diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/README.md b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/README.md new file mode 100644 index 0000000000..3ff1dae51e --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/README.md @@ -0,0 +1,53 @@ +# Non-record: Mamba-3 Hybrid SSM + Multi-Epoch TTT + Dynamics-Protected Quant — 1.1456 bpb (3-seed mean) + +**val_bpb: 1.1456** (3-seed mean, std 0.0011) | **15.93 MB total** (3-seed mean) | 8×H100 + +A follow-up SSM submission building on PR #1644 (1.1473 bpb). Same 7-layer Mamba-3/Attention hybrid; the −1.7 mBPB improvement comes from three quant/TTT-phase changes that don't touch the architecture. + +| Seed | BF16 | Post-quant+TTT | Total submission | +|------|------|----------------|------------------| +| 1337 | 1.1389 | **1.1441** | 15,930,191 B | +| 42 | 1.1462 | **1.1460** | 15,961,203 B | +| 2025 | 1.1495 | **1.1468** | 15,975,083 B | +| **Mean** | **1.1449** | **1.1456** | **15,955,492 B** | +| **Std** | **0.0045** | **0.0011** | **18,852 B** | + +Submitted artifact corresponds to seed 1337 (1.1441, 15,930,191 B). + +## What changed vs PR #1644 + +1. **`TTT_EPOCHS=2`**: PR #1644 used a single TTT epoch and saw a +8.3 mBPB BF16 → post-quant regression. With ep=2, the regression flips to approximately neutral (mean +0.7 mBPB across 3 seeds). The second epoch gives the model enough adaptation budget to recover the quant noise injected by INT6. Cost: 132s vs 76s for the TTT phase, both within the 600s eval budget. + +2. **Mixed-precision SSM dynamics protection**: the `dd_A` and `dd_dt` rows of each Mamba-3 `in_proj.weight` (32 of 2232 rows per SSM block) are quantized at INT8 instead of INT6. Q-Mamba (ICLR 2025) showed that uniform 6-bit PTQ collapses Mamba perplexity from 5.5 to >21 because A/Ā errors compound through the recurrence. Promoting just these semantic-specific rows to INT8 costs ~0.01 MiB at this scale and recovers ~0.8 mBPB of quality. Implemented as per-row bit widths threaded through both the GPTQ path and the percentile-search path. New env var `QUANT_BITS_SSM_DYNAMICS=8` (default in `Hyperparameters`). + +3. **Scale-floor quant bug fix**: an earlier mixed-precision commit accidentally hardcoded `scale.clamp_min(1.0/127)` (INT8 floor) for ALL rows, including INT6 rows that should floor at `1/31`. Consequence: INT6 q-values spread across [-31, 31] more uniformly, inflating LZMA entropy and starving selective ±1 pruning. Fixed to use per-row `1/qmax`. Net effect: ~1.4 MiB of spurious size inflation on prior runs disappears. + +## Architecture (unchanged from PR #1644) + +7-layer Mamba-3 SISO hybrid: 5 SSM blocks + 2 FlashAttention layers at positions 2 and 5, dim=512, d_state=64, expand=2, headdim=64, chunk_size=64, mlp_mult=3, 25.16M params. SP8192 BPE tokenizer trained from scratch on FineWeb. See PR #1644 for the full architectural rationale and Triton kernel analysis (no kernel-level changes here). + +## Reproduction + +```bash +SEED=1337 VOCAB_SIZE=8192 NUM_LAYERS=7 NUM_ATTN_LAYERS=2 \ + TRAIN_SEQ_LEN=4096 WARMDOWN_ITERS=2600 WARMDOWN_SHAPE=linear \ + MUON_EQ_R=1 LATE_QAT_THRESHOLD=0.15 \ + USE_GPTQ=1 QUANT_BITS=6 QUANT_BITS_EMBED=8 GPTQ_NUM_SEQS=32 \ + EVAL_OVERLAP=1024 USE_LZMA=1 EVAL_TEMP=0.9 TTT_EPOCHS=2 \ + WEIGHT_DECAY=0.04 MUON_MOMENTUM=0.99 MATRIX_LR=0.025 \ + torchrun --nproc_per_node=8 train_mamba3_hybrid.py +``` + +`QUANT_BITS_SSM_DYNAMICS=8` is the default in `Hyperparameters` and does not need to be set explicitly. Repeat with `SEED=42` and `SEED=2025` for the 3-seed mean. + +## Data + +Same as PR #1644: SP8192 BPE tokenizer trained from scratch on FineWeb-10B because the `kevclark/parameter-golf` SP8192 tokenizer was not consistent with this submission's tokenizer config. Tokenized shards and tokenizer artifacts available on a private HF dataset on request. + +## What I tested and removed + +This is a non-record submission and represents the cleaned production path from a much larger experimental sprint. The training script in this PR is the lean submission version. Many techniques that did not survive empirical validation at 25M / 10min / 16MB / SP8192 are not represented in this PR — including 1-attention ratio (works at SP4096, fails at SP8192 by +7.5 mBPB BF16), low-rank `in_proj` factorization (fails because random factored init destroys upstream's structured init for `dd_A`/`dd_dt` rows), depth recurrence at SP8192 (fails by +13.9 mBPB BF16 at expand=1.5), MLP INT5 quantization (+8 mBPB quality), and several others. Two patterns emerged worth flagging: + +- **LZMA compression penalty for SSM weights**: across three runs I measured SSM-heavy hybrids compressing ~33% under LZMA vs ~40% for attention-heavier hybrids — roughly a 3× higher compressed-bytes-per-raw-byte cost for swapping an attention block for an SSM block. The candidate mechanism (untested) is that Mamba-3's `in_proj` rows have heterogeneous distributions (z, xv, B, C, dd_dt, dd_A, trap, angles) and so quantize to higher-entropy byte streams than attention's uniform QKV. I did not run the experiment that would isolate this from other SSM-vs-attention differences. + +- **SP4096 architectural sweeps don't transfer to SP8192**: replacing 2-attn with 1-attn at SP8192 7L costs +7.5 mBPB BF16, even though the same swap at SP4096 8L was a clean −9.8 mBPB win. Depth recurrence at expand=1.5 has a similar sign flip across vocabularies. I don't have a tested explanation; one suspect I considered but didn't isolate is that Muon's Newton-Schulz orthogonalization may interact with the heterogeneous magnitude structure of SSM `in_proj` rows differently than with attention's uniform QKV. Mainly worth flagging as a methodology warning: don't extrapolate small-vocab sweep results to larger-vocab submissions. diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/requirements.txt b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/requirements.txt new file mode 100644 index 0000000000..7928acf0ab --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/requirements.txt @@ -0,0 +1,6 @@ +torch>=2.9.1 +triton>=3.5.0 +mamba-ssm>=2.3.1 +sentencepiece +einops +numpy diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/submission.json b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/submission.json new file mode 100644 index 0000000000..0979add958 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/submission.json @@ -0,0 +1,11 @@ +{ + "author": "mradassaad", + "github_id": "mradassaad", + "name": "Mamba-3 Hybrid SSM + SP8192 + Multi-Epoch TTT + Dynamics-Protected Quant", + "blurb": "Same 7L Mamba-3 SISO hybrid as PR #1644 with three additions: (1) TTT_EPOCHS=2 (multi-epoch chunk TTT recovers most quant damage), (2) mixed-precision quant protecting dd_A + dd_dt rows of in_proj at INT8, (3) scale-floor bug fix in mixed-precision pipeline. 3-seed mean post-quant+TTT 1.1456 bpb (std 0.0011).", + "date": "2026-04-22", + "val_loss": 2.95477626, + "val_bpb": 1.14408575, + "bytes_total": 15930191, + "bytes_code": 116783 +} diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_mamba3_hybrid.py b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_mamba3_hybrid.py new file mode 100644 index 0000000000..80c8024fe3 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_mamba3_hybrid.py @@ -0,0 +1,2121 @@ +"""Mamba-3 Hybrid for Parameter Golf. + +Sequential hybrid: Mamba-3 SSD blocks + a small number of standalone attention +layers (~2 attention + 5 SSM is our best-measured config at SP8192 7L). Env-var +driven; see `Hyperparameters` for all knobs. + +Feature surface in this file is the banked submission only. Many ideas were +tested and *removed* because they underperformed or required structural +changes unsuitable for an SSM at 25M / 10min / 16MB. Full list with verdicts +and the three structural findings (SSM LZMA penalty, Muon-SSM discordance, +SP4096 → SP8192 non-transfer) are documented in +`docs/ssm_structural_findings_writeup.md`. Git history holds the removed code +on commits prior to the cleanup. +""" + +from __future__ import annotations + +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +import lzma +import zlib +from pathlib import Path + +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +# ----------------------------- +# HYPERPARAMETERS +# ----------------------------- + +class Hyperparameters: + # Data paths are shard globs produced by the existing preprocessing pipeline. + # Defaults auto-select based on VOCAB_SIZE (e.g. 4096 → fineweb10B_sp4096). + _vs = os.environ.get("VOCAB_SIZE", "1024") + data_path = os.environ.get("DATA_PATH", f"./data/datasets/fineweb10B_sp{_vs}") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", f"./data/tokenizers/fineweb_{_vs}_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 1337)) + + # Validation cadence and batch size. Validation always uses the full fineweb_val split. + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) + + # Training length. + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000)) + warmdown_shape = os.environ.get("WARMDOWN_SHAPE", "cosine") # "linear" or "cosine" + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 1_048_576)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 4096)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + sweep_mode = bool(int(os.environ.get("SWEEP_MODE", "0"))) # skip post-training (quant, serialize, TTT) + + # Evaluation. + eval_stride = int(os.environ.get("EVAL_STRIDE", 16)) # sliding window stride (0 = disabled) + eval_overlap = int(os.environ.get("EVAL_OVERLAP", 0)) # stateful-overlap eval (0 = disabled, e.g. 1024) + eval_state_reset = int(os.environ.get("EVAL_STATE_RESET", 0)) # reset SSM state every N windows during eval (0 = never) + eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32)) # batch size for sliding eval + # Test-Time Training (TTT): chunk-based score-first adaptation on val data. + # Score each chunk (32×4096) under no_grad, then SGD adapt on the same chunk. + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.010)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "sgd") # "sgd" or "adamw" + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 1)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + + # Quantization. + quant_bits = int(os.environ.get("QUANT_BITS", 6)) + quant_bits_embed = int(os.environ.get("QUANT_BITS_EMBED", 0)) # 0 = same as quant_bits + # Mixed-precision SSM: dd_A and dd_dt rows of mamba3.in_proj.weight at higher bits. + # A/dt errors compound through the recurrence; only ~32 rows per SSM block, cheap to protect. + quant_bits_ssm_dynamics = int(os.environ.get("QUANT_BITS_SSM_DYNAMICS", 8)) + gptq_lite = bool(int(os.environ.get("GPTQ_LITE", "1"))) # search optimal clip percentile per tensor + use_lzma = bool(int(os.environ.get("USE_LZMA", "1"))) + eval_temp = float(os.environ.get("EVAL_TEMP", "1.0")) + use_gptq = bool(int(os.environ.get("USE_GPTQ", "0"))) + gptq_num_seqs = int(os.environ.get("GPTQ_NUM_SEQS", "32")) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", "0.0")) + target_mb = float(os.environ.get("TARGET_MB", "15.25")) # target compressed size in MiB + muon_eq_r = bool(int(os.environ.get("MUON_EQ_R", "0"))) # row-normalize before Newton-Schulz + + # Model shape. + vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) + num_layers = int(os.environ.get("NUM_LAYERS", 8)) + + # OrthoInit + SWA. + use_ortho_init = bool(int(os.environ.get("USE_ORTHO_INIT", "1"))) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4)) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + mlp_mult = float(os.environ.get("MLP_MULT", 3)) + mamba3_d_state = int(os.environ.get("MAMBA3_D_STATE", 64)) + mamba3_expand = float(os.environ.get("MAMBA3_EXPAND", 2)) + mamba3_headdim = int(os.environ.get("MAMBA3_HEADDIM", 64)) + mamba3_chunk_size = int(os.environ.get("MAMBA3_CHUNK_SIZE", 64)) + mamba3_ngroups = int(os.environ.get("MAMBA3_NGROUPS", 1)) + mamba3_rope_fraction = float(os.environ.get("MAMBA3_ROPE_FRACTION", 0.5)) + # Attention layers (evenly spaced among SSD layers). + num_attn_layers = int(os.environ.get("NUM_ATTN_LAYERS", 1)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + + # Optimizer hyperparameters. + weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.02)) + embed_wd = float(os.environ.get("EMBED_WD", 0.0)) # embedding weight decay (SOTA: 0.085) + ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "1"))) + ema_decay = float(os.environ.get("EMA_DECAY", 0.997)) + 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", 0.05)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.02)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) + +# MUON OPTIMIZER + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7, eq_r: bool = False) -> Tensor: + a, b, c = (3.4445, -4.7750, 2.0315) + X = G.bfloat16() + if eq_r: + row_norms = X.norm(dim=-1, keepdim=True).clamp(min=eps) + X = X / row_norms + X /= X.norm() + eps + transposed = G.size(0) > G.size(1) + if transposed: + X = X.T + for _ in range(steps): + A = X @ X.T + B = b * A + c * A @ A + X = a * X + B @ X + return X.T if transposed else X + + +class Muon(torch.optim.Optimizer): + def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True, weight_decay: float = 0.0, eq_r: bool = False): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay, eq_r=eq_r), + ) + + @torch.no_grad() + def step(self, closure=None): + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + distributed = dist.is_available() and dist.is_initialized() + world_size = dist.get_world_size() if distributed else 1 + rank = dist.get_rank() if distributed else 0 + + for group in self.param_groups: + params = group["params"] + if not params: + continue + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + weight_decay = group.get("weight_decay", 0.0) + eq_r = group.get("eq_r", False) + + total_params = sum(int(p.numel()) for p in params) + updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) + + curr = 0 + for i, p in enumerate(params): + if i % world_size == rank and p.grad is not None: + g = p.grad + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + buf.mul_(momentum).add_(g) + if nesterov: + g = g.add(buf, alpha=momentum) + g = zeropower_via_newtonschulz5(g, steps=backend_steps, eq_r=eq_r) + g *= max(1, g.size(0) / g.size(1)) ** 0.5 + updates_flat[curr : curr + p.numel()] = g.reshape(-1) + curr += p.numel() + + if distributed: + dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) + + curr = 0 + for p in params: + if weight_decay > 0: + p.data.mul_(1.0 - lr * weight_decay) + g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) + p.add_(g, alpha=-lr) + curr += p.numel() + + return loss + + +# TOKENIZER + EVALUATION + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) + + +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] + + +def eval_val( + args: Hyperparameters, model: nn.Module, rank: int, world_size: int, + device: torch.device, grad_accum_steps: int, val_tokens: Tensor, + base_bytes_lut: Tensor, has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < args.train_seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + ) + local_batch_seqs = local_batch_tokens // args.train_seq_len + total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * args.train_seq_len + raw_end = batch_seq_end * args.train_seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, args.train_seq_len) + y = local[1:].reshape(-1, args.train_seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + + +def eval_val_sliding( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, +) -> tuple[float, float]: + """Sliding window evaluation: score each token with maximal left-context.""" + seq_len = args.train_seq_len + stride = args.eval_stride + batch_size = args.eval_batch_seqs + total_tokens = val_tokens.numel() - 1 + + window_starts = [ws for ws in range(0, total_tokens - seq_len + 1, stride)] + if window_starts[-1] + seq_len < total_tokens: + window_starts.append(total_tokens - seq_len) + + my_starts = window_starts[rank::world_size] + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + with torch.inference_mode(): + for batch_start in range(0, len(my_starts), batch_size): + batch_ws = my_starts[batch_start:batch_start + batch_size] + bsz = len(batch_ws) + + x_list, y_list = [], [] + for ws in batch_ws: + chunk = val_tokens[ws:ws + seq_len + 1].to(dtype=torch.int64) + x_list.append(chunk[:-1]) + y_list.append(chunk[1:]) + x_batch = torch.stack(x_list).to(device=device, non_blocking=True) + y_batch = torch.stack(y_list).to(device=device, non_blocking=True) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = base_model.forward_logits(x_batch) + scaled_logits = logits.float() + if args.eval_temp != 1.0: + scaled_logits = scaled_logits / args.eval_temp + nll = F.cross_entropy( + scaled_logits.reshape(-1, scaled_logits.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = min(seq_len, total_tokens - ws) + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + + prev_ids = x_batch[i, s:wlen] + tgt_ids = y_batch[i, s:wlen] + tbytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + tbytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + byte_count += tbytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return float(val_loss), float(bits_per_token * tokens_per_byte) + + +def eval_val_stateful_overlap( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + initial_states: list | None = None, +) -> tuple[float, float]: + """Stateful eval with overlapping windows: SSM state carries forward, + attention gets `overlap` tokens of prior context per window.""" + seq_len = args.train_seq_len + overlap = args.eval_overlap + score_len = seq_len - overlap + total_tokens = val_tokens.numel() - 1 + + segment_tokens = (total_tokens // world_size // score_len) * score_len + seg_start = rank * segment_tokens + num_windows = segment_tokens // score_len + + log0 = (lambda msg: print(msg)) if rank == 0 else (lambda msg: None) + state_reset = args.eval_state_reset + log0(f"stateful_overlap_eval: {num_windows} windows, overlap={overlap}, " + f"score_len={score_len}, segment={segment_tokens} tokens, state_reset={state_reset}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + base_model.eval() + layer_states = initial_states + + with torch.inference_mode(): + # Pre-warm SSM state for non-first segments so only rank 0 pays cold-start + if layer_states is None and seg_start > 0: + warmup_len = min(seq_len, seg_start) + warmup_chunk = val_tokens[seg_start - warmup_len : seg_start + 1].to(dtype=torch.int64) + wx = warmup_chunk[:-1].unsqueeze(0).to(device=device, non_blocking=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + _, layer_states = base_model.forward_logits_stateful(wx, None) + layer_states = [tuple(s.detach() for s in st) for st in layer_states] + log0(f"stateful_overlap_eval: pre-warmed={seg_start > 0} warmup_len={min(seq_len, seg_start) if seg_start > 0 else 0} initial_states={'provided' if initial_states is not None else 'None'}") + + for wi in range(num_windows): + if state_reset > 0 and wi % state_reset == 0: + layer_states = None + + score_start = seg_start + wi * score_len + tok_start = max(score_start - overlap, seg_start) + tok_end = score_start + score_len + actual_overlap = score_start - tok_start + + chunk = val_tokens[tok_start:tok_end + 1].to(dtype=torch.int64) + x = chunk[:-1].unsqueeze(0).to(device=device, non_blocking=True) + y = chunk[1:].unsqueeze(0).to(device=device, non_blocking=True) + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits, layer_states = base_model.forward_logits_stateful(x, layer_states) + + layer_states = [tuple(s.detach() for s in st) for st in layer_states] + + scored_logits = logits[:, actual_overlap:, :].float() + scored_y = y[:, actual_overlap:] + + if args.eval_temp != 1.0: + scored_logits = scored_logits / args.eval_temp + + nll = F.cross_entropy( + scored_logits.reshape(-1, scored_logits.size(-1)), + scored_y.reshape(-1), + reduction="none", + ) + + loss_sum += nll.to(torch.float64).sum() + token_count += score_len + + prev_ids = x.squeeze(0)[actual_overlap:] + tgt_ids = y.squeeze(0)[actual_overlap:] + tbytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + tbytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + byte_count += tbytes.to(torch.float64).sum() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return float(val_loss), float(bits_per_token * tokens_per_byte) + +# QUANTIZATION + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "m3_scale,mlp_scale,attn_scale,resid_mix,skip_weight,skip_weights,dt_bias,B_bias,C_bias,.D,A_log,q_gain", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 + +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) + + +def compute_ssm_dynamics_row_mask(name: str, rows: int, cfg: dict) -> Tensor | None: + """For mamba3.in_proj.weight (dense) OR mamba3.in_proj.1.weight (factored up-projection), + return a boolean mask of shape (rows,) marking dd_A and dd_dt rows (the recurrence-sensitive + subset). Returns None for other tensors. + in_proj output layout: [z | xv | B | C | dd_dt | dd_A | trap | angles]. Factored up-proj + keeps the same row layout since the output dim matches.""" + is_dense_in_proj = "mamba3.in_proj.weight" in name + is_factored_up = "mamba3.in_proj.1.weight" in name + if not (is_dense_in_proj or is_factored_up): + return None + d_inner = int(cfg["model_dim"] * cfg["mamba3_expand"]) + d_state = int(cfg["mamba3_d_state"]) + ngroups = int(cfg["mamba3_ngroups"]) + mimo_rank = 1 # SISO (is_mimo=False in Mamba3Layer) + nheads = d_inner // int(cfg["mamba3_headdim"]) + off = 2 * d_inner + 2 * d_state * ngroups * mimo_rank + dt_start, dt_end = off, off + nheads + A_start, A_end = dt_end, dt_end + nheads + if A_end > rows: + return None # shape mismatch — bail silently + mask = torch.zeros(rows, dtype=torch.bool) + mask[dt_start:A_end] = True # dt + A span + return mask + +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t + +def _quantize_with_clip(t32: Tensor, clip_abs: Tensor | float, qmax: Tensor | int) -> tuple[Tensor, Tensor, Tensor]: + # qmax may be a per-row tensor (mixed precision). When per-row, we quantize with + # per-row step size but still store into an int8 tensor — the higher-bit rows use + # the wider range of int8, the lower-bit rows stay within their narrower clip. + if t32.ndim == 2 and isinstance(clip_abs, Tensor): + qmax_col = qmax[:, None] if isinstance(qmax, Tensor) else qmax + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / (qmax.float() if isinstance(qmax, Tensor) else float(qmax))) + # Per-row scale floor must match the per-row bit width. 1/qmax = smallest scale + # such that the minimum nonzero quantized magnitude (±1 * scale) ≈ 1/qmax range. + # Using a uniform 1/127 floor (INT8) was a bug: it let INT6 rows use scales ~4× too + # small, inflating quantized magnitudes and defeating selective ±1 pruning. + if isinstance(qmax, Tensor): + scale = torch.maximum(scale, 1.0 / qmax.float()) + else: + scale = scale.clamp_min(1.0 / float(qmax)) + q_unclamped = torch.round(clipped / scale[:, None]) + if isinstance(qmax_col, Tensor): + qmax_f = qmax_col.float() + q = torch.maximum(torch.minimum(q_unclamped, qmax_f), -qmax_f).to(torch.int8) + else: + q = torch.clamp(q_unclamped, -qmax, qmax).to(torch.int8) + recon = q.float() * scale[:, None] + return q.contiguous(), scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous(), recon + clip_abs_f = float(clip_abs) if isinstance(clip_abs, Tensor) else clip_abs + qmax_i = int(qmax) if not isinstance(qmax, Tensor) else int(qmax.item()) + scale_f = clip_abs_f / qmax_i if clip_abs_f > 0 else 1.0 + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs_f, clip_abs_f) / scale_f), -qmax_i, qmax_i).to(torch.int8) + recon = q.float() * scale_f + return q.contiguous(), torch.tensor(scale_f, dtype=torch.float32), recon + +def quantize_float_tensor(t: Tensor, bits: int | Tensor = 8, search_clip: bool = False) -> tuple[Tensor, Tensor]: + # bits may be a per-row tensor of shape (rows,) for mixed-precision 2D weights. + if isinstance(bits, Tensor): + qmax = ((1 << (bits.long() - 1)) - 1).to(torch.int64) + else: + qmax = (1 << (int(bits) - 1)) - 1 + t32 = t.float() + + if not search_clip: + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + q, scale, _ = _quantize_with_clip(t32, clip_abs, qmax) + return q, scale + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + q, scale, _ = _quantize_with_clip(t32, clip_abs, qmax) + return q, scale + + candidates = [0.999, 0.9995, 0.9999, 0.99995, 0.99999, 0.999999, 1.0] + best_q, best_scale, best_mse = None, None, float("inf") + + for pct in candidates: + if t32.ndim == 2: + if pct >= 1.0: + clip_abs = t32.abs().amax(dim=1) + else: + clip_abs = torch.quantile(t32.abs(), pct, dim=1) + q, scale, recon = _quantize_with_clip(t32, clip_abs, qmax) + else: + if pct >= 1.0: + clip_abs = float(t32.abs().max().item()) + else: + clip_abs = float(torch.quantile(t32.abs().flatten(), pct).item()) if t32.numel() else 0.0 + q, scale, recon = _quantize_with_clip(t32, clip_abs, qmax) + mse = (t32 - recon).pow(2).mean().item() + if mse < best_mse: + best_mse = mse + best_q, best_scale = q, scale + + return best_q, best_scale + +def generate_autoregressive_calib( + model: nn.Module, device: torch.device, num_seqs: int = 64, seq_len: int = 2048, + vocab_size: int = 1024, temperature: float = 0.8, batch_size: int = 8, seed: int = 42, +) -> list[Tensor]: + """Generate sequences autoregressively from the model for GPTQ calibration. + No external data accessed — fully self-contained and legal.""" + model.eval() + rng = torch.Generator(device=device) + rng.manual_seed(seed) + all_tokens = [] + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for batch_start in range(0, num_seqs, batch_size): + bs = min(batch_size, num_seqs - batch_start) + tokens = torch.randint(0, vocab_size, (bs, 1), device=device, generator=rng) + for _ in range(seq_len - 1): + logits = model.forward_logits(tokens) + next_logit = logits[:, -1, :] + probs = torch.softmax(next_logit / temperature, dim=-1) + next_tok = torch.multinomial(probs, 1, generator=rng) + tokens = torch.cat([tokens, next_tok], dim=1) + for i in range(bs): + all_tokens.append(tokens[i:i+1]) + return all_tokens + + +def collect_hessians_from_tokens( + model: nn.Module, token_seqs: list[Tensor], device: torch.device, + gptq_embed: bool = False, +) -> dict[str, Tensor]: + """Collect H = X^T X from pre-generated token sequences via forward hooks on CastedLinear layers.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + param_name = name + ".weight" + cols = module.weight.shape[1] + hessians[param_name] = torch.zeros(cols, cols, dtype=torch.float32, device=device) + def make_hook(pname: str): + def hook_fn(mod: nn.Module, inp: tuple, out: Tensor) -> None: + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pname].addmm_(x.T, x) + return hook_fn + hooks.append(module.register_forward_hook(make_hook(param_name))) + + if gptq_embed and model.tie_embeddings: + def make_output_hook(pname: str): + def hook_fn(mod: nn.Module, inp: tuple, out: Tensor) -> None: + x = out.detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if pname not in hessians: + hessians[pname] = torch.zeros(x.shape[1], x.shape[1], dtype=torch.float32, device=device) + hessians[pname].addmm_(x.T, x) + return hook_fn + hooks.append(model.final_norm.register_forward_hook(make_output_hook("tok_emb.weight"))) + + model.eval() + with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for seq in token_seqs: + x = seq[:, :-1].to(device) + y = seq[:, 1:].to(device) + model(x, y) + for h in hooks: + h.remove() + num_batches = len(token_seqs) + for name in hessians: + H = hessians[name] + H /= num_batches + damp = 0.01 * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0], device=device) + hessians[name] = H.cpu() + return hessians + + +def collect_hessians_from_train_data( + model: nn.Module, train_loader, device: torch.device, + train_batch_tokens: int, seq_len: int, grad_accum_steps: int, + n_batches: int = 64, gptq_embed: bool = True, +) -> dict[str, Tensor]: + """Collect GPTQ Hessians from training data (like SOTA). Much faster than AR self-gen. + Also collects Hessian for tied embeddings via output hook on final_norm.""" + hessians: dict[str, Tensor] = {} + hooks = [] + + def make_hook(pname: str): + def hook_fn(mod: nn.Module, inp: tuple, out: Tensor) -> None: + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if pname not in hessians: + hessians[pname] = torch.zeros(x.shape[1], x.shape[1], dtype=torch.float32, device=device) + hessians[pname].addmm_(x.T, x) + return hook_fn + + for name, module in model.named_modules(): + if isinstance(module, CastedLinear) and module.weight.numel() > INT8_KEEP_FLOAT_MAX_NUMEL: + hooks.append(module.register_forward_hook(make_hook(name + ".weight"))) + + # Embedding Hessian: hook the output of final_norm (input to logit projection = tok_emb.weight^T) + if gptq_embed and model.tie_embeddings: + def make_output_hook(pname: str): + def hook_fn(mod: nn.Module, inp: tuple, out: Tensor) -> None: + x = out.detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if pname not in hessians: + hessians[pname] = torch.zeros(x.shape[1], x.shape[1], dtype=torch.float32, device=device) + hessians[pname].addmm_(x.T, x) + return hook_fn + hooks.append(model.final_norm.register_forward_hook(make_output_hook("tok_emb.weight"))) + + model.eval() + with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(n_batches): + x, _ = train_loader.next_batch(train_batch_tokens, seq_len, grad_accum_steps) + model.forward_logits(x) + + for h in hooks: + h.remove() + for name in hessians: + H = hessians[name] + H /= n_batches + damp = 0.01 * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0], device=device) + hessians[name] = H.cpu() + return hessians + + +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor | None = None, clip_range: int | Tensor = 31, block_size: int = 128, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation and column reordering. + Falls back to percentile search if hessian is None. + clip_range may be a per-row tensor of shape (rows,) for mixed-precision quantization.""" + t32 = weight.float() + if t32.ndim != 2 or hessian is None: + return _quantize_int6_percentile(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * torch.mean(torch.diag(H)) + H[torch.arange(cols), torch.arange(cols)] += damp + # Column reordering: quantize most-activated (sensitive) columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Compute upper Cholesky of H_inv for the error propagation sweep + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except torch.linalg.LinAlgError: + print(f"gptq:cholesky_fallback shape={tuple(weight.shape)} dead={int(dead.sum())}/{cols}", flush=True) + return _quantize_int6_percentile(weight, clip_range=clip_range) + + # Per-row clip_range support: when a Tensor, broadcast as (rows,) for 1D ops + # and (rows, 1) for 2D ops. Scale clamp_min is per-row when clip_range is per-row. + cr_is_tensor = isinstance(clip_range, Tensor) + if cr_is_tensor: + cr_1d = clip_range.float() # (rows,) + cr_min = (1.0 / cr_1d).to(torch.float32) + else: + cr_1d = float(clip_range) + cr_min = 1.0 / float(clip_range) + + def _gptq_sweep(s: Tensor) -> tuple[Tensor, float]: + """Run GPTQ block-wise quantization with given per-row scale s.""" + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + if cr_is_tensor: + q = torch.maximum(torch.minimum(torch.round(w / sf), cr_1d), -cr_1d).to(torch.int8) + else: + q = torch.clamp(torch.round(w / sf), -cr_1d, cr_1d).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + return Q, mse + + def _scale_from_row_clip(row_clip: Tensor) -> Tensor: + # s = max(row_clip / clip_range, 1 / clip_range). Works scalar or per-row. + raw = row_clip / cr_1d if cr_is_tensor else row_clip / cr_1d + if cr_is_tensor: + return torch.maximum(raw, cr_min).to(torch.float16) + return raw.clamp_min(cr_min).to(torch.float16) + + # Percentile search: pick the row-wise clip percentile that minimizes MSE. + best_q, best_scale, best_err = None, None, float("inf") + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + row_clip = torch.quantile(t32.abs(), pct, dim=1) if pct < 1.0 else t32.abs().amax(dim=1) + s = _scale_from_row_clip(row_clip) + Q, mse = _gptq_sweep(s) + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale + + +def _quantize_int6_percentile(t32: Tensor, clip_range: int | Tensor = 31) -> tuple[Tensor, Tensor]: + """Fallback percentile-search quantization (no Hessian). + clip_range may be a per-row tensor of shape (rows,) for mixed precision.""" + cr_is_tensor = isinstance(clip_range, Tensor) + if t32.ndim == 2: + cr_col = clip_range[:, None].float() if cr_is_tensor else float(clip_range) + cr_1d = clip_range.float() if cr_is_tensor else float(clip_range) + cr_min = (1.0 / cr_1d) if cr_is_tensor else 1.0 / cr_1d + best_q, best_s, best_err = None, None, float("inf") + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + row_clip = torch.quantile(t32.abs(), pct, dim=1) if pct < 1.0 else t32.abs().amax(dim=1) + raw = row_clip / cr_1d + s = (torch.maximum(raw, cr_min) if cr_is_tensor else raw.clamp_min(cr_min)).to(torch.float16) + if cr_is_tensor: + q = torch.maximum(torch.minimum(torch.round(t32 / s.float()[:, None]), cr_col), -cr_col).to(torch.int8) + else: + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -cr_col, cr_col).to(torch.int8) + err = (t32 - q.float() * s.float()[:, None]).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + # 1D fallback — scalar clip_range only + cr_scalar = int(clip_range) if not cr_is_tensor else int(clip_range.max().item()) + amax = t32.abs().max().item() + scale = torch.tensor(amax / cr_scalar if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -cr_scalar, cr_scalar).to(torch.int8) + return q, scale + + +def quantize_state_dict_int8( + state_dict: dict[str, Tensor], quant_bits: int = 8, + quant_bits_embed: int = 0, search_clip: bool = False, + quant_bits_ssm_dynamics: int = 0, ssm_cfg: dict | None = None, +): + quantized: dict[str, Tensor] = {} + scales: dict[str, Tensor] = {} + dtypes: dict[str, str] = {} + passthrough: dict[str, Tensor] = {} + passthrough_orig_dtypes: dict[str, str] = {} + qmeta: dict[str, dict[str, object]] = {} + stats = dict.fromkeys( + ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), + 0, + ) + + for name, tensor in state_dict.items(): + t = tensor.detach().to("cpu").contiguous() + stats["param_count"] += int(t.numel()) + stats["num_tensors"] += 1 + stats["baseline_tensor_bytes"] += tensor_nbytes(t) + + if not t.is_floating_point(): + stats["num_nonfloat_tensors"] += 1 + passthrough[name] = t + stats["int8_payload_bytes"] += tensor_nbytes(t) + continue + + if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: + kept = keep_float_tensor(name, t, passthrough_orig_dtypes) + passthrough[name] = kept + stats["int8_payload_bytes"] += tensor_nbytes(kept) + continue + + stats["num_float_tensors"] += 1 + bits: int | Tensor = quant_bits + if quant_bits_embed > 0 and "tok_emb" in name: + bits = quant_bits_embed + # Mixed-precision in_proj: promote dd_A + dd_dt rows to higher bits. + if quant_bits_ssm_dynamics > 0 and ssm_cfg is not None and t.ndim == 2: + dyn_mask = compute_ssm_dynamics_row_mask(name, t.shape[0], ssm_cfg) + if dyn_mask is not None: + base = int(bits) if not isinstance(bits, Tensor) else int(quant_bits) + per_row = torch.full((t.shape[0],), base, dtype=torch.int64) + per_row[dyn_mask] = quant_bits_ssm_dynamics + bits = per_row + q, s = quantize_float_tensor(t, bits=bits, search_clip=search_clip) + if s.ndim > 0: + qmeta[name] = {"scheme": "per_row", "axis": 0} + quantized[name] = q + scales[name] = s + dtypes[name] = str(t.dtype).removeprefix("torch.") + stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) + + obj: dict[str, object] = { + "__quant_format__": "int8_clean_per_row_v1", + "quantized": quantized, + "scales": scales, + "dtypes": dtypes, + "passthrough": passthrough, + } + if qmeta: + obj["qmeta"] = qmeta + if passthrough_orig_dtypes: + obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes + return obj, stats + +def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + qmeta = obj.get("qmeta", {}) + passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) + passthrough_data = obj["passthrough"] + for name, q in obj["quantized"].items(): + dtype = getattr(torch, obj["dtypes"][name]) + s = obj["scales"][name] + if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: + s = s.to(dtype=torch.float32) + deq = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() + else: + scale = float(s.item()) + deq = (q.float() * scale).to(dtype=dtype).contiguous() + out[name] = deq + for name, t in passthrough_data.items(): + out_t = t.detach().to("cpu").contiguous() + orig_dtype = passthrough_orig_dtypes.get(name) + if isinstance(orig_dtype, str): + out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() + out[name] = out_t + return out + + +# ----------------------------- +# DATA LOADING +# ----------------------------- + +def load_data_shard(file: Path) -> Tensor: + header_bytes = 256 * np.dtype(" None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) + + +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + + +# ----------------------------- +# TRANSFORMER MODULES +# ----------------------------- + + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) + + +class FakeQuantizeSTE(torch.autograd.Function): + """Simulated quantization with Straight-Through Estimator for QAT.""" + @staticmethod + def forward(ctx, w: Tensor, bits: int) -> Tensor: + qmax = (1 << (bits - 1)) - 1 + if w.ndim == 2: + scale = w.detach().abs().amax(dim=1, keepdim=True) / qmax + scale = scale.clamp_min(1.0 / qmax) + return (torch.clamp(torch.round(w / scale), -qmax, qmax) * scale).to(w.dtype) + scale = w.detach().abs().amax() / qmax + scale = scale.clamp_min(1.0 / qmax) + return (torch.clamp(torch.round(w / scale), -qmax, qmax) * scale).to(w.dtype) + + @staticmethod + def backward(ctx, grad: Tensor) -> tuple[Tensor, None]: + return grad, None + + +class CastedLinear(nn.Linear): + _qat_bits: int = 0 + + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if self._qat_bits > 0 and self.weight.numel() > 65536: + w = FakeQuantizeSTE.apply(w, self._qat_bits) + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) + + +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() + + +@torch._dynamo.disable +def _mamba3_ssd_kernel(Q, K, V, ADT, DT, Trap, Q_bias, K_bias, Angles, D, Z, + chunk_size, Input_States=None, return_final_states=False): + from mamba_ssm.ops.triton.mamba3.mamba3_siso_combined import mamba3_siso_combined + return mamba3_siso_combined( + Q=Q, K=K, V=V, ADT=ADT, DT=DT, Trap=Trap, + Q_bias=Q_bias, K_bias=K_bias, Angles=Angles, D=D, Z=Z, + chunk_size=chunk_size, Input_States=Input_States, + return_final_states=return_final_states, + ) + + +class Mamba3Layer(nn.Module): + """Pure Mamba-3 SISO layer. Wraps upstream Mamba3 with CastedLinear projections + so QAT fake-quant and fp32 master weights apply to the outlier-heavy projections.""" + def __init__(self, dim: int, d_state: int = 64, expand: float = 2, + headdim: int = 64, chunk_size: int = 64, + ngroups: int = 1, rope_fraction: float = 0.5): + super().__init__() + from mamba_ssm.modules.mamba3 import Mamba3 + self.mamba3 = Mamba3( + d_model=dim, d_state=d_state, expand=expand, + headdim=headdim, is_mimo=False, chunk_size=chunk_size, + ngroups=ngroups, rope_fraction=rope_fraction, + is_outproj_norm=False, + ) + for attr in ("in_proj", "out_proj"): + src = getattr(self.mamba3, attr) + dst = CastedLinear(src.in_features, src.out_features, bias=src.bias is not None) + dst.weight = src.weight + if src.bias is not None: + dst.bias = src.bias + setattr(self.mamba3, attr, dst) + + def _pre_ssd(self, x): + """Pre-SSD ops: in_proj, split, reshape, compute ADT/DT, norms.""" + from einops import rearrange + m = self.mamba3 + zxBCdtAtrap = m.in_proj(x) + z, xv, B, C, dd_dt, dd_A, trap, angles = torch.split( + zxBCdtAtrap, + [m.d_inner, m.d_inner, + m.d_state * m.num_bc_heads * m.mimo_rank, + m.d_state * m.num_bc_heads * m.mimo_rank, + m.nheads, m.nheads, m.nheads, m.num_rope_angles], + dim=-1) + z = rearrange(z, "b l (h p) -> b l h p", p=m.headdim) + xv = rearrange(xv, "b l (h p) -> b l h p", p=m.headdim) + B = rearrange(B, "b l (r g n) -> b l r g n", r=m.mimo_rank, g=m.num_bc_heads) + C = rearrange(C, "b l (r g n) -> b l r g n", r=m.mimo_rank, g=m.num_bc_heads) + trap = rearrange(trap, "b l h -> b h l") + _A = -F.softplus(dd_A.to(torch.float32)) + _A = torch.clamp(_A, max=-m.A_floor) + DT = F.softplus(dd_dt + m.dt_bias) + ADT = _A * DT + DT = rearrange(DT, "b l n -> b n l") + ADT = rearrange(ADT, "b l n -> b n l") + angles = angles.unsqueeze(-2).expand(-1, -1, m.nheads, -1) + B = m.B_norm(B).squeeze(2) + C = m.C_norm(C).squeeze(2) + return z, xv, B, C, ADT, DT, trap, angles + + def _post_ssd(self, y): + from einops import rearrange + m = self.mamba3 + y = rearrange(y, "b l h p -> b l (h p)") + return m.out_proj(y) + + def forward(self, x: Tensor) -> Tensor: + m = self.mamba3 + z, xv, B, C, ADT, DT, trap, angles = self._pre_ssd(x) + y = _mamba3_ssd_kernel( + Q=C, K=B, V=xv, + ADT=ADT, DT=DT, Trap=trap, + Q_bias=m.C_bias.squeeze(1), K_bias=m.B_bias.squeeze(1), + Angles=angles, D=m.D, + Z=z, + chunk_size=m.chunk_size, + ) + return self._post_ssd(y) + + def forward_stateful(self, x: Tensor, input_states=None): + """Forward with SSM state carry. Returns (output, final_states).""" + m = self.mamba3 + z, xv, B, C, ADT, DT, trap, angles = self._pre_ssd(x) + result = _mamba3_ssd_kernel( + Q=C, K=B, V=xv, + ADT=ADT, DT=DT, Trap=trap, + Q_bias=m.C_bias.squeeze(1), K_bias=m.B_bias.squeeze(1), + Angles=angles, D=m.D, + Z=z, + chunk_size=m.chunk_size, + Input_States=input_states, + return_final_states=True, + ) + y, last_angle, last_state, last_k, last_v, *_rest = result + final_states = (last_angle, last_state, last_k, last_v) + return self._post_ssd(y), final_states + + +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) + freqs = torch.outer(t, self.inv_freq.to(device)) + self._cos_cached = freqs.cos()[None, None, :, :] + self._sin_cached = freqs.sin()[None, None, :, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) + + +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: + 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 AttentionLayer(nn.Module): + """Causal self-attention with GQA and RoPE. QK-norm + per-head q_gain.""" + def __init__(self, dim: int, num_heads: int, num_kv_heads: int, + rope_base: float, qk_gain_init: float): + super().__init__() + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + kv_dim = num_kv_heads * self.head_dim + + self.c_q = CastedLinear(dim, dim, bias=False) + self.c_k = CastedLinear(dim, kv_dim, bias=False) + self.c_v = CastedLinear(dim, kv_dim, bias=False) + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rotary = Rotary(self.head_dim, base=rope_base) + self.proj = CastedLinear(dim, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + bsz, seqlen, dim = x.shape + q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + v = v * v.sigmoid() + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin) + k = apply_rotary_emb(k, cos, sin) + q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] + y = F.scaled_dot_product_attention( + q, k, v, attn_mask=None, is_causal=True, + enable_gqa=(self.num_kv_heads != self.num_heads), + ) + return self.proj(y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)) + + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: float): + super().__init__() + hidden = int(mlp_mult * dim) + self.fc = CastedLinear(dim, hidden, bias=False) + self.proj = CastedLinear(hidden, dim, bias=False) + self.proj._zero_init = True + + def forward(self, x: Tensor) -> Tensor: + x = F.leaky_relu(self.fc(x), negative_slope=0.5) + return self.proj(x.square()) + + +class Block(nn.Module): + def __init__( + self, dim: int, mlp_mult: int, + mamba3_d_state: int = 64, mamba3_expand: float = 2, + mamba3_headdim: int = 64, mamba3_chunk_size: int = 64, + mamba3_ngroups: int = 1, mamba3_rope_fraction: float = 0.5, + layer_idx: int = 0, + ): + super().__init__() + self.m3_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.mamba3 = Mamba3Layer( + dim, d_state=mamba3_d_state, expand=mamba3_expand, + headdim=mamba3_headdim, chunk_size=mamba3_chunk_size, + ngroups=mamba3_ngroups, rope_fraction=mamba3_rope_fraction, + ) + self.mlp = MLP(dim, mlp_mult) + self.m3_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + m3_out = self.mamba3(self.m3_norm(x)) + x = x + self.m3_scale.to(dtype=x.dtype)[None, None, :] * m3_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + def forward_stateful(self, x: Tensor, x0: Tensor, ssm_states=None): + """Forward with SSM state carry. Returns (output, final_states).""" + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + m3_out, fstate = self.mamba3.forward_stateful(self.m3_norm(x), input_states=ssm_states) + x = x + self.m3_scale.to(dtype=x.dtype)[None, None, :] * m3_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x, fstate + + +class AttnBlock(nn.Module): + """Block with standalone attention (no SSM).""" + def __init__( + self, dim: int, mlp_mult: int, + num_heads: int = 8, num_kv_heads: int = 4, + rope_base: float = 10000.0, qk_gain_init: float = 1.0, + layer_idx: int = 0, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = AttentionLayer(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.mlp = MLP(dim, mlp_mult) + self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) + self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + + def forward(self, x: Tensor, x0: Tensor) -> Tensor: + mix = self.resid_mix.to(dtype=x.dtype) + x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn(self.attn_norm(x)) + x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out + x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) + return x + + +class GPT(nn.Module): + def __init__( + self, vocab_size: int, num_layers: int, model_dim: int, + mlp_mult: int, tie_embeddings: bool, tied_embed_init_std: float, + logit_softcap: float, + use_ortho_init: bool = False, + mamba3_d_state: int = 64, mamba3_expand: float = 2, + mamba3_headdim: int = 64, mamba3_chunk_size: int = 64, + mamba3_ngroups: int = 1, mamba3_rope_fraction: float = 0.5, + num_attn_layers: int = 1, num_heads: int = 8, num_kv_heads: int = 4, + rope_base: float = 10000.0, qk_gain_init: float = 1.0, + ): + super().__init__() + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.use_ortho_init = use_ortho_init + self.tok_emb = nn.Embedding(vocab_size, model_dim) + + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter( + torch.ones(1, self.num_skip_weights, model_dim, dtype=torch.float32) + ) + self.encoder_indices = list(range(self.num_encoder_layers)) + self.decoder_indices = list(range(self.num_encoder_layers, num_layers)) + + # Evenly-spaced attention layer indices among SSM layers. + attn_indices = set() + if num_attn_layers > 0: + for i in range(1, num_attn_layers + 1): + attn_indices.add(round(i * num_layers / (num_attn_layers + 1))) + self.attn_indices = sorted(attn_indices) + + self.blocks = nn.ModuleList() + for i in range(num_layers): + if i in attn_indices: + self.blocks.append(AttnBlock( + model_dim, mlp_mult, + num_heads=num_heads, num_kv_heads=num_kv_heads, + rope_base=rope_base, qk_gain_init=qk_gain_init, + layer_idx=i, + )) + else: + self.blocks.append(Block( + model_dim, mlp_mult, + mamba3_d_state=mamba3_d_state, mamba3_expand=mamba3_expand, + mamba3_headdim=mamba3_headdim, mamba3_chunk_size=mamba3_chunk_size, + mamba3_ngroups=mamba3_ngroups, mamba3_rope_fraction=mamba3_rope_fraction, + layer_idx=i, + )) + + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self._init_weights() + + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif self.use_ortho_init and module.weight.ndim == 2 and min(module.weight.shape) >= 16: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj" in name and name.split(".")[-1] in ("proj", "proj_D"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + + def _compute_logits_and_loss(self, x: Tensor, target_ids: Tensor) -> Tensor: + x = self.final_norm(x).reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + return F.cross_entropy(logits.float(), targets, reduction="mean") + + _embed_qat_bits: int = 0 + + def _embed(self, input_ids: Tensor) -> Tensor: + if self._embed_qat_bits > 0: + w = FakeQuantizeSTE.apply(self.tok_emb.weight.to(torch.bfloat16), self._embed_qat_bits) + x = F.embedding(input_ids, w) + else: + x = self.tok_emb(input_ids) + return F.rms_norm(x, (x.size(-1),)) + + def forward(self, input_ids: Tensor, target_ids: Tensor, + layer_states: list | None = None, stateful: bool = False): + x = self._embed(input_ids) + if stateful: + x, new_states = self._run_blocks_stateful(x, x, layer_states) + loss = self._compute_logits_and_loss(x, target_ids) + return loss, new_states + x = self._run_blocks(x, x) + return self._compute_logits_and_loss(x, target_ids) + + def _run_blocks(self, x: Tensor, x0: Tensor) -> Tensor: + enc_idx = list(range(self.num_encoder_layers)) + dec_idx = list(range(self.num_encoder_layers, len(self.blocks))) + skips: list[Tensor] = [] + for bi in enc_idx: + x = self.blocks[bi](x, x0) + skips.append(x) + for i, bi in enumerate(dec_idx): + if i < len(skips) and i < self.skip_weights.shape[1]: + x = x + self.skip_weights[0, i].to(dtype=x.dtype)[None, None, :] * skips.pop() + elif skips: + skips.pop() + x = self.blocks[bi](x, x0) + return x + + def forward_logits(self, input_ids: Tensor) -> Tensor: + bsz, seqlen = input_ids.shape + x = self._embed(input_ids) + x = self._run_blocks(x, x) + x = self.final_norm(x).reshape(-1, x.size(-1)) + w = self.tok_emb.weight if self.tie_embeddings else self.lm_head.weight + logits = self.logit_softcap * torch.tanh(F.linear(x, w) / self.logit_softcap) + return logits.reshape(bsz, seqlen, -1) + + def _run_blocks_stateful(self, x: Tensor, x0: Tensor, + layer_states: list | None = None): + """Like _run_blocks but carries SSM state per Mamba-3 layer. Returns (output, new_layer_states).""" + new_states: list = [] + si = 0 + enc_idx = list(range(self.num_encoder_layers)) + dec_idx = list(range(self.num_encoder_layers, len(self.blocks))) + skips: list[Tensor] = [] + for bi in enc_idx: + block = self.blocks[bi] + if isinstance(block, Block): + prev = layer_states[si] if layer_states is not None else None + x, fstate = block.forward_stateful(x, x0, ssm_states=prev) + new_states.append(fstate) + si += 1 + else: + x = block(x, x0) + skips.append(x) + for i, bi in enumerate(dec_idx): + if i < len(skips) and i < self.skip_weights.shape[1]: + x = x + self.skip_weights[0, i].to(dtype=x.dtype)[None, None, :] * skips.pop() + elif skips: + skips.pop() + block = self.blocks[bi] + if isinstance(block, Block): + prev = layer_states[si] if layer_states is not None else None + x, fstate = block.forward_stateful(x, x0, ssm_states=prev) + new_states.append(fstate) + si += 1 + else: + x = block(x, x0) + return x, new_states + + def forward_logits_stateful(self, input_ids: Tensor, layer_states: list | None = None): + """Forward returning (logits, new_layer_states) for stateful eval.""" + bsz, seqlen = input_ids.shape + x = self._embed(input_ids) + x, new_states = self._run_blocks_stateful(x, x, layer_states) + x = self.final_norm(x).reshape(-1, x.size(-1)) + w = self.tok_emb.weight if self.tie_embeddings else self.lm_head.weight + logits = self.logit_softcap * torch.tanh(F.linear(x, w) / self.logit_softcap) + return logits.reshape(bsz, seqlen, -1), new_states + + +# ----------------------------- +# TRAINING +# ----------------------------- + +def main() -> None: + global zeropower_via_newtonschulz5 + + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") + grad_accum_steps = int(os.environ.get("GRAD_ACCUM_STEPS", 8 // world_size)) + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + torch.backends.cuda.enable_flash_sdp(True) + torch.backends.cuda.enable_mem_efficient_sdp(False) + torch.backends.cuda.enable_math_sdp(False) + + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + + log0(code, console=False) + log0("=" * 100, console=False) + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + val_tokens = load_validation_tokens(args.val_files, args.train_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + + base_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + use_ortho_init=args.use_ortho_init, + mamba3_d_state=args.mamba3_d_state, mamba3_expand=args.mamba3_expand, + mamba3_headdim=args.mamba3_headdim, mamba3_chunk_size=args.mamba3_chunk_size, + mamba3_ngroups=args.mamba3_ngroups, mamba3_rope_fraction=args.mamba3_rope_fraction, + num_attn_layers=args.num_attn_layers, num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + ).to(device).bfloat16() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=False) + model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) \ + if distributed else compiled_model + + block_named_params = list(base_model.blocks.named_parameters()) + matrix_params = [ + p for name, p in block_named_params + if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + scalar_params = [ + p for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if hasattr(base_model, 'skip_weights') and base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + optimizer_tok = torch.optim.Adam( + tok_params, weight_decay=args.embed_wd, + betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, + ) + optimizer_muon = Muon( + matrix_params, lr=args.matrix_lr, momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, weight_decay=args.weight_decay, + eq_r=args.muon_eq_r, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.Adam( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, + ) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), eps=args.adam_eps, fused=True, + ) + optimizers.insert(1, optimizer_head) + + n_params = sum(p.numel() for p in base_model.parameters()) + log0(f"model_params:{n_params}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0(f"mode:mamba3_hybrid num_attn_layers:{args.num_attn_layers} attn_indices:{base_model.attn_indices}") + log0(f"ssd: d_state:{args.mamba3_d_state} expand:{args.mamba3_expand} headdim:{args.mamba3_headdim} ngroups:{args.mamba3_ngroups} rope_frac:{args.mamba3_rope_fraction}") + log0(f"attn: num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads} rope_base:{args.rope_base}") + log0(f"num_layers:{args.num_layers} mlp_mult:{args.mlp_mult}") + if args.muon_eq_r: + log0(f"muon_eq_r:enabled") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + log0(f"seed:{args.seed}") + + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + frac = max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + else: + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + frac = remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmdown_shape == "cosine" and frac < 1.0: + return 0.5 * (1.0 + math.cos(math.pi * (1.0 - frac))) + return frac + + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + if distributed: + model.require_backward_grad_sync = True + train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + + training_time_ms = 0.0 + stop_after_step: int | None = None + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + + ema_state: dict[str, Tensor] | None = None + if args.ema_enabled: + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + torch.cuda.synchronize() + t0 = time.perf_counter() + + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + + # Late QAT: trigger when lr_mul drops below threshold (SOTA approach) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold: + if any(m._qat_bits == 0 for m in base_model.modules() if isinstance(m, CastedLinear)): + qat_bits = args.quant_bits + embed_qat_bits = args.quant_bits_embed if args.quant_bits_embed > 0 else args.quant_bits + log0(f"late_qat:enabled bits={qat_bits} embed_bits={embed_qat_bits} at step {step} scale={scale:.4f}") + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._qat_bits = qat_bits + base_model._embed_qat_bits = embed_qat_bits + + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + if distributed: + model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + train_loss /= grad_accum_steps + + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + + step += 1 + + if ema_state is not None: + d = args.ema_decay + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(d).add_(t.detach().float(), alpha=1.0 - d) + + if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone().float() for name, t in base_model.state_dict().items()} + swa_count = 1 + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu().float() + swa_count += 1 + + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + tok_per_sec = step * args.train_batch_tokens / (approx_training_time_ms / 1000.0) + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms " + f"tok/s:{tok_per_sec:.0f}" + ) + + reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + if distributed and max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + + if args.sweep_mode: + log0("sweep_mode:exiting after training loop (skipping quantization/serialization)") + if distributed: + dist.destroy_process_group() + return + + for m in base_model.modules(): + if isinstance(m, CastedLinear): + m._qat_bits = 0 + base_model._embed_qat_bits = 0 + + if ema_state is not None: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + del ema_state + base_model.load_state_dict(avg_state, strict=True) + elif args.swa_enabled and swa_state is not None and swa_count > 1: + log0(f"swa:applying averaged {swa_count} checkpoints") + current_state = base_model.state_dict() + avg_state = { + name: (tensor / swa_count).to(dtype=current_state[name].dtype) + for name, tensor in swa_state.items() + } + del swa_state + base_model.load_state_dict(avg_state, strict=True) + + if master_process: + torch.save(base_model.state_dict(), "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + log0(f"Total submission size: {model_bytes + code_bytes} bytes") + + # Tear down DDP before GPTQ and eval: GPTQ hooks + TTT backward are incompatible + # with live NCCL context (CUDA illegal memory access in hook_fn on DDP teardown). + # TTT also doesn't parallelize across GPUs. Non-master ranks exit here. + if distributed: + dist.barrier() + dist.destroy_process_group() + distributed = False + torch.cuda.synchronize() # drain async NCCL event destruction before GPTQ touches memory + torch.cuda.empty_cache() + if not master_process: + return + + ssm_cfg = { + "model_dim": args.model_dim, + "mamba3_expand": args.mamba3_expand, + "mamba3_d_state": args.mamba3_d_state, + "mamba3_ngroups": args.mamba3_ngroups, + "mamba3_headdim": args.mamba3_headdim, + } + + if args.use_gptq: + t_gptq = time.perf_counter() + # AR GPTQ: torch.compile + triton.set_allocator(ContextVar.set) in _Mamba3Function + # corrupts the CUDA driver state, making train-data GPTQ crash (even in subprocesses + # on different GPUs). AR generation uses the compiled model (kernels cached), then + # Hessian collection with hooks also uses cached kernels — no autotuning needed. + # Cost: +5.5 mBPB vs train-data GPTQ (9.8 vs 4.3 gap), +220s eval time. + base_model.eval() + log0("gptq:generating autoregressive calibration data...") + ar_tokens = generate_autoregressive_calib( + base_model, device, num_seqs=args.gptq_num_seqs, + seq_len=args.train_seq_len, vocab_size=args.vocab_size, + ) + log0(f"gptq:generated {len(ar_tokens)} seqs in {time.perf_counter()-t_gptq:.1f}s") + log0("gptq:collecting hessians...") + hessians = collect_hessians_from_tokens(base_model, ar_tokens, device, gptq_embed=args.tie_embeddings) + del ar_tokens + log0(f"gptq:collected hessians for {len(hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + torch.cuda.empty_cache() + # Apply GPTQ to each quantizable layer in the state dict. + sd = base_model.state_dict() + gptq_sd = {} + for name, t in sd.items(): + t_cpu = t.detach().cpu() + H = hessians.get(name) + if H is not None and t_cpu.is_floating_point() and t_cpu.ndim == 2 and t_cpu.numel() > INT8_KEEP_FLOAT_MAX_NUMEL: + is_embed = "tok_emb" in name + bits = args.quant_bits_embed if (is_embed and args.quant_bits_embed > 0) else args.quant_bits + cr: int | Tensor = (1 << (bits - 1)) - 1 + # Per-row clip range for SSM dynamics rows: larger range = higher precision. + if args.quant_bits_ssm_dynamics > 0 and not is_embed: + dyn_mask = compute_ssm_dynamics_row_mask(name, t_cpu.shape[0], ssm_cfg) + if dyn_mask is not None: + dyn_cr = (1 << (args.quant_bits_ssm_dynamics - 1)) - 1 + cr_t = torch.full((t_cpu.shape[0],), float(cr), dtype=torch.float32) + cr_t[dyn_mask] = float(dyn_cr) + cr = cr_t + q, s = quantize_int6_gptq(t_cpu, hessian=H, clip_range=cr) + gptq_sd[name] = (q.float() * s.float()[:, None]).to(t_cpu.dtype) + else: + gptq_sd[name] = t_cpu + base_model.load_state_dict(gptq_sd, strict=True) + del hessians, gptq_sd + torch.cuda.empty_cache() + log0(f"gptq:quantization complete in {time.perf_counter()-t_gptq:.1f}s total") + + quant_obj, quant_stats = quantize_state_dict_int8( + base_model.state_dict(), quant_bits=args.quant_bits, + quant_bits_embed=args.quant_bits_embed, + search_clip=args.gptq_lite, + quant_bits_ssm_dynamics=args.quant_bits_ssm_dynamics, ssm_cfg=ssm_cfg, + ) + # Selective ±1 pruning: zero out least-impactful ±1 quantized values to fit target size + if args.use_lzma and args.target_mb > 0 and master_process: + target_bytes = int(args.target_mb * 1024 * 1024) + code_bytes_est = len(code.encode("utf-8")) + ones_info = [] + for name, q in quant_obj["quantized"].items(): + s = quant_obj["scales"].get(name) + if s is None or s.ndim == 0: + continue + ones_mask = (q.abs() == 1) + if ones_mask.any(): + row_idx = torch.arange(q.shape[0]).unsqueeze(1).expand_as(q)[ones_mask] + flat_idx = torch.arange(q.numel()).reshape(q.shape)[ones_mask] + errors = s.float()[row_idx].pow(2) + for fi, err in zip(flat_idx.tolist(), errors.tolist()): + ones_info.append((name, fi, err)) + if ones_info: + ones_info.sort(key=lambda x: x[2]) + def _compress_for_prune(raw): + return lzma.compress(raw, preset=9) + def _try_prune(n): + tmp_q = {k: v.clone() for k, v in quant_obj["quantized"].items()} + for i in range(min(n, len(ones_info))): + tmp_q[ones_info[i][0]].view(-1)[ones_info[i][1]] = 0 + tmp_obj = {**quant_obj, "quantized": tmp_q} + buf = io.BytesIO() + torch.save(tmp_obj, buf) + return len(_compress_for_prune(buf.getvalue())) + code_bytes_est, tmp_q + no_sz, _ = _try_prune(0) + log0(f"selective_prune: {len(ones_info)} ±1 candidates, unpruned={no_sz/(1024*1024):.2f}MB target={args.target_mb}MB") + if no_sz <= target_bytes: + log0("selective_prune: already fits, no pruning needed") + else: + full_sz, _ = _try_prune(len(ones_info)) + log0(f"selective_prune: full ±1 prune={full_sz/(1024*1024):.2f}MB") + if full_sz > target_bytes: + log0("selective_prune: even full prune not enough, applying all") + _, quant_obj["quantized"] = _try_prune(len(ones_info)) + else: + # Linear interpolation instead of binary search (LZMA-9 is too slow for 22 iterations) + frac = (no_sz - target_bytes) / max(no_sz - full_sz, 1) + est = int(frac * len(ones_info) * 1.1) + 1 # 10% margin + est = min(est, len(ones_info)) + log0(f"selective_prune: pruning {est}/{len(ones_info)} ±1 values ({100*est/len(ones_info):.1f}%) to fit {args.target_mb}MB") + _, quant_obj["quantized"] = _try_prune(est) + + quant_buf = io.BytesIO() + torch.save(quant_obj, quant_buf) + quant_raw = quant_buf.getvalue() + if args.use_lzma: + quant_blob = lzma.compress(quant_raw, preset=9) + compress_fmt = "lzma-9" + else: + quant_blob = zlib.compress(quant_raw, level=9) + compress_fmt = "zlib-9" + quant_raw_bytes = len(quant_raw) + quant_filename = f"final_model.int{args.quant_bits}.ptz" + if master_process: + with open(quant_filename, "wb") as f: + f.write(quant_blob) + quant_file_bytes = os.path.getsize(quant_filename) + code_bytes = len(code.encode("utf-8")) + ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) + log0( + f"Serialized model int{args.quant_bits}+{compress_fmt}: {quant_file_bytes} bytes " + f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + ) + log0(f"Total submission size int{args.quant_bits}+{compress_fmt}: {quant_file_bytes + code_bytes} bytes") + + if distributed: + dist.barrier() + with open(quant_filename, "rb") as f: + quant_blob_disk = f.read() + if args.use_lzma: + quant_decompressed = lzma.decompress(quant_blob_disk) + else: + quant_decompressed = zlib.decompress(quant_blob_disk) + quant_state = torch.load(io.BytesIO(quant_decompressed), map_location="cpu") + base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + + num_frozen = min(args.ttt_freeze_blocks, len(base_model.blocks)) + for i in range(num_frozen): + for p in base_model.blocks[i].parameters(): + p.requires_grad_(False) + ttt_params = [p for p in base_model.parameters() if p.requires_grad] + if args.ttt_optimizer == "adamw": + ttt_opt = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=0.01) + else: + ttt_opt = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + ttt_seq_len = args.train_seq_len + + total_tokens_val = val_tokens.numel() - 1 + total_seqs = total_tokens_val // ttt_seq_len + total_chunks = (total_seqs + args.eval_batch_seqs - 1) // args.eval_batch_seqs + + ttt_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + ttt_token_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_byte_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_step = 0 + + log0(f"ttt:starting optimizer={args.ttt_optimizer} lr={args.ttt_lr} freeze_blocks={num_frozen} epochs={args.ttt_epochs} chunks={total_chunks}") + + for seq_idx in range(0, total_seqs, args.eval_batch_seqs): + batch_end = min(seq_idx + args.eval_batch_seqs, total_seqs) + bsz = batch_end - seq_idx + raw_start = seq_idx * ttt_seq_len + raw_end = batch_end * ttt_seq_len + 1 + chunk = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64) + x = chunk[:-1].reshape(bsz, ttt_seq_len) + y = chunk[1:].reshape(bsz, ttt_seq_len) + + base_model.eval() + with torch.no_grad(): + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), y.reshape(-1), reduction="none", + ) + ttt_loss_sum += nll.to(torch.float64).sum() + n_tokens = float(y.numel()) + ttt_token_count += n_tokens + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + tbytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + tbytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + ttt_byte_count += tbytes.to(torch.float64).sum() + + base_model.train() + for _epoch in range(args.ttt_epochs): + ttt_opt.zero_grad() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + ttt_opt.step() + ttt_step += 1 + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(ttt_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(ttt_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(ttt_byte_count, op=dist.ReduceOp.SUM) + + q_val_loss = (ttt_loss_sum / ttt_token_count).item() + bits_per_token = q_val_loss / math.log(2.0) + tokens_per_byte = ttt_token_count.item() / ttt_byte_count.item() + q_val_bpb = float(bits_per_token * tokens_per_byte) + eval_mode = "online_ttt" + log0(f"ttt:completed steps:{ttt_step} time:{time.perf_counter() - t_ttt:.1f}s") + + else: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + + torch.cuda.synchronize() + t_qeval = time.perf_counter() + + if not args.ttt_enabled: + if args.eval_overlap > 0: + q_val_loss, q_val_bpb = eval_val_stateful_overlap( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + eval_mode = "stateful_overlap" + elif args.eval_stride > 0 and args.eval_stride < args.train_seq_len: + q_val_loss, q_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + eval_mode = "sliding" + else: + q_val_loss, q_val_bpb = eval_val( + args, model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + eval_mode = "standard" + torch.cuda.synchronize() + log0( + f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_mode:{eval_mode} eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + +if __name__ == "__main__": + main() diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed1337.log b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed1337.log new file mode 100644 index 0000000000..e012fa1dd9 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed1337.log @@ -0,0 +1,101 @@ +W0422 12:35:02.046000 2230 torch/distributed/run.py:803] +W0422 12:35:02.046000 2230 torch/distributed/run.py:803] ***************************************** +W0422 12:35:02.046000 2230 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0422 12:35:02.046000 2230 torch/distributed/run.py:803] ***************************************** +logs/fe9b563e-74ae-4968-9f1b-cea617e84dd1.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_8192_bpe.model +train_loader:dataset:fineweb10B_sp8192 train_shards:128 +val_loader:shards pattern=./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin tokens:40538112 +model_params:25159984 +world_size:8 grad_accum_steps:1 +mode:mamba3_hybrid num_attn_layers:2 attn_indices:[2, 5] +ssd: d_state:64 expand:2.0 headdim:64 ngroups:1 rope_frac:0.5 outproj_norm:False in_proj_rank:0 +attn: num_heads:8 num_kv_heads:4 rope_base:10000.0 +num_layers:7 mlp_mult:3.0 +muon_eq_r:enabled +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.02 +train_batch_tokens:1048576 train_seq_len:4096 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:1337 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:9.0136 val_bpb:3.4894 train_time:0ms step_avg:0.01ms +step:1/20000 train_loss:9.0133 train_time:75ms step_avg:75.29ms tok/s:13928042 +step:2/20000 train_loss:8.5278 train_time:167ms step_avg:83.72ms tok/s:12524551 +step:3/20000 train_loss:9.1236 train_time:279ms step_avg:93.10ms tok/s:11263485 +step:4/20000 train_loss:10.0854 train_time:392ms step_avg:98.05ms tok/s:10694320 +step:5/20000 train_loss:9.2906 train_time:505ms step_avg:100.95ms tok/s:10387440 +step:6/20000 train_loss:8.2810 train_time:617ms step_avg:102.77ms tok/s:10203020 +step:7/20000 train_loss:7.9847 train_time:730ms step_avg:104.23ms tok/s:10059800 +step:8/20000 train_loss:7.5765 train_time:842ms step_avg:105.21ms tok/s:9966173 +step:9/20000 train_loss:7.3217 train_time:954ms step_avg:105.99ms tok/s:9893557 +step:10/20000 train_loss:7.0980 train_time:1069ms step_avg:106.94ms tok/s:9805033 +step:200/20000 train_loss:3.7201 train_time:22682ms step_avg:113.41ms tok/s:9246057 +step:400/20000 train_loss:3.4413 train_time:45452ms step_avg:113.63ms tok/s:9228070 +step:600/20000 train_loss:3.4293 train_time:68305ms step_avg:113.84ms tok/s:9210851 +step:800/20000 train_loss:3.2558 train_time:91302ms step_avg:114.13ms tok/s:9187785 +step:1000/20000 train_loss:3.2840 train_time:114203ms step_avg:114.20ms tok/s:9181659 +step:1000/20000 val_loss:3.1969 val_bpb:1.2376 train_time:114263ms step_avg:114.26ms +step:1200/20000 train_loss:3.0975 train_time:137073ms step_avg:114.23ms tok/s:9179712 +step:1400/20000 train_loss:3.2329 train_time:159947ms step_avg:114.25ms tok/s:9178059 +step:1600/20000 train_loss:3.0392 train_time:182816ms step_avg:114.26ms tok/s:9177095 +step:1800/20000 train_loss:3.1350 train_time:205656ms step_avg:114.25ms tok/s:9177645 +step:2000/20000 train_loss:3.2366 train_time:228485ms step_avg:114.24ms tok/s:9178528 +step:2000/20000 val_loss:3.1055 val_bpb:1.2022 train_time:228545ms step_avg:114.27ms +step:2200/20000 train_loss:3.1134 train_time:251355ms step_avg:114.25ms tok/s:9177727 +step:2400/20000 train_loss:3.0939 train_time:274188ms step_avg:114.24ms tok/s:9178316 +step:2600/20000 train_loss:3.0681 train_time:297006ms step_avg:114.23ms tok/s:9179276 +step:2800/20000 train_loss:3.0348 train_time:319832ms step_avg:114.23ms tok/s:9179853 +step:3000/20000 train_loss:3.1452 train_time:342640ms step_avg:114.21ms tok/s:9180864 +step:3000/20000 val_loss:3.0593 val_bpb:1.1843 train_time:342700ms step_avg:114.23ms +step:3200/20000 train_loss:3.1345 train_time:365443ms step_avg:114.20ms tok/s:9181863 +step:3400/20000 train_loss:3.0950 train_time:388249ms step_avg:114.19ms tok/s:9182666 +step:3600/20000 train_loss:2.9872 train_time:411051ms step_avg:114.18ms tok/s:9183473 +step:3800/20000 train_loss:3.1372 train_time:433849ms step_avg:114.17ms tok/s:9184266 +step:4000/20000 train_loss:2.7386 train_time:456639ms step_avg:114.16ms tok/s:9185156 +step:4000/20000 val_loss:2.9973 val_bpb:1.1603 train_time:456700ms step_avg:114.17ms +step:4200/20000 train_loss:2.9587 train_time:479486ms step_avg:114.16ms tok/s:9184880 +step:4400/20000 train_loss:3.0956 train_time:502520ms step_avg:114.21ms tok/s:9181201 +step:4600/20000 train_loss:3.0484 train_time:525439ms step_avg:114.23ms tok/s:9179844 +step:4800/20000 train_loss:2.9379 train_time:548360ms step_avg:114.24ms tok/s:9178577 +late_qat:enabled bits=6 embed_bits=8 at step 4863 scale=0.1497 +step:5000/20000 train_loss:2.9755 train_time:578898ms step_avg:115.78ms tok/s:9056657 +step:5000/20000 val_loss:2.9587 val_bpb:1.1454 train_time:578903ms step_avg:115.78ms +step:5186/20000 val_loss:2.9420 val_bpb:1.1389 train_time:600168ms step_avg:115.73ms +stopping_early: wallclock_cap train_time:600168ms step:5186/20000 +peak memory allocated: 26118 MiB reserved: 28778 MiB +ema:applying EMA weights +Serialized model: 92283538 bytes +Code size: 116783 bytes +Total submission size: 92400321 bytes +gptq:generating autoregressive calibration data... +gptq:generated 32 seqs in 200.5s +gptq:collecting hessians... +gptq:collected hessians for 33 layers in 200.9s +gptq:quantization complete in 225.3s total +selective_prune: 3297212 ±1 candidates, unpruned=15.62MB target=15.25MB +selective_prune: full ±1 prune=14.65MB +selective_prune: pruning 1369704/3297212 ±1 values (41.5%) to fit 15.25MB +Serialized model int6+lzma-9: 15813408 bytes (payload:25318208 raw_torch:25355471 payload_ratio:3.64x) +Total submission size int6+lzma-9: 15930191 bytes +ttt:starting optimizer=sgd lr=0.01 freeze_blocks=0 epochs=2 chunks=310 +ttt:completed steps:620 time:132.0s +final_int8_zlib_roundtrip val_loss:2.9553 val_bpb:1.1441 eval_mode:online_ttt eval_time:0ms +final_int8_zlib_roundtrip_exact val_loss:2.95530257 val_bpb:1.14408575 diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed2025.log b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed2025.log new file mode 100644 index 0000000000..fb8de2d4ad --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed2025.log @@ -0,0 +1,102 @@ +W0422 13:39:29.214000 60871 torch/distributed/run.py:803] +W0422 13:39:29.214000 60871 torch/distributed/run.py:803] ***************************************** +W0422 13:39:29.214000 60871 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0422 13:39:29.214000 60871 torch/distributed/run.py:803] ***************************************** +logs/b2d8bc48-ae9c-42ea-8dae-8ac98b7f08c2.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_8192_bpe.model +train_loader:dataset:fineweb10B_sp8192 train_shards:128 +val_loader:shards pattern=./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin tokens:40538112 +model_params:25159984 +world_size:8 grad_accum_steps:1 +mode:mamba3_hybrid num_attn_layers:2 attn_indices:[2, 5] +ssd: d_state:64 expand:2.0 headdim:64 ngroups:1 rope_frac:0.5 +attn: num_heads:8 num_kv_heads:4 rope_base:10000.0 +num_layers:7 mlp_mult:3.0 +muon_eq_r:enabled +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.02 +train_batch_tokens:1048576 train_seq_len:4096 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:2025 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:9.0132 val_bpb:3.4893 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:9.0128 train_time:75ms step_avg:75.23ms tok/s:13937884 +step:2/20000 train_loss:8.7239 train_time:168ms step_avg:83.80ms tok/s:12512523 +step:3/20000 train_loss:10.0956 train_time:280ms step_avg:93.49ms tok/s:11215337 +step:4/20000 train_loss:10.0427 train_time:393ms step_avg:98.15ms tok/s:10683432 +step:5/20000 train_loss:9.1297 train_time:505ms step_avg:100.90ms tok/s:10392196 +step:6/20000 train_loss:8.3513 train_time:618ms step_avg:102.92ms tok/s:10188546 +step:7/20000 train_loss:7.9340 train_time:731ms step_avg:104.49ms tok/s:10035242 +step:8/20000 train_loss:7.4585 train_time:844ms step_avg:105.55ms tok/s:9934793 +step:9/20000 train_loss:7.1618 train_time:958ms step_avg:106.41ms tok/s:9853687 +step:10/20000 train_loss:7.0469 train_time:1071ms step_avg:107.07ms tok/s:9793298 +step:200/20000 train_loss:3.7183 train_time:22727ms step_avg:113.63ms tok/s:9227687 +step:400/20000 train_loss:3.4341 train_time:45536ms step_avg:113.84ms tok/s:9210922 +step:600/20000 train_loss:3.4274 train_time:68423ms step_avg:114.04ms tok/s:9194879 +step:800/20000 train_loss:3.2566 train_time:91337ms step_avg:114.17ms tok/s:9184207 +step:1000/20000 train_loss:3.2865 train_time:114226ms step_avg:114.23ms tok/s:9179812 +step:1000/20000 val_loss:3.1965 val_bpb:1.2375 train_time:114286ms step_avg:114.29ms +step:1200/20000 train_loss:3.0974 train_time:137098ms step_avg:114.25ms tok/s:9178070 +step:1400/20000 train_loss:3.2344 train_time:159938ms step_avg:114.24ms tok/s:9178596 +step:1600/20000 train_loss:3.0414 train_time:182767ms step_avg:114.23ms tok/s:9179556 +step:1800/20000 train_loss:3.1345 train_time:205587ms step_avg:114.21ms tok/s:9180731 +step:2000/20000 train_loss:3.2392 train_time:228406ms step_avg:114.20ms tok/s:9181675 +step:2000/20000 val_loss:3.1068 val_bpb:1.2027 train_time:228466ms step_avg:114.23ms +step:2200/20000 train_loss:3.1159 train_time:251287ms step_avg:114.22ms tok/s:9180223 +step:2400/20000 train_loss:3.0969 train_time:274110ms step_avg:114.21ms tok/s:9180934 +step:2600/20000 train_loss:3.0701 train_time:296933ms step_avg:114.20ms tok/s:9181533 +step:2800/20000 train_loss:3.0408 train_time:319754ms step_avg:114.20ms tok/s:9182107 +step:3000/20000 train_loss:3.1474 train_time:342589ms step_avg:114.20ms tok/s:9182235 +step:3000/20000 val_loss:3.0622 val_bpb:1.1855 train_time:342649ms step_avg:114.22ms +step:3200/20000 train_loss:3.1335 train_time:365401ms step_avg:114.19ms tok/s:9182911 +step:3400/20000 train_loss:3.0982 train_time:388218ms step_avg:114.18ms tok/s:9183402 +step:3600/20000 train_loss:2.9885 train_time:411036ms step_avg:114.18ms tok/s:9183794 +step:3800/20000 train_loss:3.1435 train_time:433831ms step_avg:114.17ms tok/s:9184651 +step:4000/20000 train_loss:2.7420 train_time:456624ms step_avg:114.16ms tok/s:9185456 +step:4000/20000 val_loss:2.9997 val_bpb:1.1613 train_time:456684ms step_avg:114.17ms +step:4200/20000 train_loss:2.9621 train_time:479479ms step_avg:114.16ms tok/s:9185018 +step:4400/20000 train_loss:3.0991 train_time:502492ms step_avg:114.20ms tok/s:9181714 +step:4600/20000 train_loss:3.0465 train_time:525447ms step_avg:114.23ms tok/s:9179705 +step:4800/20000 train_loss:2.9416 train_time:548358ms step_avg:114.24ms tok/s:9178617 +late_qat:enabled bits=6 embed_bits=8 at step 4863 scale=0.1497 +step:5000/20000 train_loss:3.0226 train_time:574647ms step_avg:114.93ms tok/s:9123651 +step:5000/20000 val_loss:3.0046 val_bpb:1.1632 train_time:574654ms step_avg:114.93ms +step:5200/20000 train_loss:3.0532 train_time:597587ms step_avg:114.92ms tok/s:9124361 +step:5222/20000 val_loss:2.9693 val_bpb:1.1495 train_time:600097ms step_avg:114.92ms +stopping_early: wallclock_cap train_time:600097ms step:5222/20000 +peak memory allocated: 26117 MiB reserved: 28752 MiB +ema:applying EMA weights +Serialized model: 92283474 bytes +Code size: 97714 bytes +Total submission size: 92381188 bytes +gptq:generating autoregressive calibration data... +gptq:generated 32 seqs in 192.8s +gptq:collecting hessians... +gptq:collected hessians for 33 layers in 193.2s +gptq:quantization complete in 217.4s total +selective_prune: 3309159 ±1 candidates, unpruned=15.57MB target=15.25MB +selective_prune: full ±1 prune=14.59MB +selective_prune: pruning 1179333/3309159 ±1 values (35.6%) to fit 15.25MB +Serialized model int6+lzma-9: 15858300 bytes (payload:25318208 raw_torch:25355471 payload_ratio:3.64x) +Total submission size int6+lzma-9: 15956014 bytes +ttt:starting optimizer=sgd lr=0.01 freeze_blocks=0 epochs=2 chunks=310 +ttt:completed steps:620 time:132.2s +final_int8_zlib_roundtrip val_loss:2.9624 val_bpb:1.1468 eval_mode:online_ttt eval_time:0ms +final_int8_zlib_roundtrip_exact val_loss:2.96243731 val_bpb:1.14684782 diff --git a/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed42.log b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed42.log new file mode 100644 index 0000000000..6a4927c451 --- /dev/null +++ b/records/track_non_record_16mb/2026-04-22_Mamba3Hybrid_MultiEpochTTT_DynamicsProtect/train_seed42.log @@ -0,0 +1,102 @@ +W0422 13:19:55.562000 54666 torch/distributed/run.py:803] +W0422 13:19:55.562000 54666 torch/distributed/run.py:803] ***************************************** +W0422 13:19:55.562000 54666 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. +W0422 13:19:55.562000 54666 torch/distributed/run.py:803] ***************************************** +logs/3fb53079-016d-4f9f-a289-b81ea3bc0e5b.txt +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_8192_bpe.model +train_loader:dataset:fineweb10B_sp8192 train_shards:128 +val_loader:shards pattern=./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin tokens:40538112 +model_params:25159984 +world_size:8 grad_accum_steps:1 +mode:mamba3_hybrid num_attn_layers:2 attn_indices:[2, 5] +ssd: d_state:64 expand:2.0 headdim:64 ngroups:1 rope_frac:0.5 +attn: num_heads:8 num_kv_heads:4 rope_base:10000.0 +num_layers:7 mlp_mult:3.0 +muon_eq_r:enabled +tie_embeddings:True embed_lr:0.05 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.02 +train_batch_tokens:1048576 train_seq_len:4096 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +seed:42 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +step:0/20000 val_loss:9.0140 val_bpb:3.4896 train_time:0ms step_avg:0.02ms +step:1/20000 train_loss:9.0131 train_time:75ms step_avg:75.40ms tok/s:13907658 +step:2/20000 train_loss:8.5292 train_time:168ms step_avg:83.96ms tok/s:12489144 +step:3/20000 train_loss:10.2977 train_time:280ms step_avg:93.44ms tok/s:11222076 +step:4/20000 train_loss:8.9720 train_time:392ms step_avg:98.10ms tok/s:10689374 +step:5/20000 train_loss:8.3807 train_time:506ms step_avg:101.18ms tok/s:10363198 +step:6/20000 train_loss:8.4818 train_time:618ms step_avg:103.03ms tok/s:10177510 +step:7/20000 train_loss:8.2917 train_time:731ms step_avg:104.37ms tok/s:10046790 +step:8/20000 train_loss:7.8515 train_time:843ms step_avg:105.41ms tok/s:9947804 +step:9/20000 train_loss:7.3592 train_time:956ms step_avg:106.23ms tok/s:9870478 +step:10/20000 train_loss:7.1152 train_time:1069ms step_avg:106.87ms tok/s:9811657 +step:200/20000 train_loss:3.7471 train_time:22705ms step_avg:113.53ms tok/s:9236391 +step:400/20000 train_loss:3.4519 train_time:45539ms step_avg:113.85ms tok/s:9210423 +step:600/20000 train_loss:3.4338 train_time:68437ms step_avg:114.06ms tok/s:9193032 +step:800/20000 train_loss:3.2590 train_time:91358ms step_avg:114.20ms tok/s:9182080 +step:1000/20000 train_loss:3.2867 train_time:114266ms step_avg:114.27ms tok/s:9176597 +step:1000/20000 val_loss:3.2023 val_bpb:1.2397 train_time:114326ms step_avg:114.33ms +step:1200/20000 train_loss:3.1029 train_time:137154ms step_avg:114.29ms tok/s:9174312 +step:1400/20000 train_loss:3.2379 train_time:160031ms step_avg:114.31ms tok/s:9173272 +step:1600/20000 train_loss:3.0436 train_time:182893ms step_avg:114.31ms tok/s:9173218 +step:1800/20000 train_loss:3.1381 train_time:205732ms step_avg:114.30ms tok/s:9174240 +step:2000/20000 train_loss:3.2378 train_time:228568ms step_avg:114.28ms tok/s:9175194 +step:2000/20000 val_loss:3.1102 val_bpb:1.2041 train_time:228627ms step_avg:114.31ms +step:2200/20000 train_loss:3.1182 train_time:251447ms step_avg:114.29ms tok/s:9174360 +step:2400/20000 train_loss:3.1014 train_time:274266ms step_avg:114.28ms tok/s:9175711 +step:2600/20000 train_loss:3.0719 train_time:297081ms step_avg:114.26ms tok/s:9176960 +step:2800/20000 train_loss:3.0409 train_time:319901ms step_avg:114.25ms tok/s:9177871 +step:3000/20000 train_loss:3.1521 train_time:342709ms step_avg:114.24ms tok/s:9179005 +step:3000/20000 val_loss:3.0652 val_bpb:1.1866 train_time:342770ms step_avg:114.26ms +step:3200/20000 train_loss:3.1367 train_time:365520ms step_avg:114.23ms tok/s:9179910 +step:3400/20000 train_loss:3.1031 train_time:388354ms step_avg:114.22ms tok/s:9180184 +step:3600/20000 train_loss:2.9948 train_time:411167ms step_avg:114.21ms tok/s:9180870 +step:3800/20000 train_loss:3.1441 train_time:433960ms step_avg:114.20ms tok/s:9181920 +step:4000/20000 train_loss:2.7445 train_time:456769ms step_avg:114.19ms tok/s:9182553 +step:4000/20000 val_loss:3.0026 val_bpb:1.1624 train_time:456829ms step_avg:114.21ms +step:4200/20000 train_loss:2.9601 train_time:479612ms step_avg:114.19ms tok/s:9182472 +step:4400/20000 train_loss:3.0967 train_time:502624ms step_avg:114.23ms tok/s:9179291 +step:4600/20000 train_loss:3.0519 train_time:525517ms step_avg:114.24ms tok/s:9178484 +step:4800/20000 train_loss:2.9411 train_time:548491ms step_avg:114.27ms tok/s:9176393 +late_qat:enabled bits=6 embed_bits=8 at step 4862 scale=0.1496 +step:5000/20000 train_loss:3.0014 train_time:574653ms step_avg:114.93ms tok/s:9123551 +step:5000/20000 val_loss:2.9854 val_bpb:1.1557 train_time:574662ms step_avg:114.93ms +step:5200/20000 train_loss:3.0411 train_time:597553ms step_avg:114.91ms tok/s:9124866 +step:5222/20000 val_loss:2.9609 val_bpb:1.1462 train_time:600060ms step_avg:114.91ms +stopping_early: wallclock_cap train_time:600060ms step:5222/20000 +peak memory allocated: 26117 MiB reserved: 28752 MiB +ema:applying EMA weights +Serialized model: 92283474 bytes +Code size: 97714 bytes +Total submission size: 92381188 bytes +gptq:generating autoregressive calibration data... +gptq:generated 32 seqs in 196.8s +gptq:collecting hessians... +gptq:collected hessians for 33 layers in 197.2s +gptq:quantization complete in 221.4s total +selective_prune: 3316131 ±1 candidates, unpruned=15.57MB target=15.25MB +selective_prune: full ±1 prune=14.58MB +selective_prune: pruning 1184547/3316131 ±1 values (35.7%) to fit 15.25MB +Serialized model int6+lzma-9: 15844420 bytes (payload:25318208 raw_torch:25355471 payload_ratio:3.64x) +Total submission size int6+lzma-9: 15942134 bytes +ttt:starting optimizer=sgd lr=0.01 freeze_blocks=0 epochs=2 chunks=310 +ttt:completed steps:620 time:132.0s +final_int8_zlib_roundtrip val_loss:2.9603 val_bpb:1.1460 eval_mode:online_ttt eval_time:0ms +final_int8_zlib_roundtrip_exact val_loss:2.96033913 val_bpb:1.14603555