - "github_id": "EthanYangTW", - "name": "Record: int5 GPTQ + 33.6M model + Soft-Round QAT + Legal Score-First TTT", - "blurb": "int5 GPTQ quantization ([-15,15], 31 levels) with Hessian-aware error compensation enables 33.6M params in 16MB. Soft-Round QAT (differentiable tanh rounding, alpha 1→16) replaces STE for better training quality. MHA 8/8, BigramHash 8192, MLP 3.5x (1792), XSA all 11 layers, Early QAT 0.5, EMA 0.997, legal score-first AdamW TTT with cosine LR decay.", - "date": "2026-03-24T00:00:00Z", - "val_bpb": 1.1162, - "val_bpb_std": 0.0006, - "val_loss_seed1337": 1.88347869, - "val_bpb_seed1337": 1.11550587, - "val_loss_seed42": 1.88480123, - "val_bpb_seed42": 1.11628915, - "val_loss_seed7": 1.88543543, - "val_bpb_seed7": 1.11666477, - "bytes_seed1337": 15822078, - "bytes_seed42": 15415405, - "bytes_seed7": 15368627, - "model_params": 33580124, - "quantization": "int5 GPTQ ([-15,15], 31 levels) + Soft-Round QAT", - "compression": "zstd-22", - "ttt": "legal score-first AdamW, 3 epochs, cosine LR across chunks" -} diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py deleted file mode 100644 index 757c51358..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_gpt.py +++ /dev/null @@ -1,2273 +0,0 @@ -"""V25: LeakyReLU^2 + TempCal + Mixed int5/int6 + 33.6M model.""" -from __future__ import annotations -import copy -import glob -import io -import math -import os -import random -import subprocess -import sys -import time -import uuid -import zlib -from pathlib import Path -try: - import zstandard - _COMPRESSOR = "zstd" -except ImportError: - _COMPRESSOR = "zlib" -import numpy as np -import sentencepiece as spm -import torch -import torch.distributed as dist -import torch.nn.functional as F -from torch import Tensor, nn -from torch.nn.parallel import DistributedDataParallel as DDP -try: - from flash_attn_interface import flash_attn_func as flash_attn_3_func - _HAS_FA3 = True -except ImportError: - try: - from flash_attn import flash_attn_func as flash_attn_3_func - _HAS_FA3 = True - except ImportError: - _HAS_FA3 = False - flash_attn_3_func = None - -# ===================== PPM N-gram Model ===================== -import collections - -class PPMModel: - """Prediction by Partial Matching — online n-gram model. - - Builds token-level n-gram statistics from already-scored validation tokens. - At prediction time, backs off from order K down to order 0 (unigram), - blending probabilities via escape mechanism. - - This is 100% legal: it only uses tokens that have already been scored. - """ - - def __init__(self, max_order: int = 6, vocab_size: int = 1024): - self.max_order = max_order - self.vocab_size = vocab_size - # counts[order][context_tuple] = Counter of next tokens - self.counts: list[dict[tuple, collections.Counter]] = [ - {} for _ in range(max_order + 1) - ] - self.total_tokens = 0 - - def update(self, tokens): - """Add observed tokens to the model. tokens = 1D list/array of token IDs.""" - tokens = list(tokens) - self.total_tokens += len(tokens) - for i, tok in enumerate(tokens): - for order in range(min(i, self.max_order) + 1): - ctx = tuple(tokens[i - order:i]) if order > 0 else () - if ctx not in self.counts[order]: - self.counts[order][ctx] = collections.Counter() - self.counts[order][ctx][tok] += 1 - - def predict_probs(self, context_tokens, device=None): - """Return probability distribution over vocab given context. - - Uses PPM Method C escape mechanism: - At each order, allocate escape probability = num_unique / (num_unique + total_count) - and distribute remaining probability proportional to counts. - """ - import torch - context = list(context_tokens) - - # Start with uniform distribution (order -1) - probs = torch.ones(self.vocab_size, dtype=torch.float32) / self.vocab_size - if device is not None: - probs = probs.to(device) - - # Blend from lowest order up to highest - for order in range(min(len(context), self.max_order) + 1): - ctx = tuple(context[-order:]) if order > 0 else () - if ctx in self.counts[order]: - counter = self.counts[order][ctx] - total = sum(counter.values()) - unique = len(counter) - # Escape probability (PPM Method C) - escape = unique / (unique + total) - # Build distribution for this order - order_probs = torch.zeros(self.vocab_size, dtype=torch.float32) - if device is not None: - order_probs = order_probs.to(device) - for tok, cnt in counter.items(): - if tok < self.vocab_size: - order_probs[tok] = cnt / total - # Blend: (1-escape) * order_probs + escape * lower_order_probs - probs = (1.0 - escape) * order_probs + escape * probs - - return probs - - def predict_batch(self, context_batch, targets, device=None): - """Compute log-probs for a batch of (context, target) pairs. - - context_batch: list of token lists (each is the context for one prediction) - targets: tensor of target token IDs - - Returns tensor of log-probabilities (same shape as targets). - """ - import torch - log_probs = torch.zeros(len(targets), dtype=torch.float32) - if device is not None: - log_probs = log_probs.to(device) - - for i, (ctx, tgt) in enumerate(zip(context_batch, targets)): - probs = self.predict_probs(ctx, device=device) - log_probs[i] = torch.log(probs[tgt] + 1e-10) - - return log_probs - - -class FastPPMModel: - """Faster PPM using batch numpy operations and hash-based lookup. - - Trades memory for speed. Maintains only orders 0-4 for tractability. - Uses a single forward pass over scored tokens to update. - """ - - def __init__(self, max_order: int = 4, vocab_size: int = 1024): - self.max_order = max_order - self.vocab_size = vocab_size - # For each order, store: hash(context) -> array of counts - self.counts = [{} for _ in range(max_order + 1)] - self.total_tokens = 0 - self._unigram = None # cached unigram distribution - - def update_chunk(self, tokens): - """Update with a chunk of tokens. tokens = list or 1D numpy/torch array.""" - if hasattr(tokens, 'cpu'): - tokens = tokens.cpu().tolist() - elif hasattr(tokens, 'tolist'): - tokens = tokens.tolist() - - n = len(tokens) - self.total_tokens += n - - for i in range(n): - tok = tokens[i] - for order in range(min(i, self.max_order) + 1): - if order == 0: - ctx_key = 0 # empty context - else: - ctx_key = hash(tuple(tokens[i-order:i])) - - if ctx_key not in self.counts[order]: - self.counts[order][ctx_key] = {} - d = self.counts[order][ctx_key] - d[tok] = d.get(tok, 0) + 1 - - self._unigram = None # invalidate cache - - def score_sequence(self, tokens, start_pos=0): - """Score a sequence, returning NLL for each position from start_pos. - - Returns list of -log2(prob) for each token (bits, not nats). - """ - import math - if hasattr(tokens, 'cpu'): - tokens = tokens.cpu().tolist() - elif hasattr(tokens, 'tolist'): - tokens = tokens.tolist() - - scores = [] - for i in range(start_pos, len(tokens)): - prob = self._predict_one(tokens, i) - scores.append(-math.log2(max(prob, 1e-10))) - return scores - - def get_log_probs_tensor(self, tokens, start_pos, device): - """Get log probabilities as a tensor for interpolation with neural model.""" - import torch, math - if hasattr(tokens, 'cpu'): - tokens = tokens.cpu().tolist() - elif hasattr(tokens, 'tolist'): - tokens = tokens.tolist() - - n = len(tokens) - start_pos - log_probs = torch.zeros(n, dtype=torch.float32, device=device) - - for i in range(start_pos, len(tokens)): - prob = self._predict_one(tokens, i) - log_probs[i - start_pos] = math.log(max(prob, 1e-10)) - - return log_probs - - def _predict_one(self, tokens, pos): - """Predict probability of tokens[pos] given tokens[:pos].""" - target = tokens[pos] - - # Start with uniform - prob = 1.0 / self.vocab_size - - for order in range(min(pos, self.max_order) + 1): - if order == 0: - ctx_key = 0 - else: - ctx_key = hash(tuple(tokens[pos-order:pos])) - - if ctx_key in self.counts[order]: - d = self.counts[order][ctx_key] - total = sum(d.values()) - unique = len(d) - escape = unique / (unique + total) - - count = d.get(target, 0) - if count > 0: - order_prob = count / total - prob = (1.0 - escape) * order_prob + escape * prob - else: - prob = escape * prob - - return prob - - - - -class ExactMatchCache: - """Hash-based exact-match n-gram cache. - - Stores (context_hash → Counter of next tokens) from already-scored tokens. - For repeated patterns in val data, gives near-perfect predictions. - """ - - def __init__(self, orders=(3, 4, 5, 6, 7, 8), vocab_size=1024): - self.orders = orders - self.vocab_size = vocab_size - # For each order: hash(context) -> {token: count} - self.tables = {o: {} for o in orders} - self.total_tokens = 0 - - def update_chunk(self, tokens): - """Add a chunk of tokens to the cache.""" - if hasattr(tokens, 'cpu'): - tokens = tokens.cpu().tolist() - elif hasattr(tokens, 'tolist'): - tokens = tokens.tolist() - - n = len(tokens) - self.total_tokens += n - - for i in range(n): - tok = tokens[i] - for order in self.orders: - if i >= order: - ctx = hash(tuple(tokens[i-order:i])) - if ctx not in self.tables[order]: - self.tables[order][ctx] = {} - d = self.tables[order][ctx] - d[tok] = d.get(tok, 0) + 1 - - def predict_one(self, tokens, pos): - """Get probability of tokens[pos] given exact context matches. - - Returns (prob, matched_order) or (None, -1) if no match. - Uses highest-order match available. - """ - target = tokens[pos] - - # Try highest order first - for order in sorted(self.orders, reverse=True): - if pos >= order: - ctx = hash(tuple(tokens[pos-order:pos])) - if ctx in self.tables[order]: - d = self.tables[order][ctx] - total = sum(d.values()) - if total >= 2: # require at least 2 observations - prob = d.get(target, 0) / total - return prob, order - - return None, -1 - - def get_interpolation_nll(self, tokens, pos, neural_nll, alpha_cache=0.3): - """Interpolate cache prediction with neural model NLL. - - Args: - tokens: full token sequence (list) - pos: position to predict - neural_nll: neural model NLL for this position (float) - alpha_cache: weight for cache (0.3 = 30% cache, 70% neural) - - Returns: interpolated NLL - """ - import math - cache_prob, order = self.predict_one(tokens, pos) - - if cache_prob is not None and order >= 4: - neural_prob = math.exp(-neural_nll) - # Higher weight for longer matches - weight = min(alpha_cache * (order / max(self.orders)), 0.5) - mixed_prob = (1 - weight) * neural_prob + weight * cache_prob - return -math.log(max(mixed_prob, 1e-10)) - - return neural_nll - - -# ===================== Interpolation Helper ===================== - -def interpolate_with_ppm(neural_logits, ppm_model, tokens, window_start, seq_len, - stride, alpha=0.85, device=None): - """Interpolate neural model logits with PPM predictions. - - Args: - neural_logits: (batch, seq_len, vocab) from neural model - ppm_model: FastPPMModel instance - tokens: full val_tokens tensor - window_start: starting position of this window - seq_len: sequence length - stride: stride for scoring - alpha: weight for neural model (1-alpha for PPM) - device: torch device - - Returns: - interpolated NLL values for scored positions - """ - import torch - # For now, just return neural logits — PPM interpolation happens at the NLL level - # We compute PPM log-probs and do log-space interpolation - pass - - -class LogisticContextMixer: - """GPU-vectorized logistic context mixing (inspired by PAQ compression). - - Maintains GPU-resident n-gram count tables and learns online mixing weights - using the Hedge/multiplicative-weights algorithm. All operations are batched - tensor ops — no Python per-token loops. - - Experts: - 0: Neural model (logits passed in) - 1: Unigram frequencies from scored tokens - 2: Bigram frequencies (prev_token → next_token) - """ - - def __init__(self, vocab_size: int = 1024, device: str = 'cuda', eta: float = 0.1): - self.V = vocab_size - self.device = device - self.eta = eta # Hedge learning rate - self.K = 3 # number of experts - - # Expert weights (log-domain for numerical stability) - self.log_weights = torch.zeros(self.K, device=device) - - # N-gram count tables (GPU-resident) - self.uni_counts = torch.zeros(vocab_size, device=device) - self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device) - self.total_tokens = 0 - - def update(self, tokens): - """Update n-gram tables with newly scored tokens. Fully vectorized.""" - if hasattr(tokens, 'cpu'): - t = tokens.to(self.device).long() - else: - t = torch.tensor(tokens, device=self.device, dtype=torch.long) - - n = t.numel() - if n == 0: - return - self.total_tokens += n - - # Unigram: bincount - self.uni_counts.scatter_add_(0, t, torch.ones(n, device=self.device)) - - # Bigram: scatter_add into [V, V] table - if n >= 2: - ctx = t[:-1] - nxt = t[1:] - bi_idx = ctx * self.V + nxt - flat = torch.zeros(self.V * self.V, device=self.device) - flat.scatter_add_(0, bi_idx, torch.ones(n - 1, device=self.device)) - self.bi_counts += flat.reshape(self.V, self.V) - - def get_expert_log_probs(self, neural_logits, x_batch, y_batch, wlens): - """Get log-probability of targets from each expert. All GPU-vectorized. - - Args: - neural_logits: [bsz, seq_len, V] neural model logits - x_batch: [bsz, seq_len] input tokens (context) - y_batch: [bsz, seq_len] target tokens - wlens: list of actual lengths per sequence - - Returns: - expert_nll: [bsz, seq_len, K] NLL from each expert - """ - bsz, slen, V = neural_logits.shape - - # Expert 0: Neural model - neural_lp = F.log_softmax(neural_logits, dim=-1) - neural_nll = -neural_lp.gather(2, y_batch.unsqueeze(2)).squeeze(2) # [bsz, slen] - - # Expert 1: Unigram - uni_total = self.uni_counts.sum() - if uni_total > 0: - uni_probs = (self.uni_counts + 0.1) / (uni_total + 0.1 * self.V) - uni_lp = uni_probs.log() - uni_nll = -uni_lp[y_batch] # [bsz, slen] - else: - uni_nll = torch.full((bsz, slen), math.log(self.V), device=self.device) - - # Expert 2: Bigram P(next | prev) - bi_total = self.bi_counts.sum(dim=1, keepdim=True) # [V, 1] - if bi_total.sum() > 0: - bi_probs = (self.bi_counts + 0.1) / (bi_total + 0.1 * self.V) # [V, V] - bi_lp = bi_probs.log() - # Lookup: for each position, prev=x_batch, next=y_batch - prev_flat = x_batch.reshape(-1) # [bsz*slen] - next_flat = y_batch.reshape(-1) - bi_nll_flat = -bi_lp[prev_flat, next_flat] - bi_nll = bi_nll_flat.reshape(bsz, slen) - else: - bi_nll = torch.full((bsz, slen), math.log(self.V), device=self.device) - - # Stack: [bsz, slen, K] - return torch.stack([neural_nll, uni_nll, bi_nll], dim=-1) - - def mix_and_score(self, neural_logits, x_batch, y_batch, wlens): - """Compute mixed NLL using current expert weights. Returns [bsz, slen] NLL. - - Uses log-domain mixing: NLL_mixed = -log(sum_k w_k * exp(-NLL_k)) - """ - if self.total_tokens < 10000: - # Not enough data for n-grams — just use neural - return F.cross_entropy( - neural_logits.reshape(-1, neural_logits.size(-1)), - y_batch.reshape(-1), reduction="none" - ).reshape(neural_logits.shape[0], neural_logits.shape[1]) - - expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] - - # Log-domain mixing: log(sum_k w_k * p_k) = logsumexp(log_w_k + log_p_k) - log_w = self.log_weights - self.log_weights.logsumexp(0) # normalize - # expert_lp = -expert_nll [bsz, slen, K] - mixed_lp = (-expert_nll + log_w.unsqueeze(0).unsqueeze(0)).logsumexp(dim=-1) # [bsz, slen] - - return -mixed_lp # mixed NLL - - def update_weights(self, neural_logits, x_batch, y_batch, wlens): - """Update expert weights using Hedge algorithm on this batch's losses.""" - if self.total_tokens < 10000: - return - - with torch.no_grad(): - expert_nll = self.get_expert_log_probs(neural_logits, x_batch, y_batch, wlens) # [bsz, slen, K] - - # Mean loss per expert across valid positions - mask = torch.zeros(expert_nll.shape[0], expert_nll.shape[1], device=self.device) - for i, wl in enumerate(wlens): - mask[i, :wl] = 1.0 - - # Masked mean NLL per expert - masked_nll = expert_nll * mask.unsqueeze(-1) - expert_mean_loss = masked_nll.sum(dim=(0, 1)) / mask.sum().clamp(min=1) # [K] - - # Hedge update: log_w -= eta * loss - self.log_weights -= self.eta * expert_mean_loss - - -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", 3500)) - warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) - train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) - train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) - eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) - max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) - qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) - vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) - num_layers = int(os.environ.get("NUM_LAYERS", 11)) - num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 8)) - model_dim = int(os.environ.get("MODEL_DIM", 512)) - num_heads = int(os.environ.get("NUM_HEADS", 8)) - mlp_mult = float(os.environ.get("MLP_MULT", 3.5)) - 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", 0.025)) - scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) - muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) - muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) - muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) - muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) - 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", 32)) - int6_last_n = int(os.environ.get("INT6_LAST_N", 2)) # last N layers use int6, rest use int5 - ttt_temperature = float(os.environ.get("TTT_TEMPERATURE", 0.98)) # post-TTT temperature calibration - 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", 8192)) - 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") - prune_pct = float(os.environ.get("PRUNE_PCT", 0.02)) - -def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: - a, b, c = (3.4445, -4.7750, 2.0315) - X = G.bfloat16() - X /= X.norm() + eps - transposed = G.size(0) > G.size(1) - if transposed: - X = X.T - for _ in range(steps): - A = X @ X.T - B = b * A + c * A @ A - X = a * X + B @ X - return X.T if transposed else X - -class Muon(torch.optim.Optimizer): - def __init__(self, params, lr: float, momentum: float, backend_steps: int, - nesterov: bool = True, weight_decay: float = 0.0): - super().__init__( - params, - dict(lr=lr, momentum=momentum, backend_steps=backend_steps, - nesterov=nesterov, weight_decay=weight_decay), - ) - @torch.no_grad() - def step(self, closure=None): - loss = None - if closure is not None: - with torch.enable_grad(): - loss = closure() - distributed = dist.is_available() and dist.is_initialized() - world_size = dist.get_world_size() if distributed else 1 - rank = dist.get_rank() if distributed else 0 - for group in self.param_groups: - params = group["params"] - if not params: - continue - lr = group["lr"] - momentum = group["momentum"] - backend_steps = group["backend_steps"] - nesterov = group["nesterov"] - total_params = sum(int(p.numel()) for p in params) - updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) - curr = 0 - for i, p in enumerate(params): - if i % world_size == rank and p.grad is not None: - g = p.grad - state = self.state[p] - if "momentum_buffer" not in state: - state["momentum_buffer"] = torch.zeros_like(g) - buf = state["momentum_buffer"] - buf.mul_(momentum).add_(g) - if nesterov: - g = g.add(buf, alpha=momentum) - g = zeropower_via_newtonschulz5(g, steps=backend_steps) - g *= max(1, g.size(0) / g.size(1)) ** 0.5 - updates_flat[curr : curr + p.numel()] = g.reshape(-1) - curr += p.numel() - if distributed: - dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) - wd = group.get("weight_decay", 0.0) - curr = 0 - for p in params: - if wd > 0.0: - p.data.mul_(1.0 - lr * wd) - g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) - p.add_(g, alpha=-lr) - curr += p.numel() - return loss - -def build_sentencepiece_luts( - sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device -) -> tuple[Tensor, Tensor, Tensor]: - sp_vocab_size = int(sp.vocab_size()) - table_size = max(sp_vocab_size, vocab_size) - base_bytes_np = np.zeros((table_size,), dtype=np.int16) - has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) - is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) - for token_id in range(sp_vocab_size): - if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): - continue - is_boundary_token_np[token_id] = False - if sp.is_byte(token_id): - base_bytes_np[token_id] = 1 - continue - piece = sp.id_to_piece(token_id) - if piece.startswith("▁"): - has_leading_space_np[token_id] = True - piece = piece[1:] - base_bytes_np[token_id] = len(piece.encode("utf-8")) - return ( - torch.tensor(base_bytes_np, dtype=torch.int16, device=device), - torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), - torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), - ) - -def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: - files = [Path(p) for p in sorted(glob.glob(pattern))] - if not files: - raise FileNotFoundError(f"No files found for pattern: {pattern}") - 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) -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", - ).split(",") - if pattern -) -INT8_PER_ROW_SCALE_DTYPE = torch.float16 -INT8_CLIP_Q = 0.9999984 - -def tensor_nbytes(t: Tensor) -> int: - return int(t.numel()) * int(t.element_size()) - -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 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) - -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 - _soft_round_alpha: float = 1.0 # temperature for soft-round (annealed during training) - _use_soft_round: bool = False # enable soft-round QAT instead of STE - - def __init__(self, *args, **kwargs): - super().__init__(*args, **kwargs) - self._clip_range = 15 # default int5, set to 31 for int6 layers - - @staticmethod - def soft_round(y: Tensor, alpha: float) -> Tensor: - """Differentiable approximation to round() from Agustsson & Theis (NeurIPS 2020). - s_alpha(y) = floor(y) + 0.5 * tanh(alpha * r) / tanh(alpha/2) + 0.5 - where r = y - floor(y) - 0.5 (centered fractional part) - """ - fl = torch.floor(y) - r = y - fl - 0.5 - return fl + 0.5 * torch.tanh(alpha * r) / (math.tanh(alpha / 2) + 1e-10) + 0.5 - - def forward(self, x: Tensor) -> Tensor: - w = self.weight.to(x.dtype) - cr = self._clip_range - if CastedLinear._qat_enabled and self.training and w.ndim == 2: - if CastedLinear._use_soft_round: - # Soft-Round QAT: differentiable rounding with temperature annealing - w32 = self.weight.float() - row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) - scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) - w_scaled = w32 / scale[:, None] - w_rounded = CastedLinear.soft_round(w_scaled, CastedLinear._soft_round_alpha) - w_q = (torch.clamp(w_rounded, -(cr+1), cr) * scale[:, None]).to(x.dtype) - w = w_q # fully differentiable path - else: - # Original STE QAT - with torch.no_grad(): - w32 = self.weight.float() - row_clip = torch.quantile(w32.abs(), 0.9995, dim=1) - scale = (row_clip / float(cr)).clamp_min(1.0 / float(cr)) - w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -(cr+1), cr) * 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): - super().__init__() - if dim % num_heads != 0: - raise ValueError("model_dim must be divisible by num_heads") - if num_heads % num_kv_heads != 0: - raise ValueError("num_heads must be divisible by num_kv_heads") - self.num_heads = num_heads - self.num_kv_heads = num_kv_heads - self.head_dim = dim // num_heads - if self.head_dim % 2 != 0: - raise ValueError("head_dim must be even for RoPE") - kv_dim = self.num_kv_heads * self.head_dim - self.c_q = CastedLinear(dim, dim, bias=False) - self.c_k = CastedLinear(dim, kv_dim, bias=False) - self.c_v = CastedLinear(dim, kv_dim, bias=False) - self.proj = CastedLinear(dim, dim, bias=False) - self.proj._zero_init = True - self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) - self.rope_dims = 0 - self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) - self.use_xsa = False - - def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: - B, T, H, D = y.shape - Hkv = v.size(-2) - y_g = y.reshape(B, T, Hkv, H // Hkv, D) - vn = F.normalize(v, dim=-1).unsqueeze(-2) - proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn - return (y_g - proj).reshape(B, T, H, D) - - def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor: - bsz, seqlen, dim = x.shape - q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim) - k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) - v = self.c_v(x) - if v_embed is not None: - v = v + v_embed - v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) - q = F.rms_norm(q, (q.size(-1),)) - k = F.rms_norm(k, (k.size(-1),)) - cos, sin = self.rotary(seqlen, x.device, q.dtype) - q = apply_rotary_emb(q, cos, sin, 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_FA3: - y = flash_attn_3_func(q, k, v, causal=True).contiguous() - else: - y = F.scaled_dot_product_attention( - q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2), - attn_mask=None, is_causal=True, - enable_gqa=(self.num_kv_heads != self.num_heads), - ).transpose(1, 2) - if self.use_xsa: - y = self._xsa_efficient(y, v) - y = y.reshape(bsz, seqlen, dim) - return self.proj(y) - -class SmearGate(nn.Module): - def __init__(self, dim: int): - super().__init__() - self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) - - def forward(self, x: Tensor) -> Tensor: - g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] - x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) - return (1 - g) * x + g * x_prev - -class BigramHashEmbedding(nn.Module): - def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): - super().__init__() - self.bigram_vocab_size = bigram_vocab_size - self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) - nn.init.zeros_(self.embed.weight) - self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None - if self.proj is not None: - nn.init.zeros_(self.proj.weight) - self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) - - def bigram_hash(self, tokens: Tensor) -> Tensor: - t = tokens.to(torch.int32) - mod = self.bigram_vocab_size - 1 - out = torch.empty_like(t) - out[..., 0] = mod - out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod - return out.long() - - def forward(self, token_ids: Tensor) -> Tensor: - h = self.embed(self.bigram_hash(token_ids)) - if self.proj is not None: - h = self.proj(h) - return h * self.scale.to(dtype=h.dtype) - -class ValueEmbedding(nn.Module): - def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): - super().__init__() - self.embed = nn.Embedding(vocab_size, ve_dim) - nn.init.normal_(self.embed.weight, std=0.01) - self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None - if self.proj is not None: - nn.init.zeros_(self.proj.weight) - self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) - - def forward(self, token_ids: Tensor) -> Tensor: - h = self.embed(token_ids) - if self.proj is not None: - h = self.proj(h) - return h * self.scale.to(dtype=h.dtype) - -class MLP(nn.Module): - def __init__(self, dim: int, mlp_mult: int): - super().__init__() - hidden = int(mlp_mult * dim) - self.fc = CastedLinear(dim, hidden, bias=False) - self.proj = CastedLinear(hidden, dim, bias=False) - self.proj._zero_init = True - - def forward(self, x: Tensor) -> Tensor: - return self.proj(F.leaky_relu(self.fc(x), negative_slope=0.5).square()) - -class Block(nn.Module): - def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: int, - rope_base: float, qk_gain_init: float, layer_idx: int = 0, - ln_scale: bool = False, dtg: bool = False): - super().__init__() - self.attn_norm = RMSNorm() - self.mlp_norm = RMSNorm() - self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) - self.mlp = MLP(dim, mlp_mult) - self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) - self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) - self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) - self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 - 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, v_embed: Tensor | None = None) -> Tensor: - 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, v_embed=v_embed) - x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out - x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) - 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, - 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"): - 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.tok_emb = nn.Embedding(vocab_size, model_dim) - self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None - self.smear = SmearGate(model_dim) - self.num_encoder_layers = num_layers // 2 - self.num_decoder_layers = num_layers - self.num_encoder_layers - self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) - self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) - self.blocks = nn.ModuleList([ - Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, - qk_gain_init, layer_idx=i, ln_scale=ln_scale, dtg=dtg) - for i in range(num_layers) - ]) - if rope_dims > 0: - head_dim = model_dim // num_heads - for block in self.blocks: - block.attn.rope_dims = rope_dims - block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) - self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] - kv_dim = self._ve_target_dim - if self.ve_layer_indices: - self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) - self.ve_layer_scales = nn.ParameterList( - [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] - ) - else: - self.ve_shared = None - self.ve_layer_scales = nn.ParameterList() - self.value_embeds = nn.ModuleList() - 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 - if xsa_last_n > 0: - for i in range(max(0, num_layers - xsa_last_n), num_layers): - self.blocks[i].attn.use_xsa = True - self._init_weights() - - def _init_weights(self) -> None: - if self.tie_embeddings: - nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) - num_layers = len(self.blocks) - for name, module in self.named_modules(): - if isinstance(module, nn.Linear): - if getattr(module, "_zero_init", False): - nn.init.zeros_(module.weight) - elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: - nn.init.orthogonal_(module.weight, gain=1.0) - if ".proj." in name or name.endswith(".proj"): - with torch.no_grad(): - module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) - - def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: - if self.ve_shared is None or layer_idx not in self.ve_layer_indices: - return None - if ve_cache is not None and 've' not in ve_cache: - ve_cache['ve'] = self.ve_shared(input_ids) - ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) - ve_idx = self.ve_layer_indices.index(layer_idx) - return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) - - def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: - x = self.tok_emb(input_ids) - if self.bigram is not None: - x = x + self.bigram(input_ids) - x = F.rms_norm(x, (x.size(-1),)) - x = self.smear(x) - x0 = x - skips: list[Tensor] = [] - ve_cache: dict = {} - for i in range(self.num_encoder_layers): - ve = self._get_ve(i, input_ids, ve_cache) - x = self.blocks[i](x, x0, v_embed=ve) - skips.append(x) - for i in range(self.num_decoder_layers): - bi = self.num_encoder_layers + i - if skips: - x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - ve = self._get_ve(bi, input_ids, ve_cache) - x = self.blocks[bi](x, x0, v_embed=ve) - x = self.final_norm(x) - x_flat = x.reshape(-1, x.size(-1)) - targets = target_ids.reshape(-1) - if self.tie_embeddings: - logits_proj = F.linear(x_flat, self.tok_emb.weight) - else: - if self.lm_head is None: - raise RuntimeError("lm_head is required when tie_embeddings=False") - logits_proj = self.lm_head(x_flat) - logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - return F.cross_entropy(logits.float(), targets, reduction="mean") - - def forward_logits(self, input_ids: Tensor) -> Tensor: - x = self.tok_emb(input_ids) - if self.bigram is not None: - x = x + self.bigram(input_ids) - x = F.rms_norm(x, (x.size(-1),)) - x = self.smear(x) - x0 = x - skips: list[Tensor] = [] - ve_cache: dict = {} - for i in range(self.num_encoder_layers): - ve = self._get_ve(i, input_ids, ve_cache) - x = self.blocks[i](x, x0, v_embed=ve) - skips.append(x) - for i in range(self.num_decoder_layers): - bi = self.num_encoder_layers + i - if skips: - x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - ve = self._get_ve(bi, input_ids, ve_cache) - x = self.blocks[bi](x, x0, v_embed=ve) - x = self.final_norm(x) - if self.tie_embeddings: - logits_proj = F.linear(x, self.tok_emb.weight) - else: - logits_proj = self.lm_head(x) - return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - -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() - compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) - - # Pre-compile: dummy forward+backward with TTT shapes to warm the compile cache - if rank == 0: - print(" ttt: pre-compiling forward+backward kernels...", flush=True) - _dummy_x = torch.zeros(1, seq_len, dtype=torch.int64, device=device) - _dummy_y = torch.zeros(1, seq_len, dtype=torch.int64, device=device) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - _dummy_logits = base_model.forward_logits(_dummy_x) - _dummy_loss = F.cross_entropy(_dummy_logits.reshape(-1, _dummy_logits.size(-1)), _dummy_y.reshape(-1)) - _dummy_loss.backward() - base_model.zero_grad(set_to_none=True) - if rank == 0: - print(" ttt: pre-compile done", flush=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, ttt_epochs: int = 3, ttt_lr: float = 0.001, - ttt_momentum: float = 0.9, ttt_freeze_blocks: int = 2, - batch_seqs: int = 32, eval_seq_len: int | None = None, - ttt_chunk_tokens: int = 32768, ttt_optimizer: str = "adamw", - ttt_temp: float = 1.0, - ppm_alpha: float = 0.85, - byte_weighted_ttt: bool = True, - use_cache: bool = True, - cache_alpha: float = 0.3, - adaptive_lr: bool = True, - adaptive_lr_max_mult: float = 3.0, -) -> tuple[float, float]: - """Legal score-first TTT: score each chunk, then train on it. - Every token scored BEFORE any update that could use it.""" - seq_len = eval_seq_len or args.train_seq_len - total_tokens = val_tokens.numel() - 1 - - # Initialize GPU-vectorized logistic context mixer - use_mixer = os.environ.get("USE_MIXER", "1") == "1" - mixer = LogisticContextMixer( - vocab_size=val_tokens.to(torch.int32).max().item() + 1, - device=device, - eta=float(os.environ.get("MIXER_ETA", "0.1")), - ) if use_mixer else None - if use_mixer and rank == 0: - print(f" Logistic context mixer enabled: eta={mixer.eta}") - if adaptive_lr and rank == 0: - print(f" Adaptive LR enabled: max_mult={adaptive_lr_max_mult}") - - # Pre-compute all window starts - window_starts = [ws for ws in range(0, total_tokens, stride) - if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] - - # Assign each window to a chunk based on scored token position - num_chunks = (total_tokens + ttt_chunk_tokens - 1) // ttt_chunk_tokens - chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] - for ws in window_starts: - end = min(ws + seq_len, total_tokens) - wlen = end - ws - s = 0 if ws == 0 else max(wlen - stride, 0) - scored_start = ws + s - ci = min(scored_start // ttt_chunk_tokens, num_chunks - 1) - chunk_windows[ci].append(ws) - - if rank == 0: - print(f"ttt:start chunks={num_chunks} chunk_tokens={ttt_chunk_tokens} " - f"windows={len(window_starts)} stride={stride} " - f"lr={ttt_lr} epochs={ttt_epochs} opt={ttt_optimizer} " - f"freeze_first={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) - - # Freeze everything, then selectively unfreeze for TTT - num_blocks = len(base_model.blocks) - for p in base_model.parameters(): - p.requires_grad_(False) - ttt_params = [] - ttt_param_ids = set() - use_qttt = os.environ.get("QTTT", "0") == "1" - if use_qttt: - # qTTT: only unfreeze Q projections in last N blocks + norms + head - for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): - for name, p in base_model.blocks[i].named_parameters(): - if "c_q" in name: - p.requires_grad_(True) - ttt_params.append(p) - ttt_param_ids.add(id(p)) - else: - # Standard: unfreeze all params in last N blocks - for i in range(max(0, num_blocks - ttt_freeze_blocks), num_blocks): - for p in base_model.blocks[i].parameters(): - p.requires_grad_(True) - ttt_params.append(p) - ttt_param_ids.add(id(p)) - # Unfreeze norms, scales, lm_head - for name, p in base_model.named_parameters(): - if "norm" in name or "scale" in name or "lm_head" in name: - p.requires_grad_(True) - if id(p) not in ttt_param_ids: - ttt_params.append(p) - ttt_param_ids.add(id(p)) - - if rank == 0: - n_unfrozen = sum(p.numel() for p in ttt_params) - n_frozen = sum(p.numel() for p in base_model.parameters() if not p.requires_grad) - print(f"ttt:params unfrozen={n_unfrozen} frozen={n_frozen}") - - if ttt_optimizer == "adamw": - optimizer = torch.optim.AdamW(ttt_params, lr=ttt_lr, weight_decay=0.0, betas=(0.9, 0.999)) - else: - optimizer = torch.optim.SGD(ttt_params, lr=ttt_lr, momentum=ttt_momentum) - - # Polyak averaging (TTT weight EMA) for smoother scoring - use_polyak = os.environ.get("USE_POLYAK", "1") == "1" - polyak_decay = float(os.environ.get("POLYAK_DECAY", "0.998")) - if use_polyak: - polyak_state = {id(p): p.data.clone() for p in ttt_params} - if rank == 0: - print(f" Polyak averaging enabled: decay={polyak_decay}") - - t0 = time.perf_counter() - - for ci in range(num_chunks): - windows = chunk_windows[ci] - if not windows: - continue - - # --- Phase 1: SCORE this chunk (inference_mode, no grad) --- - my_s = (len(windows) * rank) // world_size - my_e = (len(windows) * (rank + 1)) // world_size - my_windows = windows[my_s:my_e] - - # Swap in Polyak-averaged weights for scoring - if use_polyak and ci > 0: - _saved_weights = {} - for p in ttt_params: - _saved_weights[id(p)] = p.data.clone() - p.data.copy_(polyak_state[id(p)]) - - 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) - logits_scaled = logits.float() / ttt_temp - - # Adaptive temperature: sharpen confident predictions more - if ttt_temp != 1.0: - with torch.no_grad(): - probs_for_entropy = F.softmax(logits.float(), dim=-1) - token_entropy = -(probs_for_entropy * (probs_for_entropy + 1e-10).log()).sum(-1) - max_ent = math.log(logits.size(-1)) - # Confident tokens (low entropy) get more sharpening - adaptive_temp = 1.0 - (1.0 - ttt_temp) * (1.0 - token_entropy / max_ent) - adaptive_temp = adaptive_temp.clamp(min=0.9, max=1.05) - logits_scaled = logits.float() / adaptive_temp.unsqueeze(-1) - - # Logistic context mixing (GPU-vectorized) or plain CE - if mixer is not None: - nll = mixer.mix_and_score(logits_scaled, x_batch, y_batch, wlens) - mixer.update_weights(logits_scaled, x_batch, y_batch, wlens) - else: - nll = F.cross_entropy( - logits_scaled.reshape(-1, logits_scaled.size(-1)), - 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() - - # --- Update context mixer with scored chunk tokens (GPU-vectorized) --- - chunk_start_tok = ci * ttt_chunk_tokens - chunk_end_tok = min((ci + 1) * ttt_chunk_tokens, total_tokens) - if mixer is not None: - mixer.update(val_tokens[chunk_start_tok:chunk_end_tok + 1]) - - # Swap back training weights after scoring - if use_polyak and ci > 0: - for p in ttt_params: - p.data.copy_(_saved_weights[id(p)]) - - # --- Phase 2: TRAIN on this chunk (already scored = legal) --- - is_last_chunk = (ci == num_chunks - 1) - if not is_last_chunk and ttt_epochs > 0: - chunk_start = ci * ttt_chunk_tokens - chunk_end = min((ci + 1) * ttt_chunk_tokens, total_tokens) - chunk_seqs = (chunk_end - chunk_start) // seq_len - if rank == 0 and ci < 3: - print(f" ttt_train [{ci+1}] seqs={chunk_seqs} start_train...", flush=True) - if chunk_seqs > 0: - # Cosine LR across chunks + adaptive scaling - cos_lr = ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) - if adaptive_lr: - # Increase LR as we've seen more data (more confident adaptation) - progress = min(ci / max(num_chunks * 0.3, 1), 1.0) # ramp over first 30% of chunks - lr_mult = 1.0 + (adaptive_lr_max_mult - 1.0) * progress - cos_lr = cos_lr * lr_mult - 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(ttt_epochs): - if rank == 0 and ci < 3: - print(f" ttt_train [{ci+1}] epoch={_ep+1}/{ttt_epochs} batches={my_chunk_seqs} ...", flush=True) - for bs in range(0, my_chunk_seqs, batch_seqs): - be = min(bs + batch_seqs, my_chunk_seqs) - actual_bs = my_seq_s + bs - start_tok = chunk_start + actual_bs * seq_len - end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 - if end_tok > val_tokens.numel(): - continue - local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) - x = local[:-1].reshape(-1, seq_len) - y = local[1:].reshape(-1, seq_len) - optimizer.zero_grad(set_to_none=True) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - if byte_weighted_ttt: - # Byte-weighted loss: tokens covering more bytes matter more - ttt_logits = base_model.forward_logits(x) - per_token_loss = F.cross_entropy( - ttt_logits.reshape(-1, ttt_logits.size(-1)), - y.reshape(-1), reduction='none' - ).reshape(y.shape) - byte_weights = base_bytes_lut[y].float() - byte_weights = byte_weights + (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).float() - ttt_loss = (per_token_loss * byte_weights).sum() / byte_weights.sum() - else: - ttt_loss = base_model(x, y) - ttt_loss.backward() - if world_size > 1: - for p in ttt_params: - if p.grad is not None: - dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) - torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) - optimizer.step() - # Update Polyak EMA after each step - if use_polyak: - for p in ttt_params: - polyak_state[id(p)].lerp_(p.data, 1.0 - polyak_decay) - if rank == 0 and ci < 3: - print(f" step done ep={_ep+1} bs={bs} loss={ttt_loss.item():.4f}", flush=True) - - if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 5): - elapsed = time.perf_counter() - t0 - rl = loss_sum.item() / max(token_count.item(), 1) - rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 - print(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s", flush=True) - - if dist.is_available() and dist.is_initialized(): - dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) - dist.all_reduce(token_count, op=dist.ReduceOp.SUM) - dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) - - val_loss = (loss_sum / token_count).item() - val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) - - for p in base_model.parameters(): - p.requires_grad_(True) - base_model.eval() - - if rank == 0: - print(f"ttt:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " - f"elapsed={time.perf_counter() - t0:.1f}s") - return val_loss, val_bpb - -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 = 15) -> 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 _find_best_row_scales(W: Tensor, clip_range: int = 15) -> Tensor: - """Find optimal per-row scales by searching percentile clipping thresholds.""" - t32 = W.float() - best_s = t32.abs().amax(dim=1) / clip_range - best_s = best_s.clamp_min(1.0 / clip_range) - best_err = torch.full((t32.shape[0],), 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) - q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) - recon = q * s[:, None] - err = (t32 - recon).pow(2).mean(dim=1) - improved = err < best_err - best_s[improved] = s[improved] - best_err[improved] = err[improved] - return best_s - -def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 15, - block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: - """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation.""" - W = W.float().clone() - rows, cols = W.shape - row_scale = _find_best_row_scales(W, clip_range) - H = H.float().clone() - damp = percdamp * H.diag().mean() - H.diagonal().add_(damp) - perm = torch.argsort(H.diag()) - invperm = torch.argsort(perm) - W = W[:, perm] - H = H[perm][:, perm] - try: - L = torch.linalg.cholesky(H) - Hinv = torch.cholesky_inverse(L) - except torch._C._LinAlgError: - Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) - Q = torch.zeros(rows, cols, dtype=torch.int8) - for i1 in range(0, cols, block_size): - i2 = min(i1 + block_size, cols) - W_block = W[:, i1:i2].clone() - Hinv_block = Hinv[i1:i2, i1:i2] - Err = torch.zeros_like(W_block) - for j in range(i2 - i1): - w_col = W_block[:, j] - h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) - q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) - deq_col = q_col * row_scale - Q[:, i1 + j] = q_col.to(torch.int8) - err = (w_col - deq_col) / h_inv_jj - Err[:, j] = err - if j + 1 < i2 - i1: - W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) - if i2 < cols: - W[:, i2:] -= Err @ Hinv[i1:i2, i2:] - Q = Q[:, invperm] - return Q, row_scale.to(torch.float16) - -def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, - n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: - """Collect Hessian H = X^T X for each linear layer using training data.""" - hessians: dict[str, Tensor] = {} - n_seen: dict[str, int] = {} - hooks = [] - def make_hook(name: str): - def hook_fn(module, inp, out): - x = inp[0].detach().float() - if x.ndim == 3: - x = x.reshape(-1, x.shape[-1]) - if name not in hessians: - hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) - n_seen[name] = 0 - hessians[name].addmm_(x.t(), x) - n_seen[name] += x.shape[0] - return hook_fn - for name, module in model.named_modules(): - if isinstance(module, (nn.Linear, CastedLinear)): - hooks.append(module.register_forward_hook(make_hook(name))) - stream = TokenStream(train_pattern) - model.eval() - with torch.no_grad(): - for _ in range(n_samples): - tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) - x = tokens[:-1].unsqueeze(0) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16): - model.forward_logits(x) - for h in hooks: - h.remove() - for name in hessians: - hessians[name] /= max(n_seen[name], 1) - return hessians - -def _get_layer_clip_range(name: str, num_layers: int, int6_last_n: int) -> int: - """Return clip_range based on which layer the param belongs to.""" - import re - m = re.search(r'blocks\.(\d+)\.', name) - if m: - layer_idx = int(m.group(1)) - if layer_idx >= num_layers - int6_last_n: - return 31 # int6 - return 15 # int5 - -def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], - hessians: dict[str, Tensor], - num_layers: int = 11, int6_last_n: int = 2) -> tuple[dict, dict]: - """GPTQ quantization with mixed int5/int6 precision. int6 for last int6_last_n layers, int5 for rest.""" - result: dict[str, Tensor] = {} - meta: dict[str, object] = {} - gptq_count, naive_count = 0, 0 - int5_params, int6_params = 0, 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 - cr = _get_layer_clip_range(name, num_layers, int6_last_n) - if cr == 31: - int6_params += t.numel() - else: - int5_params += t.numel() - if cat in int6_cats and t.ndim == 2: - module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name - H = hessians.get(module_name) - if H is not None and H.shape[0] == t.shape[1]: - q, s = gptq_quantize_weight(t, H.cpu(), clip_range=cr) - gptq_count += 1 - else: - q, s = quantize_int6_per_row(t, clip_range=cr) - naive_count += 1 - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} - elif cat in int6_cats and t.ndim >= 1: - q, s = quantize_int6_per_row(t, clip_range=cr) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": f"int{'6' if cr == 31 else '5'}"} - naive_count += 1 - else: - q, s = quantize_float_tensor(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int8"} - print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) - print(f"mixed_precision: {int5_params} int5 params, {int6_params} int6 params", flush=True) - return result, meta - - -def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): - result: dict[str, Tensor] = {} - meta: dict[str, object] = {} - for name, tensor in state_dict.items(): - t = tensor.detach().cpu().contiguous() - cat = _classify_param(name) - if not t.is_floating_point() or t.numel() <= 65536: - result[name] = t.to(torch.float16) if t.is_floating_point() else t - meta[name] = "passthrough" - continue - if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): - result[name] = t.float() - meta[name] = "passthrough_ctrl" - continue - if cat in int6_cats and t.ndim >= 1: - q, s = quantize_int6_per_row(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int6"} - else: - q, s = quantize_float_tensor(t) - result[name + ".q"] = q - result[name + ".scale"] = s - meta[name] = {"type": "int8"} - return result, meta - -def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], - template_sd: dict[str, Tensor]) -> dict[str, Tensor]: - out: dict[str, Tensor] = {} - for name, orig in template_sd.items(): - info = meta.get(name) - if info is None: - continue - orig_dtype = orig.dtype - if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): - t = result[name] - if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): - t = t.to(orig_dtype) - out[name] = t - continue - q, s = result[name + ".q"], result[name + ".scale"] - if s.ndim > 0: - out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) - else: - out[name] = (q.float() * float(s.item())).to(orig_dtype) - return out - -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 - 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(f"Python {sys.version} PyTorch {torch.__version__}", 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, - 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, - ).to(device).bfloat16() - for module in base_model.modules(): - if isinstance(module, CastedLinear): - module.float() - restore_low_dim_params_to_fp32(base_model) - compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) - model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model - block_named_params = list(base_model.blocks.named_parameters()) - matrix_params = [ - p - for name, p in block_named_params - if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) - ] - scalar_params = [ - p - for name, p in block_named_params - if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) - ] - if base_model.skip_weights.numel() > 0: - scalar_params.append(base_model.skip_weights) - scalar_params.append(base_model.smear.gate) - if base_model.bigram is not None: - scalar_params.append(base_model.bigram.scale) - token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr - tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] - if base_model.bigram is not None: - tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) - if base_model.bigram.proj is not None: - matrix_params.append(base_model.bigram.proj.weight) - 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) - optimizer_tok = torch.optim.AdamW( - tok_params, - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - weight_decay=args.adam_wd, - fused=True, - ) - optimizer_muon = Muon( - matrix_params, - lr=args.matrix_lr, - momentum=args.muon_momentum, - backend_steps=args.muon_backend_steps, - weight_decay=args.muon_wd, - ) - for group in optimizer_muon.param_groups: - group["base_lr"] = args.matrix_lr - optimizer_scalar = torch.optim.AdamW( - [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - weight_decay=args.adam_wd, - fused=True, - ) - optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] - if base_model.lm_head is not None: - optimizer_head = torch.optim.Adam( - [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - fused=True, - ) - optimizers.insert(1, optimizer_head) - n_params = sum(p.numel() for p in base_model.parameters()) - # Set int6 clip_range for last N layers (mixed precision) - int6_start = args.num_layers - args.int6_last_n - for i, block in enumerate(base_model.blocks): - if i >= int6_start: - for m in block.modules(): - if isinstance(m, CastedLinear): - m._clip_range = 31 # int6 - if master_process: - int5_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 15) - int6_count = sum(1 for m in base_model.modules() if isinstance(m, CastedLinear) and m._clip_range == 31) - log0(f"mixed_precision: {int5_count} int5 layers, {int6_count} int6 layers (last {args.int6_last_n} blocks)") - log0(f"model_params:{n_params}") - xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] - log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}") - log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s 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 - # TTT_ONLY mode: skip training, load saved model, run TTT eval - if os.environ.get("TTT_ONLY", "0") == "1": - log0("TTT_ONLY mode: skipping training, loading saved model...") - sd_cpu = {k: v.cpu() for k, v in torch.load("final_model.pt", map_location="cpu").items()} - if args.prune_pct > 0: - for k, v in sd_cpu.items(): - if v.ndim == 2 and v.numel() > 65536: - thresh = torch.quantile(v.abs().float(), args.prune_pct) - v[v.abs() < thresh] = 0.0 - log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") - with open("final_model.int6.ptz", "rb") as f: - quant_blob_disk = f.read() - quant_state = torch.load( - io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), - map_location="cpu", - ) - deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) - eval_model = GPT( - vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, - num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, - tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, - logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, - bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, 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, - ).to(device).bfloat16() - for m in eval_model.modules(): - if isinstance(m, CastedLinear): - m.float() - restore_low_dim_params_to_fp32(eval_model) - eval_model.load_state_dict(deq_state, strict=True) - sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) - log0(f"TTT_ONLY: model loaded, starting TTT eval...") - torch.cuda.synchronize() - t_ttt = time.perf_counter() - ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) - ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) - ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) - ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) - ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") - log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") - ttt_temp = args.ttt_temperature - log0(f"TTT temperature: {ttt_temp}") - ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( - args, eval_model, rank, world_size, device, - val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, - ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, - ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, - ttt_temp=ttt_temp, - ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), - byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", - use_cache=os.environ.get("USE_CACHE", "1") == "1", - cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), - adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", - adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), - ) - torch.cuda.synchronize() - log0( - f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " - f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" - ) - log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") - if distributed: - dist.destroy_process_group() - return - - if args.warmup_steps > 0: - initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} - initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] - model.train() - for warmup_step in range(args.warmup_steps): - zero_grad_all() - for micro_step in range(grad_accum_steps): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): - warmup_loss = model(x, y) - (warmup_loss * grad_scale).backward() - for opt in optimizers: - opt.step() - zero_grad_all() - if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: - log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") - base_model.load_state_dict(initial_model_state, strict=True) - for opt, state in zip(optimizers, initial_optimizer_states, strict=True): - opt.load_state_dict(state) - zero_grad_all() - if distributed: - model.require_backward_grad_sync = True - train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) - 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) - # Anneal soft-round alpha based on QAT progress - if CastedLinear._use_soft_round and CastedLinear._qat_enabled: - qat_progress = max(0.0, 1.0 - scale / max(args.late_qat_threshold, 0.01)) - CastedLinear._soft_round_alpha = 1.0 + 15.0 * qat_progress - if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: - CastedLinear._qat_enabled = True - CastedLinear._use_soft_round = os.environ.get("SOFT_ROUND_QAT", "0") == "1" - if CastedLinear._use_soft_round and master_process: - log0(f"soft_round_qat:enabled initial_alpha=1.0") - log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") - zero_grad_all() - train_loss = torch.zeros((), device=device) - for micro_step in range(grad_accum_steps): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) - with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): - loss = model(x, y) - train_loss += loss.detach() - (loss * grad_scale).backward() - train_loss /= grad_accum_steps - frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 - muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum - for group in optimizer_muon.param_groups: - group["momentum"] = muon_momentum - for opt in optimizers: - for group in opt.param_groups: - group["lr"] = group["base_lr"] * scale - if args.grad_clip_norm > 0: - torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) - for opt in optimizers: - opt.step() - zero_grad_all() - 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" - ) - raw_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} - best_bpb = float('inf') - best_label = "raw" - best_state = raw_state - log0("ema:applying EMA weights") - current_state = base_model.state_dict() - ema_sd = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} - base_model.load_state_dict(ema_sd, strict=True) - torch.cuda.synchronize() - t_diag = time.perf_counter() - ema_val_loss, ema_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:{ema_val_loss:.4f} val_bpb:{ema_val_bpb:.4f} " - f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" - ) - if ema_val_bpb < best_bpb: - best_bpb = ema_val_bpb - best_label = "ema" - best_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} - if swa_state is not None and swa_count > 0: - log0(f"swa:applying SWA weights (count={swa_count})") - swa_sd = {} - for name in current_state: - swa_avg = (swa_state[name].float() / swa_count).to(dtype=current_state[name].dtype) - swa_sd[name] = swa_avg - base_model.load_state_dict(swa_sd, strict=True) - torch.cuda.synchronize() - t_diag = time.perf_counter() - swa_val_loss, swa_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_swa val_loss:{swa_val_loss:.4f} val_bpb:{swa_val_bpb:.4f} " - f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" - ) - if swa_val_bpb < best_bpb: - best_bpb = swa_val_bpb - best_label = "swa" - best_state = {name: t.detach().clone() for name, t in base_model.state_dict().items()} - - log0(f"best_averaging:{best_label} val_bpb:{best_bpb:.4f}") - base_model.load_state_dict(best_state, strict=True) - export_sd = base_model.state_dict() - if master_process: - torch.save(export_sd, "final_model.pt") - model_bytes = os.path.getsize("final_model.pt") - code_bytes = len(code.encode("utf-8")) - log0(f"Serialized model: {model_bytes} bytes") - log0(f"Code size: {code_bytes} bytes") - sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} - if args.prune_pct > 0: - for k, v in sd_cpu.items(): - if v.ndim == 2 and v.numel() > 65536: - thresh = torch.quantile(v.abs().float(), args.prune_pct) - v[v.abs() < thresh] = 0.0 - if master_process: - log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied") - # GPTQ calibration - log0("gptq:calibrating with training data...") - t_gptq = time.perf_counter() - gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) - log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") - quant_result, quant_meta = mixed_quantize_int6_gptq(sd_cpu, {"mlp", "attn"}, gptq_hessians, num_layers=args.num_layers, int6_last_n=args.int6_last_n) - quant_buf = io.BytesIO() - torch.save({"w": quant_result, "m": quant_meta}, quant_buf) - quant_raw = quant_buf.getvalue() - quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9) - if master_process: - with open("final_model.int6.ptz", "wb") as f: - f.write(quant_blob) - quant_file_bytes = len(quant_blob) - code_bytes = len(code.encode("utf-8")) - log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes") - log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") - if distributed: - dist.barrier() - with open("final_model.int6.ptz", "rb") as f: - quant_blob_disk = f.read() - quant_state = torch.load( - io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)), - map_location="cpu", - ) - deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) - eval_model = GPT( - vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, - num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, - tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, - logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, - bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, 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, - ).to(device).bfloat16() - for m in eval_model.modules(): - if isinstance(m, CastedLinear): - m.float() - restore_low_dim_params_to_fp32(eval_model) - eval_model.load_state_dict(deq_state, strict=True) - sw_seq_len = int(os.environ.get("EVAL_SEQ_LEN", str(effective_eval_seq_len))) - if sw_seq_len != effective_eval_seq_len and rank == 0: - log0(f"Eval seq_len override: {effective_eval_seq_len} -> {sw_seq_len}") - if args.eval_stride > 0 and args.eval_stride < sw_seq_len and not os.environ.get("SKIP_SLIDING"): - 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}") - torch.cuda.synchronize() - t_ttt = time.perf_counter() - ttt_epochs = int(os.environ.get("TTT_EPOCHS", "3")) - ttt_lr = float(os.environ.get("TTT_LR", "0.0005")) - ttt_freeze = int(os.environ.get("TTT_FREEZE_BLOCKS", "2")) - ttt_chunk = int(os.environ.get("TTT_CHUNK_TOKENS", "32768")) - ttt_opt = os.environ.get("TTT_OPTIMIZER", "adamw") - log0(f"TTT: epochs={ttt_epochs} lr={ttt_lr} freeze_first={ttt_freeze} chunk={ttt_chunk} opt={ttt_opt}") - ttt_temp = args.ttt_temperature - log0(f"TTT temperature: {ttt_temp}") - ppm_alpha_val = float(os.environ.get("PPM_ALPHA", "0.85")) - bw_ttt = os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1" - log0(f"PPM alpha: {ppm_alpha_val}, Byte-weighted TTT: {bw_ttt}") - ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( - args, eval_model, rank, world_size, device, - val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, - stride=args.eval_stride, ttt_epochs=ttt_epochs, ttt_lr=ttt_lr, - ttt_freeze_blocks=ttt_freeze, eval_seq_len=sw_seq_len, - ttt_chunk_tokens=ttt_chunk, ttt_optimizer=ttt_opt, - ttt_temp=ttt_temp, - ppm_alpha=float(os.environ.get("PPM_ALPHA", "0.85")), - byte_weighted_ttt=os.environ.get("BYTE_WEIGHTED_TTT", "1") == "1", - use_cache=os.environ.get("USE_CACHE", "1") == "1", - cache_alpha=float(os.environ.get("CACHE_ALPHA", "0.3")), - adaptive_lr=os.environ.get("ADAPTIVE_LR", "1") == "1", - adaptive_lr_max_mult=float(os.environ.get("ADAPTIVE_LR_MAX", "3.0")), - ) - torch.cuda.synchronize() - log0( - f"final_int6_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " - f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" - ) - log0(f"final_int6_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") - if distributed: - dist.destroy_process_group() -if __name__ == "__main__": - main() diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log deleted file mode 100644 index 4596180df..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed1337.log +++ /dev/null @@ -1,172 +0,0 @@ -W0324 06:47:48.570000 94727 torch/distributed/run.py:851] -W0324 06:47:48.570000 94727 torch/distributed/run.py:851] ***************************************** -W0324 06:47:48.570000 94727 torch/distributed/run.py:851] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0324 06:47:48.570000 94727 torch/distributed/run.py:851] ***************************************** -logs/0986e1a2-99db-4b93-ba13-1082fe463b5d.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33580124 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s 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:6.9304 val_bpb:4.1046 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9324 train_time:155ms step_avg:154.62ms -step:2/20000 train_loss:8.6499 train_time:244ms step_avg:122.18ms -step:3/20000 train_loss:7.7400 train_time:339ms step_avg:113.00ms -step:4/20000 train_loss:7.2905 train_time:434ms step_avg:108.40ms -step:5/20000 train_loss:7.0203 train_time:528ms step_avg:105.66ms -step:6/20000 train_loss:6.8351 train_time:623ms step_avg:103.87ms -step:7/20000 train_loss:6.7947 train_time:718ms step_avg:102.57ms -step:8/20000 train_loss:6.7258 train_time:812ms step_avg:101.51ms -step:9/20000 train_loss:6.4110 train_time:907ms step_avg:100.79ms -step:10/20000 train_loss:6.0618 train_time:1002ms step_avg:100.17ms -step:500/20000 train_loss:2.3545 train_time:48339ms step_avg:96.68ms -step:1000/20000 train_loss:2.2365 train_time:96843ms step_avg:96.84ms -step:1500/20000 train_loss:2.1818 train_time:145370ms step_avg:96.91ms -step:2000/20000 train_loss:2.0262 train_time:194003ms step_avg:97.00ms -step:2500/20000 train_loss:2.1279 train_time:242644ms step_avg:97.06ms -step:3000/20000 train_loss:2.1145 train_time:291318ms step_avg:97.11ms -step:3500/20000 train_loss:2.1254 train_time:340011ms step_avg:97.15ms -step:4000/20000 train_loss:1.9115 train_time:388714ms step_avg:97.18ms -step:4000/20000 val_loss:2.0024 val_bpb:1.1860 train_time:388719ms step_avg:97.18ms -soft_round_qat:enabled initial_alpha=1.0 -late_qat:enabled step:4424 scale:0.4997 -step:4500/20000 train_loss:2.0582 train_time:437424ms step_avg:97.21ms -step:5000/20000 train_loss:2.0351 train_time:486102ms step_avg:97.22ms -swa:start step:5500 -step:5500/20000 train_loss:1.9473 train_time:534768ms step_avg:97.23ms -step:6000/20000 train_loss:1.8706 train_time:583974ms step_avg:97.33ms -step:6163/20000 val_loss:1.8983 val_bpb:1.1243 train_time:600024ms step_avg:97.36ms -stopping_early: wallclock_cap train_time:600024ms step:6163/20000 -peak memory allocated: 26201 MiB reserved: 26418 MiB -ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:1.8966 val_bpb:1.1233 eval_time:2368ms -swa:applying SWA weights (count=14) -DIAGNOSTIC post_swa val_loss:1.8982 val_bpb:1.1242 eval_time:2360ms -best_averaging:ema val_bpb:1.1233 -Serialized model: 130956873 bytes -Code size: 106734 bytes -pruning:2.0% magnitude pruning applied -gptq:calibrating with training data... -gptq:calibrated 68 layers in 3.6s -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -Serialized model int6+zstd: 15715344 bytes -Total submission size int6+zstd: 15822078 bytes -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 1.0, Byte-weighted TTT: False -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27799624 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0755 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0637 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0631 - ttt_chunk [1/474] bpb=1.199717 time=1.3s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.9046 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.9036 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.9017 - ttt_chunk [2/474] bpb=1.153948 time=2.4s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.8512 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.8502 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.8490 - ttt_chunk [3/474] bpb=1.127169 time=3.6s - ttt_chunk [4/474] bpb=1.134789 time=4.7s - ttt_chunk [5/474] bpb=1.133690 time=5.8s - ttt_chunk [11/474] bpb=1.116604 time=12.7s - ttt_chunk [21/474] bpb=1.111760 time=24.1s - ttt_chunk [31/474] bpb=1.109901 time=35.5s - ttt_chunk [41/474] bpb=1.118054 time=46.9s - ttt_chunk [51/474] bpb=1.125284 time=58.3s - ttt_chunk [61/474] bpb=1.123258 time=69.7s - ttt_chunk [71/474] bpb=1.124623 time=81.1s - ttt_chunk [81/474] bpb=1.125042 time=92.5s - ttt_chunk [91/474] bpb=1.127019 time=103.9s - ttt_chunk [101/474] bpb=1.123588 time=115.3s - ttt_chunk [111/474] bpb=1.123831 time=126.8s - ttt_chunk [121/474] bpb=1.127096 time=138.2s - ttt_chunk [131/474] bpb=1.127790 time=149.6s - ttt_chunk [141/474] bpb=1.127537 time=161.0s - ttt_chunk [151/474] bpb=1.125756 time=172.4s - ttt_chunk [161/474] bpb=1.126665 time=183.8s - ttt_chunk [171/474] bpb=1.125481 time=195.2s - ttt_chunk [181/474] bpb=1.126243 time=206.6s - ttt_chunk [191/474] bpb=1.125132 time=218.0s - ttt_chunk [201/474] bpb=1.124308 time=229.4s - ttt_chunk [211/474] bpb=1.122924 time=240.9s - ttt_chunk [221/474] bpb=1.123005 time=252.3s - ttt_chunk [231/474] bpb=1.122391 time=263.7s - ttt_chunk [241/474] bpb=1.121305 time=275.1s - ttt_chunk [251/474] bpb=1.122402 time=286.5s - ttt_chunk [261/474] bpb=1.123029 time=297.9s - ttt_chunk [271/474] bpb=1.121517 time=309.3s - ttt_chunk [281/474] bpb=1.121123 time=320.7s - ttt_chunk [291/474] bpb=1.119558 time=332.1s - ttt_chunk [301/474] bpb=1.119987 time=343.5s - ttt_chunk [311/474] bpb=1.119381 time=354.9s - ttt_chunk [321/474] bpb=1.117767 time=366.4s - ttt_chunk [331/474] bpb=1.116735 time=377.8s - ttt_chunk [341/474] bpb=1.115996 time=389.2s - ttt_chunk [351/474] bpb=1.114351 time=400.6s - ttt_chunk [361/474] bpb=1.114833 time=412.0s - ttt_chunk [371/474] bpb=1.114520 time=423.4s - ttt_chunk [381/474] bpb=1.115285 time=434.8s - ttt_chunk [391/474] bpb=1.116303 time=446.2s - ttt_chunk [401/474] bpb=1.116709 time=457.6s - ttt_chunk [411/474] bpb=1.117120 time=469.0s - ttt_chunk [421/474] bpb=1.118532 time=480.4s - ttt_chunk [431/474] bpb=1.117100 time=491.8s - ttt_chunk [441/474] bpb=1.116713 time=503.2s - ttt_chunk [451/474] bpb=1.116048 time=514.7s - ttt_chunk [461/474] bpb=1.116249 time=526.1s - ttt_chunk [471/474] bpb=1.116338 time=537.5s - ttt_chunk [474/474] bpb=1.116206 time=540.0s -ttt:done val_loss=1.883479 val_bpb=1.115506 elapsed=540.0s -final_int6_ttt val_loss:1.8835 val_bpb:1.1155 stride:32 eval_time:540962ms -final_int6_ttt_exact val_loss:1.88347869 val_bpb:1.11550587 diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log deleted file mode 100644 index 6ea144408..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed42.log +++ /dev/null @@ -1,172 +0,0 @@ -W0324 07:15:10.419000 100096 torch/distributed/run.py:851] -W0324 07:15:10.419000 100096 torch/distributed/run.py:851] ***************************************** -W0324 07:15:10.419000 100096 torch/distributed/run.py:851] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0324 07:15:10.419000 100096 torch/distributed/run.py:851] ***************************************** -logs/026bfb42-da81-45bf-b73b-6ae924fd88fa.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33580124 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:42 -warmup_step:1/20 -warmup_step:2/20 -warmup_step:3/20 -warmup_step:4/20 -warmup_step:5/20 -warmup_step:6/20 -warmup_step:7/20 -warmup_step:8/20 -warmup_step:9/20 -warmup_step:10/20 -warmup_step:11/20 -warmup_step:12/20 -warmup_step:13/20 -warmup_step:14/20 -warmup_step:15/20 -warmup_step:16/20 -warmup_step:17/20 -warmup_step:18/20 -warmup_step:19/20 -warmup_step:20/20 -step:0/20000 val_loss:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9335 train_time:154ms step_avg:153.60ms -step:2/20000 train_loss:8.6987 train_time:244ms step_avg:121.91ms -step:3/20000 train_loss:7.7606 train_time:339ms step_avg:112.85ms -step:4/20000 train_loss:7.2812 train_time:434ms step_avg:108.41ms -step:5/20000 train_loss:7.0422 train_time:529ms step_avg:105.73ms -step:6/20000 train_loss:6.9445 train_time:623ms step_avg:103.84ms -step:7/20000 train_loss:6.8297 train_time:718ms step_avg:102.52ms -step:8/20000 train_loss:6.6897 train_time:812ms step_avg:101.55ms -step:9/20000 train_loss:6.3850 train_time:907ms step_avg:100.77ms -step:10/20000 train_loss:5.9825 train_time:1002ms step_avg:100.21ms -step:500/20000 train_loss:2.3564 train_time:48431ms step_avg:96.86ms -step:1000/20000 train_loss:2.2389 train_time:97006ms step_avg:97.01ms -step:1500/20000 train_loss:2.1831 train_time:145627ms step_avg:97.08ms -step:2000/20000 train_loss:2.0279 train_time:194337ms step_avg:97.17ms -step:2500/20000 train_loss:2.1312 train_time:243086ms step_avg:97.23ms -step:3000/20000 train_loss:2.1151 train_time:291847ms step_avg:97.28ms -step:3500/20000 train_loss:2.1228 train_time:340613ms step_avg:97.32ms -step:4000/20000 train_loss:1.9145 train_time:389395ms step_avg:97.35ms -step:4000/20000 val_loss:2.0040 val_bpb:1.1869 train_time:389401ms step_avg:97.35ms -soft_round_qat:enabled initial_alpha=1.0 -late_qat:enabled step:4413 scale:0.4999 -step:4500/20000 train_loss:2.0597 train_time:438163ms step_avg:97.37ms -step:5000/20000 train_loss:2.0373 train_time:486893ms step_avg:97.38ms -swa:start step:5500 -step:5500/20000 train_loss:1.9452 train_time:535639ms step_avg:97.39ms -step:6000/20000 train_loss:1.8708 train_time:584998ms step_avg:97.50ms -step:6152/20000 val_loss:1.8998 val_bpb:1.1252 train_time:600042ms step_avg:97.54ms -stopping_early: wallclock_cap train_time:600042ms step:6152/20000 -peak memory allocated: 26201 MiB reserved: 26418 MiB -ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:1.8982 val_bpb:1.1242 eval_time:2370ms -swa:applying SWA weights (count=14) -DIAGNOSTIC post_swa val_loss:1.8997 val_bpb:1.1251 eval_time:2371ms -best_averaging:ema val_bpb:1.1242 -Serialized model: 130956873 bytes -Code size: 106734 bytes -pruning:2.0% magnitude pruning applied -gptq:calibrating with training data... -gptq:calibrated 68 layers in 3.6s -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -Serialized model int6+zstd: 15308671 bytes -Total submission size int6+zstd: 15415405 bytes -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 1.0, Byte-weighted TTT: False -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27799624 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0787 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0670 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0662 - ttt_chunk [1/474] bpb=1.201291 time=1.3s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.9070 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.9062 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.9039 - ttt_chunk [2/474] bpb=1.154634 time=2.5s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.8655 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.8648 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.8629 - ttt_chunk [3/474] bpb=1.130046 time=3.6s - ttt_chunk [4/474] bpb=1.137970 time=4.7s - ttt_chunk [5/474] bpb=1.136011 time=5.9s - ttt_chunk [11/474] bpb=1.117481 time=12.7s - ttt_chunk [21/474] bpb=1.112444 time=24.1s - ttt_chunk [31/474] bpb=1.110774 time=35.5s - ttt_chunk [41/474] bpb=1.118885 time=47.0s - ttt_chunk [51/474] bpb=1.126265 time=58.4s - ttt_chunk [61/474] bpb=1.124173 time=69.8s - ttt_chunk [71/474] bpb=1.125678 time=81.2s - ttt_chunk [81/474] bpb=1.126147 time=92.6s - ttt_chunk [91/474] bpb=1.128060 time=104.0s - ttt_chunk [101/474] bpb=1.124603 time=115.5s - ttt_chunk [111/474] bpb=1.124793 time=126.9s - ttt_chunk [121/474] bpb=1.128136 time=138.3s - ttt_chunk [131/474] bpb=1.128812 time=149.7s - ttt_chunk [141/474] bpb=1.128665 time=161.1s - ttt_chunk [151/474] bpb=1.126959 time=172.5s - ttt_chunk [161/474] bpb=1.127822 time=183.9s - ttt_chunk [171/474] bpb=1.126462 time=195.3s - ttt_chunk [181/474] bpb=1.127216 time=206.8s - ttt_chunk [191/474] bpb=1.126189 time=218.2s - ttt_chunk [201/474] bpb=1.125335 time=229.6s - ttt_chunk [211/474] bpb=1.124007 time=241.0s - ttt_chunk [221/474] bpb=1.124128 time=252.4s - ttt_chunk [231/474] bpb=1.123537 time=263.8s - ttt_chunk [241/474] bpb=1.122433 time=275.3s - ttt_chunk [251/474] bpb=1.123539 time=286.7s - ttt_chunk [261/474] bpb=1.124188 time=298.1s - ttt_chunk [271/474] bpb=1.122695 time=309.5s - ttt_chunk [281/474] bpb=1.122278 time=320.9s - ttt_chunk [291/474] bpb=1.120755 time=332.3s - ttt_chunk [301/474] bpb=1.121114 time=343.7s - ttt_chunk [311/474] bpb=1.120507 time=355.2s - ttt_chunk [321/474] bpb=1.118816 time=366.6s - ttt_chunk [331/474] bpb=1.117759 time=378.0s - ttt_chunk [341/474] bpb=1.117018 time=389.4s - ttt_chunk [351/474] bpb=1.115370 time=400.8s - ttt_chunk [361/474] bpb=1.115858 time=412.2s - ttt_chunk [371/474] bpb=1.115523 time=423.6s - ttt_chunk [381/474] bpb=1.116283 time=435.0s - ttt_chunk [391/474] bpb=1.117326 time=446.5s - ttt_chunk [401/474] bpb=1.117734 time=457.9s - ttt_chunk [411/474] bpb=1.118144 time=469.3s - ttt_chunk [421/474] bpb=1.119563 time=480.7s - ttt_chunk [431/474] bpb=1.118129 time=492.1s - ttt_chunk [441/474] bpb=1.117766 time=503.5s - ttt_chunk [451/474] bpb=1.117119 time=515.0s - ttt_chunk [461/474] bpb=1.117338 time=526.4s - ttt_chunk [471/474] bpb=1.117400 time=537.8s - ttt_chunk [474/474] bpb=1.117261 time=540.3s -ttt:done val_loss=1.884801 val_bpb=1.116289 elapsed=540.3s -final_int6_ttt val_loss:1.8848 val_bpb:1.1163 stride:32 eval_time:541278ms -final_int6_ttt_exact val_loss:1.88480123 val_bpb:1.11628915 diff --git a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log b/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log deleted file mode 100644 index 0fe246d62..000000000 --- a/records/track_10min_16mb/2026-03-24_Int5GPTQ_33M_LegalTTT/train_seed7.log +++ /dev/null @@ -1,172 +0,0 @@ -W0324 07:20:32.173000 139509 torch/distributed/run.py:803] -W0324 07:20:32.173000 139509 torch/distributed/run.py:803] ***************************************** -W0324 07:20:32.173000 139509 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. -W0324 07:20:32.173000 139509 torch/distributed/run.py:803] ***************************************** -logs/5d3ebe13-8183-4305-965c-55651fc9638b.txt -val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model -train_loader:dataset:fineweb10B_sp1024 train_shards:80 -val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632 -mixed_precision: 68 int5 layers, 0 int6 layers (last 0 blocks) -model_params:33580124 -XSA:[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10] ws:8 gqa:8/8 -lr:embed=0.035 matrix=0.025 scalar=0.025 batch:786432 wall:600s seed:7 -warmup_step:1/20 -warmup_step:2/20 -warmup_step:3/20 -warmup_step:4/20 -warmup_step:5/20 -warmup_step:6/20 -warmup_step:7/20 -warmup_step:8/20 -warmup_step:9/20 -warmup_step:10/20 -warmup_step:11/20 -warmup_step:12/20 -warmup_step:13/20 -warmup_step:14/20 -warmup_step:15/20 -warmup_step:16/20 -warmup_step:17/20 -warmup_step:18/20 -warmup_step:19/20 -warmup_step:20/20 -step:0/20000 val_loss:6.9311 val_bpb:4.1050 train_time:0ms step_avg:0.02ms -step:1/20000 train_loss:6.9326 train_time:153ms step_avg:152.55ms -step:2/20000 train_loss:8.7576 train_time:246ms step_avg:123.04ms -step:3/20000 train_loss:7.7488 train_time:342ms step_avg:114.08ms -step:4/20000 train_loss:7.2030 train_time:438ms step_avg:109.53ms -step:5/20000 train_loss:7.0022 train_time:534ms step_avg:106.79ms -step:6/20000 train_loss:6.8717 train_time:630ms step_avg:104.99ms -step:7/20000 train_loss:6.7821 train_time:726ms step_avg:103.66ms -step:8/20000 train_loss:6.6344 train_time:822ms step_avg:102.72ms -step:9/20000 train_loss:6.3142 train_time:918ms step_avg:101.98ms -step:10/20000 train_loss:5.9783 train_time:1014ms step_avg:101.40ms -step:500/20000 train_loss:2.3552 train_time:49006ms step_avg:98.01ms -step:1000/20000 train_loss:2.2387 train_time:98370ms step_avg:98.37ms -step:1500/20000 train_loss:2.1831 train_time:147750ms step_avg:98.50ms -step:2000/20000 train_loss:2.0243 train_time:197149ms step_avg:98.57ms -step:2500/20000 train_loss:2.1289 train_time:246526ms step_avg:98.61ms -step:3000/20000 train_loss:2.1142 train_time:295868ms step_avg:98.62ms -step:3500/20000 train_loss:2.1203 train_time:345180ms step_avg:98.62ms -step:4000/20000 train_loss:1.9099 train_time:394458ms step_avg:98.61ms -step:4000/20000 val_loss:2.0013 val_bpb:1.1853 train_time:394463ms step_avg:98.62ms -soft_round_qat:enabled initial_alpha=1.0 -late_qat:enabled step:4334 scale:0.4999 -step:4500/20000 train_loss:2.0596 train_time:443781ms step_avg:98.62ms -step:5000/20000 train_loss:2.0348 train_time:493046ms step_avg:98.61ms -swa:start step:5400 -step:5500/20000 train_loss:1.9439 train_time:542488ms step_avg:98.63ms -step:6000/20000 train_loss:1.8689 train_time:592060ms step_avg:98.68ms -step:6081/20000 val_loss:1.8999 val_bpb:1.1252 train_time:600095ms step_avg:98.68ms -stopping_early: wallclock_cap train_time:600095ms step:6081/20000 -peak memory allocated: 26199 MiB reserved: 26784 MiB -ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:1.8983 val_bpb:1.1243 eval_time:2377ms -swa:applying SWA weights (count=14) -DIAGNOSTIC post_swa val_loss:1.9001 val_bpb:1.1253 eval_time:2379ms -best_averaging:ema val_bpb:1.1243 -Serialized model: 130956873 bytes -Code size: 106734 bytes -pruning:2.0% magnitude pruning applied -gptq:calibrating with training data... -gptq:calibrated 68 layers in 3.7s -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -gptq_quantize: 66 GPTQ layers, 0 naive layers -mixed_precision: 33423360 int5 params, 0 int6 params -Serialized model int6+zstd: 15261893 bytes -Total submission size int6+zstd: 15368627 bytes -TTT: epochs=3 lr=0.0001 freeze_first=2 chunk=131072 opt=adamw -TTT temperature: 0.98 -PPM alpha: 1.0, Byte-weighted TTT: False -ttt:start chunks=474 chunk_tokens=131072 windows=1938176 stride=32 lr=0.0001 epochs=3 opt=adamw freeze_first=2 -ttt:params unfrozen=5780500 frozen=27799624 - ttt_train [1] seqs=64 start_train... - ttt_train [1] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=2.0805 - ttt_train [1] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=2.0689 - ttt_train [1] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=2.0683 - ttt_chunk [1/474] bpb=1.200925 time=1.3s - ttt_train [2] seqs=64 start_train... - ttt_train [2] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.9082 - ttt_train [2] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.9075 - ttt_train [2] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.9048 - ttt_chunk [2/474] bpb=1.155071 time=2.5s - ttt_train [3] seqs=64 start_train... - ttt_train [3] epoch=1/3 batches=8 ... - step done ep=1 bs=0 loss=1.8586 - ttt_train [3] epoch=2/3 batches=8 ... - step done ep=2 bs=0 loss=1.8574 - ttt_train [3] epoch=3/3 batches=8 ... - step done ep=3 bs=0 loss=1.8559 - ttt_chunk [3/474] bpb=1.129053 time=3.6s - ttt_chunk [4/474] bpb=1.137770 time=4.8s - ttt_chunk [5/474] bpb=1.136142 time=5.9s - ttt_chunk [11/474] bpb=1.117081 time=12.8s - ttt_chunk [21/474] bpb=1.112525 time=24.2s - ttt_chunk [31/474] bpb=1.110601 time=35.6s - ttt_chunk [41/474] bpb=1.118836 time=47.0s - ttt_chunk [51/474] bpb=1.126156 time=58.4s - ttt_chunk [61/474] bpb=1.124130 time=69.8s - ttt_chunk [71/474] bpb=1.125735 time=81.2s - ttt_chunk [81/474] bpb=1.126269 time=92.7s - ttt_chunk [91/474] bpb=1.128083 time=104.1s - ttt_chunk [101/474] bpb=1.124677 time=115.5s - ttt_chunk [111/474] bpb=1.125049 time=126.9s - ttt_chunk [121/474] bpb=1.128310 time=138.3s - ttt_chunk [131/474] bpb=1.129031 time=149.7s - ttt_chunk [141/474] bpb=1.129021 time=161.1s - ttt_chunk [151/474] bpb=1.127264 time=172.5s - ttt_chunk [161/474] bpb=1.128178 time=184.0s - ttt_chunk [171/474] bpb=1.126890 time=195.4s - ttt_chunk [181/474] bpb=1.127673 time=206.8s - ttt_chunk [191/474] bpb=1.126603 time=218.2s - ttt_chunk [201/474] bpb=1.125778 time=229.6s - ttt_chunk [211/474] bpb=1.124423 time=241.0s - ttt_chunk [221/474] bpb=1.124509 time=252.4s - ttt_chunk [231/474] bpb=1.123899 time=263.8s - ttt_chunk [241/474] bpb=1.122760 time=275.3s - ttt_chunk [251/474] bpb=1.123798 time=286.7s - ttt_chunk [261/474] bpb=1.124407 time=298.1s - ttt_chunk [271/474] bpb=1.122942 time=309.5s - ttt_chunk [281/474] bpb=1.122514 time=320.9s - ttt_chunk [291/474] bpb=1.120992 time=332.3s - ttt_chunk [301/474] bpb=1.121428 time=343.7s - ttt_chunk [311/474] bpb=1.120836 time=355.1s - ttt_chunk [321/474] bpb=1.119179 time=366.5s - ttt_chunk [331/474] bpb=1.118111 time=377.9s - ttt_chunk [341/474] bpb=1.117318 time=389.4s - ttt_chunk [351/474] bpb=1.115673 time=400.8s - ttt_chunk [361/474] bpb=1.116152 time=412.2s - ttt_chunk [371/474] bpb=1.115831 time=423.6s - ttt_chunk [381/474] bpb=1.116577 time=435.0s - ttt_chunk [391/474] bpb=1.117608 time=446.4s - ttt_chunk [401/474] bpb=1.118021 time=457.9s - ttt_chunk [411/474] bpb=1.118453 time=469.2s - ttt_chunk [421/474] bpb=1.119901 time=480.6s - ttt_chunk [431/474] bpb=1.118463 time=491.9s - ttt_chunk [441/474] bpb=1.118084 time=503.3s - ttt_chunk [451/474] bpb=1.117423 time=514.7s - ttt_chunk [461/474] bpb=1.117641 time=526.0s - ttt_chunk [471/474] bpb=1.117706 time=537.4s - ttt_chunk [474/474] bpb=1.117562 time=539.8s -ttt:done val_loss=1.885435 val_bpb=1.116665 elapsed=539.8s -final_int6_ttt val_loss:1.8854 val_bpb:1.1167 stride:32 eval_time:540891ms -final_int6_ttt_exact val_loss:1.88543543 val_bpb:1.11666477 From da7823e1a267a40f2440cc4c108de79070bdfc9f Mon Sep 17 00:00:00 2001 From: ethan Date: Wed, 25 Mar 2026 15:25:58 +0800 Subject: [PATCH 4/5] Record: CROWN-Q + Full GPTQ + SWA/EMA (3-seed mean val_bpb=1.1186) --- .../README.md | 16 +++++----- .../submission.json | 31 +++++++++++++------ .../train_gpt.py | 2 +- 3 files changed, 31 insertions(+), 18 deletions(-) diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md index 53ebf74dd..0f68611d1 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/README.md @@ -2,16 +2,15 @@ ## Summary -- **CROWN-Q**: Curvature-weighted quantization variance penalty applied during warmdown. Encourages weights to settle in flat minima where int6 quantization causes less damage. Penalty: `lambda * sum(h_j * delta_j^2 / 12)` where `h_j = w^2` (curvature proxy) and `delta_j = row_max / 15` (quantization step size). -- **Full Cholesky GPTQ**: Hessian-aware quantization with act-order column permutation, block_size=128, 256-sample calibration from training data. All within 585s training budget. +- **CROWN-Q**: Curvature-weighted quantization variance penalty applied during warmdown. Encourages weights to settle in flat minima where int6 quantization causes less damage. Penalty: `lambda * mean(h_j) * delta_j^2 / 12` per row, where `h_j = w^2` (curvature proxy) and `delta_j = row_max / 15` (CROWN-Q step size). Note: the GPTQ/QAT quantizer uses clip_range=31; CROWN-Q intentionally uses a larger step size (row_max/15) to over-penalize and push weights further into flat basins. +- **Full Cholesky GPTQ**: Hessian-aware quantization with act-order column permutation, block_size=128, 256-sample calibration from training data. GPTQ runs after the 585s training phase as part of model export. - **SWA/EMA 50/50 blend**: Stochastic Weight Averaging (every 50 steps during warmdown) blended 50/50 with EMA (decay=0.997). -- **Architecture**: 11L, 512d, GQA 8H/4KV, MLP 3x LeakyReLU(0.5)^2, XSA on all 11 layers, VRL, BigramHash 3072, partial RoPE 16/64. +- **Architecture**: 11L, 512d, GQA 8H/4KV, MLP 3x LeakyReLU(0.5)^2, XSA on last 4 layers (7-10), VRL, BigramHash 3072, partial RoPE 16/64. - **Eval**: Sliding window with stride=64. No test-time training. ## Configuration ```bash -# Training (585s wallclock, includes GPTQ calibration) torchrun --standalone --nproc_per_node=8 train_gpt.py # Key env vars (all defaults in code): @@ -20,6 +19,7 @@ torchrun --standalone --nproc_per_node=8 train_gpt.py # LATE_QAT_THRESHOLD=0.15 — QAT activation point # MAX_WALLCLOCK_SECONDS=585 — training budget # WARMDOWN_ITERS=4000 — warmdown length +# TTT_ENABLED=0 — TTT disabled for this submission ``` ## Results @@ -28,7 +28,7 @@ torchrun --standalone --nproc_per_node=8 train_gpt.py |------|-------|-------------|-------------|----------| | 1337 | 6613 | 1.1387 | **1.1189** | 15,945,134 | | 42 | 6612 | 1.1382 | **1.1189** | 15,947,742 | -| 7 | 6612 | 1.1378 | **1.1179** | 15,938,790 | +| 7 | 6613 | 1.1378 | **1.1179** | 15,938,790 | | **Mean** | | 1.1382 | **1.1186** | | | **Std** | | | 0.0006 | | @@ -41,11 +41,13 @@ torchrun --standalone --nproc_per_node=8 train_gpt.py CROWN-Q (Curvature-Regularized Optimization for Weight Noise Quantization) adds a training-time penalty that makes weights more robust to quantization noise: -1. For each weight matrix, compute the per-row quantization step size `delta = row_max / 15` (int6 range [-15, 15]) +1. For each weight matrix, compute the per-row quantization step size `delta = row_max / 15` 2. Compute quantization variance `delta^2 / 12` (uniform rounding noise) -3. Weight by curvature proxy `h = w^2` (large weights in high-curvature directions) +3. Weight by curvature proxy `h = mean(w^2)` per row (mean of squared weights) 4. Penalty: `lambda * sum(h * quant_var)` encourages the optimizer to reduce weights in directions where quantization noise is most damaging +The CROWN-Q step size (row_max/15) is intentionally larger than the actual quantizer step size (row_max/31, clip_range=31). This over-penalization pushes weights further into flat basins, providing extra robustness margin against quantization damage. + Applied only during warmdown when QAT is active. Zero eval-time cost. ## Included Files diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json index 0a4b5540f..280cc779f 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/submission.json @@ -2,19 +2,30 @@ "author": "Ethan Yang", "github_id": "EthanYangTW", "name": "CROWN-Q + Full GPTQ + SWA/EMA Blend", - "blurb": "Curvature-weighted quantization variance penalty (CROWN-Q) during warmdown reduces quantization damage. Full Cholesky GPTQ with act-order, SWA/EMA 50/50 blend, VRL, XSA-all 11 layers, LeakyReLU(0.5)^2. Sliding window eval only, no TTT.", + "blurb": "Curvature-weighted quantization variance penalty (CROWN-Q) during warmdown reduces quantization damage. Full Cholesky GPTQ with act-order, SWA/EMA 50/50 blend, VRL, XSA last 4 layers, LeakyReLU(0.5)^2. Sliding window eval only, no TTT.", "date": "2026-03-25T06:30:00Z", + "val_loss": 1.8886, + "val_loss_std": 0.0009, "val_bpb": 1.1186, "val_bpb_std": 0.0006, - "val_bpb_seed1337": 1.1189, - "val_loss_seed1337": 1.8891, - "bytes_seed1337": 15945134, - "val_bpb_seed42": 1.1189, - "val_loss_seed42": 1.8891, - "bytes_seed42": 15947742, - "val_bpb_seed7": 1.1179, - "val_loss_seed7": 1.8876, - "bytes_seed7": 15938790, + "seeds": [1337, 42, 7], + "seed_results": { + "1337": { + "val_bpb": 1.1189, + "val_loss": 1.8891, + "bytes": 15945134 + }, + "42": { + "val_bpb": 1.1189, + "val_loss": 1.8891, + "bytes": 15947742 + }, + "7": { + "val_bpb": 1.1179, + "val_loss": 1.8876, + "bytes": 15938790 + } + }, "bytes_total": 15947742, "bytes_code": 95390 } diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py index d8fd02865..0b04757e2 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_gpt.py @@ -89,7 +89,7 @@ class Hyperparameters: ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) - ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + 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)) # Sliding window TTT (full-parameter, PR#461/549 recipe) From f4636fd77c1342b93ae5a9ed80bf922775c54be3 Mon Sep 17 00:00:00 2001 From: ethan Date: Wed, 25 Mar 2026 15:28:32 +0800 Subject: [PATCH 5/5] Trim logs to sliding window eval (no TTT) --- .../train_seed1337.log | 48 ----------------- .../train_seed42.log | 53 ------------------- .../train_seed7.log | 48 ----------------- 3 files changed, 149 deletions(-) diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log index ea896f40f..3f5f35014 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed1337.log @@ -87,51 +87,3 @@ final_int6_roundtrip_exact val_loss:1.92899174 val_bpb:1.14245756 final_int6_sliding_window val_loss:1.8891 val_bpb:1.1189 stride:64 eval_time:74911ms final_int6_sliding_window_exact val_loss:1.88914884 val_bpb:1.11886332 final_int8_zlib_roundtrip_exact val_loss:1.88914884 val_bpb:1.11886332 -ttt:start lr=0.002 epochs=3 chunks=131072 -ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 -ttt_sliding:params unfrozen=27124848 frozen=0 - ttt_chunk [1/474] bpb=1.198568 time=0.7s - ttt_chunk [11/474] bpb=1.116523 time=6.1s - ttt_chunk [21/474] bpb=1.112338 time=11.4s - ttt_chunk [31/474] bpb=1.111323 time=16.7s - ttt_chunk [41/474] bpb=1.119522 time=22.0s - ttt_chunk [51/474] bpb=1.126882 time=27.4s - ttt_chunk [61/474] bpb=1.124971 time=32.7s - ttt_chunk [71/474] bpb=1.126426 time=38.0s -W0325 05:52:14.355000 324884 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers -W0325 05:52:14.358000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324952 closing signal SIGTERM -W0325 05:52:14.360000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324953 closing signal SIGTERM -W0325 05:52:14.361000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324954 closing signal SIGTERM -W0325 05:52:14.363000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324955 closing signal SIGTERM -W0325 05:52:14.369000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324956 closing signal SIGTERM -W0325 05:52:14.385000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324957 closing signal SIGTERM -W0325 05:52:14.395000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324958 closing signal SIGTERM -W0325 05:52:14.396000 324884 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 324959 closing signal SIGTERM -Traceback (most recent call last): - File "/usr/local/bin/torchrun", line 7, in - sys.exit(main()) - ^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper - return f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main - run(args) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run - elastic_launch( - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ - return launch_agent(self._config, self._entrypoint, list(args)) - ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent - result = agent.run() - ^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper - result = f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run - result = self._invoke_run(role) - ^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run - time.sleep(monitor_interval) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler - raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) -torch.distributed.elastic.multiprocessing.api.SignalException: Process 324884 got signal: 15 diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log index 81ce4b461..101ca01ea 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed42.log @@ -87,56 +87,3 @@ final_int6_roundtrip_exact val_loss:1.92896314 val_bpb:1.14244063 final_int6_sliding_window val_loss:1.8891 val_bpb:1.1189 stride:64 eval_time:75758ms final_int6_sliding_window_exact val_loss:1.88913665 val_bpb:1.11885610 final_int8_zlib_roundtrip_exact val_loss:1.88913665 val_bpb:1.11885610 -ttt:start lr=0.002 epochs=3 chunks=131072 -ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 -ttt_sliding:params unfrozen=27124848 frozen=0 - ttt_chunk [1/474] bpb=1.196551 time=0.7s - ttt_chunk [11/474] bpb=1.116343 time=6.1s - ttt_chunk [21/474] bpb=1.112385 time=11.4s - ttt_chunk [31/474] bpb=1.111105 time=16.7s - ttt_chunk [41/474] bpb=1.119287 time=22.0s - ttt_chunk [51/474] bpb=1.126759 time=27.3s - ttt_chunk [61/474] bpb=1.124851 time=32.8s - ttt_chunk [71/474] bpb=1.126283 time=38.4s - ttt_chunk [81/474] bpb=1.126907 time=44.0s - ttt_chunk [91/474] bpb=1.128848 time=49.6s - ttt_chunk [101/474] bpb=1.125491 time=55.2s - ttt_chunk [111/474] bpb=1.125899 time=60.6s - ttt_chunk [121/474] bpb=1.129329 time=65.9s -W0325 06:31:28.212000 329471 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers -W0325 06:31:28.216000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329539 closing signal SIGTERM -W0325 06:31:28.219000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329540 closing signal SIGTERM -W0325 06:31:28.221000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329541 closing signal SIGTERM -W0325 06:31:28.225000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329542 closing signal SIGTERM -W0325 06:31:28.227000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329543 closing signal SIGTERM -W0325 06:31:28.228000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329544 closing signal SIGTERM -W0325 06:31:28.230000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329545 closing signal SIGTERM -W0325 06:31:28.231000 329471 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 329546 closing signal SIGTERM -Traceback (most recent call last): - File "/usr/local/bin/torchrun", line 7, in - sys.exit(main()) - ^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper - return f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main - run(args) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run - elastic_launch( - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ - return launch_agent(self._config, self._entrypoint, list(args)) - ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent - result = agent.run() - ^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper - result = f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run - result = self._invoke_run(role) - ^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run - time.sleep(monitor_interval) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler - raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) -torch.distributed.elastic.multiprocessing.api.SignalException: Process 329471 got signal: 15 diff --git a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log index d6c7b54b6..daf4432fb 100644 --- a/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log +++ b/records/track_10min_16mb/2026-03-25_CROWNQ_GPTQ_SlidingWindow/train_seed7.log @@ -85,51 +85,3 @@ final_int6_roundtrip_exact val_loss:1.92750067 val_bpb:1.14157447 final_int6_sliding_window val_loss:1.8876 val_bpb:1.1179 stride:64 eval_time:74943ms final_int6_sliding_window_exact val_loss:1.88755846 val_bpb:1.11792140 final_int8_zlib_roundtrip_exact val_loss:1.88755846 val_bpb:1.11792140 -ttt:start lr=0.002 epochs=3 chunks=131072 -ttt_sliding:start chunks=474 chunk_tokens=131072 total_windows=969088 stride=64 ttt_lr=0.002 ttt_epochs=3 freeze_blocks=0 -ttt_sliding:params unfrozen=27124848 frozen=0 - ttt_chunk [1/474] bpb=1.194589 time=0.7s - ttt_chunk [11/474] bpb=1.114842 time=6.1s - ttt_chunk [21/474] bpb=1.111008 time=11.4s - ttt_chunk [31/474] bpb=1.110006 time=16.8s - ttt_chunk [41/474] bpb=1.118383 time=22.1s - ttt_chunk [51/474] bpb=1.126049 time=27.5s - ttt_chunk [61/474] bpb=1.123955 time=33.0s - ttt_chunk [71/474] bpb=1.125390 time=38.5s -W0325 06:53:35.253000 330566 torch/distributed/elastic/agent/server/api.py:725] Received 15 death signal, shutting down workers -W0325 06:53:35.257000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330634 closing signal SIGTERM -W0325 06:53:35.258000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330635 closing signal SIGTERM -W0325 06:53:35.259000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330636 closing signal SIGTERM -W0325 06:53:35.260000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330637 closing signal SIGTERM -W0325 06:53:35.261000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330638 closing signal SIGTERM -W0325 06:53:35.262000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330639 closing signal SIGTERM -W0325 06:53:35.264000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330640 closing signal SIGTERM -W0325 06:53:35.265000 330566 torch/distributed/elastic/multiprocessing/api.py:908] Sending process 330641 closing signal SIGTERM -Traceback (most recent call last): - File "/usr/local/bin/torchrun", line 7, in - sys.exit(main()) - ^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper - return f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 936, in main - run(args) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/run.py", line 927, in run - elastic_launch( - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 156, in __call__ - return launch_agent(self._config, self._entrypoint, list(args)) - ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/launcher/api.py", line 284, in launch_agent - result = agent.run() - ^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/metrics/api.py", line 138, in wrapper - result = f(*args, **kwargs) - ^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 717, in run - result = self._invoke_run(role) - ^^^^^^^^^^^^^^^^^^^^^^ - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/agent/server/api.py", line 881, in _invoke_run - time.sleep(monitor_interval) - File "/usr/local/lib/python3.12/dist-packages/torch/distributed/elastic/multiprocessing/api.py", line 85, in _terminate_process_handler - raise SignalException(f"Process {os.getpid()} got signal: {sigval}", sigval=sigval) -torch.distributed.elastic.multiprocessing.api.SignalException: Process 330566 got signal: 15