diff --git a/train_gpt.py b/train_gpt.py index 651beb2b89..a75a4b662d 100644 --- a/train_gpt.py +++ b/train_gpt.py @@ -1,14 +1,8 @@ -""" -The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. - -Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. -""" - from __future__ import annotations - import copy import glob import io +import lzma import math import os import random @@ -18,164 +12,285 @@ import uuid import zlib from pathlib import Path - +try: + import zstandard + _COMPRESSOR = "zstd" +except ImportError: + _COMPRESSOR = "zlib" import numpy as np +try: + from numba import njit + _HAS_NUMBA = True +except ImportError: + _HAS_NUMBA = False 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 -# ----------------------------- -# Default Simple Baseline run: -# - 9 transformer blocks at width 512 -# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion -# - vocab size 1024, sequence length 1024, tied embeddings -# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap - +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + try: + from flash_attn import flash_attn_func as flash_attn_3_func + except ImportError: + # SDPA fallback for cards without flash_attn (e.g., Blackwell) + def flash_attn_3_func(q, k, v, causal=True): + # q: (B, T, Hq, D), k: (B, T, Hk, D), v: (B, T, Hk, D) + # GQA: repeat KV heads to match Q heads + Hq, Hk = q.size(2), k.size(2) + if Hq != Hk: + rep = Hq // Hk + k = k.repeat_interleave(rep, dim=2) + v = v.repeat_interleave(rep, dim=2) + # SDPA expects (B, H, T, D) + q = q.transpose(1, 2) + k = k.transpose(1, 2) + v = v.transpose(1, 2) + y = torch.nn.functional.scaled_dot_product_attention(q, k, v, is_causal=causal) + return y.transpose(1, 2) # back to (B, T, H, D) class Hyperparameters: - # Data paths are shard globs produced by the existing preprocessing pipeline. data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") train_files = os.path.join(data_path, "fineweb_train_*.bin") val_files = os.path.join(data_path, "fineweb_val_*.bin") tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model") run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) seed = int(os.environ.get("SEED", 1337)) - - # 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. + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) iterations = int(os.environ.get("ITERATIONS", 20000)) - warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) - train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) - train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) - - # Model shape. vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) - num_layers = int(os.environ.get("NUM_LAYERS", 9)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) model_dim = int(os.environ.get("MODEL_DIM", 512)) num_heads = int(os.environ.get("NUM_HEADS", 8)) - mlp_mult = int(os.environ.get("MLP_MULT", 2)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) - - # Optimizer hyperparameters. 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_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) - matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) - scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) - muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) - muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) - muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) beta1 = float(os.environ.get("BETA1", 0.9)) beta2 = float(os.environ.get("BETA2", 0.95)) adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) - grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0)) - -# ----------------------------- -# MUON OPTIMIZER -# ----------------------------- -# -# As borrowed from modded-nanogpt -# Background on Muon: https://kellerjordan.github.io/posts/muon/ - -def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: - # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. - # Muon uses this to normalize matrix-shaped gradients before applying them. + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "0"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 4)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.95)) + ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) X = G.bfloat16() - X /= X.norm() + eps - transposed = G.size(0) > G.size(1) + transposed = X.size(-2) > X.size(-1) if transposed: - X = X.T + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) for _ in range(steps): - A = X @ X.T - B = b * A + c * A @ A + A = X @ X.mT + B = b * A + c * (A @ A) X = a * X + B @ X - return X.T if transposed else X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X +# --- Parallel Muon optimizer --- class Muon(torch.optim.Optimizer): - def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): super().__init__( params, - dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) @torch.no_grad() def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" 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 + if not self._built: + self._build() for group in self.param_groups: - params = group["params"] - if not params: - continue lr = group["lr"] momentum = group["momentum"] backend_steps = group["backend_steps"] nesterov = group["nesterov"] - - total_params = sum(int(p.numel()) for p in params) - updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) - - curr = 0 - for i, p in enumerate(params): - if i % world_size == rank and p.grad is not None: - g = p.grad + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() state = self.state[p] if "momentum_buffer" not in state: state["momentum_buffer"] = torch.zeros_like(g) buf = state["momentum_buffer"] - buf.mul_(momentum).add_(g) - if nesterov: - g = g.add(buf, alpha=momentum) - g = zeropower_via_newtonschulz5(g, steps=backend_steps) - # Scale correction from Muon reference implementations. - 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: - g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) - p.add_(g, alpha=-lr) - curr += p.numel() + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures return loss - -# ----------------------------- -# TOKENIZER-AGNOSTIC EVALUATION SETUP -# ----------------------------- -# -# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. -# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. -# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. -# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. +# --- Tokenizer evaluation helpers --- def build_sentencepiece_luts( sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device @@ -193,7 +308,7 @@ def build_sentencepiece_luts( base_bytes_np[token_id] = 1 continue piece = sp.id_to_piece(token_id) - if piece.startswith("▁"): + if piece.startswith("\u2581"): has_leading_space_np[token_id] = True piece = piece[1:] base_bytes_np[token_id] = len(piece.encode("utf-8")) @@ -202,20 +317,15 @@ def build_sentencepiece_luts( 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}") - # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. 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, @@ -227,34 +337,32 @@ def eval_val( base_bytes_lut: Tensor, has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, ) -> tuple[float, float]: - # Validation computes two metrics: - # - val_loss: token cross-entropy (natural log) - # - val_bpb: tokenizer-agnostic compression metric used by the challenge + seq_len = eval_seq_len or args.train_seq_len local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) - if local_batch_tokens < args.train_seq_len: + if local_batch_tokens < seq_len: raise ValueError( "VAL_BATCH_SIZE must provide at least one sequence per rank; " f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " - f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}" + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" ) - local_batch_seqs = local_batch_tokens // args.train_seq_len - total_seqs = (val_tokens.numel() - 1) // args.train_seq_len + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len seq_start = (total_seqs * rank) // world_size seq_end = (total_seqs * (rank + 1)) // world_size val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) val_token_count = torch.zeros((), device=device, dtype=torch.float64) val_byte_count = torch.zeros((), device=device, dtype=torch.float64) - model.eval() with torch.inference_mode(): for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) - raw_start = batch_seq_start * args.train_seq_len - raw_end = batch_seq_end * args.train_seq_len + 1 + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) - x = local[:-1].reshape(-1, args.train_seq_len) - y = local[1:].reshape(-1, args.train_seq_len) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): batch_loss = model(x, y).detach() batch_token_count = float(y.numel()) @@ -265,31 +373,23 @@ def eval_val( 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) -# ----------------------------- -# POST-TRAINING QUANTIZATION -# ----------------------------- -# -# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. -# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. -# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. +# --- Quantization helpers --- CONTROL_TENSOR_NAME_PATTERNS = tuple( pattern for pattern in os.environ.get( "CONTROL_TENSOR_NAME_PATTERNS", - "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", ).split(",") if pattern ) @@ -306,10 +406,8 @@ def eval_val( INT8_PER_ROW_SCALE_DTYPE = torch.float16 INT8_CLIP_PERCENTILE = 99.99984 INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 - def tensor_nbytes(t: Tensor) -> int: return int(t.numel()) * int(t.element_size()) - def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): return t.float().contiguous() @@ -317,12 +415,9 @@ def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, s passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() return t - def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: t32 = t.float() if t32.ndim == 2: - # Matrices get one scale per row, which usually tracks output-channel - # ranges much better than a single tensor-wide scale. clip_abs = ( torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) if t32.numel() @@ -332,19 +427,11 @@ def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() - - # Vectors / scalars use a simpler per-tensor scale. clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() return q, scale - def quantize_state_dict_int8(state_dict: dict[str, Tensor]): - # Single supported clean-script export format: - # - per-row int8 for 2D float tensors - # - per-tensor int8 for other float tensors - # - exact passthrough for non-floats - # - passthrough for small float tensors, stored as fp16 to save bytes quantized: dict[str, Tensor] = {} scales: dict[str, Tensor] = {} dtypes: dict[str, str] = {} @@ -355,27 +442,21 @@ def quantize_state_dict_int8(state_dict: dict[str, Tensor]): ("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 - - # Small float tensors are cheap enough to keep directly. We still downcast - # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: kept = keep_float_tensor(name, t, passthrough_orig_dtypes) passthrough[name] = kept stats["int8_payload_bytes"] += tensor_nbytes(kept) continue - stats["num_float_tensors"] += 1 q, s = quantize_float_tensor(t) if s.ndim > 0: @@ -384,7 +465,6 @@ def quantize_state_dict_int8(state_dict: dict[str, Tensor]): 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, @@ -397,7 +477,6 @@ def quantize_state_dict_int8(state_dict: dict[str, Tensor]): 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", {}) @@ -407,13 +486,11 @@ def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: s = obj["scales"][name] if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: s = s.to(dtype=torch.float32) - # Broadcast the saved row scale back across trailing dimensions. out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() else: scale = float(s.item()) out[name] = (q.float() * scale).to(dtype=dtype).contiguous() for name, t in obj["passthrough"].items(): - # Restore small tensors, undoing the temporary fp16 storage cast if needed. out_t = t.detach().to("cpu").contiguous() orig_dtype = passthrough_orig_dtypes.get(name) if isinstance(orig_dtype, str): @@ -421,16 +498,12 @@ def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: out[name] = out_t return out - -# ----------------------------- -# DATA LOADING -# ----------------------------- +# --- Data loading --- def load_data_shard(file: Path) -> Tensor: header_bytes = 256 * np.dtype(" Tensor: if tokens_np.size != num_tokens: raise ValueError(f"Short read for {file}") return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) - - class TokenStream: - # Reads shards sequentially and wraps around forever. The training loop therefore - # has deterministic, simple streaming behavior with no sampling or workers. def __init__(self, pattern: str): self.files = [Path(p) for p in sorted(glob.glob(pattern))] if not self.files: @@ -453,12 +522,10 @@ def __init__(self, pattern: str): self.file_idx = 0 self.tokens = load_data_shard(self.files[0]) self.pos = 0 - def _advance_file(self) -> 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 @@ -472,17 +539,12 @@ def take(self, n: int) -> Tensor: self.pos += k remaining -= k return chunks[0] if len(chunks) == 1 else torch.cat(chunks) - - class DistributedTokenLoader: - # Each call consumes a contiguous chunk from the shared token stream, then slices out - # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. 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 @@ -493,44 +555,44 @@ def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> y = local[1:].reshape(-1, seq_len) return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) -# ----------------------------- -# TRANSFORMER MODULES -# ----------------------------- +# --- Transformer modules --- class RMSNorm(nn.Module): def __init__(self, eps: float | None = None): super().__init__() self.eps = eps - def forward(self, x: Tensor) -> Tensor: return F.rms_norm(x, (x.size(-1),), eps=self.eps) - - class CastedLinear(nn.Linear): - # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + _qat_enabled: bool = False def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() bias = self.bias.to(x.dtype) if self.bias is not None else None - return F.linear(x, self.weight.to(x.dtype), bias) - - + return F.linear(x, w, bias) def restore_low_dim_params_to_fp32(module: nn.Module) -> None: - # Keep small/control parameters in fp32 even when the model body runs in bf16. with torch.no_grad(): for name, param in module.named_parameters(): if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: param.data = param.data.float() - - class Rotary(nn.Module): - # Caches cos/sin tables per sequence length on the current device. - def __init__(self, dim: int, base: float = 10000.0): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): super().__init__() - inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) self.register_buffer("inv_freq", inv_freq, persistent=False) self._seq_len_cached = 0 self._cos_cached: Tensor | None = None self._sin_cached: Tensor | None = None - def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: if ( self._cos_cached is None @@ -538,20 +600,30 @@ def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tup 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, :, :] + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] self._seq_len_cached = seq_len return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) - - -def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) half = x.size(-1) // 2 x1, x2 = x[..., :half], x[..., half:] return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) - class CausalSelfAttention(nn.Module): def __init__( self, @@ -560,6 +632,8 @@ def __init__( num_kv_heads: int, rope_base: float, qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, ): super().__init__() if dim % num_heads != 0: @@ -571,51 +645,114 @@ def __init__( self.head_dim = dim // num_heads if self.head_dim % 2 != 0: raise ValueError("head_dim must be even for RoPE") - kv_dim = self.num_kv_heads * self.head_dim - self.c_q = CastedLinear(dim, dim, bias=False) - self.c_k = CastedLinear(dim, kv_dim, bias=False) - self.c_v = CastedLinear(dim, kv_dim, bias=False) - self.proj = CastedLinear(dim, dim, bias=False) - self.proj._zero_init = True + # No CastedLinear -- weights come from banks self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) - self.rotary = Rotary(self.head_dim, base=rope_base) - - def forward(self, x: Tensor) -> Tensor: + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vr_lambda = nn.Parameter(torch.tensor([0.5, 0.5], dtype=torch.float32)) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: 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) + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + lam = self.vr_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v q = F.rms_norm(q, (q.size(-1),)) k = F.rms_norm(k, (k.size(-1),)) cos, sin = self.rotary(seqlen, x.device, q.dtype) - q = apply_rotary_emb(q, cos, sin) - 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), - ) - y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) - return self.proj(y) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) class MLP(nn.Module): - # relu^2 MLP from the original modded-nanogpt setup def __init__(self, dim: int, mlp_mult: int): super().__init__() - hidden = 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 = torch.relu(self.fc(x)) - return self.proj(x.square()) - + # No CastedLinear -- weights come from banks + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.leaky_relu(F.linear(x, up_w.to(x.dtype)), negative_slope=0.5) + return F.linear(x.square(), down_w.to(x.dtype)) class Block(nn.Module): def __init__( @@ -626,24 +763,38 @@ def __init__( mlp_mult: int, rope_base: float, qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, ): super().__init__() self.attn_norm = RMSNorm() self.mlp_norm = RMSNorm() - self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) 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: + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: 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 - + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v class GPT(nn.Module): def __init__( @@ -659,18 +810,46 @@ def __init__( logit_softcap: float, rope_base: float, qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, ): super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection if logit_softcap <= 0.0: raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") self.tie_embeddings = tie_embeddings self.tied_embed_init_std = tied_embed_init_std self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) self.num_encoder_layers = num_layers // 2 self.num_decoder_layers = num_layers - self.num_encoder_layers self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) self.blocks = nn.ModuleList( [ Block( @@ -680,65 +859,835 @@ def __init__( mlp_mult, rope_base, qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, ) for i in range(num_layers) ] ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat self.final_norm = RMSNorm() self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) if self.lm_head is not None: self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True self._init_weights() - def _init_weights(self) -> None: if self.tie_embeddings: nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) - for module in self.modules(): - if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): - nn.init.zeros_(module.weight) - - def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor, return_logits: bool = False) -> Tensor: + n = self.num_layers x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) x0 = x + v0 = None skips: list[Tensor] = [] - - # First half stores skips; second half reuses them in reverse order. + ve_cache: dict = {} for i in range(self.num_encoder_layers): - x = self.blocks[i](x, x0) + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v skips.append(x) for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i if skips: x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - x = self.blocks[self.num_encoder_layers + i](x, x0) - - x = self.final_norm(x).reshape(-1, x.size(-1)) + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) targets = target_ids.reshape(-1) if self.tie_embeddings: - logits_proj = F.linear(x, self.tok_emb.weight) + logits_proj = F.linear(x_flat, self.tok_emb.weight) else: if self.lm_head is None: raise RuntimeError("lm_head is required when tie_embeddings=False") - logits_proj = self.lm_head(x) + logits_proj = self.lm_head(x_flat) logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - return F.cross_entropy(logits.float(), targets, reduction="mean") + if return_logits: + return logits, targets + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- Sliding window evaluation --- + +def _make_ngram_numba_fns(): + """Create numba-compiled n-gram scoring and update functions.""" + if not _HAS_NUMBA: + return None, None + + @njit(cache=False) + def ngram_score_window(val_np, ws, wlen, n_orders, ngram_min_order, + ctx_tables, full_tables, ng_mask, ng_primes): + """Score all tokens in a window using n-gram logit boost. Returns boost array.""" + GOLDEN = np.uint64(2654435761) + MASK64 = np.uint64(0xFFFFFFFFFFFFFFFF) + seg_boost = np.zeros(wlen, dtype=np.float64) + for j in range(wlen): + gpos = ws + j + 1 + best_boost = 0.0 + for order_idx in range(n_orders - 1, -1, -1): + order = ngram_min_order + order_idx + ctx_start = gpos - order + if ctx_start < 0: + continue + ctx_hash = np.uint64(0) + for k in range(order - 1): + ctx_hash ^= np.uint64(val_np[ctx_start + k]) * ng_primes[k % len(ng_primes)] + ctx_hash = (ctx_hash * GOLDEN) & MASK64 + ctx_bucket = int(ctx_hash & ng_mask) + ctx_count = ctx_tables[order_idx][ctx_bucket] + if ctx_count == 0: + continue + full_hash = ctx_hash ^ (np.uint64(val_np[gpos]) * ng_primes[(order - 1) % len(ng_primes)]) + full_hash = (full_hash * GOLDEN) & MASK64 + full_bucket = int(full_hash & ng_mask) + full_count = full_tables[order_idx][full_bucket] + if full_count > 0: + remainder = max(int(ctx_count) - int(full_count), 1) + boost = np.log(1.0 + float(full_count)) - np.log(1.0 + float(remainder)) + if boost > best_boost: + best_boost = boost + seg_boost[j] = best_boost + return seg_boost + + @njit(cache=False) + def ngram_update_window(val_np, ws, wlen, n_orders, ngram_min_order, + ctx_tables, full_tables, ng_mask, ng_primes): + """Update n-gram tables for all tokens in a window (score-first: call AFTER scoring).""" + GOLDEN = np.uint64(2654435761) + MASK64 = np.uint64(0xFFFFFFFFFFFFFFFF) + for j in range(wlen): + gpos = ws + j + 1 + for order_idx in range(n_orders): + order = ngram_min_order + order_idx + ctx_start = gpos - order + if ctx_start < 0: + continue + ctx_hash = np.uint64(0) + for k in range(order - 1): + ctx_hash ^= np.uint64(val_np[ctx_start + k]) * ng_primes[k % len(ng_primes)] + ctx_hash = (ctx_hash * GOLDEN) & MASK64 + ctx_bucket = int(ctx_hash & ng_mask) + ctx_tables[order_idx][ctx_bucket] += 1 + full_hash = ctx_hash ^ (np.uint64(val_np[gpos]) * ng_primes[(order - 1) % len(ng_primes)]) + full_hash = (full_hash * GOLDEN) & MASK64 + full_bucket = int(full_hash & ng_mask) + full_tables[order_idx][full_bucket] += 1 + + return ngram_score_window, ngram_update_window + +_ngram_score_fn, _ngram_update_fn = _make_ngram_numba_fns() if _HAS_NUMBA else (None, None) + + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context. + Optionally blends model predictions with n-gram backoff cache (entropy-adaptive). + Set NGRAM_CACHE=0 to disable.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # N-gram cache + use_ngram = bool(int(os.environ.get("NGRAM_CACHE", "1"))) + if use_ngram: + val_np = val_tokens.cpu().numpy().astype(np.int64) + ngram_min_order = 2 + ngram_order = 10 + n_orders = ngram_order - ngram_min_order + 1 + ng_buckets = 4 * 1024 * 1024 + # Use 2D arrays for numba compatibility + ctx_tables_2d = np.zeros((n_orders, ng_buckets), dtype=np.uint32) + full_tables_2d = np.zeros((n_orders, ng_buckets), dtype=np.uint32) + # Keep list views for Python fallback + ctx_tables = [ctx_tables_2d[i] for i in range(n_orders)] + full_tables = [full_tables_2d[i] for i in range(n_orders)] + ng_mask = np.uint64(ng_buckets - 1) + ng_primes = np.array([np.uint64(36313), np.uint64(27191), np.uint64(51647), + np.uint64(81929), np.uint64(131071), np.uint64(175447), + np.uint64(209591)], dtype=np.uint64) + ent_base, ent_range, ent_scale, ent_thresh = 0.05, 0.55, 2.0, 4.0 + # HedgeMixer: online learning weights for neural vs n-gram experts + use_hedge = bool(int(os.environ.get("HEDGE_MIXER", "1"))) + hedge_eta = float(os.environ.get("HEDGE_ETA", "0.1")) + hedge_w = np.array([0.5, 0.5], dtype=np.float64) # [neural, ngram_boosted] + + # Online bias correction (Nacrith 2026): per-document logit bias vector + use_bias_correction = bool(int(os.environ.get("ONLINE_BIAS", "1"))) + if use_bias_correction: + bias_lr = float(os.environ.get("BIAS_LR", "0.001")) + online_bias = torch.zeros(args.vocab_size, device=device, dtype=torch.float32) + + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + + # Online bias correction on logits (Nacrith 2026) + # NOTE: Ideally bias should be on post-blend output, but n-gram blend + # happens per-token in numpy. Applying to raw logits may partially + # undo complementary training in n-gram regions. Acceptable tradeoff + # since n-gram boost dominates in those regions anyway. + if use_bias_correction: + logits = logits.float() + online_bias.unsqueeze(0).unsqueeze(0) + + if not use_ngram: + # Original path: plain cross-entropy + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + else: + # N-gram blended path + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + + # Model log-probs and entropy + log_probs = F.log_softmax(logits[i, :wlen].float(), dim=-1) + targets_i = y_batch[i, :wlen] + seg_model_p = log_probs[range(wlen), targets_i].exp().cpu().numpy() + probs = log_probs.exp() + seg_ent = -(probs * log_probs).sum(dim=-1).cpu().numpy() + + # N-gram logit boost (properly normalized via softmax) + if _ngram_score_fn is not None: + # Numba-accelerated path (~50-100x faster) + seg_logit_boost = _ngram_score_fn( + val_np, ws, wlen, n_orders, ngram_min_order, + ctx_tables_2d, full_tables_2d, ng_mask, ng_primes) + else: + # Pure Python fallback + seg_logit_boost = np.zeros(wlen, dtype=np.float64) + for j in range(wlen): + gpos = ws + j + 1 + best_boost = 0.0 + for order_idx in range(n_orders - 1, -1, -1): + order = ngram_min_order + order_idx + ctx_start = gpos - order + if ctx_start < 0: + continue + ctx_hash = np.uint64(0) + for k in range(order - 1): + ctx_hash ^= np.uint64(val_np[ctx_start + k]) * ng_primes[k % len(ng_primes)] + ctx_hash = (ctx_hash * np.uint64(2654435761)) & np.uint64(0xFFFFFFFFFFFFFFFF) + ctx_bucket = int(ctx_hash & ng_mask) + ctx_count = ctx_tables[order_idx][ctx_bucket] + if ctx_count == 0: + continue + full_hash = ctx_hash ^ (np.uint64(val_np[gpos]) * ng_primes[(order - 1) % len(ng_primes)]) + full_hash = (full_hash * np.uint64(2654435761)) & np.uint64(0xFFFFFFFFFFFFFFFF) + full_bucket = int(full_hash & ng_mask) + full_count = full_tables[order_idx][full_bucket] + if full_count > 0: + remainder = max(int(ctx_count) - int(full_count), 1) + boost = np.log1p(float(full_count)) - np.log1p(float(remainder)) + best_boost = max(best_boost, boost) + seg_logit_boost[j] = best_boost + + # Apply boost to model log-probs (scored positions only) + scored_model_lp = np.log(np.clip(seg_model_p[s:wlen], 1e-12, 1.0)) + scored_boost = seg_logit_boost[s:wlen] + scored_ent = seg_ent[s:wlen] + + if use_hedge: + # HedgeMixer: two experts - neural only vs neural+ngram + neural_nll = -scored_model_lp + # For ngram expert: use entropy-adaptive boost as before + boost_alpha_raw = ent_base + ent_range / (1.0 + np.exp(ent_scale * (scored_ent - ent_thresh))) + boost_alpha_raw = boost_alpha_raw * (scored_boost > 0).astype(np.float64) + boosted_lp = scored_model_lp + boost_alpha_raw * scored_boost + ngram_nll = -boosted_lp + # Weighted mixture in log-prob space + w_neural, w_ngram = hedge_w[0], hedge_w[1] + # Mixture: p_mix = w0*p_neural + w1*p_ngram (in prob space) + # NLL_mix = -log(w0*exp(-nll_neural) + w1*exp(-nll_ngram)) + max_neg_nll = np.maximum(-neural_nll, -ngram_nll) + mix_lp = max_neg_nll + np.log( + w_neural * np.exp(-neural_nll - max_neg_nll) + + w_ngram * np.exp(-ngram_nll - max_neg_nll) + 1e-30) + nll_blended = -mix_lp + # Hedge update: reduce weights of experts that did poorly + avg_neural_nll = neural_nll.mean() + avg_ngram_nll = ngram_nll.mean() + hedge_w[0] *= np.exp(-hedge_eta * avg_neural_nll) + hedge_w[1] *= np.exp(-hedge_eta * avg_ngram_nll) + hedge_w /= hedge_w.sum() # renormalize + else: + # Original entropy-adaptive boost + boost_alpha = ent_base + ent_range / (1.0 + np.exp(ent_scale * (scored_ent - ent_thresh))) + boost_alpha = boost_alpha * (scored_boost > 0).astype(np.float64) + boosted_lp = scored_model_lp + boost_alpha * scored_boost + nll_blended = -boosted_lp + + loss_sum += float(nll_blended.sum()) + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # Score-first: update n-gram tables AFTER scoring, for ALL positions in the window + if _ngram_update_fn is not None: + _ngram_update_fn(val_np, ws, wlen, n_orders, ngram_min_order, + ctx_tables_2d, full_tables_2d, ng_mask, ng_primes) + else: + for j in range(wlen): + gpos = ws + j + 1 + for order_idx in range(n_orders): + order = ngram_min_order + order_idx + ctx_start = gpos - order + if ctx_start < 0: + continue + ctx_hash = np.uint64(0) + for k in range(order - 1): + ctx_hash ^= np.uint64(val_np[ctx_start + k]) * ng_primes[k % len(ng_primes)] + ctx_hash = (ctx_hash * np.uint64(2654435761)) & np.uint64(0xFFFFFFFFFFFFFFFF) + ctx_bucket = int(ctx_hash & ng_mask) + ctx_tables[order_idx][ctx_bucket] += 1 + full_hash = ctx_hash ^ (np.uint64(val_np[gpos]) * ng_primes[(order - 1) % len(ng_primes)]) + full_hash = (full_hash * np.uint64(2654435761)) & np.uint64(0xFFFFFFFFFFFFFFFF) + full_bucket = int(full_hash & ng_mask) + full_tables[order_idx][full_bucket] += 1 + + # Online bias update (score-first: update AFTER all scoring in this batch) + if use_bias_correction: + with torch.no_grad(): + probs = F.softmax(logits.float(), dim=-1) + all_targets = y_batch[:bsz, :seq_len].reshape(-1) + mean_probs = probs[:bsz].reshape(-1, probs.size(-1)).mean(dim=0) + target_counts = torch.zeros_like(online_bias) + target_counts.scatter_add_(0, all_targets, torch.ones(all_targets.size(0), device=device, dtype=torch.float32)) + n_tokens = float(bsz * seq_len) + online_bias -= bias_lr * mean_probs + online_bias += bias_lr * target_counts / max(n_tokens, 1.0) + + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +def eval_val_sliding_ttt( + args: Hyperparameters, base_model: nn.Module, rank: int, world_size: int, + device: torch.device, val_tokens: Tensor, base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, is_boundary_token_lut: Tensor, + stride: int, batch_seqs: int = 32, log0=print, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461 recipe): score each chunk with sliding windows, + then train on it. Every token scored BEFORE any update that could use it.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + log0(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = False + for bi in frozen_block_ids: + if f"blocks.{bi}." in name: + freeze = True + break + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + # Per-layer LR groups (PR #995 recipe: output proj 3x, fc/input 0.5x, rest 1x) + ttt_proj_params, ttt_fc_params, ttt_other_params = [], [], [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + if "c_proj" in name or "out_proj" in name or "lm_head" in name: + ttt_proj_params.append(p) + elif "c_fc" in name or "embed" in name or "wte" in name: + ttt_fc_params.append(p) + else: + ttt_other_params.append(p) + ttt_proj_lr = args.ttt_lr * float(os.environ.get("TTT_PROJ_LR_MULT", "3.0")) + ttt_fc_lr = args.ttt_lr * float(os.environ.get("TTT_FC_LR_MULT", "0.5")) + param_groups = [] + if ttt_proj_params: + param_groups.append({"params": ttt_proj_params, "lr": ttt_proj_lr, "_base_lr": ttt_proj_lr}) + if ttt_fc_params: + param_groups.append({"params": ttt_fc_params, "lr": ttt_fc_lr, "_base_lr": ttt_fc_lr}) + if ttt_other_params: + param_groups.append({"params": ttt_other_params, "lr": args.ttt_lr, "_base_lr": args.ttt_lr}) + log0(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)} " + f"proj_params={sum(p.numel() for p in ttt_proj_params)}(lr={ttt_proj_lr:.4f}) " + f"fc_params={sum(p.numel() for p in ttt_fc_params)}(lr={ttt_fc_lr:.4f}) " + f"other_params={sum(p.numel() for p in ttt_other_params)}(lr={args.ttt_lr:.4f})") + + optimizer = torch.optim.SGD(param_groups, lr=args.ttt_lr, momentum=args.ttt_momentum) + t0 = time.perf_counter() -# ----------------------------- -# TRAINING -# ----------------------------- + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_scale = 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg['lr'] = pg.get('_base_lr', args.ttt_lr) * cos_scale + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_batch_seqs): + be = min(bs + args.ttt_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + # Polyak averaging after TTT chunk (PR #995: decay=0.998) + # Smooths weights in-place — next chunk scoring uses smoothed weights + polyak_decay = float(os.environ.get("TTT_POLYAK_DECAY", "0.998")) + if 0 < polyak_decay < 1: + with torch.no_grad(): + for p in ttt_params: + if not hasattr(p, '_ema'): + p._ema = p.data.clone() + p._ema.mul_(polyak_decay).add_(p.data, alpha=1 - polyak_decay) + p.data.copy_(p._ema) + + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rl = loss_sum.item() / max(token_count.item(), 1) + rbpb = rl / math.log(2.0) * (token_count.item() / max(byte_count.item(), 1)) if token_count.item() > 0 else 0.0 + log0(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + + # Clean up Polyak EMA state + for p in ttt_params: + if hasattr(p, '_ema'): + del p._ema -def main() -> None: - global zeropower_via_newtonschulz5 + 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) - code = Path(__file__).read_text(encoding="utf-8") - args = Hyperparameters() - zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + log0(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + + +# --- GPTQ-lite int6 quantization --- + +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale + +def _unbank_state_dict(sd: dict[str, Tensor], num_layers: int) -> dict[str, Tensor]: + """Convert 3D bank tensors into individual 2D tensors with standard names.""" + out: dict[str, Tensor] = {} + n = num_layers + for name, tensor in sd.items(): + if name == "qo_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_q.weight"] = tensor[i] + out[f"blocks.{i}.attn.proj.weight"] = tensor[n + i] + elif name == "kv_bank": + for i in range(n): + out[f"blocks.{i}.attn.c_k.weight"] = tensor[i] + out[f"blocks.{i}.attn.c_v.weight"] = tensor[n + i] + elif name == "mlp_up_bank": + for i in range(n): + out[f"blocks.{i}.mlp.fc.weight"] = tensor[i] + elif name == "mlp_down_bank": + for i in range(n): + out[f"blocks.{i}.mlp.proj.weight"] = tensor[i] + else: + out[name] = tensor + return out + +def _rebank_state_dict(sd: dict[str, Tensor], num_layers: int, template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + """Convert individual 2D tensors back into 3D bank tensors.""" + out: dict[str, Tensor] = {} + n = num_layers + # Reconstruct banks from individual weight keys + qo_slices = [None] * (2 * n) + kv_slices = [None] * (2 * n) + up_slices = [None] * n + down_slices = [None] * n + consumed = set() + for i in range(n): + qk = f"blocks.{i}.attn.c_q.weight" + if qk in sd: + qo_slices[i] = sd[qk] + consumed.add(qk) + ok = f"blocks.{i}.attn.proj.weight" + if ok in sd: + qo_slices[n + i] = sd[ok] + consumed.add(ok) + kk = f"blocks.{i}.attn.c_k.weight" + if kk in sd: + kv_slices[i] = sd[kk] + consumed.add(kk) + vk = f"blocks.{i}.attn.c_v.weight" + if vk in sd: + kv_slices[n + i] = sd[vk] + consumed.add(vk) + fk = f"blocks.{i}.mlp.fc.weight" + if fk in sd: + up_slices[i] = sd[fk] + consumed.add(fk) + dk = f"blocks.{i}.mlp.proj.weight" + if dk in sd: + down_slices[i] = sd[dk] + consumed.add(dk) + out["qo_bank"] = torch.stack(qo_slices).to(dtype=template_sd["qo_bank"].dtype) + out["kv_bank"] = torch.stack(kv_slices).to(dtype=template_sd["kv_bank"].dtype) + out["mlp_up_bank"] = torch.stack(up_slices).to(dtype=template_sd["mlp_up_bank"].dtype) + out["mlp_down_bank"] = torch.stack(down_slices).to(dtype=template_sd["mlp_down_bank"].dtype) + for name, tensor in sd.items(): + if name not in consumed: + out[name] = tensor + return out + +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out - # ----------------------------- - # DISTRIBUTED + CUDA SETUP - # ----------------------------- +# --- Training --- +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile 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")) @@ -757,23 +1706,18 @@ def main() -> None: dist.init_process_group(backend="nccl", device_id=device) dist.barrier() master_process = rank == 0 - - # Fast math knobs torch.backends.cuda.matmul.allow_tf32 = True torch.backends.cudnn.allow_tf32 = True from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp - enable_cudnn_sdp(False) enable_flash_sdp(True) enable_mem_efficient_sdp(False) enable_math_sdp(False) - logfile = None if master_process: os.makedirs("logs", exist_ok=True) logfile = f"logs/{args.run_id}.txt" print(logfile) - def log0(msg: str, console: bool = True) -> None: if not master_process: return @@ -782,7 +1726,6 @@ def log0(msg: str, console: bool = True) -> None: 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) @@ -792,37 +1735,38 @@ def log0(msg: str, console: bool = True) -> None: console=False, ) log0("=" * 100, console=False) - - # ----------------------------- - # TOKENIZER + VALIDATION METRIC SETUP - # ----------------------------- - 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())}" - ) + has_tokenizer = args.tokenizer_path.endswith(".model") and Path(args.tokenizer_path).exists() + if has_tokenizer: + 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())}" + ) + else: + sp = None + log0(f"WARNING: tokenizer not found at {args.tokenizer_path}, BPB will use estimate") 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}") + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + if sp is not None: + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + else: + # Estimate: avg ~3.5 bytes per token for 1024 vocab BPE on English text + base_bytes_lut = torch.full((args.vocab_size,), 3.5, device=device) + has_leading_space_lut = torch.zeros(args.vocab_size, dtype=torch.bool, device=device) + is_boundary_token_lut = torch.zeros(args.vocab_size, dtype=torch.bool, device=device) + log0(f"val_bpb:{'enabled' if sp else 'estimated'} 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}") - - # ----------------------------- - # MODEL + OPTIMIZER SETUP - # ----------------------------- - + CastedLinear._qat_enabled = args.qat_enabled base_model = GPT( vocab_size=args.vocab_size, num_layers=args.num_layers, @@ -835,25 +1779,44 @@ def log0(msg: str, console: bool = True) -> None: logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() for module in base_model.modules(): if isinstance(module, CastedLinear): module.float() restore_low_dim_params_to_fp32(base_model) + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) - model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model + model = compiled_model # Optimizer split: - # - token embedding (Adam) uses EMBED_LR - # - untied lm_head (Adam) uses HEAD_LR - # - matrix params in transformer blocks use MATRIX_LR via Muon - # - vectors/scalars use SCALAR_LR via Adam - block_named_params = list(base_model.blocks.named_parameters()) + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) 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) + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, ] + block_named_params = list(base_model.blocks.named_parameters()) scalar_params = [ p for name, p in block_named_params @@ -861,11 +1824,27 @@ def log0(msg: str, console: bool = True) -> None: ] if base_model.skip_weights.numel() > 0: scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr - optimizer_tok = torch.optim.Adam( - [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.adam_wd, fused=True, ) optimizer_muon = Muon( @@ -873,16 +1852,24 @@ def log0(msg: str, console: bool = True) -> None: lr=args.matrix_lr, momentum=args.muon_momentum, backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, ) for group in optimizer_muon.param_groups: group["base_lr"] = args.matrix_lr - optimizer_scalar = torch.optim.Adam( + optimizer_scalar = torch.optim.AdamW( [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.adam_wd, fused=True, ) - optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None if base_model.lm_head is not None: optimizer_head = torch.optim.Adam( [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], @@ -890,10 +1877,16 @@ def log0(msg: str, console: bool = True) -> None: eps=args.adam_eps, fused=True, ) - optimizers.insert(1, optimizer_head) - + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") @@ -908,19 +1901,11 @@ def log0(msg: str, console: bool = True) -> None: f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" ) log0(f"seed:{args.seed}") - - # ----------------------------- - # DATA LOADER & MODEL WARMUP - # ----------------------------- - 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 @@ -931,9 +1916,6 @@ def lr_mul(step: int, elapsed_ms: float) -> float: warmdown_ms = args.warmdown_iters * step_ms remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 - - # Warmup primes the compiled forward/backward/optimizer paths, then we restore the - # initial weights/optimizer state so measured training starts from the true init. 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] @@ -941,12 +1923,15 @@ def lr_mul(step: int, elapsed_ms: float) -> float: 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() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) for opt in optimizers: opt.step() zero_grad_all() @@ -956,23 +1941,37 @@ def lr_mul(step: int, elapsed_ms: float) -> float: 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) - - # ----------------------------- - # MAIN TRAINING LOOP - # ----------------------------- - + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 training_time_ms = 0.0 stop_after_step: int | None = None + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + if complement_alpha > 0: + bi_counts = torch.zeros(args.vocab_size, args.vocab_size, device=device) + bi_totals = torch.zeros(args.vocab_size, device=device) + def get_complement_weights(x, y): + xf = x.reshape(-1) + yf = y.reshape(-1) + total = bi_totals[xf] + count = bi_counts[xf, yf] + ngram_prob = count / (total + 1) + return (1.0 - complement_alpha * ngram_prob).clamp(min=0.1) + def update_bigram_counts(x, y): + xf = x.reshape(-1) + yf = y.reshape(-1) + bi_counts[xf, yf] += 1 + bi_totals[xf] += 1 + log0(f"complement_alpha:{complement_alpha}") 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() @@ -995,7 +1994,6 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) torch.cuda.synchronize() t0 = time.perf_counter() - if last_step: if stop_after_step is not None and step < args.iterations: log0( @@ -1003,38 +2001,68 @@ def lr_mul(step: int, elapsed_ms: float) -> float: f"step:{step}/{args.iterations}" ) break - elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") zero_grad_all() train_loss = torch.zeros((), device=device) for micro_step in range(grad_accum_steps): - 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) + if complement_alpha > 0: + logits, targets = model(x, y, return_logits=True) + w = get_complement_weights(x, y) + per_token_loss = F.cross_entropy(logits.float(), targets, reduction="none") + loss = (per_token_loss * w).sum() / w.sum() + update_bigram_counts(x, y) + else: + 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() + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() zero_grad_all() - + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) step += 1 approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) 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) @@ -1044,8 +2072,6 @@ def lr_mul(step: int, elapsed_ms: float) -> float: 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" ) - - # Needed to sync whether we've reached the wallclock cap. 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) @@ -1053,74 +2079,159 @@ def lr_mul(step: int, elapsed_ms: float) -> float: 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" ) - - # ----------------------------- - # SERIALIZATION + ROUNDTRIP VALIDATION - # ----------------------------- - # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce - # the compressed int8+zlib artifact and validate the round-tripped weights. - + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + elif step >= 1000: + log0(f"ema:applying EMA weights (step={step})") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + else: + log0(f"ema:SKIPPING EMA (step={step} < 1000, too few steps for convergence)") + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") if master_process: - torch.save(base_model.state_dict(), "final_model.pt") + torch.save(export_sd, "final_model.pt") model_bytes = os.path.getsize("final_model.pt") code_bytes = len(code.encode("utf-8")) log0(f"Serialized model: {model_bytes} bytes") log0(f"Code size: {code_bytes} bytes") - log0(f"Total submission size: {model_bytes + code_bytes} bytes") - - quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) + # Unbank 3D tensors into individual 2D tensors for quantization + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + unbanked_sd = _unbank_state_dict(sd_cpu, args.num_layers) + quant_result, quant_meta = mixed_quantize_int6(unbanked_sd, {"mlp", "attn"}) quant_buf = io.BytesIO() - torch.save(quant_obj, quant_buf) + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) quant_raw = quant_buf.getvalue() - quant_blob = zlib.compress(quant_raw, level=9) - quant_raw_bytes = len(quant_raw) + quant_blob = lzma.compress(quant_raw, preset=6) if master_process: - with open("final_model.int8.ptz", "wb") as f: + with open("final_model.int6.ptz", "wb") as f: f.write(quant_blob) - quant_file_bytes = os.path.getsize("final_model.int8.ptz") + quant_file_bytes = len(quant_blob) code_bytes = len(code.encode("utf-8")) - ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) - log0( - f"Serialized model int8+zlib: {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 int8+zlib: {quant_file_bytes + code_bytes} bytes") - + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") if distributed: dist.barrier() - with open("final_model.int8.ptz", "rb") as f: + with open("final_model.int6.ptz", "rb") as f: quant_blob_disk = f.read() - quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") - base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) + quant_state = torch.load( + io.BytesIO(lzma.decompress(quant_blob_disk)), + map_location="cpu", + ) + deq_unbanked = dequantize_mixed_int6(quant_state["w"], quant_state["m"], unbanked_sd) + # Re-bank the dequantized tensors + deq_state = _rebank_state_dict(deq_unbanked, args.num_layers, sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) torch.cuda.synchronize() t_qeval = time.perf_counter() q_val_loss, q_val_bpb = eval_val( - args, - model, - rank, - world_size, - device, - grad_accum_steps, - val_tokens, - base_bytes_lut, - has_leading_space_lut, - is_boundary_token_lut, + args, compiled_eval, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + eval_seq_len=effective_eval_seq_len, ) torch.cuda.synchronize() log0( - f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" ) - log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + sw_seq_len = effective_eval_seq_len + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_int6_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_int6_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + log0(f"final_int8_zlib_roundtrip_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + # Legal score-first TTT (PR #461 recipe) + if args.ttt_enabled: + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_loss, ttt_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, log0=log0, + ) + torch.cuda.synchronize() + log0(f"legal_ttt val_loss:{ttt_loss:.4f} val_bpb:{ttt_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms") + log0(f"legal_ttt_exact val_loss:{ttt_loss:.8f} val_bpb:{ttt_bpb:.8f}") if distributed: dist.destroy_process_group() - - if __name__ == "__main__": main()