From d5b7a30a8c222f0d523855e79878dc72680c5ff3 Mon Sep 17 00:00:00 2001 From: "Dixing (Dex) Xu" Date: Sat, 11 Apr 2026 06:50:19 +0000 Subject: [PATCH 1/2] =?UTF-8?q?Record:=20SP8192=20+=20VarLen=20Attention?= =?UTF-8?q?=20+=20Doc-Independent=20LoRA=20TTT=20+=20Banking=20+=20Muon=20?= =?UTF-8?q?0.97=20=E2=80=94=20val=5Fbpb=201.07747=20(3-seed=20mean)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - 3-seed mean: 1.07747 BPP (std 0.00064) / 2.78321 nats - ~15.99 MB artifact, 8×H100 SXM, 600s - VarLen attention (within-document only), doc-independent LoRA TTT - Parameter banking + triple depth recurrence + parallel residuals - PyTorch MLP fallback (no Triton/CUTLASS dependency) - Based on PR #1530, PR #1523, PR #1514 --- .../README.md | 144 + .../submission.json | 38 + .../train_gpt.py | 2756 +++++++++++++++++ .../train_seed0.log | 284 ++ .../train_seed1337.log | 274 ++ .../train_seed42.log | 281 ++ 6 files changed, 3777 insertions(+) create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/README.md create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/submission.json create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed0.log create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed1337.log create mode 100644 records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed42.log diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/README.md b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/README.md new file mode 100644 index 0000000000..4ae3741f95 --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/README.md @@ -0,0 +1,144 @@ +# Record: SP8192 + VarLen Attention + Doc-Independent LoRA TTT + Banking + Muon 0.97 — val_bpb 1.07747 (3-seed mean) + +**val_bpb: 1.07747** (3-seed mean, std 0.00064) | **2.78321 nats** | **~15.99 MB** | 8xH100 SXM, 600s | Doc-Independent LoRA TTT + +## Results (8xH100 80GB SXM, PyTorch 2.9.1+cu128, Doc-Independent LoRA TTT) + +### Core Results + +| Seed | Steps | ms/step | Pre-TTT BPB | **Post-TTT BPB** | TTT gain | TTT time | Artifact | +|------|-------|---------|-------------|------------------|----------|----------|----------| +| 42 | 4826 | 121.9 | 1.08732 | **1.07687** | -0.01045 | 252.2s | 15,992,482 | +| 0 | 4820 | 122.0 | 1.08756 | **1.07719** | -0.01037 | 251.0s | 15,995,179 | +| 1337 | 4817 | 122.1 | 1.08885 | **1.07835** | -0.01050 | 249.6s | 15,995,796 | +| **Mean** | **4821** | **122.0** | **1.08791** | **1.07747** | **-0.01044** | **250.9s** | **15,994,486** | +| **Std** | | | | **0.00064** | | | | + +### Supplemental Diagnostics + +| Seed | Post-EMA BPB | Quantized BPB | TTT BPB | val_loss (nats) | Code size | Total submission | Train time | Eval time | +|------|-------------|---------------|---------|-----------------|-----------|------------------|------------|-----------| +| 42 | 1.07591 | 1.08732 | 1.07687 | 2.78166 | 25,659 | 15,992,482 | 588.1s | 559.5s | +| 0 | 1.07671 | 1.08756 | 1.07719 | 2.78249 | 25,659 | 15,995,179 | 588.1s | 431.9s | +| 1337 | 1.07734 | 1.08885 | 1.07835 | 2.78549 | 25,659 | 15,995,796 | 588.1s | 427.1s | + +Merged SOTA (PR #1493): **1.0810 BPB**. Delta: **-0.0035 BPB / -0.0091 nats**. Clears the 0.005-nat threshold. + +## Key Innovation + +### 1. VarLen Flash Attention (Within-Document Only) + +Uses `flash_attn_varlen_func` to compute attention within document boundaries only, eliminating cross-document attention leakage during training. Documents are packed with `cu_seqlens` tracking document boundaries, so attention is strictly masked to the current document. + +### 2. Doc-Independent LoRA TTT + +At eval time, each document gets its own independent LoRA adaptation: +- **Rank 96 LoRA** on K, O, and MLP projections +- Each document scored independently — LoRA weights reset between documents +- Score-before-update: tokens are scored, then LoRA is updated from the loss +- Adam optimizer with `lr=0.0001`, `beta2=0.999`, `weight_decay=0.5` +- Documents sorted by length (longest first) for efficient batching + +### 3. Parameter Banking (Depth Recurrence) + +Layers 3-5 are reused with `num_loops=2`, creating an encoder-decoder pattern: +- Encoder path: `[0, 1, 2, 3, 4, 5, 3, 4]` +- Decoder path: `[5, 3, 4, 5, 6, 7, 8, 9, 10]` +- Banking activated at training fraction 0.35 with gradual warmup + +### 4. PyTorch MLP Fallback (No Triton/CUTLASS) + +Replaces the Triton fused MLP kernel from PR #1530 with a pure PyTorch implementation using `F.silu` gating. This eliminates the Triton/CUTLASS dependency while maintaining competitive throughput (~122 ms/step). + +## Changes from PR #1530 Baseline + +| Aspect | PR #1530 (@samacqua) | This submission | +|--------|---------------------|-----------------| +| MLP kernel | Triton fused TMA | PyTorch `F.silu` fallback | +| Muon momentum | 0.95 (default) | 0.97 (from PR #1514) | +| Triton dependency | Required | None | + +## Architecture + +11L x 512d x 8H / 4KV, MLP 4x, SiLU gating (PyTorch), Partial RoPE (16/64 dims), layerwise LN scale, tied embeddings, logit softcap=30.0. Depth recurrence: encoder [0,1,2,3,4,5,3,4] decoder [5,3,4,5,6,7,8,9,10] (loops layers 3-5, activated at frac=0.35). Parallel residuals from layer 7. Skip gates (sigmoid-gated U-Net connections). VarLen Flash Attention with document boundary tracking. + +## Training + +MuonEq-R optimizer (row-normalized Muon, momentum=0.97, Newton-Schulz 5 steps), AdamW for embeddings/scalars. ~4821 steps in 588s on 8xH100 SXM. Linear warmdown to LR=0 over final 66.7% of training. EMA decay 0.997. + +## Quantization + +Full-Hessian GPTQ with SDClip: `clip = k * std(row)` for principled rate-distortion. int6 for attention/MLP matrices (k=12.85), int8 for token embeddings (k=20.0). Brotli-11 compression. 64 calibration batches. + +## TTT (Doc-Independent LoRA) + +Doc-independent LoRA adaptation at eval time: +- Sort val documents by length (longest first), batch by size +- For each document: apply LoRA (rank 96) to K, O, MLP projections +- Single gradient step per chunk (chunk_size=64 tokens) +- Adam optimizer: lr=0.0001, beta2=0.999, weight_decay=0.5 +- LoRA weights reset between documents — no information leaks across docs +- Total TTT eval time: ~251s (within 600s eval budget) + +## Rule Compliance + +Per Issue #1017 (Track B -- legal eval-time adaptation): + +- **Condition 1 (Causality):** VarLen attention is strictly causal within documents. No cross-document attention. +- **Condition 2 (Normalized distribution):** Standard softmax over full vocab. No n-gram cache, no logit biasing. +- **Condition 3 (Score before update):** Doc-independent LoRA scores each chunk before updating. No same-token adaptation. +- **Condition 4 (Single pass):** Each token scored exactly once. No rescoring, no multi-pass selection. + +Additional: +- No SLOT (standard or causal) +- No pre-quant TTT on val data (model quantized once during training, LoRA adapts at eval time) +- No ETLB (eval-time logit bias) +- No n-gram cache or tilt +- All artifacts under 16,000,000 bytes on all 3 seeds +- Training under 600s on all 3 seeds (~588s actual) +- Eval (TTT LoRA) under 600s on all 3 seeds (max 559.5s with compile warmup) + +## Requirements + +``` +torch>=2.9.0 +flash-attn-3 (flash_attn_interface with varlen support) +sentencepiece +brotli +numpy +``` + +No Triton or CUTLASS required. + +## Run Command (3-seed loop) + +```bash +MATCHED_FINEWEB_REPO_ID=kevclark/parameter-golf python3 data/cached_challenge_fineweb.py --variant sp8192 + +for SEED in 42 0 1337; do + SEED=$SEED torchrun --standalone --nproc_per_node=8 train_gpt.py +done +``` + +## Lineage + +PR #1530 (@samacqua, varlen + doc TTT) + PR #1523 (@EthanYangTW, banking/recurrence) + PR #1514 (@dexhunter, Muon 0.97) + PR #1394 (@clarkkev, SP8192 GPTQ SDClip baseline) + +## Credits + +- **@samacqua** — VarLen Flash Attention + Doc-Independent LoRA TTT framework (PR #1530) +- **@EthanYangTW** — Parameter banking / depth recurrence pattern (PR #1523) +- **@dexhunter** — Muon momentum 0.97 (PR #1514), depth recurrence (PR #1331, #1437) +- **@clarkkev** — SP8192 + GPTQ SDClip + MuonEq-R baseline (PR #1394) +- **@abaybektursun** — Original TTT framework (PR #549, merged precedent) +- **@Robby955** — Parallel residuals concept (PR #1412) +- **@msisovic** — Parallel residuals (PR #1204) + +## Included Files + +- `README.md` (this file) +- `submission.json` +- `train_gpt.py` +- `train_seed42.log` +- `train_seed0.log` +- `train_seed1337.log` diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/submission.json b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/submission.json new file mode 100644 index 0000000000..8aa7a4708f --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/submission.json @@ -0,0 +1,38 @@ +{ + "author": "dexhunter", + "github_id": "dexhunter", + "name": "SP8192 + VarLen Attention + Doc-Independent LoRA TTT + Banking + Muon 0.97 + PyTorch MLP Fallback", + "date": "2026-04-11", + "track": "10min_16mb", + "val_bpb": 1.07747, + "val_bpb_std": 0.00064, + "seeds": [42, 0, 1337], + "seed_results": { + "42": {"val_bpb": 1.07687, "val_loss": 2.78166, "artifact_bytes": 15992482}, + "0": {"val_bpb": 1.07719, "val_loss": 2.78249, "artifact_bytes": 15995179}, + "1337": {"val_bpb": 1.07835, "val_loss": 2.78549, "artifact_bytes": 15995796} + }, + "hardware": "8xH100 80GB SXM", + "pytorch_version": "2.9.1+cu128", + "technique_summary": "SP8192 + VarLen Flash Attention (within-document only) + Doc-Independent LoRA TTT (rank 96) + Parameter Banking (layers 3-5) + Muon 0.97 + Parallel Residuals + GPTQ int6 SDClip + Brotli + PyTorch MLP (no Triton)", + "compliance": { + "train_under_600s": true, + "artifact_under_16mb": true, + "eval_under_600s": true, + "no_slot": true, + "no_pre_quant_ttt": true, + "no_etlb": true, + "no_ngram_cache": true, + "doc_independent_ttt": true, + "three_seeds": true + }, + "attribution": { + "varlen_attention_doc_ttt": "@samacqua (PR #1530)", + "parameter_banking": "@EthanYangTW (PR #1523)", + "muon_097": "@dexhunter (PR #1514)", + "sp8192_gptq_sdclip": "@clarkkev (PR #1394)", + "depth_recurrence": "@dexhunter (PR #1331, #1437)", + "parallel_residuals": "@Robby955 (PR #1412), @msisovic (PR #1204)", + "legal_ttt_framework": "@abaybektursun (PR #549)" + } +} diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py new file mode 100644 index 0000000000..b09af88ef8 --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py @@ -0,0 +1,2756 @@ +import base64, collections, copy, fcntl, glob, io, json, lzma, math, os +from pathlib import Path +import random, re, subprocess, sys, time, uuid, numpy as np, sentencepiece as spm, torch, torch.distributed as dist, torch.nn.functional as F +from torch import nn +from flash_attn_interface import ( + flash_attn_func as flash_attn_3_func, + flash_attn_varlen_func, +) +from concurrent.futures import ThreadPoolExecutor +# Triton imports disabled — using PyTorch fallback for fused MLP +# import triton +# import triton.language as tl +# from triton.tools.tensor_descriptor import TensorDescriptor + + +class Hyperparameters: + data_dir = os.environ.get("DATA_DIR", "./data/") + seed = int(os.environ.get("SEED", 1337)) + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_frac = float(os.environ.get("WARMDOWN_FRAC", 0.667)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 6e2)) + val_batch_tokens = int(os.environ.get("VAL_BATCH_TOKENS", 524288)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + sliding_window_enabled = bool(int(os.environ.get("SLIDING_WINDOW_ENABLED", "0"))) + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + embedding_dim = int(os.environ.get("EMBEDDING_DIM", 512)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 4.0)) + skip_gates_enabled = bool(int(os.environ.get("SKIP_GATES_ENABLED", "1"))) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 3e1)) + rope_base = float(os.environ.get("ROPE_BASE", 1e4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + rope_train_seq_len = int(os.environ.get("ROPE_TRAIN_SEQ_LEN", 2048)) + rope_yarn = bool(int(os.environ.get("ROPE_YARN", "0"))) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 5.0)) + num_loops = int(os.environ.get("NUM_LOOPS", 2)) + loop_start = int(os.environ.get("LOOP_START", 3)) + loop_end = int(os.environ.get("LOOP_END", 5)) + enable_looping_at = float(os.environ.get("ENABLE_LOOPING_AT", 0.35)) + parallel_start_layer = int(os.environ.get("PARALLEL_START_LAYER", 7)) + min_lr = float(os.environ.get("MIN_LR", 0.0)) + 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", 0.03)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.022)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.97)) + 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)) + muon_row_normalize = bool(int(os.environ.get("MUON_ROW_NORMALIZE", "1"))) + 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-08)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + adam_wd = float(os.environ.get("ADAM_WD", 0.02)) + muon_wd = float(os.environ.get("MUON_WD", 0.095)) + embed_wd = float(os.environ.get("EMBED_WD", 0.095)) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 96)) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.0001)) + ttt_chunk_size = int(os.environ.get("TTT_CHUNK_SIZE", 64)) + ttt_eval_seq_len = int(os.environ.get("TTT_EVAL_SEQ_LEN", 2048)) + ttt_batch_size = int(os.environ.get("TTT_BATCH_SIZE", 64)) + ttt_grad_steps = int(os.environ.get("TTT_GRAD_STEPS", 1)) + ttt_weight_decay = float(os.environ.get("TTT_WEIGHT_DECAY", 0.5)) + ttt_beta1 = float(os.environ.get("TTT_BETA1", 0)) + ttt_beta2 = float(os.environ.get("TTT_BETA2", 0.999)) + ttt_k_lora = bool(int(os.environ.get("TTT_K_LORA", "1"))) + ttt_mlp_lora = bool(int(os.environ.get("TTT_MLP_LORA", "1"))) + ttt_o_lora = bool(int(os.environ.get("TTT_O_LORA", "1"))) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adam") + ttt_eval_batches = os.environ.get("TTT_EVAL_BATCHES", "") + ttt_output_dir = os.environ.get("TTT_OUTPUT_DIR", "") + val_doc_fraction = float(os.environ.get("VAL_DOC_FRACTION", 1.0)) + etlb_lr = float(os.environ.get("ETLB_LR", 0.05)) + etlb_steps = int(os.environ.get("ETLB_STEPS", 5)) + etlb_clip = float(os.environ.get("ETLB_CLIP", 3.0)) + compressor = os.environ.get("COMPRESSOR", "brotli") + gptq_calibration_batches = int(os.environ.get("GPTQ_CALIBRATION_BATCHES", 64)) + gptq_reserve_seconds = float(os.environ.get("GPTQ_RESERVE_SECONDS", 12.0)) + matrix_bits = int(os.environ.get("MATRIX_BITS", 6)) + embed_bits = int(os.environ.get("EMBED_BITS", 8)) + matrix_clip_sigmas = float(os.environ.get("MATRIX_CLIP_SIGMAS", 12.85)) + embed_clip_sigmas = float(os.environ.get("EMBED_CLIP_SIGMAS", 2e1)) + 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")) + is_main_process = rank == 0 + grad_accum_steps = 8 // world_size + datasets_dir = os.path.join(data_dir, "datasets", f"fineweb10B_sp{vocab_size}") + train_files = os.path.join(datasets_dir, "fineweb_train_*.bin") + val_files = os.path.join(datasets_dir, "fineweb_val_*.bin") + tokenizer_path = os.path.join( + data_dir, "tokenizers", f"fineweb_{vocab_size}_bpe.model" + ) + artifact_dir = os.environ.get("ARTIFACT_DIR", "") + eval_only_path = os.environ.get("EVAL_ONLY_PATH", "") + logfile = ( + os.path.join(artifact_dir, f"{run_id}.txt") + if artifact_dir + else f"logs/{run_id}.txt" + ) + model_path = ( + os.path.join(artifact_dir, "final_model.pt") + if artifact_dir + else "final_model.pt" + ) + quantized_model_path = ( + os.path.join(artifact_dir, "final_model.int6.ptz") + if artifact_dir + else "final_model.int6.ptz" + ) + + +_logger_hparams = None + + +def set_logging_hparams(h): + global _logger_hparams + _logger_hparams = h + + +def log(msg, console=True): + if _logger_hparams is None: + print(msg) + return + if _logger_hparams.is_main_process: + if console: + print(msg) + if _logger_hparams.logfile is not None: + with open(_logger_hparams.logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + +class ValidationData: + def __init__(self, h, device): + self.sp = spm.SentencePieceProcessor(model_file=h.tokenizer_path) + if int(self.sp.vocab_size()) != h.vocab_size: + raise ValueError( + f"VOCAB_SIZE={h.vocab_size} does not match tokenizer vocab_size={int(self.sp.vocab_size())}" + ) + self.val_tokens = load_validation_tokens(h.val_files, h.eval_seq_len) + ( + self.base_bytes_lut, + self.has_leading_space_lut, + self.is_boundary_token_lut, + ) = build_sentencepiece_luts(self.sp, h.vocab_size, device) + + +def build_sentencepiece_luts(sp, vocab_size, device): + sp_vocab_size = int(sp.vocab_size()) + assert ( + sp.piece_to_id("▁") != sp.unk_id() + ), "Tokenizer must have '▁' (space) as its own token for correct BPB byte counting" + 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, seq_len): + 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 load_data_shard(file): + header_bytes = 256 * np.dtype(" 0: + pos = start + while pos < end: + seg_starts.append(pos) + pos += max_doc_len + else: + seg_starts.append(start) + boundaries = seg_starts + [total_len] + padded_len = get_next_multiple_of_n(len(boundaries), bucket_size) + cu = torch.full((padded_len,), total_len, dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + seg_ends = seg_starts[1:] + [total_len] + max_seqlen = max(end - start for start, end in zip(seg_starts, seg_ends)) + return cu, max_seqlen + +class DocumentPackingLoader: + _shard_pool = ThreadPoolExecutor(1) + + def __init__(self, h, device, cu_bucket_size=64): + self.rank = h.rank + self.world_size = h.world_size + self.device = device + self.cu_bucket_size = cu_bucket_size + self.max_seq_len = h.train_seq_len + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files + self.file_iter = iter(self.files) + self._init_shard(load_data_shard(next(self.file_iter))) + self._next_shard = self._submit_next_shard() + self._batch_pool = ThreadPoolExecutor(1) + self._next_batch = None + + def _init_shard(self, tokens): + global BOS_ID + self.tokens = tokens + self.shard_size = tokens.numel() + if BOS_ID is None: + BOS_ID = 1 + self.bos_idx = ( + (tokens == BOS_ID).nonzero(as_tuple=True)[0].to(torch.int64).cpu().numpy() + ) + if self.bos_idx.size == 0: + self.bos_idx = np.array([0], dtype=np.int64) + self.cursor = int(self.bos_idx[0]) + + def _submit_next_shard(self): + try: + path = next(self.file_iter) + return self._shard_pool.submit(load_data_shard, path) + except StopIteration: + return None + + def _advance_shard(self): + if self._next_shard is None: + self.file_iter = iter(self.files) + self._next_shard = self._shard_pool.submit( + load_data_shard, next(self.file_iter) + ) + self._init_shard(self._next_shard.result()) + self._next_shard = self._submit_next_shard() + + def _local_doc_starts(self, local_start, total_len): + lo = np.searchsorted(self.bos_idx, local_start, side="left") + hi = np.searchsorted(self.bos_idx, local_start + total_len, side="left") + return (self.bos_idx[lo:hi] - local_start).tolist() + + def _prepare_batch(self, num_tokens_local, max_seq_len): + per_rank_span = num_tokens_local + 1 + global_span = per_rank_span * self.world_size + while self.cursor + global_span > self.shard_size: + self._advance_shard() + local_start = self.cursor + self.rank * per_rank_span + buf = self.tokens[local_start : local_start + per_rank_span] + inputs = buf[:-1].to(dtype=torch.int64).pin_memory() + targets = buf[1:].to(dtype=torch.int64).pin_memory() + starts = self._local_doc_starts(local_start, inputs.numel()) + cu_seqlens, max_seqlen = _build_cu_seqlens( + starts, inputs.numel(), inputs.device, max_seq_len, self.cu_bucket_size + ) + cu_seqlens = cu_seqlens.pin_memory() + self.cursor += global_span + return inputs, targets, cu_seqlens, max_seqlen + + def next_batch(self, global_tokens, grad_accum_steps): + num_tokens_local = global_tokens // (self.world_size * grad_accum_steps) + if self._next_batch is not None: + inputs, targets, cu_seqlens, max_seqlen = self._next_batch.result() + else: + inputs, targets, cu_seqlens, max_seqlen = self._prepare_batch( + num_tokens_local, self.max_seq_len + ) + self._next_batch = self._batch_pool.submit( + self._prepare_batch, num_tokens_local, self.max_seq_len + ) + return ( + inputs[None].to(self.device, non_blocking=True), + targets[None].to(self.device, non_blocking=True), + cu_seqlens.to(self.device, non_blocking=True), + max_seqlen, + ) + + +class ShuffledSequenceLoader: + def __init__(self, h, device): + self.world_size = h.world_size + self.seq_len = h.train_seq_len + self.device = device + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files[h.rank :: h.world_size] + self.rng = np.random.Generator(np.random.PCG64(h.rank)) + self.num_tokens = [_read_num_tokens(f) for f in self.files] + self.start_inds = [[] for _ in self.files] + for si in range(len(self.files)): + self._reset_shard(si) + + def _reset_shard(self, si): + max_phase = min( + self.seq_len - 1, max(0, self.num_tokens[si] - self.seq_len - 1) + ) + phase = int(self.rng.integers(max_phase + 1)) if max_phase > 0 else 0 + num_sequences = (self.num_tokens[si] - 1 - phase) // self.seq_len + sequence_order = self.rng.permutation(num_sequences) + self.start_inds[si] = (phase + sequence_order * self.seq_len).tolist() + + def next_batch(self, global_tokens, grad_accum_steps): + device_tokens = global_tokens // (self.world_size * grad_accum_steps) + device_batch_size = device_tokens // self.seq_len + remaining = np.array([len(s) for s in self.start_inds], dtype=np.float64) + x = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + y = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + for bi in range(device_batch_size): + total = remaining.sum() + if total <= 0: + for si in range(len(self.files)): + self._reset_shard(si) + remaining = np.array( + [len(s) for s in self.start_inds], dtype=np.float64 + ) + total = remaining.sum() + probs = remaining / total + si = int(self.rng.choice(len(self.files), p=probs)) + start_ind = self.start_inds[si].pop() + remaining[si] -= 1 + mm = _get_shard_memmap(self.files[si]) + window = torch.as_tensor( + np.array(mm[start_ind : start_ind + self.seq_len + 1], dtype=np.int64) + ) + x[bi] = window[:-1] + y[bi] = window[1:] + return x.to(self.device, non_blocking=True), y.to( + self.device, non_blocking=True + ) + + +class RMSNorm(nn.Module): + def __init__(self, eps=None): + super().__init__() + self.eps = eps + + def forward(self, x): + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class CastedLinear(nn.Linear): + def forward(self, x): + w = self.weight.to(x.dtype) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +# Triton TMA kernel removed — incompatible with our Triton 3.5.1 + + +def linear_leaky_relu_square(a, b, aux=None): + # PyTorch fallback (replaces Triton TMA kernel for compatibility) + pre = a @ b.T # [M, N] + if aux is not None: + # Backward path: multiply by activation gradient + act_grad = torch.where(aux >= 0, 2 * aux, 0.5 * aux) # d/dx of leaky_relu(x,0.5)^2 + return pre * act_grad + # Forward path: apply leaky_relu(0.5) then square + activated = torch.where(pre >= 0, pre, 0.5 * pre) + post = activated * activated + return pre, post + + +# Triton launcher function removed — using PyTorch fallback above + + +class FusedLinearLeakyReLUSquareFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x, w1, w2): + x_flat = x.reshape(-1, x.shape[-1]) + pre, post = linear_leaky_relu_square(x_flat, w1) + out = F.linear(post, w2) + ctx.save_for_backward(x, w1, w2, pre, post) + return out.view(*x.shape[:-1], out.shape[-1]) + + @staticmethod + def backward(ctx, grad_output): + x, w1, w2, pre, post = ctx.saved_tensors + x_flat = x.reshape(-1, x.shape[-1]) + grad_output_flat = grad_output.reshape(-1, grad_output.shape[-1]) + dw2 = grad_output_flat.T @ post + dpre = linear_leaky_relu_square(grad_output_flat, w2.T.contiguous(), aux=pre) + dw1 = dpre.T @ x_flat + dx = dpre @ w1 + return dx.view_as(x), dw1, dw2 + + +FusedLeakyReLUSquareMLP = FusedLinearLeakyReLUSquareFunction.apply + + +class Rotary(nn.Module): + def __init__(self, dim, base=1e4, train_seq_len=1024, rope_dims=0, yarn=True): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.yarn = yarn + 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 = None + self._sin_cached = None + + def forward(self, seq_len, device, dtype): + 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 self.yarn and 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.float().to(device) + t = torch.arange(seq_len, device=device, dtype=torch.float32) + 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[:, :seq_len].to(dtype=dtype), self._sin_cached[:, :seq_len].to(dtype=dtype) + + +def apply_rotary_emb(x, cos, sin, rope_dims=0): + 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, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=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") + self.q_gain = nn.Parameter( + torch.full((num_heads,), qk_gain_init, dtype=torch.float32) + ) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=train_seq_len, yarn=yarn) + self.use_xsa = False + + def _xsa_efficient(self, y, v): + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x, q_w, k_w, v_w, out_w, cu_seqlens=None, max_seqlen=0): + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if cu_seqlens is not None: + y = flash_attn_varlen_func( + q[0], + k[0], + v[0], + cu_seqlens_q=cu_seqlens, + cu_seqlens_k=cu_seqlens, + max_seqlen_q=max_seqlen, + max_seqlen_k=max_seqlen, + causal=True, + window_size=(-1, -1), + )[None] + else: + 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) + self._last_proj_input = y.detach() if getattr(self, "_calib", False) else None + return F.linear(y, out_w.to(x.dtype)) + + +class MLP(nn.Module): + def __init__(self, dim, mlp_mult): + super().__init__() + self.use_fused = True + + def forward(self, x, up_w, down_w): + if self.training and self.use_fused: + return FusedLeakyReLUSquareMLP(x, up_w.to(x.dtype), down_w.to(x.dtype)) + hidden = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5).square() + self._last_down_input = hidden.detach() if getattr(self, "_calib", False) else None + return F.linear(hidden, down_w.to(x.dtype)) + + +class Block(nn.Module): + def __init__( + self, + dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + train_seq_len, + layer_idx=0, + ln_scale=False, + yarn=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, train_seq_len, yarn=yarn + ) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack((torch.ones(dim), torch.zeros(dim))).float() + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + + def forward(self, x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=None, max_seqlen=0): + 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, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + 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) + return x_out + + def forward_attn(self, x, x0, q_w, k_w, v_w, out_w, cu_seqlens=None, max_seqlen=0): + 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, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + return x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + + def forward_mlp(self, x, up_w, down_w): + return x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp( + self.mlp_norm(x) * self.ln_scale_factor, up_w, down_w + ) + + +class GPT(nn.Module): + def __init__(self, h): + super().__init__() + if h.logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {h.logit_softcap}") + self.tie_embeddings = h.tie_embeddings + self.tied_embed_init_std = h.tied_embed_init_std + self.logit_softcap = h.logit_softcap + self.tok_emb = nn.Embedding(h.vocab_size, h.embedding_dim) + if h.embedding_dim != h.model_dim: + self.embed_proj = CastedLinear(h.embedding_dim, h.model_dim, bias=False) + self.head_proj = CastedLinear(h.model_dim, h.embedding_dim, bias=False) + else: + self.embed_proj = None + self.head_proj = None + self.num_layers = h.num_layers + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + self.qo_bank = nn.Parameter(torch.empty(2 * h.num_layers, h.model_dim, h.model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * h.num_layers, kv_dim, h.model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(h.num_layers, hidden_dim, h.model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(h.num_layers, h.model_dim, hidden_dim)) + self.num_encoder_layers = h.num_layers // 2 + self.num_decoder_layers = h.num_layers - self.num_encoder_layers + self.blocks = nn.ModuleList( + [ + Block( + h.model_dim, + h.num_heads, + h.num_kv_heads, + h.mlp_mult, + h.rope_base, + h.qk_gain_init, + h.train_seq_len, + layer_idx=i, + ln_scale=h.ln_scale, + yarn=h.rope_yarn, + ) + for i in range(h.num_layers) + ] + ) + if h.rope_dims > 0: + head_dim = h.model_dim // h.num_heads + for block in self.blocks: + block.attn.rope_dims = h.rope_dims + block.attn.rotary = Rotary( + head_dim, + base=h.rope_base, + train_seq_len=h.train_seq_len, + rope_dims=h.rope_dims, + yarn=h.rope_yarn, + ) + self.final_norm = RMSNorm() + self.lm_head = ( + None + if h.tie_embeddings + else CastedLinear(h.embedding_dim, h.vocab_size, bias=False) + ) + if self.lm_head is not None: + self.lm_head._zero_init = True + if h.xsa_last_n > 0: + for i in range(max(0, h.num_layers - h.xsa_last_n), h.num_layers): + self.blocks[i].attn.use_xsa = True + self.looping_active = False + if h.num_loops > 0: + loop_seg = list(range(h.loop_start, h.loop_end + 1)) + all_indices = list(range(h.loop_start)) + for _ in range(h.num_loops + 1): + all_indices.extend(loop_seg) + all_indices.extend(range(h.loop_end + 1, h.num_layers)) + num_enc = len(all_indices) // 2 + self.encoder_indices = all_indices[:num_enc] + self.decoder_indices = all_indices[num_enc:] + else: + self.encoder_indices = list(range(self.num_encoder_layers)) + self.decoder_indices = list(range(self.num_encoder_layers, h.num_layers)) + self.num_skip_weights = min( + len(self.encoder_indices), len(self.decoder_indices) + ) + self.skip_weights = nn.Parameter( + torch.ones(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + self.skip_gates = ( + nn.Parameter( + torch.zeros(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + if h.skip_gates_enabled + else None + ) + self.parallel_start_layer = h.parallel_start_layer + self.lane_merge = ( + nn.Parameter(torch.tensor(0.5)) if self.parallel_start_layer > 0 else None + ) + self._init_weights() + + def _init_weights(self): + 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) + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) + nn.init.zeros_(self.qo_bank.data[n + i]) + self.qo_bank.data[n + i].mul_(proj_scale) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) + nn.init.zeros_(self.mlp_down_bank.data[i]) + self.mlp_down_bank.data[i].mul_(proj_scale) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif ( + module.weight.ndim == 2 + and module.weight.shape[0] >= 64 + and module.weight.shape[1] >= 64 + ): + nn.init.orthogonal_(module.weight, gain=1.0) + + def _bank_weights(self, i): + n = self.num_layers + return ( + 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], + ) + + def forward_logits(self, input_ids, cu_seqlens=None, max_seqlen=0): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + if self.embed_proj is not None: + x = self.embed_proj(x) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else range(self.num_encoder_layers) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if lane0 is None: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[ + None, None, : + ] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + if i >= psl and psl > 0: + lane0 = x + lane1 = x.clone() + lane0 = self.blocks[i].forward_attn(lane0, x0, q_w, k_w, v_w, out_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + lane1 = self.blocks[i].forward_mlp(lane1, up_w, down_w) + else: + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[ + None, None, : + ] + lane0 = torch.lerp(scaled_skip, lane0, g) + else: + lane0 = lane0 + scaled_skip + lane0 = self.blocks[i].forward_attn(lane0, x0, q_w, k_w, v_w, out_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + lane1 = self.blocks[i].forward_mlp(lane1, up_w, down_w) + if lane0 is not None: + lm = self.lane_merge.to(dtype=lane0.dtype) + x = lm * lane0 + (1.0 - lm) * lane1 + x = self.final_norm(x) + if self.head_proj is not None: + x = self.head_proj(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 forward(self, input_ids, target_ids, cu_seqlens=None, max_seqlen=0): + logits = self.forward_logits( + input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen + ) + return F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + target_ids.reshape(-1), + reduction="mean", + ) + + def forward_ttt(self, input_ids, target_ids, lora): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + if self.embed_proj is not None: + x = self.embed_proj(x) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else list(range(self.num_encoder_layers)) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else list( + range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + ) + slot = 0 + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if lane0 is None: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[ + None, None, : + ] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + if i >= psl and psl > 0: + lane0 = x + lane1 = x.clone() + lane0 = self._block_with_lora_attn(self.blocks[i], lane0, x0, lora, slot, q_w, k_w, v_w, out_w) + lane1 = self._block_with_lora_mlp(self.blocks[i], lane1, lora, slot, up_w, down_w) + else: + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[ + None, None, : + ] + lane0 = torch.lerp(scaled_skip, lane0, g) + else: + lane0 = lane0 + scaled_skip + lane0 = self._block_with_lora_attn(self.blocks[i], lane0, x0, lora, slot, q_w, k_w, v_w, out_w) + lane1 = self._block_with_lora_mlp(self.blocks[i], lane1, lora, slot, up_w, down_w) + slot += 1 + if lane0 is not None: + lm = self.lane_merge.to(dtype=lane0.dtype) + x = lm * lane0 + (1.0 - lm) * lane1 + x = self.final_norm(x) + if self.head_proj is not None: + x = self.head_proj(x) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = logits + lora.lm_head_lora(x) + logits = self.logit_softcap * torch.tanh(logits / self.logit_softcap) + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + + def _block_with_lora(self, block, x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w): + mix = block.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = block.attn_norm(x_in) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q = (F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n)).reshape( + bsz, seqlen, attn.num_heads, attn.head_dim + ) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + x_out = x_in + block.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + mlp_n = block.mlp_norm(x_out) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + x_out = x_out + block.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * mlp_out + return x_out + + def _block_with_lora_attn(self, block, x, x0, lora, slot, q_w, k_w, v_w, out_w): + mix = block.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = block.attn_norm(x_in) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q = (F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n)).reshape( + bsz, seqlen, attn.num_heads, attn.head_dim + ) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + return x_in + block.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + + def _block_with_lora_mlp(self, block, x, lora, slot, up_w, down_w): + mlp_n = block.mlp_norm(x) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + return x + block.mlp_scale.to(dtype=x.dtype)[None, None, :] * mlp_out + + +class BatchedLinearLoRA(nn.Module): + def __init__(self, bsz, in_features, out_features, rank): + super().__init__() + self._bound = 1.0 / math.sqrt(in_features) + self.A = nn.Parameter( + torch.empty(bsz, rank, in_features).uniform_(-self._bound, self._bound) + ) + self.B = nn.Parameter(torch.zeros(bsz, out_features, rank)) + + def reset(self): + with torch.no_grad(): + self.A.uniform_(-self._bound, self._bound) + self.B.zero_() + + def forward(self, x): + return (x @ self.A.transpose(1, 2)) @ self.B.transpose(1, 2) + + +class BatchedTTTLoRA(nn.Module): + def __init__(self, bsz, model, rank, k_lora=True, mlp_lora=True, o_lora=True): + super().__init__() + self.bsz = bsz + dim = model.qo_bank.shape[-1] + vocab = model.tok_emb.num_embeddings + if getattr(model, "looping_active", False): + num_slots = len(model.encoder_indices) + len(model.decoder_indices) + else: + num_slots = len(model.blocks) + kv_dim = model.blocks[0].attn.num_kv_heads * ( + dim // model.blocks[0].attn.num_heads + ) + embed_dim = model.tok_emb.embedding_dim + self.lm_head_lora = BatchedLinearLoRA(bsz, embed_dim, vocab, rank) + self.q_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + self.v_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + self.k_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + if k_lora + else None + ) + self.mlp_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if mlp_lora + else None + ) + self.o_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if o_lora + else None + ) + + def reset(self): + with torch.no_grad(): + self.lm_head_lora.reset() + for loras in [self.q_loras, self.v_loras, self.k_loras, + self.mlp_loras, self.o_loras]: + if loras is not None: + for lora in loras: + lora.reset() + + +def classify_param(name): + 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" + + +@torch.compile +def zeropower_via_newtonschulz5(G, steps=10, eps=1e-07): + a, b, c = 3.4445, -4.775, 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): + def __init__( + self, + params, + lr, + momentum, + backend_steps, + nesterov=True, + weight_decay=0.0, + row_normalize=False, + ): + super().__init__( + params, + dict( + lr=lr, + momentum=momentum, + backend_steps=backend_steps, + nesterov=nesterov, + weight_decay=weight_decay, + row_normalize=row_normalize, + ), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + "p": p, + "B": B, + "padded_grad": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "shard": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "shard_mom": torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + "full_update": torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + "scale": max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + self._bank_meta.sort(key=lambda m: -m["p"].numel()) + self._built = True + + def launch_reduce_scatters(self): + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m["p"] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m["padded_grad"] + pg[: m["B"]].copy_(p.grad.bfloat16()) + if pg.shape[0] > m["B"]: + pg[m["B"] :].zero_() + fut = dist.reduce_scatter_tensor( + m["shard"], pg, op=dist.ReduceOp.AVG, async_op=True + ) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + if not self._built: + self._build() + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + row_normalize = group.get("row_normalize", False) + prev_ag_handle = None + prev_m = None + sharded = self._distributed and hasattr(self, "_rs_futures") + for idx, 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[idx] is not None: + self._rs_futures[idx].wait() + g = m["shard"] + buf = m["shard_mom"] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + if row_normalize: + rn = update.float().norm(dim=-1, keepdim=True).clamp_min(1e-07) + update = update / rn.to(update.dtype) + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m["full_update"], update, async_op=True + ) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m["scale"]) + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m["p"] + upd = prev_m["full_update"][: prev_m["B"]] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m["scale"]) + if hasattr(self, "_rs_futures"): + del self._rs_futures + return loss + + +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,skip_gates,lane_merge", + ).split(",") + if pattern +) + + +PACKED_REPLICATED_GRAD_MAX_NUMEL = 1 << 15 + + +class Optimizers: + def __init__(self, h, base_model): + matrix_params = [ + base_model.qo_bank, + base_model.kv_bank, + base_model.mlp_up_bank, + base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for (name, p) in block_named_params + if p.ndim < 2 + or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + if base_model.skip_gates is not None and base_model.skip_gates.numel() > 0: + scalar_params.append(base_model.skip_gates) + if base_model.lane_merge is not None: + scalar_params.append(base_model.lane_merge) + token_lr = h.tied_embed_lr if h.tie_embeddings else h.embed_lr + tok_params = [ + {"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr} + ] + self.optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.embed_wd, + fused=True, + ) + self.optimizer_muon = Muon( + matrix_params, + lr=h.matrix_lr, + momentum=h.muon_momentum, + backend_steps=h.muon_backend_steps, + weight_decay=h.muon_wd, + row_normalize=h.muon_row_normalize, + ) + for group in self.optimizer_muon.param_groups: + group["base_lr"] = h.matrix_lr + self.optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": h.scalar_lr, "base_lr": h.scalar_lr}], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.adam_wd, + fused=True, + ) + self.optimizers = [ + self.optimizer_tok, + self.optimizer_muon, + self.optimizer_scalar, + ] + if base_model.lm_head is not None: + self.optimizer_head = torch.optim.Adam( + [ + { + "params": [base_model.lm_head.weight], + "lr": h.head_lr, + "base_lr": h.head_lr, + } + ], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + fused=True, + ) + self.optimizers.insert(1, self.optimizer_head) + else: + self.optimizer_head = None + self.replicated_params = list(tok_params[0]["params"]) + self.replicated_params.extend(scalar_params) + if base_model.lm_head is not None: + self.replicated_params.append(base_model.lm_head.weight) + self.replicated_large_params = [] + self.replicated_packed_params = [] + for p in self.replicated_params: + if p.numel() <= PACKED_REPLICATED_GRAD_MAX_NUMEL: + self.replicated_packed_params.append(p) + else: + self.replicated_large_params.append(p) + + def __iter__(self): + return iter(self.optimizers) + + def zero_grad_all(self): + for opt in self.optimizers: + opt.zero_grad(set_to_none=True) + + def _all_reduce_packed_grads(self): + grads_by_key = collections.defaultdict(list) + for p in self.replicated_packed_params: + if p.grad is not None: + grads_by_key[(p.grad.device, p.grad.dtype)].append(p.grad) + for grads in grads_by_key.values(): + flat = torch.empty( + sum(g.numel() for g in grads), + device=grads[0].device, + dtype=grads[0].dtype, + ) + offset = 0 + for g in grads: + n = g.numel() + flat[offset : offset + n].copy_(g.contiguous().view(-1)) + offset += n + dist.all_reduce(flat, op=dist.ReduceOp.AVG) + offset = 0 + for g in grads: + n = g.numel() + g.copy_(flat[offset : offset + n].view_as(g)) + offset += n + + def step(self, distributed=False): + self.optimizer_muon.launch_reduce_scatters() + if distributed: + reduce_handles = [ + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG, async_op=True) + for p in self.replicated_large_params + if p.grad is not None + ] + self._all_reduce_packed_grads() + for handle in reduce_handles: + handle.wait() + self.optimizer_tok.step() + self.optimizer_scalar.step() + if self.optimizer_head is not None: + self.optimizer_head.step() + self.optimizer_muon.step() + self.zero_grad_all() + + +def restore_fp32_params(model): + for module in model.modules(): + if isinstance(module, CastedLinear): + module.float() + for name, param in model.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() + if hasattr(model, "qo_bank"): + model.qo_bank.data = model.qo_bank.data.float() + model.kv_bank.data = model.kv_bank.data.float() + model.mlp_up_bank.data = model.mlp_up_bank.data.float() + model.mlp_down_bank.data = model.mlp_down_bank.data.float() + + +def collect_hessians(model, train_loader, h, device, n_calibration_batches=64): + hessians = {} + hooks = [] + for i, block in enumerate(model.blocks): + block.attn._calib = True + block.mlp._calib = True + block.mlp.use_fused = False + + def make_attn_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + for suffix in ["c_q", "c_k", "c_v"]: + name = f"blocks.{layer_idx}.attn.{suffix}.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + y = module._last_proj_input + if y is not None: + y = y.float() + if y.ndim == 3: + y = y.reshape(-1, y.shape[-1]) + name = f"blocks.{layer_idx}.attn.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + y.shape[1], y.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(y.T, y) + return hook_fn + + def make_mlp_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + name = f"blocks.{layer_idx}.mlp.fc.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + h_act = module._last_down_input + if h_act is not None: + h_act = h_act.float() + if h_act.ndim == 3: + h_act = h_act.reshape(-1, h_act.shape[-1]) + name = f"blocks.{layer_idx}.mlp.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + h_act.shape[1], h_act.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(h_act.T, h_act) + return hook_fn + + for i, block in enumerate(model.blocks): + hooks.append(block.attn.register_forward_hook(make_attn_hook(i))) + hooks.append(block.mlp.register_forward_hook(make_mlp_hook(i))) + if model.tie_embeddings: + hook_module = ( + model.head_proj if model.head_proj is not None else model.final_norm + ) + + def make_output_hook(name): + def hook_fn(module, inp, out): + x = out.detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + return hook_fn + + hooks.append( + hook_module.register_forward_hook(make_output_hook("tok_emb.weight")) + ) + model.eval() + with torch.no_grad(): + for _ in range(n_calibration_batches): + x, _ = train_loader.next_batch(h.train_batch_tokens, h.grad_accum_steps) + model.forward_logits(x) + for hook in hooks: + hook.remove() + for i, block in enumerate(model.blocks): + block.attn._calib = False + block.mlp._calib = False + block.mlp.use_fused = True + for name in hessians: + hessians[name] = hessians[name].cpu() / n_calibration_batches + return hessians + + +def gptq_quantize_weight(w, H, clip_sigmas=3.0, clip_range=63, block_size=128): + W_orig = w.float().clone() + rows, cols = W_orig.shape + H = H.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * H.diag().mean() + H.diagonal().add_(damp) + 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] + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + row_std = W_orig.std(dim=1) + s = (clip_sigmas * row_std / clip_range).clamp_min(1e-10).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:] + return Q[:, invperm], s + + +def gptq_mixed_quantize(state_dict, hessians, h): + result = {} + meta = {} + for (name, tensor) in state_dict.items(): + t = tensor.detach().cpu().contiguous() + 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 (float16)" + continue + cs = h.embed_clip_sigmas if "tok_emb" in name else h.matrix_clip_sigmas + bits = h.embed_bits if "tok_emb" in name else h.matrix_bits + q, s = gptq_quantize_weight( + t, hessians[name], clip_sigmas=cs, clip_range=2 ** (bits - 1) - 1 + ) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = f"gptq (int{bits})" + categories = collections.defaultdict(set) + for (name, cat) in meta.items(): + short = re.sub("\\.\\d+$", "", re.sub("blocks\\.\\d+", "blocks", name)) + categories[cat].add(short) + log("Quantized weights:") + for cat in sorted(categories): + log(f" {cat}: {', '.join(sorted(categories[cat]))}") + return result, meta + + +def dequantize_mixed(result, meta, template_sd): + out = {} + for (name, orig) in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if "passthrough" in info: + 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 + + +_BSHF_MAGIC = b"BSHF" + + +def _byte_shuffle(data, stride=2): + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off : dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() + + +def _byte_unshuffle(data): + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off : src_off + chunk_len] + src_off += chunk_len + return out.tobytes() + + +def _compress(data, compressor): + data = _byte_shuffle(data) + if compressor == "lzma": + return lzma.compress(data, preset=6) + elif compressor == "brotli": + import brotli + + return brotli.compress(data, quality=11) + raise ValueError(f"Unknown compressor: {compressor!r}") + + +def _decompress(data, compressor): + if compressor == "lzma": + raw = lzma.decompress(data) + elif compressor == "brotli": + import brotli + + raw = brotli.decompress(data) + else: + raise ValueError(f"Unknown compressor: {compressor!r}") + raw = _byte_unshuffle(raw) + return raw + + +def _unbank_state_dict(state_dict, num_layers): + sd = {} + n = num_layers + for k, v in state_dict.items(): + t = v.detach().cpu() + if k == "qo_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_q.weight"] = t[i] + sd[f"blocks.{i}.attn.proj.weight"] = t[n + i] + elif k == "kv_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_k.weight"] = t[i] + sd[f"blocks.{i}.attn.c_v.weight"] = t[n + i] + elif k == "mlp_up_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.fc.weight"] = t[i] + elif k == "mlp_down_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.proj.weight"] = t[i] + else: + sd[k] = t + return sd + + +def _rebank_state_dict(flat_sd, num_layers, model_dim, kv_dim, hidden_dim): + sd = {} + n = num_layers + sd["qo_bank"] = torch.zeros(2 * n, model_dim, model_dim) + sd["kv_bank"] = torch.zeros(2 * n, kv_dim, model_dim) + sd["mlp_up_bank"] = torch.zeros(n, hidden_dim, model_dim) + sd["mlp_down_bank"] = torch.zeros(n, model_dim, hidden_dim) + for i in range(n): + sd["qo_bank"][i] = flat_sd[f"blocks.{i}.attn.c_q.weight"] + sd["qo_bank"][n + i] = flat_sd[f"blocks.{i}.attn.proj.weight"] + sd["kv_bank"][i] = flat_sd[f"blocks.{i}.attn.c_k.weight"] + sd["kv_bank"][n + i] = flat_sd[f"blocks.{i}.attn.c_v.weight"] + sd["mlp_up_bank"][i] = flat_sd[f"blocks.{i}.mlp.fc.weight"] + sd["mlp_down_bank"][i] = flat_sd[f"blocks.{i}.mlp.proj.weight"] + for k, v in flat_sd.items(): + if not ( + k.startswith("blocks.") + and any( + p in k + for p in [ + ".attn.c_q.", ".attn.c_k.", ".attn.c_v.", + ".attn.proj.", ".mlp.fc.", ".mlp.proj.", + ] + ) + ): + sd[k] = v + return sd + + +def _compressed_code_size(code): + code_raw = code.encode("utf-8") + minified = subprocess.run( + ["pyminify", "--no-rename-locals", "--no-hoist-literals", "--remove-literal-statements", "-"], + input=code_raw, capture_output=True, check=True, + ).stdout + compressed = lzma.compress(minified) + encoded = base64.b85encode(compressed) + wrapper = b'import lzma as L,base64 as B\nexec(L.decompress(B.b85decode("' + encoded + b'")))\n' + return len(code_raw), len(wrapper) + + +def serialize(h, base_model, code): + code_bytes_uncompressed, code_bytes = _compressed_code_size(code) + if h.is_main_process: + torch.save(base_model.state_dict(), h.model_path) + model_bytes = os.path.getsize(h.model_path) + log(f"Serialized model: {model_bytes} bytes") + log(f"Code size (uncompressed): {code_bytes_uncompressed} bytes") + log(f"Code size (compressed): {code_bytes} bytes") + sd_cpu = _unbank_state_dict(base_model.state_dict(), h.num_layers) + device = torch.device("cuda", h.local_rank) + log("GPTQ:collecting Hessians from calibration data...") + t0 = time.perf_counter() + calib_loader = ShuffledSequenceLoader(h, device) + hessians = collect_hessians( + base_model, + calib_loader, + h, + device, + n_calibration_batches=h.gptq_calibration_batches, + ) + log(f"GPTQ:collected {len(hessians)} Hessians in {time.perf_counter()-t0:.1f}s") + quant_result, quant_meta = gptq_mixed_quantize(sd_cpu, hessians, h) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress(quant_raw, h.compressor) + quant_file_bytes = len(quant_blob) + bytes_total = quant_file_bytes + code_bytes + if h.is_main_process: + with open(h.quantized_model_path, "wb") as f: + f.write(quant_blob) + log(f"Serialized model quantized+{h.compressor}: {quant_file_bytes} bytes") + log(f"Total submission size quantized+{h.compressor}: {bytes_total} bytes") + return bytes_total, quant_file_bytes + + +def deserialize(h, device): + eval_model = GPT(h).to(device).bfloat16() + restore_fp32_params(eval_model) + flat_template = _unbank_state_dict(eval_model.state_dict(), h.num_layers) + with open(h.quantized_model_path, "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress(quant_blob_disk, h.compressor)), map_location="cpu" + ) + deq_flat = dequantize_mixed(quant_state["w"], quant_state["m"], flat_template) + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + deq_state = _rebank_state_dict(deq_flat, h.num_layers, h.model_dim, kv_dim, hidden_dim) + eval_model.load_state_dict(deq_state, strict=True) + return eval_model + + +def _loss_bpb(loss_sum, token_count, byte_count): + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + return val_loss, val_bpb + + +def eval_val(h, device, val_data, model, forward_logits_fn=None): + seq_len = h.eval_seq_len + local_batch_tokens = h.val_batch_tokens // (h.world_size * h.grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + f"VAL_BATCH_SIZE must provide at least one sequence per rank; got VAL_BATCH_SIZE={h.val_batch_tokens}, WORLD_SIZE={h.world_size}, GRAD_ACCUM_STEPS={h.grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_data.val_tokens.numel() - 1) // seq_len + seq_start = total_seqs * h.rank // h.world_size + seq_end = total_seqs * (h.rank + 1) // h.world_size + + # TODO: Don't truncate this. + seq_end = seq_start + ((seq_end - seq_start) // local_batch_seqs) * local_batch_seqs + + 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) + run_forward_logits = ( + (model.module.forward_logits if hasattr(model, "module") else model.forward_logits) + if forward_logits_fn is None + else forward_logits_fn + ) + model.eval() + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + with torch.no_grad(): + 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_data.val_tokens[raw_start:raw_end].to( + device=device, dtype=torch.int64, non_blocking=True + ) + x = local[:-1] + y = local[1:] + bos_pos = (x == BOS_ID).nonzero(as_tuple=True)[0].tolist() + cu_seqlens, max_seqlen = _build_cu_seqlens( + bos_pos, x.numel(), x.device, h.eval_seq_len, 64 + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = run_forward_logits( + x[None], cu_seqlens=cu_seqlens, max_seqlen=max_seqlen + ).detach() + per_token_loss = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y.reshape(-1), + reduction="none", + ) + val_loss_sum += per_token_loss.to(torch.float64).sum() + val_token_count += float(y.numel()) + prev_ids = x + tgt_ids = y + token_bytes = val_data.base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += ( + val_data.has_leading_space_lut[tgt_ids] + & ~val_data.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) + model.train() + return _loss_bpb(val_loss_sum, val_token_count, val_byte_count) + + +def eval_val_sliding(h, device, val_data, base_model, forward_logits_fn=None, batch_seqs=32): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + run_forward_logits = base_model.forward_logits if forward_logits_fn is None else forward_logits_fn + seq_len = h.eval_seq_len + stride = h.eval_stride + total_tokens = val_data.val_tokens.numel() - 1 + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + total_windows = len(window_starts) + my_s = (total_windows * h.rank) // h.world_size + my_e = (total_windows * (h.rank + 1)) // h.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) + total_batches = (len(my_windows) + batch_seqs - 1) // batch_seqs + is_master = h.rank == 0 + cu_bucket = 64 + t_sw_start = time.perf_counter() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_idx = bi // batch_seqs + if is_master and (batch_idx % 50 == 0 or batch_idx == total_batches - 1): + elapsed = time.perf_counter() - t_sw_start + rl = float(loss_sum.item() / token_count.item()) if token_count.item() > 0 else 0.0 + rb = float((rl / math.log(2.0)) * token_count.item() / byte_count.item()) if byte_count.item() > 0 else 0.0 + log(f"sliding_progress: batch {batch_idx+1}/{total_batches} " + f"tokens:{int(token_count.item())} running_loss:{rl:.4f} running_bpb:{rb:.4f} " + f"elapsed:{elapsed:.1f}s") + batch_ws = my_windows[bi:bi + batch_seqs] + x_parts = [] + y_parts = [] + cu_starts = [] + score_ranges = [] + offset = 0 + for ws in batch_ws: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + chunk_cpu = val_data.val_tokens[ws:end + 1] + bos_pos = (chunk_cpu[:-1] == BOS_ID).nonzero(as_tuple=True)[0].tolist() + if not bos_pos or bos_pos[0] != 0: + bos_pos = [0] + bos_pos + cu_starts.extend(offset + pos for pos in bos_pos) + chunk = chunk_cpu.to(dtype=torch.int64, device=device) + x_parts.append(chunk[:-1]) + y_parts.append(chunk[1:]) + score_ranges.append((offset, wlen, ws)) + offset += wlen + x_cat = torch.cat(x_parts, dim=0)[None] + y_cat = torch.cat(y_parts, dim=0) + boundaries = cu_starts + [offset] + padded_len = get_next_multiple_of_n(len(boundaries), cu_bucket) + cu_seqlens = torch.full((padded_len,), offset, dtype=torch.int32, device=device) + cu_seqlens[:len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = run_forward_logits(x_cat, cu_seqlens=cu_seqlens, max_seqlen=seq_len) + flat_nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_cat, + reduction="none", + ) + flat_x = x_cat.reshape(-1) + for off, wlen, ws in score_ranges: + s = 0 if ws == 0 else context_size + lo = off + s + hi = off + wlen + scored_nll = flat_nll[lo:hi].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(hi - lo) + tgt = y_cat[lo:hi] + prev = flat_x[lo:hi] + tb = val_data.base_bytes_lut[tgt].to(torch.float64) + tb += (val_data.has_leading_space_lut[tgt] & ~val_data.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) + base_model.train() + return _loss_bpb(loss_sum, token_count, byte_count) + + +def _find_docs(all_tokens): + bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].numpy() + docs = [] + for i in range(len(bos_positions)): + start = int(bos_positions[i]) + end = ( + int(bos_positions[i + 1]) + if i + 1 < len(bos_positions) + else all_tokens.numel() + ) + if i + 1 < len(bos_positions): + end += 1 + assert end - start >= 2 + docs.append((start, end - start)) + return docs + + +def _build_ttt_global_batches(doc_entries, h, ascending=False): + batch_size = h.ttt_batch_size + global_doc_entries = sorted(doc_entries, key=lambda x: x[1][1]) + global_batches = [ + global_doc_entries[i : i + batch_size] + for i in range(0, len(global_doc_entries), batch_size) + ] + indexed = list(enumerate(global_batches)) + if not ascending: + indexed.sort(key=lambda ib: -max(dl for _, (_, dl) in ib[1])) + return indexed + + +def _init_batch_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(4, "little")) + + +def _claim_next_batch(counter_path, queue_len): + try: + with open(counter_path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + idx = int.from_bytes(f.read(4), "little") + f.seek(0) + f.write((idx + 1).to_bytes(4, "little")) + f.flush() + except FileNotFoundError: + return queue_len + return idx + + +def _compute_chunk_window(ci, pred_len, num_chunks, chunk_size, eval_seq_len): + chunk_end = pred_len if ci == num_chunks - 1 else (ci + 1) * chunk_size + win_start = max(0, chunk_end - eval_seq_len) + win_len = chunk_end - win_start + chunk_start = ci * chunk_size + chunk_offset = chunk_start - win_start + chunk_len = chunk_end - chunk_start + return win_start, win_len, chunk_offset, chunk_len + + +def _accumulate_bpb( + ptl, + x, + y, + chunk_offsets, + chunk_lens, + pos_idx, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, +): + pos = pos_idx[: x.size(1)].unsqueeze(0) + mask = ( + (chunk_lens.unsqueeze(1) > 0) + & (pos >= chunk_offsets.unsqueeze(1)) + & (pos < (chunk_offsets + chunk_lens).unsqueeze(1)) + ) + mask_f64 = mask.to(torch.float64) + tok_bytes = base_bytes_lut[y].to(torch.float64) + tok_bytes += (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).to( + torch.float64 + ) + loss_sum += (ptl.to(torch.float64) * mask_f64).sum() + byte_sum += (tok_bytes * mask_f64).sum() + token_count += chunk_lens.to(torch.float64).sum() + +def eval_val_ttt_lora(h, base_model, device, val_data, forward_ttt_train): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(False) + all_tokens = val_data.val_tokens + all_tokens_idx = all_tokens.to(torch.int32) + docs = _find_docs(all_tokens) + doc_entries = list(enumerate(docs)) + if h.val_doc_fraction < 1.0: + sample_n = max(1, int(round(len(docs) * h.val_doc_fraction))) + sampled_indices = sorted( + random.Random(h.seed).sample(range(len(docs)), sample_n) + ) + doc_entries = [(i, docs[i]) for i in sampled_indices] + log( + f"ttt_lora:docs:{len(doc_entries)} rank:{h.ttt_lora_rank} lr:{h.ttt_lora_lr} chunk:{h.ttt_chunk_size}" + ) + if os.environ.get("TTT_DEBUG_BYPASS") and h.rank == 0: + test_doc = doc_entries[0][1] + ds, dl = test_doc + log(f"DEBUG: test doc start={ds} len={dl}") + toks = all_tokens_idx[ds : ds + dl].to(device=device, dtype=torch.int64) + x_d = toks[:-1].unsqueeze(0) + y_d = toks[1:].unsqueeze(0) + with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_d = base_model.forward_logits(x_d) + ptl_d = F.cross_entropy( + logits_d.float().reshape(-1, logits_d.size(-1)), + y_d.reshape(-1), reduction="none", + ) + direct_loss = ptl_d.mean().item() + direct_bpb = direct_loss / math.log(2.0) + log(f"DEBUG: direct forward_logits loss={direct_loss:.6f} bpb={direct_bpb:.6f} ntokens={y_d.numel()}") + toks_first5 = toks[:5].tolist() + ptl_first5 = ptl_d[:5].tolist() + log(f"DEBUG: first 5 tokens={toks_first5} ptl={[f'{v:.4f}' for v in ptl_first5]}") + chunk_size, eval_seq_len = h.ttt_chunk_size, h.ttt_eval_seq_len + eval_batch_set = None + if h.ttt_eval_batches: + eval_batch_set = set(int(x) for x in h.ttt_eval_batches.split(",") if x.strip()) + use_ascending = eval_batch_set is not None + global_batches_sorted = _build_ttt_global_batches(doc_entries, h, ascending=use_ascending) + queue_len = len(global_batches_sorted) + counter_path = f"/tmp/ttt_counter_{h.run_id}" + if h.rank == 0: + _init_batch_counter(counter_path) + if dist.is_available() and dist.is_initialized(): + path_list = [counter_path] + dist.broadcast_object_list(path_list, src=0) + counter_path = path_list[0] + dist.barrier() + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + byte_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + t_start = time.perf_counter() + local_batch_count = 0 + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + + def _build_opt(lora): + if h.ttt_optimizer == "sgd": + return torch.optim.SGD( + lora.parameters(), lr=h.ttt_lora_lr, + momentum=h.ttt_beta1, weight_decay=h.ttt_weight_decay, + ) + return torch.optim.AdamW( + lora.parameters(), lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, weight_decay=h.ttt_weight_decay, fused=True, + ) + + reusable_opt = _build_opt(reusable_lora) + progress_f = None + if h.ttt_output_dir and h.rank == 0: + os.makedirs(h.ttt_output_dir, exist_ok=True) + progress_f = open(os.path.join(h.ttt_output_dir, "progress.jsonl"), "w") + try: + while True: + queue_idx = _claim_next_batch(counter_path, queue_len) + if queue_idx >= queue_len: + break + orig_batch_idx, batch_entries = global_batches_sorted[queue_idx] + batch = [doc for _, doc in batch_entries] + bsz = len(batch) + prev_loss = loss_sum.item() + prev_bytes = byte_sum.item() + prev_tokens = token_count.item() + if bsz == reusable_lora.bsz: + reusable_lora.reset() + for s in reusable_opt.state.values(): + for k, v in s.items(): + if isinstance(v, torch.Tensor): + v.zero_() + elif k == "step": + s[k] = 0 + cur_lora = reusable_lora + cur_opt = reusable_opt + else: + cur_lora = BatchedTTTLoRA( + bsz, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + cur_opt = _build_opt(cur_lora) + pred_lens = [doc_len - 1 for _, doc_len in batch] + num_chunks = [(pl + chunk_size - 1) // chunk_size for pl in pred_lens] + max_nc = max(num_chunks) + num_chunks_t = torch.tensor(num_chunks, dtype=torch.int64, device=device) + for ci in range(max_nc): + active = [ci < nc for nc in num_chunks] + needs_train = any(ci < nc - 1 for nc in num_chunks) + tok_starts = torch.zeros(bsz, dtype=torch.int64) + tok_wls = torch.zeros(bsz, dtype=torch.int64) + chunk_offsets_cpu = torch.zeros(bsz, dtype=torch.int64) + chunk_lens_cpu = torch.zeros(bsz, dtype=torch.int64) + for b in range(bsz): + if not active[b]: + continue + doc_start, doc_len = batch[b] + win_start, win_len, chunk_offset, chunk_len = _compute_chunk_window( + ci, pred_lens[b], num_chunks[b], chunk_size, eval_seq_len + ) + tok_starts[b] = doc_start + win_start + tok_wls[b] = win_len + chunk_offsets_cpu[b] = chunk_offset + chunk_lens_cpu[b] = chunk_len + _, context_size, chunk_offset, _ = _compute_chunk_window( + ci, (ci + 1) * chunk_size, ci + 1, chunk_size, eval_seq_len + ) + col_idx = torch.arange(context_size + 1) + idx = tok_starts.unsqueeze(1) + col_idx.unsqueeze(0) + idx.clamp_(max=all_tokens.numel() - 1) + gathered_gpu = all_tokens_idx[idx].to( + device=device, dtype=torch.int64, non_blocking=True + ) + valid = (col_idx[:context_size].unsqueeze(0) < tok_wls.unsqueeze(1)).to( + device, non_blocking=True + ) + chunk_offsets = chunk_offsets_cpu.to(device, non_blocking=True) + chunk_lens = chunk_lens_cpu.to(device, non_blocking=True) + x = torch.where(valid, gathered_gpu[:, :context_size], 0) + y = torch.where(valid, gathered_gpu[:, 1 : context_size + 1], 0) + ctx_pos = torch.arange(context_size, device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + with torch.no_grad(): + _accumulate_bpb( + per_tok_loss, + x, + y, + chunk_offsets, + chunk_lens, + ctx_pos, + val_data.base_bytes_lut, + val_data.has_leading_space_lut, + val_data.is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + ) + if needs_train: + activate_chunk_mask = (num_chunks_t - 1 > ci).float() + for gi in range(h.ttt_grad_steps): + if gi > 0: + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + per_doc = per_tok_loss[ + :, chunk_offset : chunk_offset + chunk_size + ].mean(dim=-1) + cur_opt.zero_grad(set_to_none=True) + (per_doc * activate_chunk_mask).sum().backward() + cur_opt.step() + else: + del per_tok_loss + local_batch_count += 1 + batch_num = orig_batch_idx + 1 + doc_lens = [dl for _, dl in batch] + should_report = False + if eval_batch_set is not None: + should_report = batch_num in eval_batch_set + else: + # should_report = local_batch_count % 10 == 0 + should_report = True + if should_report: + cur_tokens = token_count.item() + cur_loss_val = loss_sum.item() + cur_bytes_val = byte_sum.item() + dt = cur_tokens - prev_tokens + if dt > 0: + b_loss = (cur_loss_val - prev_loss) / dt + b_bpb = b_loss / math.log(2.0) * (dt / (cur_bytes_val - prev_bytes)) + else: + b_loss = b_bpb = 0.0 + r_loss = cur_loss_val / max(cur_tokens, 1) + r_bpb = r_loss / math.log(2.0) * (cur_tokens / max(cur_bytes_val, 1)) + elapsed = time.perf_counter() - t_start + log( + f"ttt_progress: batch {batch_num}/{queue_len} batch_loss:{b_loss:.4f} " + f"batch_bpb:{b_bpb:.4f} running_loss:{r_loss:.4f} running_bpb:{r_bpb:.4f} " + f"doc_len:{min(doc_lens)}-{max(doc_lens)}" + ) + if progress_f is not None: + progress_f.write( + json.dumps({ + "batch": batch_num, "total_batches": queue_len, + "batch_loss": round(b_loss, 8), "batch_bpb": round(b_bpb, 8), + "running_loss": round(r_loss, 8), "running_bpb": round(r_bpb, 8), + "doc_len_min": min(doc_lens), "doc_len_max": max(doc_lens), + "chunk_size": chunk_size, + "elapsed_s": round(elapsed, 3), + "batch_t_s": round(elapsed, 3), + }) + "\n" + ) + progress_f.flush() + del cur_lora, cur_opt + finally: + if progress_f is not None: + progress_f.close() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.train() + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_sum.item()) + return val_loss, val_bpb + + +def timed_eval(label, fn, *args, **kwargs): + torch.cuda.synchronize() + t0 = time.perf_counter() + val_loss, val_bpb = fn(*args, **kwargs) + torch.cuda.synchronize() + elapsed_ms = 1e3 * (time.perf_counter() - t0) + log( + f"{label} val_loss:{val_loss:.8f} val_bpb:{val_bpb:.8f} eval_time:{elapsed_ms:.0f}ms" + ) + return val_loss, val_bpb + + +def train_model(h, device, val_data): + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True + ) + model = compiled_model + log(f"model_params:{sum(p.numel()for p in base_model.parameters())}") + optimizers = Optimizers(h, base_model) + train_loader = DocumentPackingLoader(h, device) + max_wallclock_ms = ( + 1e3 * h.max_wallclock_seconds if h.max_wallclock_seconds > 0 else None + ) + if max_wallclock_ms is not None: + max_wallclock_ms -= h.gptq_reserve_seconds * 1e3 + log( + f"gptq:reserving {h.gptq_reserve_seconds:.0f}s, effective={max_wallclock_ms:.0f}ms" + ) + + def training_frac(step, elapsed_ms): + if max_wallclock_ms is None: + return step / max(h.iterations, 1) + return elapsed_ms / max(max_wallclock_ms, 1e-09) + + def lr_mul(frac): + if h.warmdown_frac <= 0: + return 1.0 + if frac >= 1.0 - h.warmdown_frac: + return max((1.0 - frac) / h.warmdown_frac, h.min_lr) + return 1.0 + + def step_fn(step, lr_scale): + optimizers.zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(h.grad_accum_steps): + x, y, cu_seqlens, _max_seqlen = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y, cu_seqlens=cu_seqlens, max_seqlen=h.train_seq_len) + train_loss += loss.detach() + (loss / h.grad_accum_steps).backward() + train_loss /= h.grad_accum_steps + frac = ( + min(step / h.muon_momentum_warmup_steps, 1.0) + if h.muon_momentum_warmup_steps > 0 + else 1.0 + ) + muon_momentum = ( + 1 - frac + ) * h.muon_momentum_warmup_start + frac * h.muon_momentum + for group in optimizers.optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * lr_scale + if h.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), h.grad_clip_norm) + optimizers.step(distributed=h.distributed) + return train_loss + + if h.warmup_steps > 0: + initial_model_state = { + name: tensor.detach().cpu().clone() + for (name, tensor) in base_model.state_dict().items() + } + initial_optimizer_states = [ + copy.deepcopy(opt.state_dict()) for opt in optimizers + ] + model.train() + num_tokens_local = h.train_batch_tokens // h.world_size + for blk in base_model.blocks: + blk.attn.rotary(num_tokens_local, device, torch.bfloat16) + cu_bucket_size = train_loader.cu_bucket_size + warmup_cu_buckets = tuple(cu_bucket_size * i for i in range(1, 5)) + warmup_cu_iters = 3 + x, y, cu_seqlens, _ = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + log(f"warmup_cu_buckets:{','.join(str(b) for b in warmup_cu_buckets)} iters_each:{warmup_cu_iters}") + def _run_cu_bucket_warmup(): + for bucket_len in warmup_cu_buckets: + boundaries = list(range(0, x.size(1), max(h.train_seq_len, 1))) + if boundaries[-1] != x.size(1): + boundaries.append(x.size(1)) + cu = torch.full((bucket_len,), x.size(1), dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + for _ in range(warmup_cu_iters): + optimizers.zero_grad_all() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + wloss = model(x, y, cu_seqlens=cu, max_seqlen=h.train_seq_len) + (wloss / h.grad_accum_steps).backward() + optimizers.zero_grad_all() + _run_cu_bucket_warmup() + if h.num_loops > 0: + base_model.looping_active = True + _run_cu_bucket_warmup() + base_model.looping_active = False + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"warmup_step: {warmup_step+1}/{h.warmup_steps}") + if h.num_loops > 0: + base_model.looping_active = True + log( + f"loop_warmup:enabled encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"loop_warmup_step: {warmup_step+1}/{h.warmup_steps}") + base_model.looping_active = False + 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) + optimizers.zero_grad_all() + train_loader = DocumentPackingLoader(h, device) + ema_state = { + name: t.detach().float().clone() + for (name, t) in base_model.state_dict().items() + } + ema_decay = h.ema_decay + training_time_ms = 0.0 + stop_after_step = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = ( + step == h.iterations + or stop_after_step is not None + and step >= stop_after_step + ) + should_validate = ( + last_step or h.val_loss_every > 0 and step % h.val_loss_every == 0 + ) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1e3 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + h, device, val_data, model, compiled_forward_logits + ) + log( + f"{step}/{h.iterations} val_loss: {val_loss:.4f} val_bpb: {val_bpb:.4f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < h.iterations: + log( + f"stopping_early: wallclock_cap train_time: {training_time_ms:.0f}ms step: {step}/{h.iterations}" + ) + break + elapsed_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + frac = training_frac(step, elapsed_ms) + scale = lr_mul(frac) + if ( + h.num_loops > 0 + and not base_model.looping_active + and frac >= h.enable_looping_at + ): + base_model.looping_active = True + log( + f"layer_loop:enabled step:{step} frac:{frac:.3f} encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + train_loss = step_fn(step, scale) + 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 + 1e3 * (time.perf_counter() - t0) + should_log_train = h.train_log_every > 0 and ( + step <= 5 or step % h.train_log_every == 0 or stop_after_step is not None + ) + if should_log_train: + tok_per_sec = step * h.train_batch_tokens / (approx_training_time_ms / 1e3) + log( + f"{step}/{h.iterations} train_loss: {train_loss.item():.4f} train_time: {approx_training_time_ms/60000:.1f}m tok/s: {tok_per_sec:.0f}" + ) + reached_cap = ( + max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + ) + if h.distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log( + f"peak memory allocated: {torch.cuda.max_memory_allocated()//1024//1024} MiB reserved: {torch.cuda.max_memory_reserved()//1024//1024} MiB" + ) + log("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) + return base_model, compiled_model, compiled_forward_logits + + +def train_and_eval(h, device): + random.seed(h.seed) + np.random.seed(h.seed) + torch.manual_seed(h.seed) + torch.cuda.manual_seed_all(h.seed) + if h.artifact_dir and h.is_main_process: + os.makedirs(h.artifact_dir, exist_ok=True) + val_data = ValidationData(h, device) + if h.eval_only_path: + log(f"eval_only:loading checkpoint from {h.eval_only_path}") + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + base_model.load_state_dict(torch.load(h.eval_only_path, map_location=device)) + if h.num_loops > 0: + base_model.looping_active = True + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True + ) + else: + log( + f"train_shards: {len(list(Path(h.datasets_dir).resolve().glob('fineweb_train_*.bin')))}" + ) + log(f"val_tokens: {val_data.val_tokens.numel()-1}") + base_model, compiled_model, compiled_forward_logits = train_model( + h, device, val_data + ) + _skip_training = bool(h.eval_only_path) + log("\nbeginning eval timer") + t_all_eval = time.perf_counter() + + torch._dynamo.reset() + timed_eval( + "pre-quantization post-ema", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + if not _skip_training: + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + else: + log("eval_only: skipping serialize (already have quantized model)") + if not os.path.exists(h.quantized_model_path): + log("eval_only: no quantized model found, running serialize anyway") + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + if h.distributed: + dist.barrier() + eval_model = deserialize(h, device) + if h.num_loops > 0: + eval_model.looping_active = True + compiled_model = torch.compile(eval_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + eval_model.forward_logits, dynamic=False, fullgraph=True + ) + timed_eval( + "quantized", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + if h.sliding_window_enabled: + timed_eval( + "quantized_sliding_window", + eval_val_sliding, + h, + device, + val_data, + eval_model, + forward_logits_fn=compiled_forward_logits, + ) + ttt_compile_time = 0.0 + if h.ttt_enabled: + del eval_model, compiled_model + torch._dynamo.reset() + torch.cuda.empty_cache() + ttt_model = deserialize(h, device) + if h.num_loops > 0: + ttt_model.looping_active = True + for p in ttt_model.parameters(): + p.requires_grad_(False) + + if h.rope_yarn: + _yarn_seqlen = h.train_batch_tokens // h.grad_accum_steps + for block in ttt_model.blocks: + block.attn.rotary(_yarn_seqlen, device, torch.bfloat16) + else: + for block in ttt_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + block.attn.rotary(h.ttt_eval_seq_len, device, torch.bfloat16) + + def _fwd_ttt(input_ids, target_ids, lora): + return ttt_model.forward_ttt(input_ids, target_ids, lora=lora) + + _ttt_debug_bypass = bool(os.environ.get("TTT_DEBUG_BYPASS")) + if _ttt_debug_bypass: + def _fwd_ttt_bypass(input_ids, target_ids, lora): + logits = ttt_model.forward_logits(input_ids) + dummy = lora.q_loras[0].B.sum() * 0 + logits = logits + dummy + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + fwd_ttt_compiled = _fwd_ttt_bypass + log("ttt_lora:DEBUG BYPASS active - using forward_logits directly (no compile warmup)") + ttt_compile_time = 0.0 + else: + fwd_ttt_compiled = torch.compile(_fwd_ttt, dynamic=True) + log(f"ttt_lora:warming up compile") + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + ds0 = 0 + val_tokens_idx = val_data.val_tokens.to(torch.int32) + t_warmup = time.perf_counter() + warmup_bszes = [h.ttt_batch_size] + for bsz in warmup_bszes: + wl = BatchedTTTLoRA( + bsz, ttt_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + wo = torch.optim.AdamW( + wl.parameters(), + lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, + weight_decay=h.ttt_weight_decay, + fused=True, + ) + for ctx_len in (h.ttt_chunk_size, h.ttt_eval_seq_len): + col_w = torch.arange(ctx_len + 1) + idx_w = (ds0 + col_w).clamp_(max=val_data.val_tokens.numel() - 1) + row_w = val_tokens_idx[idx_w].to(device=device, dtype=torch.int64) + xw = row_w[:ctx_len].unsqueeze(0).expand(bsz, -1).contiguous() + yw = row_w[1 : ctx_len + 1].unsqueeze(0).expand(bsz, -1).contiguous() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = fwd_ttt_compiled(xw, yw, lora=wl) + ptl[:, : min(h.ttt_chunk_size, ctx_len)].mean(dim=-1).sum().backward() + wo.step() + wo.zero_grad(set_to_none=True) + del wl, wo + del val_tokens_idx + torch.cuda.empty_cache() + ttt_compile_time = time.perf_counter() - t_warmup + log(f"ttt_lora:compile warmup done ({ttt_compile_time:.1f}s)") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt_lora( + h, ttt_model, device, val_data, forward_ttt_train=fwd_ttt_compiled + ) + torch.cuda.synchronize() + log( + f"quantized_ttt_lora val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f} eval_time:{1e3*(time.perf_counter()-t_ttt):.0f}ms" + ) + del ttt_model + + total_eval = time.perf_counter() - t_all_eval + log(f"total_eval_time:{total_eval - ttt_compile_time:.1f}s") + log(f"total_eval_time_with_compile:{total_eval:.1f}s") + + +def main(): + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + 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" + ) + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + torch.set_float32_matmul_precision("high") + 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) + torch._dynamo.config.optimize_ddp = False + torch._dynamo.config.cache_size_limit = 16 + h = Hyperparameters() + set_logging_hparams(h) + if h.is_main_process: + os.makedirs(h.artifact_dir if h.artifact_dir else "logs", exist_ok=True) + log(100 * "=", console=False) + log("Hyperparameters:", console=True) + for (k, v) in sorted(vars(type(h)).items()): + if not k.startswith("_"): + log(f" {k}: {v}", console=True) + log("=" * 100, console=False) + log(f"Running Python {sys.version}", console=False) + log(f"Running PyTorch {torch.__version__}", console=False) + log( + subprocess.run( + ["nvidia-smi"], + stdout=subprocess.PIPE, + stderr=subprocess.PIPE, + text=True, + check=False, + ).stdout, + console=False, + ) + log("=" * 100, console=False) + train_and_eval(h, device) + if distributed: + dist.destroy_process_group() + + +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed0.log b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed0.log new file mode 100644 index 0000000000..96eb6cb276 --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed0.log @@ -0,0 +1,284 @@ + +***************************************** +Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: + beta1: 0.9 + beta2: 0.95 + compressor: brotli + data_dir: /home/dex/parameter-golf-with-cc/data + datasets_dir: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192 + distributed: True + ema_decay: 0.997 + embed_bits: 8 + embed_clip_sigmas: 20.0 + embed_lr: 0.6 + embed_wd: 0.095 + embedding_dim: 512 + enable_looping_at: 0.35 + etlb_clip: 3.0 + etlb_lr: 0.05 + etlb_steps: 5 + eval_only_path: + eval_seq_len: 2048 + eval_stride: 64 + gptq_calibration_batches: 64 + gptq_reserve_seconds: 12.0 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + head_lr: 0.008 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: logs/PR1530_notriton_v3_s0.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.022 + max_wallclock_seconds: 600.0 + min_lr: 0.0 + mlp_mult: 4.0 + model_dim: 512 + model_path: final_model.pt + muon_backend_steps: 5 + muon_beta2: 0.95 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_start_layer: 7 + qk_gain_init: 5.0 + quantized_model_path: final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: PR1530_notriton_v3_s0 + scalar_lr: 0.02 + seed: 0 + skip_gates_enabled: True + sliding_window_enabled: False + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /home/dex/parameter-golf-with-cc/data/tokenizers/fineweb_8192_bpe.model + train_batch_tokens: 786432 + train_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.999 + ttt_chunk_size: 64 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 96 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_output_dir: + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_doc_fraction: 1.0 + val_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.667 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 40540160 +model_params:35944537 +gptq:reserving 12s, effective=588000ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +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: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0085 val_bpb: 3.4874 +1/20000 train_loss: 9.0089 train_time: 0.0m tok/s: 16239426 +2/20000 train_loss: 12.2349 train_time: 0.0m tok/s: 12723200 +3/20000 train_loss: 11.1726 train_time: 0.0m tok/s: 10874459 +4/20000 train_loss: 9.5569 train_time: 0.0m tok/s: 10084457 +5/20000 train_loss: 8.1986 train_time: 0.0m tok/s: 9708425 +500/20000 train_loss: 3.2889 train_time: 0.8m tok/s: 8161633 +1000/20000 train_loss: 3.0354 train_time: 1.6m tok/s: 8139088 +1500/20000 train_loss: 3.0423 train_time: 2.4m tok/s: 8130632 +2000/20000 train_loss: 3.0086 train_time: 3.2m tok/s: 8129048 +layer_loop:enabled step:2127 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 3.0891 train_time: 4.3m tok/s: 7594822 +3000/20000 train_loss: 2.9236 train_time: 5.5m tok/s: 7151380 +3500/20000 train_loss: 2.9870 train_time: 6.7m tok/s: 6866727 +4000/20000 train_loss: 2.9132 train_time: 7.9m tok/s: 6665542 +4000/20000 val_loss: 2.8878 val_bpb: 1.1179 +4500/20000 train_loss: 2.8566 train_time: 9.0m tok/s: 6519538 +4820/20000 val_loss: 2.7807 val_bpb: 1.0765 +stopping_early: wallclock_cap train_time: 588142ms step: 4820/20000 +peak memory allocated: 40019 MiB reserved: 44090 MiB +ema:applying EMA weights + +beginning eval timer +pre-quantization post-ema val_loss:2.78135262 val_bpb:1.07671358 eval_time:6572ms +Serialized model: 135408623 bytes +Code size (uncompressed): 25659 bytes +Code size (compressed): 26505 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 12.4s +Quantized weights: + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int8): tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights +Serialized model quantized+brotli: 15968674 bytes +Total submission size quantized+brotli: 15995179 bytes +quantized val_loss:2.80936176 val_bpb:1.08755643 eval_time:8271ms +ttt_lora:warming up compile +ttt_lora:compile warmup done (85.6s) +ttt_lora:docs:50000 rank:96 lr:0.0001 chunk:64 +ttt_progress: batch 779/782 batch_loss:2.6598 batch_bpb:1.0834 running_loss:2.6598 running_bpb:1.0834 doc_len:9037-11049 +ttt_progress: batch 771/782 batch_loss:2.7707 batch_bpb:1.0833 running_loss:2.6961 running_bpb:1.0834 doc_len:4701-4937 +ttt_progress: batch 766/782 batch_loss:2.5785 batch_bpb:1.0095 running_loss:2.6715 running_bpb:1.0676 doc_len:3846-3962 +ttt_progress: batch 761/782 batch_loss:2.7626 batch_bpb:1.0686 running_loss:2.6855 running_bpb:1.0678 doc_len:3336-3430 +ttt_progress: batch 755/782 batch_loss:2.7033 batch_bpb:1.0474 running_loss:2.6876 running_bpb:1.0653 doc_len:2899-2972 +ttt_progress: batch 749/782 batch_loss:2.8417 batch_bpb:1.0933 running_loss:2.7022 running_bpb:1.0680 doc_len:2580-2638 +ttt_progress: batch 742/782 batch_loss:2.7952 batch_bpb:1.0706 running_loss:2.7095 running_bpb:1.0682 doc_len:2319-2353 +ttt_progress: batch 738/782 batch_loss:2.7553 batch_bpb:1.0571 running_loss:2.7126 running_bpb:1.0675 doc_len:2194-2227 +ttt_progress: batch 732/782 batch_loss:2.8324 batch_bpb:1.1023 running_loss:2.7198 running_bpb:1.0696 doc_len:2041-2062 +ttt_progress: batch 722/782 batch_loss:2.7801 batch_bpb:1.0629 running_loss:2.7229 running_bpb:1.0692 doc_len:1846-1861 +ttt_progress: batch 716/782 batch_loss:2.8213 batch_bpb:1.0411 running_loss:2.7275 running_bpb:1.0678 doc_len:1739-1754 +ttt_progress: batch 710/782 batch_loss:2.7667 batch_bpb:1.0727 running_loss:2.7291 running_bpb:1.0680 doc_len:1661-1673 +ttt_progress: batch 700/782 batch_loss:2.6876 batch_bpb:1.0490 running_loss:2.7275 running_bpb:1.0673 doc_len:1552-1562 +ttt_progress: batch 697/782 batch_loss:2.7714 batch_bpb:1.0441 running_loss:2.7291 running_bpb:1.0665 doc_len:1522-1534 +ttt_progress: batch 692/782 batch_loss:2.7819 batch_bpb:1.0555 running_loss:2.7309 running_bpb:1.0661 doc_len:1477-1484 +ttt_progress: batch 681/782 batch_loss:2.8281 batch_bpb:1.0737 running_loss:2.7339 running_bpb:1.0663 doc_len:1383-1393 +ttt_progress: batch 676/782 batch_loss:2.8020 batch_bpb:1.0706 running_loss:2.7358 running_bpb:1.0664 doc_len:1347-1353 +ttt_progress: batch 667/782 batch_loss:2.8278 batch_bpb:1.1078 running_loss:2.7383 running_bpb:1.0676 doc_len:1288-1295 +ttt_progress: batch 660/782 batch_loss:2.8645 batch_bpb:1.0960 running_loss:2.7415 running_bpb:1.0683 doc_len:1245-1250 +ttt_progress: batch 653/782 batch_loss:2.7694 batch_bpb:1.0390 running_loss:2.7422 running_bpb:1.0676 doc_len:1203-1209 +ttt_progress: batch 650/782 batch_loss:2.8001 batch_bpb:1.0782 running_loss:2.7435 running_bpb:1.0678 doc_len:1188-1193 +ttt_progress: batch 643/782 batch_loss:2.8022 batch_bpb:1.0683 running_loss:2.7448 running_bpb:1.0678 doc_len:1150-1155 +ttt_progress: batch 632/782 batch_loss:2.7480 batch_bpb:1.0322 running_loss:2.7449 running_bpb:1.0671 doc_len:1096-1101 +ttt_progress: batch 626/782 batch_loss:2.8204 batch_bpb:1.0480 running_loss:2.7463 running_bpb:1.0667 doc_len:1068-1073 +ttt_progress: batch 619/782 batch_loss:2.8017 batch_bpb:1.0614 running_loss:2.7474 running_bpb:1.0666 doc_len:1037-1041 +ttt_progress: batch 613/782 batch_loss:2.8323 batch_bpb:1.0660 running_loss:2.7489 running_bpb:1.0666 doc_len:1012-1016 +ttt_progress: batch 607/782 batch_loss:2.7027 batch_bpb:1.0416 running_loss:2.7481 running_bpb:1.0661 doc_len:986-990 +ttt_progress: batch 600/782 batch_loss:2.8042 batch_bpb:1.0645 running_loss:2.7490 running_bpb:1.0661 doc_len:958-963 +ttt_progress: batch 591/782 batch_loss:2.6782 batch_bpb:1.0120 running_loss:2.7479 running_bpb:1.0653 doc_len:927-930 +ttt_progress: batch 585/782 batch_loss:2.7722 batch_bpb:1.0689 running_loss:2.7483 running_bpb:1.0653 doc_len:908-911 +ttt_progress: batch 578/782 batch_loss:2.8139 batch_bpb:1.0721 running_loss:2.7492 running_bpb:1.0654 doc_len:884-887 +ttt_progress: batch 571/782 batch_loss:2.7193 batch_bpb:1.0373 running_loss:2.7488 running_bpb:1.0650 doc_len:862-865 +ttt_progress: batch 563/782 batch_loss:2.8136 batch_bpb:1.0674 running_loss:2.7496 running_bpb:1.0650 doc_len:837-840 +ttt_progress: batch 555/782 batch_loss:2.7734 batch_bpb:1.0584 running_loss:2.7499 running_bpb:1.0650 doc_len:812-815 +ttt_progress: batch 549/782 batch_loss:2.7707 batch_bpb:1.0660 running_loss:2.7502 running_bpb:1.0650 doc_len:795-798 +ttt_progress: batch 543/782 batch_loss:2.7909 batch_bpb:1.0477 running_loss:2.7507 running_bpb:1.0648 doc_len:779-782 +ttt_progress: batch 538/782 batch_loss:2.6987 batch_bpb:1.0437 running_loss:2.7501 running_bpb:1.0645 doc_len:767-769 +ttt_progress: batch 531/782 batch_loss:2.7845 batch_bpb:1.0561 running_loss:2.7505 running_bpb:1.0644 doc_len:750-752 +ttt_progress: batch 524/782 batch_loss:2.8216 batch_bpb:1.0544 running_loss:2.7512 running_bpb:1.0643 doc_len:732-735 +ttt_progress: batch 517/782 batch_loss:2.7841 batch_bpb:1.0538 running_loss:2.7516 running_bpb:1.0642 doc_len:715-717 +ttt_progress: batch 509/782 batch_loss:2.7549 batch_bpb:1.0721 running_loss:2.7516 running_bpb:1.0643 doc_len:695-698 +ttt_progress: batch 503/782 batch_loss:2.8342 batch_bpb:1.0793 running_loss:2.7524 running_bpb:1.0644 doc_len:683-685 +ttt_progress: batch 496/782 batch_loss:2.8501 batch_bpb:1.0563 running_loss:2.7533 running_bpb:1.0643 doc_len:666-668 +ttt_progress: batch 489/782 batch_loss:2.8052 batch_bpb:1.0847 running_loss:2.7538 running_bpb:1.0645 doc_len:651-653 +ttt_progress: batch 482/782 batch_loss:2.7580 batch_bpb:1.0824 running_loss:2.7538 running_bpb:1.0647 doc_len:637-639 +ttt_progress: batch 475/782 batch_loss:2.7397 batch_bpb:1.0271 running_loss:2.7537 running_bpb:1.0644 doc_len:622-623 +ttt_progress: batch 468/782 batch_loss:2.7904 batch_bpb:1.0592 running_loss:2.7540 running_bpb:1.0643 doc_len:608-610 +ttt_progress: batch 461/782 batch_loss:2.7893 batch_bpb:1.0637 running_loss:2.7543 running_bpb:1.0643 doc_len:595-597 +ttt_progress: batch 455/782 batch_loss:2.8067 batch_bpb:1.0763 running_loss:2.7547 running_bpb:1.0644 doc_len:584-586 +ttt_progress: batch 449/782 batch_loss:2.8091 batch_bpb:1.0573 running_loss:2.7551 running_bpb:1.0643 doc_len:573-575 +ttt_progress: batch 444/782 batch_loss:2.6805 batch_bpb:1.0156 running_loss:2.7546 running_bpb:1.0640 doc_len:564-566 +ttt_progress: batch 438/782 batch_loss:2.7217 batch_bpb:1.0589 running_loss:2.7543 running_bpb:1.0639 doc_len:553-555 +ttt_progress: batch 432/782 batch_loss:2.7686 batch_bpb:1.0533 running_loss:2.7544 running_bpb:1.0639 doc_len:542-544 +ttt_progress: batch 425/782 batch_loss:2.7688 batch_bpb:1.0534 running_loss:2.7545 running_bpb:1.0638 doc_len:530-532 +ttt_progress: batch 418/782 batch_loss:2.8239 batch_bpb:1.0772 running_loss:2.7550 running_bpb:1.0639 doc_len:517-519 +ttt_progress: batch 411/782 batch_loss:2.8208 batch_bpb:1.0754 running_loss:2.7554 running_bpb:1.0640 doc_len:507-508 +ttt_progress: batch 404/782 batch_loss:2.7979 batch_bpb:1.0736 running_loss:2.7557 running_bpb:1.0640 doc_len:495-497 +ttt_progress: batch 397/782 batch_loss:2.8961 batch_bpb:1.1002 running_loss:2.7565 running_bpb:1.0642 doc_len:484-486 +ttt_progress: batch 390/782 batch_loss:2.8260 batch_bpb:1.0961 running_loss:2.7570 running_bpb:1.0644 doc_len:473-475 +ttt_progress: batch 383/782 batch_loss:2.8426 batch_bpb:1.0886 running_loss:2.7575 running_bpb:1.0646 doc_len:463-464 +ttt_progress: batch 376/782 batch_loss:2.7253 batch_bpb:1.0466 running_loss:2.7573 running_bpb:1.0645 doc_len:453-454 +ttt_progress: batch 369/782 batch_loss:2.9339 batch_bpb:1.0890 running_loss:2.7582 running_bpb:1.0646 doc_len:443-444 +ttt_progress: batch 362/782 batch_loss:2.8241 batch_bpb:1.0678 running_loss:2.7586 running_bpb:1.0646 doc_len:433-434 +ttt_progress: batch 355/782 batch_loss:2.7086 batch_bpb:1.0673 running_loss:2.7583 running_bpb:1.0646 doc_len:423-424 +ttt_progress: batch 348/782 batch_loss:2.8220 batch_bpb:1.0724 running_loss:2.7586 running_bpb:1.0647 doc_len:414-415 +ttt_progress: batch 341/782 batch_loss:2.8790 batch_bpb:1.1022 running_loss:2.7592 running_bpb:1.0649 doc_len:404-406 +ttt_progress: batch 334/782 batch_loss:2.8738 batch_bpb:1.1058 running_loss:2.7598 running_bpb:1.0651 doc_len:395-396 +ttt_progress: batch 327/782 batch_loss:2.7826 batch_bpb:1.0802 running_loss:2.7599 running_bpb:1.0651 doc_len:387-388 +ttt_progress: batch 321/782 batch_loss:2.8120 batch_bpb:1.1050 running_loss:2.7601 running_bpb:1.0653 doc_len:378-380 +ttt_progress: batch 314/782 batch_loss:2.8075 batch_bpb:1.0671 running_loss:2.7603 running_bpb:1.0653 doc_len:369-370 +ttt_progress: batch 307/782 batch_loss:2.9078 batch_bpb:1.1116 running_loss:2.7610 running_bpb:1.0655 doc_len:361-362 +ttt_progress: batch 300/782 batch_loss:2.8651 batch_bpb:1.0920 running_loss:2.7614 running_bpb:1.0656 doc_len:352-353 +ttt_progress: batch 293/782 batch_loss:2.7720 batch_bpb:1.0709 running_loss:2.7614 running_bpb:1.0656 doc_len:343-345 +ttt_progress: batch 258/782 batch_loss:2.9654 batch_bpb:1.1693 running_loss:2.7622 running_bpb:1.0660 doc_len:304-305 +ttt_progress: batch 251/782 batch_loss:2.8799 batch_bpb:1.1108 running_loss:2.7626 running_bpb:1.0662 doc_len:296-297 +ttt_progress: batch 244/782 batch_loss:2.9584 batch_bpb:1.1603 running_loss:2.7632 running_bpb:1.0665 doc_len:289-290 +ttt_progress: batch 238/782 batch_loss:2.8954 batch_bpb:1.1485 running_loss:2.7636 running_bpb:1.0667 doc_len:283-284 +ttt_progress: batch 232/782 batch_loss:2.9408 batch_bpb:1.1375 running_loss:2.7642 running_bpb:1.0670 doc_len:277-278 +ttt_progress: batch 226/782 batch_loss:2.9505 batch_bpb:1.1478 running_loss:2.7648 running_bpb:1.0672 doc_len:271-272 +ttt_progress: batch 221/782 batch_loss:2.8511 batch_bpb:1.1442 running_loss:2.7650 running_bpb:1.0674 doc_len:266-267 +ttt_progress: batch 216/782 batch_loss:2.9461 batch_bpb:1.1211 running_loss:2.7656 running_bpb:1.0676 doc_len:261-262 +ttt_progress: batch 211/782 batch_loss:2.8965 batch_bpb:1.1539 running_loss:2.7660 running_bpb:1.0678 doc_len:256-257 +ttt_progress: batch 207/782 batch_loss:2.8519 batch_bpb:1.1219 running_loss:2.7662 running_bpb:1.0680 doc_len:253-254 +ttt_progress: batch 203/782 batch_loss:2.7762 batch_bpb:1.0906 running_loss:2.7662 running_bpb:1.0680 doc_len:249-250 +ttt_progress: batch 197/782 batch_loss:2.8505 batch_bpb:1.1241 running_loss:2.7665 running_bpb:1.0682 doc_len:244-245 +ttt_progress: batch 191/782 batch_loss:2.9458 batch_bpb:1.1503 running_loss:2.7670 running_bpb:1.0684 doc_len:238-239 +ttt_progress: batch 185/782 batch_loss:2.8783 batch_bpb:1.1298 running_loss:2.7672 running_bpb:1.0686 doc_len:233-234 +ttt_progress: batch 178/782 batch_loss:2.8703 batch_bpb:1.1448 running_loss:2.7675 running_bpb:1.0688 doc_len:227-228 +ttt_progress: batch 172/782 batch_loss:3.0097 batch_bpb:1.1837 running_loss:2.7681 running_bpb:1.0690 doc_len:222-223 +ttt_progress: batch 167/782 batch_loss:2.9762 batch_bpb:1.1897 running_loss:2.7686 running_bpb:1.0693 doc_len:218-218 +ttt_progress: batch 159/782 batch_loss:3.0028 batch_bpb:1.1830 running_loss:2.7692 running_bpb:1.0696 doc_len:211-212 +ttt_progress: batch 155/782 batch_loss:2.8925 batch_bpb:1.1368 running_loss:2.7694 running_bpb:1.0697 doc_len:207-208 +ttt_progress: batch 150/782 batch_loss:2.9496 batch_bpb:1.1594 running_loss:2.7698 running_bpb:1.0699 doc_len:204-204 +ttt_progress: batch 143/782 batch_loss:3.0267 batch_bpb:1.1989 running_loss:2.7704 running_bpb:1.0702 doc_len:198-199 +ttt_progress: batch 139/782 batch_loss:3.0027 batch_bpb:1.1621 running_loss:2.7709 running_bpb:1.0704 doc_len:195-195 +ttt_progress: batch 131/782 batch_loss:3.0561 batch_bpb:1.2147 running_loss:2.7715 running_bpb:1.0707 doc_len:188-189 +ttt_progress: batch 126/782 batch_loss:2.9370 batch_bpb:1.1933 running_loss:2.7718 running_bpb:1.0709 doc_len:185-185 +ttt_progress: batch 118/782 batch_loss:2.9599 batch_bpb:1.1564 running_loss:2.7722 running_bpb:1.0711 doc_len:178-179 +ttt_progress: batch 111/782 batch_loss:2.9834 batch_bpb:1.1904 running_loss:2.7726 running_bpb:1.0713 doc_len:173-174 +ttt_progress: batch 104/782 batch_loss:2.9952 batch_bpb:1.1655 running_loss:2.7730 running_bpb:1.0715 doc_len:168-169 +ttt_progress: batch 98/782 batch_loss:2.9892 batch_bpb:1.1864 running_loss:2.7734 running_bpb:1.0717 doc_len:164-164 +ttt_progress: batch 91/782 batch_loss:3.0438 batch_bpb:1.2182 running_loss:2.7738 running_bpb:1.0719 doc_len:158-159 +ttt_progress: batch 85/782 batch_loss:2.9945 batch_bpb:1.2023 running_loss:2.7742 running_bpb:1.0721 doc_len:154-154 +ttt_progress: batch 76/782 batch_loss:3.0630 batch_bpb:1.2289 running_loss:2.7747 running_bpb:1.0724 doc_len:147-148 +ttt_progress: batch 72/782 batch_loss:2.9516 batch_bpb:1.1995 running_loss:2.7749 running_bpb:1.0726 doc_len:144-144 +ttt_progress: batch 66/782 batch_loss:3.1086 batch_bpb:1.2733 running_loss:2.7754 running_bpb:1.0728 doc_len:139-140 +ttt_progress: batch 58/782 batch_loss:2.9734 batch_bpb:1.2264 running_loss:2.7757 running_bpb:1.0731 doc_len:133-134 +ttt_progress: batch 52/782 batch_loss:3.0613 batch_bpb:1.1978 running_loss:2.7761 running_bpb:1.0732 doc_len:128-129 +ttt_progress: batch 45/782 batch_loss:3.0890 batch_bpb:1.2357 running_loss:2.7765 running_bpb:1.0734 doc_len:122-123 +ttt_progress: batch 38/782 batch_loss:3.0565 batch_bpb:1.2199 running_loss:2.7769 running_bpb:1.0736 doc_len:117-118 +ttt_progress: batch 32/782 batch_loss:3.0348 batch_bpb:1.2132 running_loss:2.7772 running_bpb:1.0738 doc_len:112-113 +ttt_progress: batch 26/782 batch_loss:3.0914 batch_bpb:1.2604 running_loss:2.7775 running_bpb:1.0740 doc_len:107-107 +ttt_progress: batch 20/782 batch_loss:3.1387 batch_bpb:1.2694 running_loss:2.7779 running_bpb:1.0742 doc_len:101-102 +ttt_progress: batch 13/782 batch_loss:3.1511 batch_bpb:1.2683 running_loss:2.7783 running_bpb:1.0744 doc_len:93-94 +ttt_progress: batch 6/782 batch_loss:3.2906 batch_bpb:1.2838 running_loss:2.7788 running_bpb:1.0745 doc_len:82-84 +quantized_ttt_lora val_loss:2.78249445 val_bpb:1.07719031 eval_time:250998ms +total_eval_time:346.3s +total_eval_time_with_compile:431.9s diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed1337.log b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed1337.log new file mode 100644 index 0000000000..f941bd8025 --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed1337.log @@ -0,0 +1,274 @@ + +***************************************** +Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: + beta1: 0.9 + beta2: 0.95 + compressor: brotli + data_dir: /home/dex/parameter-golf-with-cc/data + datasets_dir: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192 + distributed: True + ema_decay: 0.997 + embed_bits: 8 + embed_clip_sigmas: 20.0 + embed_lr: 0.6 + embed_wd: 0.095 + embedding_dim: 512 + enable_looping_at: 0.35 + etlb_clip: 3.0 + etlb_lr: 0.05 + etlb_steps: 5 + eval_only_path: + eval_seq_len: 2048 + eval_stride: 64 + gptq_calibration_batches: 64 + gptq_reserve_seconds: 12.0 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + head_lr: 0.008 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: logs/PR1530_notriton_v3_s1337.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.022 + max_wallclock_seconds: 600.0 + min_lr: 0.0 + mlp_mult: 4.0 + model_dim: 512 + model_path: final_model.pt + muon_backend_steps: 5 + muon_beta2: 0.95 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_start_layer: 7 + qk_gain_init: 5.0 + quantized_model_path: final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: PR1530_notriton_v3_s1337 + scalar_lr: 0.02 + seed: 1337 + skip_gates_enabled: True + sliding_window_enabled: False + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /home/dex/parameter-golf-with-cc/data/tokenizers/fineweb_8192_bpe.model + train_batch_tokens: 786432 + train_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.999 + ttt_chunk_size: 64 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 96 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_output_dir: + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_doc_fraction: 1.0 + val_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.667 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 40540160 +model_params:35944537 +gptq:reserving 12s, effective=588000ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +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: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0094 val_bpb: 3.4877 +1/20000 train_loss: 9.0093 train_time: 0.0m tok/s: 16250640 +2/20000 train_loss: 12.1683 train_time: 0.0m tok/s: 12647599 +3/20000 train_loss: 11.1668 train_time: 0.0m tok/s: 10839522 +4/20000 train_loss: 9.5518 train_time: 0.0m tok/s: 10010689 +5/20000 train_loss: 8.2075 train_time: 0.0m tok/s: 9652243 +500/20000 train_loss: 3.2886 train_time: 0.8m tok/s: 8173568 +1000/20000 train_loss: 3.0384 train_time: 1.6m tok/s: 8142416 +1500/20000 train_loss: 3.0417 train_time: 2.4m tok/s: 8132976 +2000/20000 train_loss: 3.0063 train_time: 3.2m tok/s: 8131654 +layer_loop:enabled step:2128 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 3.0879 train_time: 4.3m tok/s: 7598665 +3000/20000 train_loss: 2.9238 train_time: 5.5m tok/s: 7153130 +3500/20000 train_loss: 2.9898 train_time: 6.7m tok/s: 6864428 +4000/20000 train_loss: 2.9173 train_time: 7.9m tok/s: 6664852 +4000/20000 val_loss: 2.8893 val_bpb: 1.1185 +4500/20000 train_loss: 2.8623 train_time: 9.1m tok/s: 6516063 +4817/20000 val_loss: 2.7823 val_bpb: 1.0771 +stopping_early: wallclock_cap train_time: 588061ms step: 4817/20000 +peak memory allocated: 40019 MiB reserved: 44090 MiB +ema:applying EMA weights + +beginning eval timer +pre-quantization post-ema val_loss:2.78297692 val_bpb:1.07734237 eval_time:6607ms +Serialized model: 135408623 bytes +Code size (uncompressed): 25659 bytes +Code size (compressed): 26505 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 12.4s +Quantized weights: + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int8): tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights +Serialized model quantized+brotli: 15969291 bytes +Total submission size quantized+brotli: 15995796 bytes +quantized val_loss:2.81269248 val_bpb:1.08884582 eval_time:9711ms +ttt_lora:warming up compile +ttt_lora:compile warmup done (80.3s) +ttt_lora:docs:50000 rank:96 lr:0.0001 chunk:64 +ttt_progress: batch 781/782 batch_loss:2.5796 batch_bpb:1.0649 running_loss:2.5796 running_bpb:1.0649 doc_len:14510-25988 +ttt_progress: batch 723/782 batch_loss:2.7953 batch_bpb:1.0664 running_loss:2.5991 running_bpb:1.0651 doc_len:1861-1885 +ttt_progress: batch 716/782 batch_loss:2.8252 batch_bpb:1.0425 running_loss:2.6167 running_bpb:1.0631 doc_len:1739-1754 +ttt_progress: batch 709/782 batch_loss:2.8023 batch_bpb:1.0648 running_loss:2.6295 running_bpb:1.0632 doc_len:1649-1661 +ttt_progress: batch 702/782 batch_loss:2.8137 batch_bpb:1.0703 running_loss:2.6408 running_bpb:1.0637 doc_len:1572-1581 +ttt_progress: batch 698/782 batch_loss:2.7950 batch_bpb:1.0359 running_loss:2.6495 running_bpb:1.0620 doc_len:1534-1543 +ttt_progress: batch 688/782 batch_loss:2.7615 batch_bpb:1.0535 running_loss:2.6552 running_bpb:1.0615 doc_len:1441-1450 +ttt_progress: batch 685/782 batch_loss:2.7842 batch_bpb:1.0671 running_loss:2.6613 running_bpb:1.0618 doc_len:1414-1422 +ttt_progress: batch 675/782 batch_loss:2.8505 batch_bpb:1.0701 running_loss:2.6694 running_bpb:1.0622 doc_len:1341-1347 +ttt_progress: batch 667/782 batch_loss:2.8333 batch_bpb:1.1099 running_loss:2.6759 running_bpb:1.0641 doc_len:1288-1295 +ttt_progress: batch 660/782 batch_loss:2.8686 batch_bpb:1.0976 running_loss:2.6830 running_bpb:1.0654 doc_len:1245-1250 +ttt_progress: batch 655/782 batch_loss:2.6965 batch_bpb:1.0257 running_loss:2.6834 running_bpb:1.0640 doc_len:1215-1220 +ttt_progress: batch 646/782 batch_loss:2.7862 batch_bpb:1.0789 running_loss:2.6867 running_bpb:1.0645 doc_len:1166-1171 +ttt_progress: batch 639/782 batch_loss:2.8658 batch_bpb:1.0856 running_loss:2.6921 running_bpb:1.0651 doc_len:1129-1134 +ttt_progress: batch 635/782 batch_loss:2.7481 batch_bpb:1.0638 running_loss:2.6938 running_bpb:1.0651 doc_len:1111-1116 +ttt_progress: batch 628/782 batch_loss:2.7792 batch_bpb:1.0510 running_loss:2.6961 running_bpb:1.0647 doc_len:1078-1082 +ttt_progress: batch 623/782 batch_loss:2.7966 batch_bpb:1.0764 running_loss:2.6987 running_bpb:1.0650 doc_len:1055-1060 +ttt_progress: batch 616/782 batch_loss:2.8651 batch_bpb:1.0926 running_loss:2.7028 running_bpb:1.0657 doc_len:1024-1027 +ttt_progress: batch 604/782 batch_loss:2.7345 batch_bpb:1.0396 running_loss:2.7035 running_bpb:1.0651 doc_len:974-978 +ttt_progress: batch 597/782 batch_loss:2.7845 batch_bpb:1.0455 running_loss:2.7053 running_bpb:1.0646 doc_len:947-950 +ttt_progress: batch 593/782 batch_loss:2.8084 batch_bpb:1.0503 running_loss:2.7074 running_bpb:1.0643 doc_len:933-937 +ttt_progress: batch 586/782 batch_loss:2.7378 batch_bpb:1.0187 running_loss:2.7080 running_bpb:1.0633 doc_len:911-914 +ttt_progress: batch 579/782 batch_loss:2.6552 batch_bpb:1.0119 running_loss:2.7070 running_bpb:1.0623 doc_len:888-891 +ttt_progress: batch 573/782 batch_loss:2.9443 batch_bpb:1.0776 running_loss:2.7114 running_bpb:1.0626 doc_len:868-871 +ttt_progress: batch 564/782 batch_loss:2.8842 batch_bpb:1.1159 running_loss:2.7144 running_bpb:1.0636 doc_len:840-843 +ttt_progress: batch 557/782 batch_loss:2.8052 batch_bpb:1.0459 running_loss:2.7159 running_bpb:1.0633 doc_len:818-821 +ttt_progress: batch 551/782 batch_loss:2.8384 batch_bpb:1.0699 running_loss:2.7179 running_bpb:1.0634 doc_len:801-804 +ttt_progress: batch 544/782 batch_loss:2.7654 batch_bpb:1.0473 running_loss:2.7186 running_bpb:1.0631 doc_len:782-785 +ttt_progress: batch 537/782 batch_loss:2.7258 batch_bpb:1.0307 running_loss:2.7188 running_bpb:1.0626 doc_len:764-767 +ttt_progress: batch 532/782 batch_loss:2.8264 batch_bpb:1.0613 running_loss:2.7203 running_bpb:1.0626 doc_len:752-754 +ttt_progress: batch 525/782 batch_loss:2.7931 batch_bpb:1.0746 running_loss:2.7213 running_bpb:1.0628 doc_len:735-737 +ttt_progress: batch 517/782 batch_loss:2.7894 batch_bpb:1.0558 running_loss:2.7222 running_bpb:1.0627 doc_len:715-717 +ttt_progress: batch 510/782 batch_loss:2.7672 batch_bpb:1.0235 running_loss:2.7228 running_bpb:1.0621 doc_len:698-700 +ttt_progress: batch 502/782 batch_loss:2.8423 batch_bpb:1.0673 running_loss:2.7243 running_bpb:1.0622 doc_len:680-682 +ttt_progress: batch 495/782 batch_loss:2.7762 batch_bpb:1.0600 running_loss:2.7249 running_bpb:1.0622 doc_len:664-666 +ttt_progress: batch 489/782 batch_loss:2.8108 batch_bpb:1.0868 running_loss:2.7259 running_bpb:1.0625 doc_len:651-653 +ttt_progress: batch 484/782 batch_loss:2.8161 batch_bpb:1.0748 running_loss:2.7269 running_bpb:1.0626 doc_len:641-643 +ttt_progress: batch 476/782 batch_loss:2.7653 batch_bpb:1.0561 running_loss:2.7274 running_bpb:1.0625 doc_len:624-626 +ttt_progress: batch 471/782 batch_loss:2.8538 batch_bpb:1.0753 running_loss:2.7287 running_bpb:1.0627 doc_len:614-616 +ttt_progress: batch 464/782 batch_loss:2.7314 batch_bpb:1.0825 running_loss:2.7287 running_bpb:1.0629 doc_len:600-602 +ttt_progress: batch 457/782 batch_loss:2.7771 batch_bpb:1.0545 running_loss:2.7292 running_bpb:1.0628 doc_len:587-589 +ttt_progress: batch 450/782 batch_loss:2.7718 batch_bpb:1.0345 running_loss:2.7296 running_bpb:1.0625 doc_len:575-576 +ttt_progress: batch 447/782 batch_loss:2.8439 batch_bpb:1.0936 running_loss:2.7307 running_bpb:1.0628 doc_len:569-571 +ttt_progress: batch 440/782 batch_loss:2.8770 batch_bpb:1.0984 running_loss:2.7320 running_bpb:1.0631 doc_len:556-559 +ttt_progress: batch 433/782 batch_loss:2.7897 batch_bpb:1.0707 running_loss:2.7325 running_bpb:1.0632 doc_len:544-545 +ttt_progress: batch 426/782 batch_loss:2.7365 batch_bpb:1.0709 running_loss:2.7326 running_bpb:1.0633 doc_len:532-533 +ttt_progress: batch 419/782 batch_loss:2.8035 batch_bpb:1.0419 running_loss:2.7332 running_bpb:1.0631 doc_len:519-521 +ttt_progress: batch 411/782 batch_loss:2.8262 batch_bpb:1.0775 running_loss:2.7339 running_bpb:1.0632 doc_len:507-508 +ttt_progress: batch 403/782 batch_loss:2.8265 batch_bpb:1.0562 running_loss:2.7346 running_bpb:1.0631 doc_len:493-495 +ttt_progress: batch 396/782 batch_loss:2.7708 batch_bpb:1.0603 running_loss:2.7349 running_bpb:1.0631 doc_len:482-484 +ttt_progress: batch 389/782 batch_loss:2.7987 batch_bpb:1.0660 running_loss:2.7354 running_bpb:1.0631 doc_len:471-473 +ttt_progress: batch 382/782 batch_loss:2.9297 batch_bpb:1.1405 running_loss:2.7367 running_bpb:1.0637 doc_len:461-463 +ttt_progress: batch 376/782 batch_loss:2.7314 batch_bpb:1.0490 running_loss:2.7367 running_bpb:1.0636 doc_len:453-454 +ttt_progress: batch 366/782 batch_loss:2.8905 batch_bpb:1.1316 running_loss:2.7377 running_bpb:1.0640 doc_len:439-440 +ttt_progress: batch 359/782 batch_loss:2.8104 batch_bpb:1.0862 running_loss:2.7382 running_bpb:1.0642 doc_len:429-430 +ttt_progress: batch 351/782 batch_loss:2.8471 batch_bpb:1.0958 running_loss:2.7389 running_bpb:1.0644 doc_len:418-419 +ttt_progress: batch 344/782 batch_loss:2.8971 batch_bpb:1.1105 running_loss:2.7399 running_bpb:1.0647 doc_len:408-410 +ttt_progress: batch 337/782 batch_loss:2.8436 batch_bpb:1.0827 running_loss:2.7405 running_bpb:1.0648 doc_len:399-400 +ttt_progress: batch 330/782 batch_loss:2.8773 batch_bpb:1.0969 running_loss:2.7413 running_bpb:1.0650 doc_len:390-392 +ttt_progress: batch 323/782 batch_loss:2.8341 batch_bpb:1.0542 running_loss:2.7418 running_bpb:1.0649 doc_len:381-382 +ttt_progress: batch 317/782 batch_loss:2.8839 batch_bpb:1.1149 running_loss:2.7425 running_bpb:1.0652 doc_len:373-374 +ttt_progress: batch 309/782 batch_loss:2.8462 batch_bpb:1.1104 running_loss:2.7431 running_bpb:1.0654 doc_len:363-364 +ttt_progress: batch 302/782 batch_loss:2.8494 batch_bpb:1.1051 running_loss:2.7436 running_bpb:1.0656 doc_len:354-355 +ttt_progress: batch 295/782 batch_loss:2.8594 batch_bpb:1.1274 running_loss:2.7442 running_bpb:1.0659 doc_len:345-347 +ttt_progress: batch 288/782 batch_loss:2.8275 batch_bpb:1.1101 running_loss:2.7446 running_bpb:1.0661 doc_len:337-339 +ttt_progress: batch 281/782 batch_loss:2.9302 batch_bpb:1.1553 running_loss:2.7455 running_bpb:1.0665 doc_len:329-330 +ttt_progress: batch 274/782 batch_loss:2.8330 batch_bpb:1.0999 running_loss:2.7459 running_bpb:1.0667 doc_len:322-323 +ttt_progress: batch 267/782 batch_loss:2.8752 batch_bpb:1.1025 running_loss:2.7464 running_bpb:1.0668 doc_len:314-315 +ttt_progress: batch 260/782 batch_loss:2.8356 batch_bpb:1.1061 running_loss:2.7468 running_bpb:1.0670 doc_len:306-307 +ttt_progress: batch 253/782 batch_loss:2.7739 batch_bpb:1.0893 running_loss:2.7469 running_bpb:1.0671 doc_len:298-299 +ttt_progress: batch 246/782 batch_loss:2.9162 batch_bpb:1.1422 running_loss:2.7476 running_bpb:1.0674 doc_len:291-292 +ttt_progress: batch 239/782 batch_loss:2.9099 batch_bpb:1.1412 running_loss:2.7483 running_bpb:1.0677 doc_len:284-285 +ttt_progress: batch 232/782 batch_loss:2.9395 batch_bpb:1.1370 running_loss:2.7490 running_bpb:1.0680 doc_len:277-278 +ttt_progress: batch 225/782 batch_loss:2.9080 batch_bpb:1.1326 running_loss:2.7496 running_bpb:1.0682 doc_len:270-271 +ttt_progress: batch 218/782 batch_loss:2.7594 batch_bpb:1.1096 running_loss:2.7496 running_bpb:1.0683 doc_len:263-264 +ttt_progress: batch 211/782 batch_loss:2.9100 batch_bpb:1.1593 running_loss:2.7502 running_bpb:1.0687 doc_len:256-257 +ttt_progress: batch 204/782 batch_loss:2.9142 batch_bpb:1.1336 running_loss:2.7507 running_bpb:1.0689 doc_len:250-251 +ttt_progress: batch 196/782 batch_loss:2.9169 batch_bpb:1.1688 running_loss:2.7513 running_bpb:1.0692 doc_len:243-244 +ttt_progress: batch 188/782 batch_loss:2.9121 batch_bpb:1.1536 running_loss:2.7518 running_bpb:1.0695 doc_len:236-237 +ttt_progress: batch 182/782 batch_loss:2.8698 batch_bpb:1.1416 running_loss:2.7522 running_bpb:1.0697 doc_len:230-231 +ttt_progress: batch 175/782 batch_loss:2.8678 batch_bpb:1.1242 running_loss:2.7525 running_bpb:1.0698 doc_len:225-225 +ttt_progress: batch 168/782 batch_loss:2.9263 batch_bpb:1.1468 running_loss:2.7530 running_bpb:1.0701 doc_len:218-219 +ttt_progress: batch 161/782 batch_loss:2.9703 batch_bpb:1.1820 running_loss:2.7536 running_bpb:1.0704 doc_len:212-213 +ttt_progress: batch 153/782 batch_loss:3.0251 batch_bpb:1.1669 running_loss:2.7544 running_bpb:1.0706 doc_len:206-207 +ttt_progress: batch 147/782 batch_loss:2.9415 batch_bpb:1.1637 running_loss:2.7549 running_bpb:1.0709 doc_len:201-202 +ttt_progress: batch 138/782 batch_loss:2.9253 batch_bpb:1.1644 running_loss:2.7553 running_bpb:1.0711 doc_len:194-195 +ttt_progress: batch 132/782 batch_loss:2.9675 batch_bpb:1.1420 running_loss:2.7558 running_bpb:1.0713 doc_len:189-189 +ttt_progress: batch 124/782 batch_loss:2.8811 batch_bpb:1.1525 running_loss:2.7561 running_bpb:1.0715 doc_len:183-184 +ttt_progress: batch 116/782 batch_loss:3.0187 batch_bpb:1.1937 running_loss:2.7567 running_bpb:1.0718 doc_len:177-178 +ttt_progress: batch 109/782 batch_loss:3.0874 batch_bpb:1.2167 running_loss:2.7575 running_bpb:1.0721 doc_len:172-173 +ttt_progress: batch 103/782 batch_loss:2.9119 batch_bpb:1.1270 running_loss:2.7578 running_bpb:1.0722 doc_len:168-168 +ttt_progress: batch 94/782 batch_loss:3.0021 batch_bpb:1.1840 running_loss:2.7583 running_bpb:1.0724 doc_len:160-161 +ttt_progress: batch 87/782 batch_loss:3.0304 batch_bpb:1.2112 running_loss:2.7589 running_bpb:1.0727 doc_len:155-156 +ttt_progress: batch 81/782 batch_loss:2.9334 batch_bpb:1.1666 running_loss:2.7592 running_bpb:1.0729 doc_len:151-151 +ttt_progress: batch 73/782 batch_loss:3.0589 batch_bpb:1.2101 running_loss:2.7598 running_bpb:1.0731 doc_len:144-145 +ttt_progress: batch 66/782 batch_loss:3.1110 batch_bpb:1.2743 running_loss:2.7604 running_bpb:1.0735 doc_len:139-140 +ttt_progress: batch 58/782 batch_loss:2.9812 batch_bpb:1.2296 running_loss:2.7608 running_bpb:1.0737 doc_len:133-134 +ttt_progress: batch 52/782 batch_loss:3.0617 batch_bpb:1.1980 running_loss:2.7613 running_bpb:1.0739 doc_len:128-129 +ttt_progress: batch 45/782 batch_loss:3.1020 batch_bpb:1.2409 running_loss:2.7618 running_bpb:1.0742 doc_len:122-123 +ttt_progress: batch 37/782 batch_loss:3.0917 batch_bpb:1.2138 running_loss:2.7623 running_bpb:1.0744 doc_len:116-117 +ttt_progress: batch 30/782 batch_loss:3.1302 batch_bpb:1.2539 running_loss:2.7628 running_bpb:1.0747 doc_len:110-111 +ttt_progress: batch 23/782 batch_loss:3.1488 batch_bpb:1.2550 running_loss:2.7634 running_bpb:1.0749 doc_len:104-105 +ttt_progress: batch 16/782 batch_loss:3.0656 batch_bpb:1.2223 running_loss:2.7637 running_bpb:1.0751 doc_len:97-98 +ttt_progress: batch 9/782 batch_loss:3.2289 batch_bpb:1.2796 running_loss:2.7643 running_bpb:1.0753 doc_len:87-89 +ttt_progress: batch 2/782 batch_loss:3.1649 batch_bpb:1.1738 running_loss:2.7646 running_bpb:1.0754 doc_len:70-75 +quantized_ttt_lora val_loss:2.78548071 val_bpb:1.07834639 eval_time:249634ms +total_eval_time:346.9s +total_eval_time_with_compile:427.1s diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed42.log b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed42.log new file mode 100644 index 0000000000..de2c03695c --- /dev/null +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_seed42.log @@ -0,0 +1,281 @@ + +***************************************** +Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +***************************************** +Hyperparameters: + adam_eps: 1e-08 + adam_wd: 0.02 + artifact_dir: + beta1: 0.9 + beta2: 0.95 + compressor: brotli + data_dir: /home/dex/parameter-golf-with-cc/data + datasets_dir: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192 + distributed: True + ema_decay: 0.997 + embed_bits: 8 + embed_clip_sigmas: 20.0 + embed_lr: 0.6 + embed_wd: 0.095 + embedding_dim: 512 + enable_looping_at: 0.35 + etlb_clip: 3.0 + etlb_lr: 0.05 + etlb_steps: 5 + eval_only_path: + eval_seq_len: 2048 + eval_stride: 64 + gptq_calibration_batches: 64 + gptq_reserve_seconds: 12.0 + grad_accum_steps: 1 + grad_clip_norm: 0.3 + head_lr: 0.008 + is_main_process: True + iterations: 20000 + ln_scale: True + local_rank: 0 + logfile: logs/PR1530_notriton_v3_s42.txt + logit_softcap: 30.0 + loop_end: 5 + loop_start: 3 + matrix_bits: 6 + matrix_clip_sigmas: 12.85 + matrix_lr: 0.022 + max_wallclock_seconds: 600.0 + min_lr: 0.0 + mlp_mult: 4.0 + model_dim: 512 + model_path: final_model.pt + muon_backend_steps: 5 + muon_beta2: 0.95 + muon_momentum: 0.97 + muon_momentum_warmup_start: 0.92 + muon_momentum_warmup_steps: 1500 + muon_row_normalize: True + muon_wd: 0.095 + num_heads: 8 + num_kv_heads: 4 + num_layers: 11 + num_loops: 2 + parallel_start_layer: 7 + qk_gain_init: 5.0 + quantized_model_path: final_model.int6.ptz + rank: 0 + rope_base: 10000.0 + rope_dims: 16 + rope_train_seq_len: 2048 + rope_yarn: False + run_id: PR1530_notriton_v3_s42 + scalar_lr: 0.02 + seed: 42 + skip_gates_enabled: True + sliding_window_enabled: False + tie_embeddings: True + tied_embed_init_std: 0.005 + tied_embed_lr: 0.03 + tokenizer_path: /home/dex/parameter-golf-with-cc/data/tokenizers/fineweb_8192_bpe.model + train_batch_tokens: 786432 + train_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_train_*.bin + train_log_every: 500 + train_seq_len: 2048 + ttt_batch_size: 64 + ttt_beta1: 0.0 + ttt_beta2: 0.999 + ttt_chunk_size: 64 + ttt_enabled: True + ttt_eval_batches: + ttt_eval_seq_len: 2048 + ttt_grad_steps: 1 + ttt_k_lora: True + ttt_lora_lr: 0.0001 + ttt_lora_rank: 96 + ttt_mlp_lora: True + ttt_o_lora: True + ttt_optimizer: adam + ttt_output_dir: + ttt_weight_decay: 0.5 + val_batch_tokens: 524288 + val_doc_fraction: 1.0 + val_files: /home/dex/parameter-golf-with-cc/data/datasets/fineweb10B_sp8192/fineweb_val_*.bin + val_loss_every: 4000 + vocab_size: 8192 + warmdown_frac: 0.667 + warmup_steps: 20 + world_size: 8 + xsa_last_n: 11 +train_shards: 80 +val_tokens: 40540160 +model_params:35944537 +gptq:reserving 12s, effective=588000ms +warmup_cu_buckets:64,128,192,256 iters_each:3 +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: 10/20 +warmup_step: 20/20 +loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +loop_warmup_step: 1/20 +loop_warmup_step: 2/20 +loop_warmup_step: 3/20 +loop_warmup_step: 4/20 +loop_warmup_step: 5/20 +loop_warmup_step: 6/20 +loop_warmup_step: 10/20 +loop_warmup_step: 20/20 +0/20000 val_loss: 9.0078 val_bpb: 3.4871 +1/20000 train_loss: 9.0072 train_time: 0.0m tok/s: 16241731 +2/20000 train_loss: 12.2941 train_time: 0.0m tok/s: 12699583 +3/20000 train_loss: 11.2384 train_time: 0.0m tok/s: 10918372 +4/20000 train_loss: 9.5967 train_time: 0.0m tok/s: 10130097 +5/20000 train_loss: 8.2349 train_time: 0.0m tok/s: 9722403 +500/20000 train_loss: 3.2783 train_time: 0.8m tok/s: 8179536 +1000/20000 train_loss: 3.0360 train_time: 1.6m tok/s: 8161371 +1500/20000 train_loss: 3.0430 train_time: 2.4m tok/s: 8153970 +2000/20000 train_loss: 3.0036 train_time: 3.2m tok/s: 8148031 +layer_loop:enabled step:2132 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10] +2500/20000 train_loss: 3.0847 train_time: 4.3m tok/s: 7619523 +3000/20000 train_loss: 2.9189 train_time: 5.5m tok/s: 7170615 +3500/20000 train_loss: 2.9832 train_time: 6.7m tok/s: 6881511 +4000/20000 train_loss: 2.9146 train_time: 7.8m tok/s: 6679305 +4000/20000 val_loss: 2.8871 val_bpb: 1.1177 +4500/20000 train_loss: 2.8574 train_time: 9.0m tok/s: 6530491 +4826/20000 val_loss: 2.7785 val_bpb: 1.0756 +stopping_early: wallclock_cap train_time: 588119ms step: 4826/20000 +peak memory allocated: 40019 MiB reserved: 44090 MiB +ema:applying EMA weights + +beginning eval timer +pre-quantization post-ema val_loss:2.77927549 val_bpb:1.07590948 eval_time:6585ms +Serialized model: 135408623 bytes +Code size (uncompressed): 25659 bytes +Code size (compressed): 26505 bytes +GPTQ:collecting Hessians from calibration data... +GPTQ:collected 67 Hessians in 12.4s +Quantized weights: + gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight + gptq (int8): tok_emb.weight + passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, lane_merge, skip_gates, skip_weights +Serialized model quantized+brotli: 15965977 bytes +Total submission size quantized+brotli: 15992482 bytes +quantized val_loss:2.80875062 val_bpb:1.08731985 eval_time:54100ms +ttt_lora:warming up compile +ttt_lora:compile warmup done (165.9s) +ttt_lora:docs:50000 rank:96 lr:0.0001 chunk:64 +ttt_progress: batch 779/782 batch_loss:2.6576 batch_bpb:1.0825 running_loss:2.6576 running_bpb:1.0825 doc_len:9037-11049 +ttt_progress: batch 769/782 batch_loss:2.7794 batch_bpb:1.0999 running_loss:2.6950 running_bpb:1.0879 doc_len:4307-4479 +ttt_progress: batch 763/782 batch_loss:2.7996 batch_bpb:1.1047 running_loss:2.7160 running_bpb:1.0914 doc_len:3536-3637 +ttt_progress: batch 758/782 batch_loss:2.8844 batch_bpb:1.0883 running_loss:2.7412 running_bpb:1.0909 doc_len:3108-3187 +ttt_progress: batch 750/782 batch_loss:2.8373 batch_bpb:1.0704 running_loss:2.7520 running_bpb:1.0885 doc_len:2638-2688 +ttt_progress: batch 746/782 batch_loss:2.6860 batch_bpb:1.0576 running_loss:2.7457 running_bpb:1.0855 doc_len:2459-2501 +ttt_progress: batch 738/782 batch_loss:2.7521 batch_bpb:1.0559 running_loss:2.7462 running_bpb:1.0832 doc_len:2194-2227 +ttt_progress: batch 733/782 batch_loss:2.7693 batch_bpb:1.0568 running_loss:2.7478 running_bpb:1.0813 doc_len:2062-2090 +ttt_progress: batch 725/782 batch_loss:2.7652 batch_bpb:1.0718 running_loss:2.7488 running_bpb:1.0807 doc_len:1900-1915 +ttt_progress: batch 720/782 batch_loss:2.8311 batch_bpb:1.0814 running_loss:2.7532 running_bpb:1.0808 doc_len:1816-1832 +ttt_progress: batch 712/782 batch_loss:2.8404 batch_bpb:1.0814 running_loss:2.7573 running_bpb:1.0808 doc_len:1684-1697 +ttt_progress: batch 705/782 batch_loss:2.7887 batch_bpb:1.0742 running_loss:2.7587 running_bpb:1.0805 doc_len:1606-1617 +ttt_progress: batch 698/782 batch_loss:2.7929 batch_bpb:1.0351 running_loss:2.7600 running_bpb:1.0786 doc_len:1534-1543 +ttt_progress: batch 690/782 batch_loss:2.8501 batch_bpb:1.0680 running_loss:2.7633 running_bpb:1.0782 doc_len:1458-1467 +ttt_progress: batch 685/782 batch_loss:2.7809 batch_bpb:1.0659 running_loss:2.7638 running_bpb:1.0778 doc_len:1414-1422 +ttt_progress: batch 675/782 batch_loss:2.8482 batch_bpb:1.0692 running_loss:2.7665 running_bpb:1.0775 doc_len:1341-1347 +ttt_progress: batch 667/782 batch_loss:2.8275 batch_bpb:1.1076 running_loss:2.7682 running_bpb:1.0784 doc_len:1288-1295 +ttt_progress: batch 660/782 batch_loss:2.8641 batch_bpb:1.0959 running_loss:2.7708 running_bpb:1.0789 doc_len:1245-1250 +ttt_progress: batch 655/782 batch_loss:2.6930 batch_bpb:1.0244 running_loss:2.7688 running_bpb:1.0774 doc_len:1215-1220 +ttt_progress: batch 646/782 batch_loss:2.7801 batch_bpb:1.0765 running_loss:2.7691 running_bpb:1.0774 doc_len:1166-1171 +ttt_progress: batch 639/782 batch_loss:2.8616 batch_bpb:1.0840 running_loss:2.7712 running_bpb:1.0776 doc_len:1129-1134 +ttt_progress: batch 634/782 batch_loss:2.7090 batch_bpb:1.0458 running_loss:2.7699 running_bpb:1.0769 doc_len:1105-1111 +ttt_progress: batch 627/782 batch_loss:2.7464 batch_bpb:1.0396 running_loss:2.7694 running_bpb:1.0761 doc_len:1073-1077 +ttt_progress: batch 621/782 batch_loss:2.8418 batch_bpb:1.0884 running_loss:2.7708 running_bpb:1.0763 doc_len:1046-1050 +ttt_progress: batch 614/782 batch_loss:2.7992 batch_bpb:1.0706 running_loss:2.7713 running_bpb:1.0762 doc_len:1016-1020 +ttt_progress: batch 608/782 batch_loss:2.7380 batch_bpb:1.0334 running_loss:2.7707 running_bpb:1.0754 doc_len:990-994 +ttt_progress: batch 601/782 batch_loss:2.7727 batch_bpb:1.0656 running_loss:2.7708 running_bpb:1.0752 doc_len:963-966 +ttt_progress: batch 591/782 batch_loss:2.6754 batch_bpb:1.0109 running_loss:2.7692 running_bpb:1.0742 doc_len:927-930 +ttt_progress: batch 584/782 batch_loss:2.7697 batch_bpb:1.0406 running_loss:2.7692 running_bpb:1.0736 doc_len:904-907 +ttt_progress: batch 577/782 batch_loss:2.7671 batch_bpb:1.0466 running_loss:2.7692 running_bpb:1.0732 doc_len:880-884 +ttt_progress: batch 570/782 batch_loss:2.7867 batch_bpb:1.0845 running_loss:2.7694 running_bpb:1.0734 doc_len:858-862 +ttt_progress: batch 563/782 batch_loss:2.8101 batch_bpb:1.0660 running_loss:2.7700 running_bpb:1.0732 doc_len:837-840 +ttt_progress: batch 556/782 batch_loss:2.8444 batch_bpb:1.0875 running_loss:2.7710 running_bpb:1.0734 doc_len:815-818 +ttt_progress: batch 549/782 batch_loss:2.7693 batch_bpb:1.0655 running_loss:2.7710 running_bpb:1.0733 doc_len:795-798 +ttt_progress: batch 542/782 batch_loss:2.8410 batch_bpb:1.0762 running_loss:2.7719 running_bpb:1.0734 doc_len:777-779 +ttt_progress: batch 535/782 batch_loss:2.7993 batch_bpb:1.0614 running_loss:2.7722 running_bpb:1.0732 doc_len:759-762 +ttt_progress: batch 528/782 batch_loss:2.7616 batch_bpb:1.0345 running_loss:2.7721 running_bpb:1.0728 doc_len:742-745 +ttt_progress: batch 521/782 batch_loss:2.7800 batch_bpb:1.0548 running_loss:2.7722 running_bpb:1.0726 doc_len:725-727 +ttt_progress: batch 514/782 batch_loss:2.9189 batch_bpb:1.1009 running_loss:2.7737 running_bpb:1.0729 doc_len:707-710 +ttt_progress: batch 506/782 batch_loss:2.8182 batch_bpb:1.0796 running_loss:2.7742 running_bpb:1.0729 doc_len:688-690 +ttt_progress: batch 499/782 batch_loss:2.7914 batch_bpb:1.0534 running_loss:2.7744 running_bpb:1.0727 doc_len:673-675 +ttt_progress: batch 492/782 batch_loss:2.8141 batch_bpb:1.0583 running_loss:2.7748 running_bpb:1.0726 doc_len:657-659 +ttt_progress: batch 485/782 batch_loss:2.7954 batch_bpb:1.0520 running_loss:2.7750 running_bpb:1.0724 doc_len:643-645 +ttt_progress: batch 478/782 batch_loss:2.8024 batch_bpb:1.0554 running_loss:2.7752 running_bpb:1.0722 doc_len:628-630 +ttt_progress: batch 471/782 batch_loss:2.8463 batch_bpb:1.0724 running_loss:2.7758 running_bpb:1.0722 doc_len:614-616 +ttt_progress: batch 464/782 batch_loss:2.7297 batch_bpb:1.0818 running_loss:2.7754 running_bpb:1.0723 doc_len:600-602 +ttt_progress: batch 457/782 batch_loss:2.7739 batch_bpb:1.0533 running_loss:2.7754 running_bpb:1.0722 doc_len:587-589 +ttt_progress: batch 450/782 batch_loss:2.7666 batch_bpb:1.0326 running_loss:2.7754 running_bpb:1.0718 doc_len:575-576 +ttt_progress: batch 447/782 batch_loss:2.8378 batch_bpb:1.0913 running_loss:2.7759 running_bpb:1.0720 doc_len:569-571 +ttt_progress: batch 440/782 batch_loss:2.8742 batch_bpb:1.0974 running_loss:2.7766 running_bpb:1.0722 doc_len:556-559 +ttt_progress: batch 433/782 batch_loss:2.7838 batch_bpb:1.0684 running_loss:2.7767 running_bpb:1.0721 doc_len:544-545 +ttt_progress: batch 426/782 batch_loss:2.7327 batch_bpb:1.0694 running_loss:2.7764 running_bpb:1.0721 doc_len:532-533 +ttt_progress: batch 419/782 batch_loss:2.7968 batch_bpb:1.0394 running_loss:2.7765 running_bpb:1.0719 doc_len:519-521 +ttt_progress: batch 412/782 batch_loss:2.7212 batch_bpb:1.0568 running_loss:2.7761 running_bpb:1.0718 doc_len:508-510 +ttt_progress: batch 405/782 batch_loss:2.8333 batch_bpb:1.0709 running_loss:2.7765 running_bpb:1.0718 doc_len:497-498 +ttt_progress: batch 398/782 batch_loss:2.8915 batch_bpb:1.0982 running_loss:2.7772 running_bpb:1.0720 doc_len:486-487 +ttt_progress: batch 391/782 batch_loss:2.8331 batch_bpb:1.1034 running_loss:2.7776 running_bpb:1.0721 doc_len:475-476 +ttt_progress: batch 384/782 batch_loss:2.8589 batch_bpb:1.0968 running_loss:2.7781 running_bpb:1.0723 doc_len:464-466 +ttt_progress: batch 377/782 batch_loss:2.8137 batch_bpb:1.0910 running_loss:2.7783 running_bpb:1.0724 doc_len:454-455 +ttt_progress: batch 369/782 batch_loss:2.9268 batch_bpb:1.0864 running_loss:2.7792 running_bpb:1.0725 doc_len:443-444 +ttt_progress: batch 362/782 batch_loss:2.8245 batch_bpb:1.0679 running_loss:2.7794 running_bpb:1.0725 doc_len:433-434 +ttt_progress: batch 355/782 batch_loss:2.7132 batch_bpb:1.0690 running_loss:2.7790 running_bpb:1.0725 doc_len:423-424 +ttt_progress: batch 348/782 batch_loss:2.8213 batch_bpb:1.0721 running_loss:2.7793 running_bpb:1.0724 doc_len:414-415 +ttt_progress: batch 341/782 batch_loss:2.8783 batch_bpb:1.1019 running_loss:2.7798 running_bpb:1.0726 doc_len:404-406 +ttt_progress: batch 334/782 batch_loss:2.8711 batch_bpb:1.1048 running_loss:2.7802 running_bpb:1.0728 doc_len:395-396 +ttt_progress: batch 327/782 batch_loss:2.7790 batch_bpb:1.0789 running_loss:2.7802 running_bpb:1.0728 doc_len:387-388 +ttt_progress: batch 317/782 batch_loss:2.8860 batch_bpb:1.1157 running_loss:2.7807 running_bpb:1.0730 doc_len:373-374 +ttt_progress: batch 310/782 batch_loss:2.8066 batch_bpb:1.0873 running_loss:2.7808 running_bpb:1.0731 doc_len:364-365 +ttt_progress: batch 302/782 batch_loss:2.8472 batch_bpb:1.1043 running_loss:2.7811 running_bpb:1.0732 doc_len:354-355 +ttt_progress: batch 295/782 batch_loss:2.8563 batch_bpb:1.1262 running_loss:2.7815 running_bpb:1.0734 doc_len:345-347 +ttt_progress: batch 288/782 batch_loss:2.8166 batch_bpb:1.1058 running_loss:2.7816 running_bpb:1.0735 doc_len:337-339 +ttt_progress: batch 281/782 batch_loss:2.9348 batch_bpb:1.1571 running_loss:2.7822 running_bpb:1.0739 doc_len:329-330 +ttt_progress: batch 274/782 batch_loss:2.8231 batch_bpb:1.0960 running_loss:2.7824 running_bpb:1.0740 doc_len:322-323 +ttt_progress: batch 267/782 batch_loss:2.8694 batch_bpb:1.1002 running_loss:2.7827 running_bpb:1.0741 doc_len:314-315 +ttt_progress: batch 261/782 batch_loss:2.8728 batch_bpb:1.1237 running_loss:2.7830 running_bpb:1.0742 doc_len:308-309 +ttt_progress: batch 254/782 batch_loss:2.9129 batch_bpb:1.1474 running_loss:2.7835 running_bpb:1.0745 doc_len:299-300 +ttt_progress: batch 247/782 batch_loss:2.7947 batch_bpb:1.0798 running_loss:2.7836 running_bpb:1.0745 doc_len:292-293 +ttt_progress: batch 240/782 batch_loss:2.9181 batch_bpb:1.1582 running_loss:2.7840 running_bpb:1.0748 doc_len:285-286 +ttt_progress: batch 233/782 batch_loss:2.8693 batch_bpb:1.1273 running_loss:2.7843 running_bpb:1.0750 doc_len:278-279 +ttt_progress: batch 226/782 batch_loss:2.9440 batch_bpb:1.1453 running_loss:2.7848 running_bpb:1.0752 doc_len:271-272 +ttt_progress: batch 219/782 batch_loss:2.9186 batch_bpb:1.1387 running_loss:2.7852 running_bpb:1.0754 doc_len:264-265 +ttt_progress: batch 212/782 batch_loss:2.9407 batch_bpb:1.1509 running_loss:2.7857 running_bpb:1.0756 doc_len:257-258 +ttt_progress: batch 204/782 batch_loss:2.9146 batch_bpb:1.1338 running_loss:2.7861 running_bpb:1.0758 doc_len:250-251 +ttt_progress: batch 197/782 batch_loss:2.8614 batch_bpb:1.1284 running_loss:2.7863 running_bpb:1.0759 doc_len:244-245 +ttt_progress: batch 190/782 batch_loss:2.8776 batch_bpb:1.0939 running_loss:2.7866 running_bpb:1.0760 doc_len:237-238 +ttt_progress: batch 183/782 batch_loss:2.8788 batch_bpb:1.1491 running_loss:2.7868 running_bpb:1.0762 doc_len:231-232 +ttt_progress: batch 176/782 batch_loss:2.8259 batch_bpb:1.1086 running_loss:2.7869 running_bpb:1.0763 doc_len:225-226 +ttt_progress: batch 169/782 batch_loss:2.9203 batch_bpb:1.1668 running_loss:2.7872 running_bpb:1.0765 doc_len:219-220 +ttt_progress: batch 162/782 batch_loss:2.9587 batch_bpb:1.1480 running_loss:2.7877 running_bpb:1.0767 doc_len:213-214 +ttt_progress: batch 155/782 batch_loss:2.8866 batch_bpb:1.1345 running_loss:2.7879 running_bpb:1.0768 doc_len:207-208 +ttt_progress: batch 148/782 batch_loss:2.9944 batch_bpb:1.1638 running_loss:2.7884 running_bpb:1.0770 doc_len:202-203 +ttt_progress: batch 141/782 batch_loss:2.9098 batch_bpb:1.1471 running_loss:2.7887 running_bpb:1.0772 doc_len:196-197 +ttt_progress: batch 134/782 batch_loss:3.0394 batch_bpb:1.2156 running_loss:2.7892 running_bpb:1.0775 doc_len:190-191 +ttt_progress: batch 129/782 batch_loss:2.9565 batch_bpb:1.1869 running_loss:2.7896 running_bpb:1.0777 doc_len:187-187 +ttt_progress: batch 120/782 batch_loss:2.9857 batch_bpb:1.1732 running_loss:2.7900 running_bpb:1.0779 doc_len:180-181 +ttt_progress: batch 113/782 batch_loss:3.0505 batch_bpb:1.1994 running_loss:2.7905 running_bpb:1.0781 doc_len:175-176 +ttt_progress: batch 106/782 batch_loss:2.9684 batch_bpb:1.1990 running_loss:2.7908 running_bpb:1.0783 doc_len:170-171 +ttt_progress: batch 100/782 batch_loss:2.9474 batch_bpb:1.1570 running_loss:2.7911 running_bpb:1.0785 doc_len:165-166 +ttt_progress: batch 93/782 batch_loss:2.9629 batch_bpb:1.1885 running_loss:2.7914 running_bpb:1.0787 doc_len:160-160 +ttt_progress: batch 84/782 batch_loss:3.0333 batch_bpb:1.2221 running_loss:2.7919 running_bpb:1.0789 doc_len:153-154 +ttt_progress: batch 78/782 batch_loss:2.9121 batch_bpb:1.1299 running_loss:2.7921 running_bpb:1.0790 doc_len:148-149 +ttt_progress: batch 72/782 batch_loss:2.9444 batch_bpb:1.1966 running_loss:2.7923 running_bpb:1.0792 doc_len:144-144 +ttt_progress: batch 65/782 batch_loss:3.0486 batch_bpb:1.2240 running_loss:2.7927 running_bpb:1.0794 doc_len:139-139 +ttt_progress: batch 57/782 batch_loss:3.0538 batch_bpb:1.2310 running_loss:2.7931 running_bpb:1.0796 doc_len:132-133 +ttt_progress: batch 50/782 batch_loss:2.9791 batch_bpb:1.2229 running_loss:2.7934 running_bpb:1.0798 doc_len:126-127 +ttt_progress: batch 43/782 batch_loss:3.0119 batch_bpb:1.1978 running_loss:2.7937 running_bpb:1.0800 doc_len:121-122 +ttt_progress: batch 37/782 batch_loss:3.1019 batch_bpb:1.2177 running_loss:2.7941 running_bpb:1.0801 doc_len:116-117 +ttt_progress: batch 30/782 batch_loss:3.1395 batch_bpb:1.2576 running_loss:2.7945 running_bpb:1.0803 doc_len:110-111 +ttt_progress: batch 23/782 batch_loss:3.1629 batch_bpb:1.2606 running_loss:2.7949 running_bpb:1.0805 doc_len:104-105 +ttt_progress: batch 16/782 batch_loss:3.0702 batch_bpb:1.2241 running_loss:2.7952 running_bpb:1.0807 doc_len:97-98 +ttt_progress: batch 9/782 batch_loss:3.2070 batch_bpb:1.2709 running_loss:2.7956 running_bpb:1.0809 doc_len:87-89 +ttt_progress: batch 2/782 batch_loss:3.1747 batch_bpb:1.1775 running_loss:2.7959 running_bpb:1.0810 doc_len:70-75 +quantized_ttt_lora val_loss:2.78165791 val_bpb:1.07686646 eval_time:252222ms +total_eval_time:393.6s +total_eval_time_with_compile:559.5s From 903d34585d7e916523223111c6acb8383439d1bc Mon Sep 17 00:00:00 2001 From: "Dixing (Dex) Xu" Date: Sat, 11 Apr 2026 14:35:25 +0000 Subject: [PATCH 2/2] fix(legality): use random tokens for TTT compile warmup instead of val data MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The compile warmup was using actual validation tokens for torch.compile JIT warmup, which could be flagged under Issue #1017 as touching val data before the official eval loop. Replace with random tokens — compile only needs correct shapes, not meaningful data. --- .../train_gpt.py | 15 ++++++--------- 1 file changed, 6 insertions(+), 9 deletions(-) diff --git a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py index b09af88ef8..7f7c7ca975 100644 --- a/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py +++ b/records/track_10min_16mb/2026-04-11_VarLenDocTTT_PyTorchFallback/train_gpt.py @@ -2638,12 +2638,10 @@ def _fwd_ttt_bypass(input_ids, target_ids, lora): ttt_compile_time = 0.0 else: fwd_ttt_compiled = torch.compile(_fwd_ttt, dynamic=True) - log(f"ttt_lora:warming up compile") + log(f"ttt_lora:warming up compile (random tokens, no val data)") global BOS_ID if BOS_ID is None: BOS_ID = 1 - ds0 = 0 - val_tokens_idx = val_data.val_tokens.to(torch.int32) t_warmup = time.perf_counter() warmup_bszes = [h.ttt_batch_size] for bsz in warmup_bszes: @@ -2660,18 +2658,17 @@ def _fwd_ttt_bypass(input_ids, target_ids, lora): fused=True, ) for ctx_len in (h.ttt_chunk_size, h.ttt_eval_seq_len): - col_w = torch.arange(ctx_len + 1) - idx_w = (ds0 + col_w).clamp_(max=val_data.val_tokens.numel() - 1) - row_w = val_tokens_idx[idx_w].to(device=device, dtype=torch.int64) - xw = row_w[:ctx_len].unsqueeze(0).expand(bsz, -1).contiguous() - yw = row_w[1 : ctx_len + 1].unsqueeze(0).expand(bsz, -1).contiguous() + # Use random tokens for compile warmup — NOT val tokens. + # This avoids touching val data before the official eval loop, + # satisfying Issue #1017 and README guidelines. + xw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + yw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) with torch.autocast(device_type="cuda", dtype=torch.bfloat16): ptl = fwd_ttt_compiled(xw, yw, lora=wl) ptl[:, : min(h.ttt_chunk_size, ctx_len)].mean(dim=-1).sum().backward() wo.step() wo.zero_grad(set_to_none=True) del wl, wo - del val_tokens_idx torch.cuda.empty_cache() ttt_compile_time = time.perf_counter() - t_warmup log(f"ttt_lora:compile warmup done ({ttt_compile_time:.1f}s)")