diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/README.md b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/README.md new file mode 100644 index 000000000..b823288a9 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/README.md @@ -0,0 +1,88 @@ +# Record: 11L XSA-all + Full GPTQ (Budget-Legal) + Parallel Muon + Selective Pruning + +**val_bpb: 1.1178** (3-seed mean, std 0.0001) | **15.95 MB** max artifact | 8xH100 SXM, ~596s total compute + +## Update (2026-03-26) + +This PR was updated to fix a GPTQ budget violation identified in [issue #677](https://github.com/openai/parameter-golf/issues/677). The previous version trained for the full 600s, then ran GPTQ calibration for ~46s on top, exceeding the 600s artifact-production budget. The fix reserves 14s from the training budget for GPTQ calibration (`gptq_reserve_ms = 14000.0`), ensuring total compute (training ~586s + GPTQ ~10s = ~596s) stays within the 600s limit. All results below use the fixed code with fresh 3-seed runs. + +## Results (3 seeds, 8xH100 SXM) + +| Seed | Steps | ms/step | Sliding BPB (s64) | val_loss | Artifact | Train Time | GPTQ Time | Total | +|------|-------|---------|--------------------|----------|----------|------------|-----------|-------| +| 1337 | 6,674 | ~88 | **1.1177** | 1.8871 | 15,929,433 bytes | 586,128ms | 9,786ms | 595,915ms | +| 42 | 6,732 | ~87 | 1.1179 | 1.8875 | 15,949,353 bytes | 586,050ms | 9,792ms | 595,842ms | +| 7 | 6,731 | ~87 | 1.1179 | 1.8875 | 15,946,145 bytes | 586,066ms | 9,823ms | 595,889ms | + +**Mean: 1.1178 | Std: 0.0001** + +## Key Techniques + +### XSA on All 11 Layers +Standard practice applies Exclusive Self-Attention to only the last 4 layers. Applying to all 11 forces cross-position information mixing from layer 0, improving representation quality. Zero new parameters — just a config change. -0.0016 BPB vs XSA-last-4 in ablation. + +### Full Hessian GPTQ (Budget-Legal) +- 64-batch GPU Hessian calibration from training data +- Column-wise int6 quantization with Cholesky error compensation, block size 128, percdamp 0.01 +- QAT STE aligned to export quantizer using row-maximum (amax) clipping with [-32, 31] range +- **Budget reservation:** `gptq_reserve_ms = 14000.0` — training stops ~14s early so GPTQ calibration fits within 600s +- Log verification: `gptq:budget_check train:586128ms + gptq:9786ms = 595915ms (budget:600000ms)` + +### Parallel Muon Optimizer with Parameter Banking +- Weight matrices stored in contiguous parameter banks (qo_bank, kv_bank, mlp_up_bank, mlp_down_bank) +- 3-phase overlapped optimizer step: async reduce-scatter -> batched Newton-Schulz orthogonalization -> async all-gather +- Eliminates DDP double-communication overhead, achieving ~87ms/step (~6,700 steps in 586s) + +### Selective Magnitude Pruning +Post-GPTQ, sort quantized values at +/-1 by reconstruction error (scale^2), zero least-impactful first until artifact fits target. Binary search for exact target size. + +### LZMA Compression +LZMA preset 6 replacing zstd-22. Better compression ratio on int6 quantized weights. + +## Architecture + +- 11 transformer layers, dim=512, 8 heads, 4 KV heads (GQA) +- 3x MLP expansion (hidden=1536) with **LeakyReLU(0.5)^2** activation +- **XSA on all 11 layers** (Exclusive Self-Attention) +- Partial RoPE (16/64 dims) + NTK-aware scaling +- LN Scale Factor 1/sqrt(layer_idx+1) +- U-Net skip connections (5 encoder, 6 decoder) +- SmearGate temporal gating +- BigramHash (2048 buckets, 128-dim) +- Shared Value Embedding (dim=128, layers 9-10) +- FlashAttention 3 (Hopper native kernels) +- Orthogonal init, logit softcap 30, tied embeddings + +## Training + +- Parallel Muon optimizer (matrices): lr=0.025, momentum=0.99, WD=0.04, 5 Newton-Schulz steps +- AdamW (embeddings): lr=0.035, (scalars): lr=0.025, WD=0.04 +- Gradient clip: 0.3 +- Batch: 786,432 tokens/step, seq_len=2048 +- Warmdown: 3,500 iters (wallclock-based) +- EMA (decay=0.997) + Tight SWA (every 50 steps, scale<0.2) +- Late QAT: STE int6 fake-quantization when LR scale<0.15 + +## Quantization & Compression + +- Full GPTQ with 64-batch GPU Hessian calibration, block_size=128, percdamp=0.01 +- Int6 per-row with amax clipping, range [-32, 31] +- Selective magnitude pruning (target 15.9MB) +- Small tensors + tok_emb.weight in fp16 +- LZMA preset 6 compression + +## Compliance + +- [x] 3 seeds, all total compute <= 600s on 8xH100 SXM (verified: max 595,915ms) +- [x] GPTQ calibration WITHIN training budget (14s reserved, verified via `gptq:budget_check`) +- [x] All artifacts <= 16,000,000 bytes (max: 15,949,353) +- [x] No TTT on validation data +- [x] No training data accessed during evaluation +- [x] No network calls during evaluation +- [x] Sliding window eval stride=64, consistent across seeds (std=0.0001) + +## Run Command + +```bash +SEED=1337 TARGET_MB=15.9 torchrun --standalone --nproc_per_node=8 train_gpt.py +``` diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/submission.json b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/submission.json new file mode 100644 index 000000000..79fda9d2e --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/submission.json @@ -0,0 +1,19 @@ +{ + "author": "Raahil Shah", + "github_id": "raahilshah", + "name": "11L XSA-all + Full GPTQ (budget-legal) + Parallel Muon + LZMA + Selective Pruning", + "blurb": "XSA on all 11 layers, Hessian-aware GPTQ with 14s budget reservation (train 586s + GPTQ 10s = 596s total), amax-aligned QAT, Parallel Muon optimizer with parameter banking, LZMA compression, selective magnitude pruning. LeakyReLU(0.5)² activation, EMA(0.997), Tight SWA, VE128, Partial RoPE 16/64, LN Scale, BigramHash(2048), U-Net skips.", + "date": "2026-03-26T00:00:00Z", + "val_loss": 1.88736798, + "val_bpb": 1.11780859, + "pre_quant_val_loss": 1.9210, + "pre_quant_val_bpb": 1.1377, + "bytes_total": 15949353, + "seeds": { + "1337": {"val_bpb": 1.11765772, "val_loss": 1.88711325, "bytes_total": 15929433}, + "42": {"val_bpb": 1.11789104, "val_loss": 1.88750721, "bytes_total": 15949353}, + "7": {"val_bpb": 1.11787700, "val_loss": 1.88748353, "bytes_total": 15946145} + }, + "mean_val_bpb": 1.11780859, + "std_val_bpb": 0.00010683 +} diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train.log b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train.log new file mode 100644 index 000000000..11593307b --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train.log @@ -0,0 +1,88 @@ +logs/88f192b5-5eef-4f0f-99ee-b94b7e4b0298.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[0, 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:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9317 train_time:5842ms step_avg:5842.44ms +step:2/20000 train_loss:8.6935 train_time:5886ms step_avg:2943.25ms +step:3/20000 train_loss:7.5958 train_time:5971ms step_avg:1990.36ms +step:4/20000 train_loss:7.3348 train_time:6057ms step_avg:1514.13ms +step:5/20000 train_loss:7.2640 train_time:6141ms step_avg:1228.14ms +step:6/20000 train_loss:7.1178 train_time:6224ms step_avg:1037.40ms +step:7/20000 train_loss:6.9214 train_time:6308ms step_avg:901.16ms +step:8/20000 train_loss:6.8019 train_time:6392ms step_avg:798.97ms +step:9/20000 train_loss:6.4159 train_time:6476ms step_avg:719.52ms +step:10/20000 train_loss:6.0449 train_time:6560ms step_avg:656.01ms +step:500/20000 train_loss:2.3890 train_time:49041ms step_avg:98.08ms +step:1000/20000 train_loss:2.2584 train_time:92205ms step_avg:92.20ms +step:1500/20000 train_loss:2.2044 train_time:135515ms step_avg:90.34ms +step:2000/20000 train_loss:2.0479 train_time:178935ms step_avg:89.47ms +step:2500/20000 train_loss:2.1542 train_time:222374ms step_avg:88.95ms +step:3000/20000 train_loss:2.1427 train_time:265854ms step_avg:88.62ms +step:3500/20000 train_loss:2.1571 train_time:309328ms step_avg:88.38ms +step:4000/20000 train_loss:1.9474 train_time:352778ms step_avg:88.19ms +step:4000/20000 val_loss:2.0405 val_bpb:1.2085 train_time:352821ms step_avg:88.21ms +step:4500/20000 train_loss:2.0987 train_time:396236ms step_avg:88.05ms +step:5000/20000 train_loss:2.0803 train_time:439652ms step_avg:87.93ms +step:5500/20000 train_loss:1.9943 train_time:483113ms step_avg:87.84ms +swa:start step:6000 +step:6000/20000 train_loss:1.9186 train_time:526547ms step_avg:87.76ms +late_qat:enabled step:6150 scale:0.1499 +step:6500/20000 train_loss:2.0554 train_time:570717ms step_avg:87.80ms +step:6674/20000 val_loss:1.9226 val_bpb:1.1387 train_time:586128ms step_avg:87.82ms +stopping_early: wallclock_cap train_time:586128ms step:6674/20000 +peak memory allocated: 22861 MiB reserved: 23032 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9210 val_bpb:1.1377 eval_time:2075ms +Serialized model: 106158518 bytes +Code size: 113761 bytes +gptq:building calibration model... +gptq:calibrating with 64 training batches... +gptq:calibrated 68 layers in 9.8s +gptq:budget_check train:586128ms + gptq:9786ms = 595915ms (budget:600000ms) +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +selective_prune: 4110461 ±1 candidates, unpruned=15.19MB target=15.9MB +Serialized model int6+lzma: 15815672 bytes +Total submission size int6+lzma: 15929433 bytes +Total submission size int8+zlib: 15929433 bytes +final_int6_roundtrip val_loss:1.9268 val_bpb:1.1411 eval_time:49161ms +final_int6_roundtrip_exact val_loss:1.92677762 val_bpb:1.14114624 +final_int6_sliding_window val_loss:1.8871 val_bpb:1.1177 stride:64 eval_time:102894ms +final_int6_sliding_window_exact val_loss:1.88711325 val_bpb:1.11765772 +final_int8_zlib_roundtrip_exact val_loss:1.88711325 val_bpb:1.11765772 diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_gpt.py b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_gpt.py new file mode 100644 index 000000000..800924a2b --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_gpt.py @@ -0,0 +1,2246 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path +try: + import zstandard +except ImportError: + pass +_COMPRESSOR = "lzma" +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP +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", 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", 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", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) # XSA on all layers (PR #609: -0.0016 BPB) + 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)) + 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") + ngram_cache = bool(int(os.environ.get("NGRAM_CACHE", "0"))) + ngram_alpha = float(os.environ.get("NGRAM_ALPHA", "0.40")) + ngram_order = int(os.environ.get("NGRAM_ORDER", "7")) + ngram_min_count = int(os.environ.get("NGRAM_MIN_COUNT", "2")) + ngram_buckets = int(os.environ.get("NGRAM_BUCKETS", "4194304")) + ngram_min_order = int(os.environ.get("NGRAM_MIN_ORDER", "7")) + ngram_entropy = bool(int(os.environ.get("NGRAM_ENTROPY", "0"))) + ngram_ent_base = float(os.environ.get("NGRAM_ENT_BASE", "0.05")) + ngram_ent_range = float(os.environ.get("NGRAM_ENT_RANGE", "0.55")) + ngram_ent_scale = float(os.environ.get("NGRAM_ENT_SCALE", "2.0")) + ngram_ent_thresh = float(os.environ.get("NGRAM_ENT_THRESH", "4.0")) + ngram_order_adaptive = bool(int(os.environ.get("NGRAM_ORDER_ADAPTIVE", "0"))) + ngram_order_ent_slope = float(os.environ.get("NGRAM_ORDER_ENT_SLOPE", "0.25")) + ngram_order_alpha_mult = os.environ.get("NGRAM_ORDER_ALPHA_MULT", "") + ngram_order_min_count = os.environ.get("NGRAM_ORDER_MIN_COUNT", "") + ngram_chunk_tokens = int(os.environ.get("NGRAM_CHUNK_TOKENS", "262144")) + eval_diagnostics = bool(int(os.environ.get("EVAL_DIAGNOSTICS", "0"))) +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X +class Muon(torch.optim.Optimizer): + """Parallel Muon: reduce-scatter -> local batched NS5 -> all-gather. + Each param is a 3D bank tensor (B, M, N).""" + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather.""" + if not self._built: + self._build() + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + sharded = self._distributed and hasattr(self, '_rs_futures') + prev_ag_handle, prev_m = None, None + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + update = g.add(buf, alpha=momentum) if nesterov else buf.clone() + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + if hasattr(self, '_rs_futures'): + del self._rs_futures + return None +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_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def quantize_state_dict_int8(state_dict: dict[str, Tensor]): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + stats["num_float_tensors"] += 1 + q, s = quantize_float_tensor(t) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + out[name] = (q.float() * scale).to(dtype=dtype).contiguous() + for name, t in obj["passthrough"].items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out +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 + 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() + with torch.no_grad(): + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + x_norm = w32 / scale[:, None] + q_hard = torch.clamp(torch.round(x_norm), -32, 31) + w_q = (q_hard.detach() * 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, + use_banks: bool = True, + ): + 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") + if not use_banks: + kv_dim = 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 + 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, q_w: Tensor | None = None, k_w: Tensor | None = None, v_w: Tensor | None = None, out_w: Tensor | None = None, v_embed: Tensor | None = None) -> Tensor: + bsz, seqlen, dim = x.shape + if q_w is not None: + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + else: + 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] + 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 F.linear(y, out_w.to(x.dtype)) if out_w is not None else 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): + """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, use_banks: bool = True): + super().__init__() + if not use_banks: + 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, up_w: Tensor | None = None, down_w: Tensor | None = None) -> Tensor: + if up_w is not None: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) + 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, + use_banks: bool = True, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, use_banks=use_banks) + self.mlp = MLP(dim, mlp_mult, use_banks=use_banks) + 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, q_w: Tensor | None = None, k_w: Tensor | None = None, v_w: Tensor | None = None, out_w: Tensor | None = None, up_w: Tensor | None = None, down_w: Tensor | None = None, 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, q_w, k_w, v_w, out_w, 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, up_w, down_w) + 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, + 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_banks: bool = True, + ): + super().__init__() + self.use_banks = use_banks + 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.num_layers = num_layers + if use_banks: + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + 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, + use_banks=use_banks, + ) + 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() # 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) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init bank parameters (only when using banked architecture) + if self.use_banks: + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Init remaining non-bank modules + 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_(proj_scale) + 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 _run_blocks(self, x: Tensor, x0: Tensor, input_ids: Tensor) -> Tensor: + n = self.num_layers + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + if self.use_banks: + x = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve) + else: + 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) + if self.use_banks: + x = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve) + else: + x = self.blocks[bi](x, x0, v_embed=ve) + return x + 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 + x = self._run_blocks(x, x0, input_ids) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + x = self._run_blocks(x, x0, input_ids) + 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, + use_ngram: bool = False, +) -> 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) + if use_ngram: + assert 2 <= args.ngram_min_order <= args.ngram_order + assert args.ngram_buckets > 0 and (args.ngram_buckets & (args.ngram_buckets - 1)) == 0, "ngram_buckets must be power of two" + val_np = val_tokens.cpu().numpy() + n_orders = args.ngram_order - args.ngram_min_order + 1 + ctx_tables = [np.zeros((args.ngram_buckets,), dtype=np.uint32) for _ in range(n_orders)] + full_tables = [np.zeros((args.ngram_buckets,), dtype=np.uint32) for _ in range(n_orders)] + ng_mask = np.uint64(args.ngram_buckets - 1) + ng_primes = np.array( + [36313, 27191, 51647, 81929, 131071, 175447, 209591], + dtype=np.uint64, + ) + # Parse per-order alpha multipliers (if provided) + order_alpha_mult = None + if args.ngram_order_alpha_mult: + order_alpha_mult = np.array([float(x) for x in args.ngram_order_alpha_mult.split(",")], dtype=np.float64) + assert len(order_alpha_mult) == n_orders, f"ngram_order_alpha_mult length {len(order_alpha_mult)} != n_orders {n_orders}" + # Parse per-order minimum count schedule (if provided) + if args.ngram_order_min_count: + order_min_count = np.array([float(x) for x in args.ngram_order_min_count.split(",")], dtype=np.float64) + assert len(order_min_count) == n_orders, f"ngram_order_min_count length {len(order_min_count)} != n_orders {n_orders}" + else: + order_min_count = np.full(n_orders, float(args.ngram_min_count), dtype=np.float64) + if rank == 0: + mc_str = f"min_count_by_order={args.ngram_order_min_count}" if args.ngram_order_min_count else f"min_count={args.ngram_min_count}" + print( + f"ngram_cache:enabled orders={args.ngram_min_order}-{args.ngram_order} " + f"entropy={int(args.ngram_entropy)} alpha={args.ngram_alpha} " + f"ent_base={args.ngram_ent_base} ent_range={args.ngram_ent_range} " + f"ent_scale={args.ngram_ent_scale} ent_thresh={args.ngram_ent_thresh} " + f"order_adaptive={int(args.ngram_order_adaptive)} order_ent_slope={args.ngram_order_ent_slope} " + f"order_alpha_mult={args.ngram_order_alpha_mult or 'none'} " + f"{mc_str} buckets={args.ngram_buckets}", + flush=True, + ) + 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) + seg_len = wlen - s + if seg_len <= 0: + continue + scored_nll = nll[i, s:wlen].to(torch.float64) + if use_ngram: + seg_nll_np = scored_nll.cpu().numpy() + seg_model_p = np.exp(-seg_nll_np) + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + + if args.ngram_entropy: + lp = F.log_softmax(logits[i, s:wlen].float(), dim=-1) + seg_ent = -(lp.exp() * lp).sum(dim=-1).cpu().numpy() + alpha_per_tok = args.ngram_ent_base + args.ngram_ent_range / ( + 1.0 + np.exp(-args.ngram_ent_scale * (seg_ent - args.ngram_ent_thresh)) + ) + + order_data = [] + for oi in range(n_orders): + ctx_w = args.ngram_min_order + oi - 1 + valid = global_j >= ctx_w + if not valid.any(): + order_data.append(None) + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_w): + tok = val_np[jv - (ctx_w - k)].astype(np.uint64) + ctx_hash ^= tok * ng_primes[k % len(ng_primes)] + ctx_key = (ctx_hash & ng_mask).astype(np.int64) + tgt_np = val_np[jv].astype(np.uint64) + full_key = ( + (ctx_hash ^ (tgt_np * ng_primes[ctx_w % len(ng_primes)])) & ng_mask + ).astype(np.int64) + order_data.append((v_idx, ctx_key, full_key)) + + best_p_ng = np.full(seg_len, -1.0, dtype=np.float64) + best_order_ng = np.full(seg_len, args.ngram_min_order - 1, dtype=np.int32) + for oi in range(n_orders - 1, -1, -1): + data = order_data[oi] + if data is None: + continue + v_idx, ctx_key, full_key = data + ctx_counts = ctx_tables[oi][ctx_key].astype(np.float64) + full_counts = full_tables[oi][full_key].astype(np.float64) + has_match = ctx_counts >= order_min_count[oi] + needs_fill = has_match & (best_p_ng[v_idx] < 0.0) + if needs_fill.any(): + fill_idx = v_idx[needs_fill] + p = np.minimum(full_counts[needs_fill], ctx_counts[needs_fill]) / np.maximum(ctx_counts[needs_fill], 1.0) + best_p_ng[fill_idx] = np.clip(p, 0.0, 1.0) + best_order_ng[fill_idx] = args.ngram_min_order + oi + + has_match = best_p_ng >= 0.0 + if has_match.any(): + if not args.ngram_entropy: + alpha = args.ngram_alpha + elif not args.ngram_order_adaptive: + alpha = alpha_per_tok[has_match] + else: + order_thresh = args.ngram_ent_thresh - args.ngram_order_ent_slope * (best_order_ng[has_match].astype(np.float64) - float(args.ngram_min_order)) + alpha = args.ngram_ent_base + args.ngram_ent_range / (1.0 + np.exp(-args.ngram_ent_scale * (seg_ent[has_match] - order_thresh))) + # Apply per-order alpha multiplier if enabled + if order_alpha_mult is not None: + matched_order_idx = best_order_ng[has_match] - args.ngram_min_order + alpha = alpha * order_alpha_mult[matched_order_idx] + alpha = np.clip(alpha, 0.0, 0.95) + seg_model_p[has_match] = (1.0 - alpha) * seg_model_p[has_match] + alpha * best_p_ng[has_match] + seg_nll_np = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + + # Legal score-first update: update every order only after scoring. + for oi in range(n_orders): + data = order_data[oi] + if data is None: + continue + _, ctx_key, full_key = data + np.add.at(ctx_tables[oi], ctx_key, 1) + np.add.at(full_tables[oi], full_key, 1) + + scored_nll = torch.from_numpy(seg_nll_np).to(dtype=torch.float64, device=device) + 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 _build_sliding_segments(total_tokens, seq_len, stride): + """Return list of (ws, wlen, s, tgt_start, tgt_end) for sliding window segments. + tgt_start/tgt_end define the scored target position range [tgt_start, tgt_end).""" + segments = [] + for ws in range(0, total_tokens, stride): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + if wlen < 1: + continue + s = 0 if ws == 0 else max(wlen - stride, 0) + if wlen - s <= 0: + continue + tgt_start = ws + s + 1 + tgt_end = ws + wlen + 1 + segments.append((ws, wlen, s, tgt_start, tgt_end)) + return segments +def eval_val_sliding_chunked_ngram( + 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]: + """Chunk-synchronized n-gram cache evaluator. All ranks share the same global-prefix cache.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + assert args.ngram_chunk_tokens >= seq_len, \ + f"ngram_chunk_tokens={args.ngram_chunk_tokens} < eval_seq_len={seq_len}" + assert (args.ngram_chunk_tokens - seq_len) % stride == 0, \ + f"(ngram_chunk_tokens - eval_seq_len) % stride != 0" + segments = _build_sliding_segments(total_tokens, seq_len, stride) + # N-gram setup (same as eval_val_sliding) + assert 2 <= args.ngram_min_order <= args.ngram_order + assert args.ngram_buckets > 0 and (args.ngram_buckets & (args.ngram_buckets - 1)) == 0 + val_np = val_tokens.cpu().numpy() + n_orders = args.ngram_order - args.ngram_min_order + 1 + ctx_tables = [np.zeros((args.ngram_buckets,), dtype=np.uint32) for _ in range(n_orders)] + full_tables = [np.zeros((args.ngram_buckets,), dtype=np.uint32) for _ in range(n_orders)] + ng_mask = np.uint64(args.ngram_buckets - 1) + ng_primes = np.array([36313, 27191, 51647, 81929, 131071, 175447, 209591], dtype=np.uint64) + order_alpha_mult = None + if args.ngram_order_alpha_mult: + order_alpha_mult = np.array([float(x) for x in args.ngram_order_alpha_mult.split(",")], dtype=np.float64) + assert len(order_alpha_mult) == n_orders + if args.ngram_order_min_count: + order_min_count = np.array([float(x) for x in args.ngram_order_min_count.split(",")], dtype=np.float64) + assert len(order_min_count) == n_orders + else: + order_min_count = np.full(n_orders, float(args.ngram_min_count), dtype=np.float64) + if rank == 0: + mc_str = f"min_count_by_order={args.ngram_order_min_count}" if args.ngram_order_min_count else f"min_count={args.ngram_min_count}" + print( + f"ngram_chunksync chunk_tokens={args.ngram_chunk_tokens} orders={args.ngram_min_order}-{args.ngram_order} buckets={args.ngram_buckets}", + flush=True, + ) + print( + f"ngram_cache:enabled orders={args.ngram_min_order}-{args.ngram_order} " + f"entropy={int(args.ngram_entropy)} alpha={args.ngram_alpha} " + f"ent_base={args.ngram_ent_base} ent_range={args.ngram_ent_range} " + f"ent_scale={args.ngram_ent_scale} ent_thresh={args.ngram_ent_thresh} " + f"order_adaptive={int(args.ngram_order_adaptive)} order_ent_slope={args.ngram_order_ent_slope} " + f"order_alpha_mult={args.ngram_order_alpha_mult or 'none'} " + f"{mc_str} buckets={args.ngram_buckets}", + flush=True, + ) + 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) + seg_cursor = 0 + with torch.inference_mode(): + for chunk_start in range(1, total_tokens + 1, args.ngram_chunk_tokens): + chunk_end = min(chunk_start + args.ngram_chunk_tokens, total_tokens + 1) + # Collect segments whose scored range is fully within this chunk + chunk_segments = [] + while seg_cursor < len(segments) and segments[seg_cursor][4] <= chunk_end: + if segments[seg_cursor][3] >= chunk_start: + chunk_segments.append(segments[seg_cursor]) + seg_cursor += 1 + # Distribute across ranks (round-robin) + rank_segments = chunk_segments[rank::world_size] + for bi in range(0, len(rank_segments), batch_seqs): + batch_segs = rank_segments[bi:bi + batch_seqs] + bsz = len(batch_segs) + 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 = [] + seg_s_list = [] + seg_ws_list = [] + for i, (ws, wlen, s, _, _) in enumerate(batch_segs): + wlens.append(wlen) + seg_s_list.append(s) + seg_ws_list.append(ws) + chunk_data = val_tokens[ws:ws + wlen + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_data[:-1] + y_batch[i, :wlen] = chunk_data[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 in range(bsz): + ws = seg_ws_list[i] + wlen = wlens[i] + s = seg_s_list[i] + seg_len = wlen - s + if seg_len <= 0: + continue + scored_nll = nll[i, s:wlen].to(torch.float64) + seg_nll_np = scored_nll.cpu().numpy() + seg_model_p = np.exp(-seg_nll_np) + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + if args.ngram_entropy: + lp = F.log_softmax(logits[i, s:wlen].float(), dim=-1) + seg_ent = -(lp.exp() * lp).sum(dim=-1).cpu().numpy() + alpha_per_tok = args.ngram_ent_base + args.ngram_ent_range / ( + 1.0 + np.exp(-args.ngram_ent_scale * (seg_ent - args.ngram_ent_thresh)) + ) + order_data = [] + for oi in range(n_orders): + ctx_w = args.ngram_min_order + oi - 1 + valid = global_j >= ctx_w + if not valid.any(): + order_data.append(None) + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_w): + tok = val_np[jv - (ctx_w - k)].astype(np.uint64) + ctx_hash ^= tok * ng_primes[k % len(ng_primes)] + ctx_key = (ctx_hash & ng_mask).astype(np.int64) + tgt_np_arr = val_np[jv].astype(np.uint64) + full_key = ( + (ctx_hash ^ (tgt_np_arr * ng_primes[ctx_w % len(ng_primes)])) & ng_mask + ).astype(np.int64) + order_data.append((v_idx, ctx_key, full_key)) + best_p_ng = np.full(seg_len, -1.0, dtype=np.float64) + best_order_ng = np.full(seg_len, args.ngram_min_order - 1, dtype=np.int32) + for oi in range(n_orders - 1, -1, -1): + data = order_data[oi] + if data is None: + continue + v_idx, ctx_key, full_key = data + ctx_counts = ctx_tables[oi][ctx_key].astype(np.float64) + full_counts = full_tables[oi][full_key].astype(np.float64) + has_match = ctx_counts >= order_min_count[oi] + needs_fill = has_match & (best_p_ng[v_idx] < 0.0) + if needs_fill.any(): + fill_idx = v_idx[needs_fill] + p = np.minimum(full_counts[needs_fill], ctx_counts[needs_fill]) / np.maximum(ctx_counts[needs_fill], 1.0) + best_p_ng[fill_idx] = np.clip(p, 0.0, 1.0) + best_order_ng[fill_idx] = args.ngram_min_order + oi + has_match = best_p_ng >= 0.0 + if has_match.any(): + if not args.ngram_entropy: + alpha = args.ngram_alpha + elif not args.ngram_order_adaptive: + alpha = alpha_per_tok[has_match] + else: + order_thresh = args.ngram_ent_thresh - args.ngram_order_ent_slope * (best_order_ng[has_match].astype(np.float64) - float(args.ngram_min_order)) + alpha = args.ngram_ent_base + args.ngram_ent_range / (1.0 + np.exp(-args.ngram_ent_scale * (seg_ent[has_match] - order_thresh))) + if order_alpha_mult is not None: + matched_order_idx = best_order_ng[has_match] - args.ngram_min_order + alpha = alpha * order_alpha_mult[matched_order_idx] + alpha = np.clip(alpha, 0.0, 0.95) + seg_model_p[has_match] = (1.0 - alpha) * seg_model_p[has_match] + alpha * best_p_ng[has_match] + seg_nll_np = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + scored_nll = torch.from_numpy(seg_nll_np).to(dtype=torch.float64, device=device) + loss_sum += scored_nll.sum() + token_count += float(seg_len) + 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() + # Chunk-synchronized cache update: every rank updates with full chunk range + for oi in range(n_orders): + ctx_w = args.ngram_min_order + oi - 1 + j_range = np.arange(max(chunk_start, ctx_w), chunk_end, dtype=np.int64) + if len(j_range) == 0: + continue + ctx_hash = np.zeros(len(j_range), dtype=np.uint64) + for k in range(ctx_w): + tok = val_np[j_range - (ctx_w - k)].astype(np.uint64) + ctx_hash ^= tok * ng_primes[k % len(ng_primes)] + ctx_key = (ctx_hash & ng_mask).astype(np.int64) + tgt_arr = val_np[j_range].astype(np.uint64) + full_key = ( + (ctx_hash ^ (tgt_arr * ng_primes[ctx_w % len(ng_primes)])) & ng_mask + ).astype(np.int64) + np.add.at(ctx_tables[oi], ctx_key, 1) + np.add.at(full_tables[oi], full_key, 1) + 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 _unbank_state_dict(banked_sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert banked 3D params to per-layer 2D params for GPTQ/serialization.""" + out = {} + n = num_layers + for name, t in banked_sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = t[i] + out[f"blocks.{i}.attn.proj.weight"] = t[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = t[i] + out[f"blocks.{i}.attn.c_v.weight"] = t[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = t[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = t[i] + else: + out[name] = t + return out +def _rebank_state_dict(unbanked_sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert per-layer 2D params back to banked 3D params.""" + out = {} + n = num_layers + qo_slices, kv_slices, up_slices, down_slices = [], [], [], [] + for name, t in unbanked_sd.items(): + if ".attn.c_q.weight" in name or ".attn.c_k.weight" in name or \ + ".attn.c_v.weight" in name or ".attn.proj.weight" in name or \ + ".mlp.fc.weight" in name or ".mlp.proj.weight" in name: + continue # handled below + out[name] = t + for i in range(n): + qo_slices.append(unbanked_sd[f"blocks.{i}.attn.c_q.weight"]) + for i in range(n): + qo_slices.append(unbanked_sd[f"blocks.{i}.attn.proj.weight"]) + for i in range(n): + kv_slices.append(unbanked_sd[f"blocks.{i}.attn.c_k.weight"]) + for i in range(n): + kv_slices.append(unbanked_sd[f"blocks.{i}.attn.c_v.weight"]) + for i in range(n): + up_slices.append(unbanked_sd[f"blocks.{i}.mlp.fc.weight"]) + for i in range(n): + down_slices.append(unbanked_sd[f"blocks.{i}.mlp.proj.weight"]) + out["qo_bank"] = torch.stack(qo_slices) + out["kv_bank"] = torch.stack(kv_slices) + out["mlp_up_bank"] = torch.stack(up_slices) + out["mlp_down_bank"] = torch.stack(down_slices) + return out +def gptq_calibrate(calib_model: nn.Module, train_pattern: str, rank: int, world_size: int, + device: torch.device, num_batches: int = 256, + batch_tokens: int = 786432, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X using training batches. Accumulates on GPU for speed.""" + hessians: dict[str, Tensor] = {} + hooks = [] + grad_accum_steps = max(1, 8 // world_size) + # Pre-initialize Hessians on GPU for fast accumulation + for name, module in calib_model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + cols = module.weight.shape[1] + hessians[name] = torch.zeros(cols, cols, dtype=torch.float32, device=device) + 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]) + hessians[name] += x.t() @ x + return hook_fn + for name, module in calib_model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + loader = DistributedTokenLoader(train_pattern, rank, world_size, device) + calib_model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = loader.next_batch(batch_tokens, seq_len, grad_accum_steps) + calib_model(x, y) + for h in hooks: + h.remove() + # Move to CPU, normalize, and add damping + for name in hessians: + hessians[name] = hessians[name].cpu() / num_batches + damp = 0.01 * torch.diag(hessians[name]).mean().clamp_min(1e-6) + hessians[name] += damp * torch.eye(hessians[name].shape[0]) + calib_model.train() + return hessians +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 128, percdamp: float = 0.01) -> tuple[Tensor, Tensor]: + """GPTQ quantize weight W using Hessian H — reference implementation. + Uses actorder (descending Hessian diagonal), Cholesky of H^{-1}, and multi-scale search.""" + W_orig = W.float().clone() + rows, cols = W_orig.shape + H = H.float().clone() + # Kill dead columns + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + # Damping + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + # Actorder: process most important (highest activation) columns first + perm = torch.argsort(H.diag(), descending=True) + invperm = torch.argsort(perm) + W_perm = W_orig[:, perm].clone() + W_perm[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} (per GPTQ Algorithm 1) + try: + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except torch.linalg.LinAlgError: + return quantize_int6_per_row(W_orig, clip_range) + 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(W_orig.abs(), pct, dim=1) + else: + row_clip = W_orig.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + sf = s.float() + Q = torch.zeros(rows, cols, dtype=torch.int8) + W_work = W_perm.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W_work[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros(rows, i2 - i1) + for j in range(i2 - i1): + w_col = W_block[:, j] + d = Hinv_block[j, j] + q_col = torch.clamp(torch.round(w_col / sf), -clip_range, clip_range) + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - q_col.float() * sf) / d + Err[:, j] = err + W_block[:, j:] -= err.unsqueeze(1) * Hinv_block[j, j:].unsqueeze(0) + if i2 < cols: + W_work[:, i2:] -= Err @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W_perm - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, invperm] + return best_q, best_scale +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + 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 _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 mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count = 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 + if cat in int6_cats and t.ndim >= 1: + # Try GPTQ if Hessian available for this layer + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) if hessians else None + if H is not None and t.ndim == 2: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + 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"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers", flush=True) + 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, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + # Bank parameters in float32 for optimizer + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + # No DDP — Parallel Muon handles bank communication via reduce-scatter, + # non-bank grads are manually all-reduced before Adam steps. + model = compiled_model + block_named_params = list(base_model.blocks.named_parameters()) + # 4 bank parameters -> Muon (batched Newton-Schulz) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + # Everything else in blocks -> Adam (all block params are scalar/control now) + extra_adam_2d = [] + if base_model.mtp_num_heads > 0: + extra_adam_2d.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) + scalar_params = [p for _, p in block_named_params] + 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: + extra_adam_2d.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: + extra_adam_2d.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) + scalar_params.extend(extra_adam_2d) + 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] + optimizer_head = None + 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) + # Non-bank params that need manual all-reduce (no DDP) + replicated_params: list[nn.Parameter] = [] + for pg in optimizer_tok.param_groups: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + if base_model.lm_head is not None: + replicated_params.append(base_model.lm_head.weight) + # Pre-allocate flat buffer for single all-reduce (avoids 60+ per-param calls) + rep_total = sum(p.numel() for p in replicated_params) + rep_grad_buf = torch.zeros(rep_total, device=device, dtype=torch.bfloat16) if distributed else None + 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}") + 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 + # Reserve time within training budget for GPTQ calibration (must fit in 600s total) + gptq_reserve_ms = 14000.0 + train_budget_ms = (max_wallclock_ms - gptq_reserve_ms) if max_wallclock_ms else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if train_budget_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) + if step < 50: + step_ms = min(step_ms, 150.0) # protect against first-step compilation spike + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(train_budget_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + skip_training = bool(int(os.environ.get("SKIP_TRAINING", "0"))) + if args.warmup_steps > 0 and not skip_training: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # Warmup: simple all-reduce for all grads (not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + 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() + 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: + if skip_training: + break + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + frac = 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) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for Muon banks + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while RS is in-flight) + if distributed: + off = 0 + for p in replicated_params: + n = p.numel() + rep_grad_buf[off:off+n].copy_(p.grad.reshape(-1).bfloat16() if p.grad is not None else rep_grad_buf[off:off+n].zero_()) + off += n + dist.all_reduce(rep_grad_buf, op=dist.ReduceOp.AVG) + off = 0 + for p in replicated_params: + n = p.numel() + if p.grad is not None: + p.grad.copy_(rep_grad_buf[off:off+n].reshape(p.shape).to(dtype=p.dtype)) + off += n + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local batched NS5, all-gather (banks last) + optimizer_muon.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 = train_budget_ms is not None and approx_training_time_ms >= train_budget_ms + if distributed and train_budget_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 + if not skip_training: + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # Apply EMA weights (better than SWA alone per PR#401) + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + 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()} + # Unbank state dict for GPTQ and serialization + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + # GPTQ calibration: build non-banked model for Hessian collection + log0("gptq:building calibration model...") + t_gptq = time.perf_counter() + calib_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, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_banks=False, + ).to(device).bfloat16() + for m in calib_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(calib_model) + calib_model.load_state_dict( + {k: v.to(device) for k, v in unbanked_sd.items() if k in calib_model.state_dict()}, + strict=False, + ) + log0("gptq:calibrating with 64 training batches...") + gptq_hessians = gptq_calibrate(calib_model, args.train_files, rank, world_size, device, + num_batches=64, batch_tokens=args.train_batch_tokens, seq_len=args.train_seq_len) + gptq_elapsed_s = time.perf_counter() - t_gptq + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {gptq_elapsed_s:.1f}s") + total_compute_ms = training_time_ms + gptq_elapsed_s * 1000.0 + log0(f"gptq:budget_check train:{training_time_ms:.0f}ms + gptq:{gptq_elapsed_s*1000:.0f}ms = {total_compute_ms:.0f}ms (budget:{max_wallclock_ms:.0f}ms)") + del calib_model + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}, gptq_hessians) + code_bytes = len(code.encode("utf-8")) + target_mb = float(os.environ.get("TARGET_MB", "15.9")) + target_bytes = int(target_mb * 1024 * 1024) + # Selective ±1 pruning: zero least-impactful |q|=1 entries to fit target (PR #609) + ones_info = [] # (tensor_key, flat_idx, error=scale²) + for name, info in quant_meta.items(): + if not (isinstance(info, dict) and info.get("type") == "int6"): + continue + qk, sk = name + ".q", name + ".scale" + if qk not in quant_result or sk not in quant_result: + continue + q, s = quant_result[qk], quant_result[sk] + if s.ndim > 0: + ones_mask = (q.abs() == 1) + if ones_mask.any(): + row_idx = torch.arange(q.shape[0]).unsqueeze(1).expand_as(q)[ones_mask] + flat_idx = torch.arange(q.numel()).reshape(q.shape)[ones_mask] + errors = s.float()[row_idx].pow(2) + for fi, err in zip(flat_idx.tolist(), errors.tolist()): + ones_info.append((qk, fi, err)) + ones_info.sort(key=lambda x: x[2]) # sort by error ascending (prune least impactful first) + def _compress_artifact(qr): + buf = io.BytesIO() + torch.save({"w": qr, "m": quant_meta}, buf) + raw = buf.getvalue() + blob = lzma.compress(raw, preset=6) + return len(blob) + code_bytes, blob + def _try_prune(n): + tmp = {k: v.clone() for k, v in quant_result.items()} + for i in range(min(n, len(ones_info))): + tmp[ones_info[i][0]].view(-1)[ones_info[i][1]] = 0 + return _compress_artifact(tmp) + unpruned_size, quant_blob = _compress_artifact(quant_result) + log0(f"selective_prune: {len(ones_info)} ±1 candidates, unpruned={unpruned_size/(1024*1024):.2f}MB target={target_mb}MB") + if unpruned_size > target_bytes and ones_info: + full_size, _ = _try_prune(len(ones_info)) + log0(f"selective_prune: full ±1 prune={full_size/(1024*1024):.2f}MB") + if full_size > target_bytes: + log0("selective_prune: even full prune not enough, applying all") + _, quant_blob = _try_prune(len(ones_info)) + for i in range(len(ones_info)): + quant_result[ones_info[i][0]].view(-1)[ones_info[i][1]] = 0 + else: + lo, hi = 0, len(ones_info) + while lo < hi: + mid = (lo + hi) // 2 + sz, _ = _try_prune(mid) + if sz <= target_bytes: + hi = mid + else: + lo = mid + 1 + log0(f"selective_prune: pruning {lo}/{len(ones_info)} ±1 values ({100*lo/len(ones_info):.1f}%) to fit {target_mb}MB") + _, quant_blob = _try_prune(lo) + for i in range(lo): + quant_result[ones_info[i][0]].view(-1)[ones_info[i][1]] = 0 + 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") + log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") + else: + # Eval-only: create template state dict for dequantization (shapes/dtypes match) + log0("SKIP_TRAINING: creating template state dict for eval") + _fresh_sd = base_model.state_dict() + _export_sd = {k: v for k, v in _fresh_sd.items() if "mtp_heads" not in k} + unbanked_sd = _unbank_state_dict({k: v.detach().cpu() for k, v in _export_sd.items()}, args.num_layers) + 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"], unbanked_sd) + 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_banks=False, # eval uses non-banked (per-layer) weights from quantized artifact + ).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) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) + sw_seq_len = effective_eval_seq_len + run_diagnostics = not args.ngram_cache or args.eval_diagnostics + if run_diagnostics: + 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}") + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + if run_diagnostics: + 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, + use_ngram=False, + ) + 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}") + if not args.ngram_cache: + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.ngram_cache: + # Save current n-gram args + saved_ng = {k: getattr(args, k) for k in [ + 'ngram_cache', 'ngram_order', 'ngram_min_order', 'ngram_min_count', + 'ngram_buckets', 'ngram_alpha', 'ngram_entropy', 'ngram_ent_base', + 'ngram_ent_range', 'ngram_ent_scale', 'ngram_ent_thresh', + 'ngram_order_adaptive', 'ngram_order_ent_slope', 'ngram_order_alpha_mult', + 'ngram_order_min_count', 'ngram_chunk_tokens', + ]} + # Pin shared n-gram settings + args.ngram_cache = True + args.ngram_order = 7 + args.ngram_min_order = 2 + args.ngram_min_count = 2 + args.ngram_buckets = 4194304 + args.ngram_alpha = 0.40 + args.ngram_ent_base = 0.05 + args.ngram_ent_range = 0.55 + args.ngram_ent_scale = 2.0 + + # Run 1: Multi-order backoff with FIXED alpha (safest legal variant) + if run_diagnostics: + args.ngram_entropy = False + args.ngram_order_adaptive = False + args.ngram_order_alpha_mult = "" + args.ngram_order_min_count = "" + torch.cuda.synchronize() + t_ng_fixed = time.perf_counter() + ng_fixed_loss, ng_fixed_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, + use_ngram=True, + ) + torch.cuda.synchronize() + log0( + f"final_ngram_fixed_alpha val_loss:{ng_fixed_loss:.4f} val_bpb:{ng_fixed_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ng_fixed):.0f}ms" + ) + log0(f"final_ngram_fixed_alpha_exact val_loss:{ng_fixed_loss:.8f} val_bpb:{ng_fixed_bpb:.8f}") + + # Run 2: Chunk-synchronized entropy-adaptive 12-gram (experimental) + args.ngram_order = 12 + args.ngram_min_order = 2 + args.ngram_entropy = True + args.ngram_ent_thresh = 3.0 + args.ngram_order_adaptive = True + args.ngram_order_ent_slope = 0.25 + args.ngram_ent_base = 0.05 + args.ngram_ent_range = 0.55 + args.ngram_ent_scale = 2.0 + args.ngram_buckets = 4194304 + args.ngram_order_alpha_mult = "0.30,0.30,0.97,2.00,2.00,2.00,1.50,1.50,1.50,1.50,1.50" + args.ngram_order_min_count = "2,2,2,2,2,2,1,1,1,1,1" + args.ngram_chunk_tokens = 16384 + torch.cuda.synchronize() + t_ng_ent = time.perf_counter() + ng_ent_loss, ng_ent_bpb = eval_val_sliding_chunked_ngram( + 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_ngram_entropy_order_adaptive_order_scale_12gram_hiorder_mc1_chunksync_c16384 val_loss:{ng_ent_loss:.4f} val_bpb:{ng_ent_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ng_ent):.0f}ms" + ) + log0(f"final_ngram_entropy_order_adaptive_order_scale_12gram_hiorder_mc1_chunksync_c16384_exact val_loss:{ng_ent_loss:.8f} val_bpb:{ng_ent_bpb:.8f}") + + # Restore saved n-gram args + for k, v in saved_ng.items(): + setattr(args, k, v) + + # Point canonical score at order-adaptive result; fixed-alpha is the invariant check + log0(f"final_int8_zlib_roundtrip_exact val_loss:{ng_ent_loss:.8f} val_bpb:{ng_ent_bpb:.8f}") + if run_diagnostics and args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed1337.log b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed1337.log new file mode 100644 index 000000000..11593307b --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed1337.log @@ -0,0 +1,88 @@ +logs/88f192b5-5eef-4f0f-99ee-b94b7e4b0298.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[0, 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:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9309 val_bpb:4.1049 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9317 train_time:5842ms step_avg:5842.44ms +step:2/20000 train_loss:8.6935 train_time:5886ms step_avg:2943.25ms +step:3/20000 train_loss:7.5958 train_time:5971ms step_avg:1990.36ms +step:4/20000 train_loss:7.3348 train_time:6057ms step_avg:1514.13ms +step:5/20000 train_loss:7.2640 train_time:6141ms step_avg:1228.14ms +step:6/20000 train_loss:7.1178 train_time:6224ms step_avg:1037.40ms +step:7/20000 train_loss:6.9214 train_time:6308ms step_avg:901.16ms +step:8/20000 train_loss:6.8019 train_time:6392ms step_avg:798.97ms +step:9/20000 train_loss:6.4159 train_time:6476ms step_avg:719.52ms +step:10/20000 train_loss:6.0449 train_time:6560ms step_avg:656.01ms +step:500/20000 train_loss:2.3890 train_time:49041ms step_avg:98.08ms +step:1000/20000 train_loss:2.2584 train_time:92205ms step_avg:92.20ms +step:1500/20000 train_loss:2.2044 train_time:135515ms step_avg:90.34ms +step:2000/20000 train_loss:2.0479 train_time:178935ms step_avg:89.47ms +step:2500/20000 train_loss:2.1542 train_time:222374ms step_avg:88.95ms +step:3000/20000 train_loss:2.1427 train_time:265854ms step_avg:88.62ms +step:3500/20000 train_loss:2.1571 train_time:309328ms step_avg:88.38ms +step:4000/20000 train_loss:1.9474 train_time:352778ms step_avg:88.19ms +step:4000/20000 val_loss:2.0405 val_bpb:1.2085 train_time:352821ms step_avg:88.21ms +step:4500/20000 train_loss:2.0987 train_time:396236ms step_avg:88.05ms +step:5000/20000 train_loss:2.0803 train_time:439652ms step_avg:87.93ms +step:5500/20000 train_loss:1.9943 train_time:483113ms step_avg:87.84ms +swa:start step:6000 +step:6000/20000 train_loss:1.9186 train_time:526547ms step_avg:87.76ms +late_qat:enabled step:6150 scale:0.1499 +step:6500/20000 train_loss:2.0554 train_time:570717ms step_avg:87.80ms +step:6674/20000 val_loss:1.9226 val_bpb:1.1387 train_time:586128ms step_avg:87.82ms +stopping_early: wallclock_cap train_time:586128ms step:6674/20000 +peak memory allocated: 22861 MiB reserved: 23032 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9210 val_bpb:1.1377 eval_time:2075ms +Serialized model: 106158518 bytes +Code size: 113761 bytes +gptq:building calibration model... +gptq:calibrating with 64 training batches... +gptq:calibrated 68 layers in 9.8s +gptq:budget_check train:586128ms + gptq:9786ms = 595915ms (budget:600000ms) +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +selective_prune: 4110461 ±1 candidates, unpruned=15.19MB target=15.9MB +Serialized model int6+lzma: 15815672 bytes +Total submission size int6+lzma: 15929433 bytes +Total submission size int8+zlib: 15929433 bytes +final_int6_roundtrip val_loss:1.9268 val_bpb:1.1411 eval_time:49161ms +final_int6_roundtrip_exact val_loss:1.92677762 val_bpb:1.14114624 +final_int6_sliding_window val_loss:1.8871 val_bpb:1.1177 stride:64 eval_time:102894ms +final_int6_sliding_window_exact val_loss:1.88711325 val_bpb:1.11765772 +final_int8_zlib_roundtrip_exact val_loss:1.88711325 val_bpb:1.11765772 diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed42.log b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed42.log new file mode 100644 index 000000000..731fad330 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed42.log @@ -0,0 +1,88 @@ +logs/ebac4837-194f-46f9-b7ce-2d1edc621560.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[0, 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:600.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9297 val_bpb:4.1042 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:6.9319 train_time:554ms step_avg:554.08ms +step:2/20000 train_loss:8.6636 train_time:601ms step_avg:300.45ms +step:3/20000 train_loss:7.6340 train_time:684ms step_avg:228.09ms +step:4/20000 train_loss:7.3769 train_time:769ms step_avg:192.37ms +step:5/20000 train_loss:7.2666 train_time:853ms step_avg:170.64ms +step:6/20000 train_loss:7.0297 train_time:937ms step_avg:156.21ms +step:7/20000 train_loss:6.8713 train_time:1021ms step_avg:145.87ms +step:8/20000 train_loss:6.8220 train_time:1105ms step_avg:138.16ms +step:9/20000 train_loss:6.4674 train_time:1189ms step_avg:132.08ms +step:10/20000 train_loss:6.0499 train_time:1273ms step_avg:127.29ms +step:500/20000 train_loss:2.4025 train_time:43447ms step_avg:86.89ms +step:1000/20000 train_loss:2.2665 train_time:86665ms step_avg:86.67ms +step:1500/20000 train_loss:2.2093 train_time:130055ms step_avg:86.70ms +step:2000/20000 train_loss:2.0536 train_time:173556ms step_avg:86.78ms +step:2500/20000 train_loss:2.1570 train_time:217082ms step_avg:86.83ms +step:3000/20000 train_loss:2.1477 train_time:260589ms step_avg:86.86ms +step:3500/20000 train_loss:2.1640 train_time:304077ms step_avg:86.88ms +step:4000/20000 train_loss:1.9533 train_time:347534ms step_avg:86.88ms +step:4000/20000 val_loss:2.0446 val_bpb:1.2109 train_time:347578ms step_avg:86.89ms +step:4500/20000 train_loss:2.1046 train_time:391010ms step_avg:86.89ms +step:5000/20000 train_loss:2.0845 train_time:434488ms step_avg:86.90ms +step:5500/20000 train_loss:1.9995 train_time:477952ms step_avg:86.90ms +step:6000/20000 train_loss:1.9209 train_time:521443ms step_avg:86.91ms +swa:start step:6050 +late_qat:enabled step:6214 scale:0.1499 +step:6500/20000 train_loss:2.0585 train_time:565564ms step_avg:87.01ms +step:6732/20000 val_loss:1.9232 val_bpb:1.1390 train_time:586050ms step_avg:87.05ms +stopping_early: wallclock_cap train_time:586050ms step:6732/20000 +peak memory allocated: 22851 MiB reserved: 23004 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9215 val_bpb:1.1380 eval_time:2076ms +Serialized model: 106158518 bytes +Code size: 113761 bytes +gptq:building calibration model... +gptq:calibrating with 64 training batches... +gptq:calibrated 68 layers in 9.8s +gptq:budget_check train:586050ms + gptq:9792ms = 595842ms (budget:600000ms) +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +selective_prune: 4082706 ±1 candidates, unpruned=15.21MB target=15.9MB +Serialized model int6+lzma: 15835592 bytes +Total submission size int6+lzma: 15949353 bytes +Total submission size int8+zlib: 15949353 bytes +final_int6_roundtrip val_loss:1.9272 val_bpb:1.1414 eval_time:17731ms +final_int6_roundtrip_exact val_loss:1.92716514 val_bpb:1.14137575 +final_int6_sliding_window val_loss:1.8875 val_bpb:1.1179 stride:64 eval_time:85533ms +final_int6_sliding_window_exact val_loss:1.88750721 val_bpb:1.11789104 +final_int8_zlib_roundtrip_exact val_loss:1.88750721 val_bpb:1.11789104 diff --git a/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed7.log b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed7.log new file mode 100644 index 000000000..d598381f0 --- /dev/null +++ b/records/track_10min_16mb/2026-03-24_11L_XSA-all_GPTQ_ParallelMuon_1.1171/train_seed7.log @@ -0,0 +1,88 @@ +logs/2261cece-fe41-441c-bac2-0fc02aa42186.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:26993756 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[0, 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:600.000 +seed:7 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:6.9296 val_bpb:4.1041 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:6.9310 train_time:557ms step_avg:556.92ms +step:2/20000 train_loss:8.7435 train_time:600ms step_avg:300.23ms +step:3/20000 train_loss:7.6071 train_time:684ms step_avg:227.92ms +step:4/20000 train_loss:7.2515 train_time:768ms step_avg:191.92ms +step:5/20000 train_loss:7.0889 train_time:851ms step_avg:170.26ms +step:6/20000 train_loss:7.0715 train_time:935ms step_avg:155.91ms +step:7/20000 train_loss:6.9931 train_time:1020ms step_avg:145.71ms +step:8/20000 train_loss:6.8863 train_time:1104ms step_avg:137.97ms +step:9/20000 train_loss:6.4406 train_time:1188ms step_avg:131.96ms +step:10/20000 train_loss:6.0429 train_time:1272ms step_avg:127.20ms +step:500/20000 train_loss:2.4032 train_time:43435ms step_avg:86.87ms +step:1000/20000 train_loss:2.2585 train_time:86652ms step_avg:86.65ms +step:1500/20000 train_loss:2.2069 train_time:130036ms step_avg:86.69ms +step:2000/20000 train_loss:2.0506 train_time:173536ms step_avg:86.77ms +step:2500/20000 train_loss:2.1543 train_time:217037ms step_avg:86.81ms +step:3000/20000 train_loss:2.1484 train_time:260561ms step_avg:86.85ms +step:3500/20000 train_loss:2.1657 train_time:304082ms step_avg:86.88ms +step:4000/20000 train_loss:1.9546 train_time:347606ms step_avg:86.90ms +step:4000/20000 val_loss:2.0439 val_bpb:1.2105 train_time:347649ms step_avg:86.91ms +step:4500/20000 train_loss:2.1016 train_time:391106ms step_avg:86.91ms +step:5000/20000 train_loss:2.0829 train_time:434616ms step_avg:86.92ms +step:5500/20000 train_loss:1.9989 train_time:478099ms step_avg:86.93ms +step:6000/20000 train_loss:1.9191 train_time:521578ms step_avg:86.93ms +swa:start step:6050 +late_qat:enabled step:6213 scale:0.1498 +step:6500/20000 train_loss:2.0618 train_time:565667ms step_avg:87.03ms +step:6731/20000 val_loss:1.9231 val_bpb:1.1389 train_time:586066ms step_avg:87.07ms +stopping_early: wallclock_cap train_time:586066ms step:6731/20000 +peak memory allocated: 22851 MiB reserved: 23004 MiB +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:1.9213 val_bpb:1.1379 eval_time:2075ms +Serialized model: 106158518 bytes +Code size: 113761 bytes +gptq:building calibration model... +gptq:calibrating with 64 training batches... +gptq:calibrated 68 layers in 9.8s +gptq:budget_check train:586066ms + gptq:9823ms = 595889ms (budget:600000ms) +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +gptq_quantize: 66 GPTQ layers, 0 naive layers +selective_prune: 4081929 ±1 candidates, unpruned=15.21MB target=15.9MB +Serialized model int6+lzma: 15832384 bytes +Total submission size int6+lzma: 15946145 bytes +Total submission size int8+zlib: 15946145 bytes +final_int6_roundtrip val_loss:1.9270 val_bpb:1.1413 eval_time:16969ms +final_int6_roundtrip_exact val_loss:1.92704654 val_bpb:1.14130551 +final_int6_sliding_window val_loss:1.8875 val_bpb:1.1179 stride:64 eval_time:85193ms +final_int6_sliding_window_exact val_loss:1.88748349 val_bpb:1.11787700 +final_int8_zlib_roundtrip_exact val_loss:1.88748349 val_bpb:1.11787700