diff --git a/train_gpt.py b/train_gpt.py index 651beb2b89..f81d95bd1b 100644 --- a/train_gpt.py +++ b/train_gpt.py @@ -1,11 +1,4 @@ -""" -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 @@ -16,9 +9,9 @@ import sys import time import uuid +import lzma import zlib from pathlib import Path - import numpy as np import sentencepiece as spm import torch @@ -26,76 +19,98 @@ 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 - +from flash_attn_interface import flash_attn_func as flash_attn_3_func 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/ - + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) # tighter: collect more recent checkpoints + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 3072)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 4)) # XSA on last 4 layers (0 = disabled) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + vrl = bool(int(os.environ.get("VRL", "1"))) # Value Residual Learning (ResFormer arXiv:2410.17897) + # TTT Burst: replay recent training batches at low LR before EMA + ttt_burst_enabled = bool(int(os.environ.get("TTT_BURST_ENABLED", "1"))) + ttt_burst_epochs = int(os.environ.get("TTT_BURST_EPOCHS", 2)) + ttt_burst_lr_factor = float(os.environ.get("TTT_BURST_LR_FACTOR", 0.1)) + ttt_burst_steps = int(os.environ.get("TTT_BURST_STEPS", 100)) + ttt_burst_trigger = float(os.environ.get("TTT_BURST_TRIGGER", 0.2)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.002)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + # Sliding window TTT (full-parameter, PR#461/549 recipe) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_freeze_blocks = int(os.environ.get("TTT_FREEZE_BLOCKS", 0)) + ttt_freeze_embeddings = bool(int(os.environ.get("TTT_FREEZE_EMBEDDINGS", "0"))) + ttt_train_batch_seqs = int(os.environ.get("TTT_TRAIN_BATCH_SEQS", 32)) + ttt_grad_clip = float(os.environ.get("TTT_GRAD_CLIP", 1.0)) + eb_ttt = bool(int(os.environ.get("EB_TTT", "0"))) # Empirical Bayes adaptive per-layer TTT LR + eb_ttt_min = float(os.environ.get("EB_TTT_MIN", "0.3")) + eb_ttt_max = float(os.environ.get("EB_TTT_MAX", "3.0")) + eb_ttt_born = bool(int(os.environ.get("EB_TTT_BORN", "0"))) # Born-rule: SNR² scaling + # GPTQ calibration + gptq_enabled = bool(int(os.environ.get("GPTQ_ENABLED", "1"))) + gptq_calib_batches = int(os.environ.get("GPTQ_CALIB_BATCHES", 256)) + gptq_block_size = int(os.environ.get("GPTQ_BLOCK_SIZE", 128)) + # TTT optimizer + ttt_adamw = bool(int(os.environ.get("TTT_ADAMW", "0"))) + ttt_wd = float(os.environ.get("TTT_WD", 0.01)) + # Eval-only mode: skip training + GPTQ, load saved quantized model + eval_only = bool(int(os.environ.get("EVAL_ONLY", "0"))) def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: - # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. - # Muon uses this to normalize matrix-shaped gradients before applying them. a, b, c = (3.4445, -4.7750, 2.0315) X = G.bfloat16() X /= X.norm() + eps @@ -107,26 +122,23 @@ def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) - B = b * A + c * A @ A X = a * X + B @ X return X.T if transposed else X - - class Muon(torch.optim.Optimizer): - def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): + 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), ) - @torch.no_grad() def step(self, closure=None): loss = None if closure is not None: with torch.enable_grad(): loss = closure() - distributed = dist.is_available() and dist.is_initialized() world_size = dist.get_world_size() if distributed else 1 rank = dist.get_rank() if distributed else 0 - for group in self.param_groups: params = group["params"] if not params: @@ -135,10 +147,8 @@ def step(self, closure=None): 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: @@ -151,32 +161,20 @@ def step(self, closure=None): 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) - + wd = group.get("weight_decay", 0.0) curr = 0 for p in params: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) p.add_(g, alpha=-lr) curr += p.numel() - return loss - - -# ----------------------------- -# 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. - def build_sentencepiece_luts( sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device ) -> tuple[Tensor, Tensor, Tensor]: @@ -193,7 +191,7 @@ 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 +200,15 @@ 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 +220,32 @@ 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()) @@ -265,64 +256,29 @@ 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. - 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,vrl_lambda", ).split(",") if pattern ) -INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( - pattern - for pattern in os.environ.get( - "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", - ",".join(CONTROL_TENSOR_NAME_PATTERNS), - ).split(",") - if pattern -) -INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 -INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 INT8_PER_ROW_SCALE_DTYPE = torch.float16 INT8_CLIP_PERCENTILE = 99.99984 INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 - -def tensor_nbytes(t: Tensor) -> int: - return int(t.numel()) * int(t.element_size()) - -def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: - if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): - return t.float().contiguous() - if t.dtype in {torch.float32, torch.bfloat16}: - passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") - return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() - return t - def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: t32 = t.float() if t32.ndim == 2: - # 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() @@ -332,105 +288,14 @@ 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] = {} - passthrough: dict[str, Tensor] = {} - passthrough_orig_dtypes: dict[str, str] = {} - qmeta: dict[str, dict[str, object]] = {} - stats = dict.fromkeys( - ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), - 0, - ) - - for name, tensor in state_dict.items(): - t = tensor.detach().to("cpu").contiguous() - stats["param_count"] += int(t.numel()) - stats["num_tensors"] += 1 - stats["baseline_tensor_bytes"] += tensor_nbytes(t) - - if not t.is_floating_point(): - stats["num_nonfloat_tensors"] += 1 - passthrough[name] = t - stats["int8_payload_bytes"] += tensor_nbytes(t) - continue - - # 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: - qmeta[name] = {"scheme": "per_row", "axis": 0} - quantized[name] = q - scales[name] = s - dtypes[name] = str(t.dtype).removeprefix("torch.") - stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) - - obj: dict[str, object] = { - "__quant_format__": "int8_clean_per_row_v1", - "quantized": quantized, - "scales": scales, - "dtypes": dtypes, - "passthrough": passthrough, - } - if qmeta: - obj["qmeta"] = qmeta - if passthrough_orig_dtypes: - obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes - return obj, stats - -def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: - out: dict[str, Tensor] = {} - qmeta = obj.get("qmeta", {}) - passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) - for name, q in obj["quantized"].items(): - dtype = getattr(torch, obj["dtypes"][name]) - s = obj["scales"][name] - if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: - s = s.to(dtype=torch.float32) - # 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): - out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() - out[name] = out_t - return out - - -# ----------------------------- -# DATA LOADING -# ----------------------------- - def load_data_shard(file: Path) -> Tensor: header_bytes = 256 * np.dtype(" 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 +314,10 @@ 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 +331,12 @@ 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 @@ -492,45 +346,52 @@ def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> x = local[:-1].reshape(-1, seq_len) y = local[1:].reshape(-1, seq_len) return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) - -# ----------------------------- -# TRANSFORMER MODULES -# ----------------------------- - class RMSNorm(nn.Module): def __init__(self, eps: float | None = None): super().__init__() self.eps = eps - def forward(self, x: Tensor) -> Tensor: return F.rms_norm(x, (x.size(-1),), eps=self.eps) - - class CastedLinear(nn.Linear): - # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. + _qat_enabled: bool = False + _soft_tau: float = 1000.0 # High = hard round; low = soft (annealed during QAT) def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1).detach() + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + x_norm = w32 / scale[:, None] + # Hard quantized value (forward) + q_hard = torch.clamp(torch.round(x_norm), -31, 31) + # Soft interpolation (backward) for gradient signal + x_floor = x_norm.detach().floor() + frac = x_norm - x_floor + p = torch.sigmoid((frac - 0.5) / max(CastedLinear._soft_tau, 0.01)) + q_soft = torch.clamp(x_floor.detach() + p, -31, 31) + # STE: hard forward, soft backward + q = q_hard.detach() + (q_soft - q_soft.detach()) + w_q = (q * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() bias = self.bias.to(x.dtype) if self.bias is not None else None - return F.linear(x, 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 +399,29 @@ 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, @@ -578,45 +448,104 @@ def __init__( self.proj = CastedLinear(dim, dim, bias=False) self.proj._zero_init = True self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) - self.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 + self.use_vrl = False # set by GPT.__init__; VRL on all layers except first + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] — broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, v_embed: Tensor | None = None, q_delta: Tensor | None = None, v_delta: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: bsz, seqlen, dim = x.shape - q = self.c_q(x).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 = self.c_q(x) + if q_delta is not None: + q = q + q_delta + q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = self.c_v(x) + if v_embed is not None: + v = v + v_embed + if v_delta is not None: + v = v + v_delta + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v # cache for VRL before blending + if self.use_vrl and v0 is not None: + lam = self.vrl_lambda.to(dtype=v.dtype) + v = lam[0] * v0 + lam[1] * v q = F.rms_norm(q, (q.size(-1),)) k = F.rms_norm(k, (k.size(-1),)) cos, sin = self.rotary(seqlen, x.device, q.dtype) - q = apply_rotary_emb(q, cos, sin) - 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) + y = y.reshape(bsz, seqlen, dim) + return self.proj(y), raw_v +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) class MLP(nn.Module): - # relu^2 MLP from the original modded-nanogpt setup def __init__(self, dim: int, mlp_mult: int): super().__init__() - hidden = mlp_mult * dim + hidden = int(mlp_mult * dim) self.fc = CastedLinear(dim, hidden, bias=False) self.proj = CastedLinear(hidden, dim, bias=False) self.proj._zero_init = True - def forward(self, x: Tensor) -> Tensor: - x = torch.relu(self.fc(x)) + x = F.leaky_relu(self.fc(x), negative_slope=0.5) return self.proj(x.square()) - - class Block(nn.Module): def __init__( self, @@ -626,6 +555,9 @@ def __init__( mlp_mult: int, rope_base: float, qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, ): super().__init__() self.attn_norm = RMSNorm() @@ -635,16 +567,26 @@ def __init__( 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, v_embed: Tensor | None = None, q_delta_fn=None, v_delta_fn=None, v0: Tensor | None = None) -> tuple[Tensor, Tensor]: mix = self.resid_mix.to(dtype=x.dtype) - x = 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 + n = self.attn_norm(x_in) * self.ln_scale_factor + qd = q_delta_fn(n) if q_delta_fn is not None else None + vd = v_delta_fn(n) if v_delta_fn is not None else None + attn_out, raw_v = self.attn(n, v_embed=v_embed, q_delta=qd, v_delta=vd, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v class GPT(nn.Module): def __init__( self, @@ -659,14 +601,32 @@ 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", + use_vrl: bool = False, ): super().__init__() + self.use_vrl = use_vrl + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection if logit_softcap <= 0.0: raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") self.tie_embeddings = tie_embeddings self.tied_embed_init_std = tied_embed_init_std self.logit_softcap = logit_softcap + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) self.num_encoder_layers = num_layers // 2 self.num_decoder_layers = num_layers - self.num_encoder_layers self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) @@ -680,65 +640,632 @@ def __init__( mlp_mult, rope_base, qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, ) for i in range(num_layers) ] ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + # VRL: Value Residual Learning — blend layer 0's V into all subsequent layers + if use_vrl: + for i, block in enumerate(self.blocks): + if i > 0: # layer 0 produces v0, all others blend + block.attn.use_vrl = True + block.attn.vrl_lambda = nn.Parameter(torch.tensor([0.01, 0.99], dtype=torch.float32)) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat self.final_norm = RMSNorm() self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) if self.lm_head is not None: self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True self._init_weights() - def _init_weights(self) -> None: if self.tie_embeddings: nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) - 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: + num_layers = len(self.blocks) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + if ".proj." in name or name.endswith(".proj"): + with torch.no_grad(): + module.weight.mul_(1.0 / math.sqrt(2 * num_layers)) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor, lora=None) -> Tensor: x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) x0 = x skips: list[Tensor] = [] - - # First half stores skips; second half reuses them in reverse order. + ve_cache: dict = {} + v0 = None # VRL: cached V from first layer for i in range(self.num_encoder_layers): - x = self.blocks[i](x, x0) + ve = self._get_ve(i, input_ids, ve_cache) + qd = lora.q_loras[i] if lora else None + vd = lora.v_loras[i] if lora else None + x, raw_v = self.blocks[i](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v skips.append(x) for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i if skips: x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - 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) + qd = lora.q_loras[bi] if lora else None + vd = lora.v_loras[bi] if lora else None + x, _ = self.blocks[bi](x, x0, v_embed=ve, q_delta_fn=qd, v_delta_fn=vd, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) targets = target_ids.reshape(-1) if self.tie_embeddings: - logits_proj = F.linear(x, 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_flat) + logits_proj = logits_proj + (lora.lm_head_lora(x).reshape(-1, logits_proj.size(-1)) if lora else 0) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if lora: + bsz, sl, V = logits_proj.shape[0] // target_ids.shape[1], target_ids.shape[1], logits_proj.shape[-1] + return F.cross_entropy(logits.float(), targets, reduction="none").reshape(bsz, sl) + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor, return_hidden: bool = False): + """Return logits (bsz, seq_len, vocab) without computing loss.""" + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + skips: list[Tensor] = [] + ve_cache: dict = {} + v0 = None # VRL: cached V from first layer + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, v_embed=ve, v0=v0) + if i == 0 and self.use_vrl: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: logits_proj = self.lm_head(x) logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - return F.cross_entropy(logits.float(), targets, reduction="mean") + if return_hidden: + return logits, x + return logits +def eval_val_sliding_ttt( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int = 64, + batch_seqs: int = 32, + log_fn=None, +) -> tuple[float, float]: + """Legal score-first TTT (PR #461/549 recipe): score each 32K chunk with + sliding windows, then train on it. Every token scored BEFORE any update + that could use it. Model synchronized across GPUs via all-reduce.""" + seq_len = args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + + # Pre-compute all window starts + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0] + + # Assign each window to a chunk based on the first token it scores + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + + if log_fn: + log_fn(f"ttt_sliding:start chunks={num_chunks} chunk_tokens={ttt_chunk} " + f"total_windows={len(window_starts)} stride={stride} " + f"ttt_lr={args.ttt_lr} ttt_epochs={args.ttt_epochs} " + f"freeze_blocks={args.ttt_freeze_blocks}") + + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + + # Freeze first N blocks + optionally embeddings + frozen_block_ids = set(range(min(args.ttt_freeze_blocks, len(base_model.blocks)))) + ttt_params = [] + for name, p in base_model.named_parameters(): + freeze = any(f"blocks.{bi}." in name for bi in frozen_block_ids) + # Freeze embeddings during TTT: adapting vocab embeddings to a local chunk + # distorts representations for tokens not in that chunk + if args.ttt_freeze_embeddings and any(k in name for k in ("tok_emb", "bigram", "lm_head")): + freeze = True + if freeze: + p.requires_grad_(False) + else: + p.requires_grad_(True) + ttt_params.append(p) + + if log_fn: + log_fn(f"ttt_sliding:params unfrozen={sum(p.numel() for p in ttt_params)} " + f"frozen={sum(p.numel() for p in base_model.parameters() if not p.requires_grad)}") + + if args.ttt_adamw: + optimizer = torch.optim.AdamW(ttt_params, lr=args.ttt_lr, weight_decay=args.ttt_wd) + else: + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + # Precompute layer keys for EB-adaptive TTT + if args.eb_ttt: + ttt_param_layer_keys: list[str] = [] + for name, p in base_model.named_parameters(): + if not p.requires_grad: + continue + parts = name.split(".") + lk = f"{parts[0]}.{parts[1]}" if len(parts) > 1 and parts[1].isdigit() else parts[0] + ttt_param_layer_keys.append(lk) + t0 = time.perf_counter() + + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + + # --- Phase 1: SCORE this chunk's windows (inference_mode) --- + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + base_model.eval() + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = base_model.forward_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt, prev = y_batch[i, s:wlen], x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + + # --- Phase 2: TRAIN on this chunk (already scored = legal) --- + is_last_chunk = (ci == num_chunks - 1) + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + # Cross-chunk cosine: base LR decays as we move through validation + chunk_base_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + steps_per_ep = max(1, (my_chunk_seqs + args.ttt_train_batch_seqs - 1) // args.ttt_train_batch_seqs) + total_steps = args.ttt_epochs * steps_per_ep + step_counter = 0 + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, args.ttt_train_batch_seqs): + # Intra-chunk cosine: decay within this chunk's epochs + progress = step_counter / max(total_steps - 1, 1) + intra_mul = 0.5 * (1.0 + math.cos(math.pi * progress)) + lr_min_ratio = 0.1 # floor at 10% of base + cur_lr = chunk_base_lr * (lr_min_ratio + (1.0 - lr_min_ratio) * intra_mul) + for pg in optimizer.param_groups: + pg['lr'] = cur_lr + step_counter += 1 + be = min(bs + args.ttt_train_batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1: + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + # Empirical Bayes adaptive TTT: scale gradients per-layer by SNR + # High SNR (consistent direction) → amplify; Low SNR → stay at prior + if args.eb_ttt: + with torch.no_grad(): + layer_grads: dict[str, list[Tensor]] = {} + for pi, p in enumerate(ttt_params): + if p.grad is None: + continue + lk = ttt_param_layer_keys[pi] + if lk not in layer_grads: + layer_grads[lk] = [] + layer_grads[lk].append(p.grad) + layer_scales: dict[str, float] = {} + for lk, grads in layer_grads.items(): + flat = torch.cat([g.float().flatten() for g in grads]) + snr = (flat.abs().mean() / (flat.std() + 1e-8)).item() + # Born-rule: probabilities scale as |ψ|², giving sharper + # discrimination between signal (high SNR) and noise (low SNR) + scale = snr ** 2 if args.eb_ttt_born else snr + layer_scales[lk] = max(args.eb_ttt_min, min(args.eb_ttt_max, scale)) + for pi, p in enumerate(ttt_params): + if p.grad is not None: + p.grad.mul_(layer_scales.get(ttt_param_layer_keys[pi], 1.0)) + torch.nn.utils.clip_grad_norm_(ttt_params, args.ttt_grad_clip) + optimizer.step() + + if log_fn and (ci % 10 == 0 or ci == num_chunks - 1): + elapsed = time.perf_counter() - t0 + rbpb = float((loss_sum / math.log(2.0)) / byte_count) if byte_count > 0 else 0.0 + log_fn(f" ttt_chunk [{ci+1}/{num_chunks}] bpb={rbpb:.6f} time={elapsed:.1f}s") + if args.eb_ttt and ci % 100 == 0 and 'layer_scales' in dir(): + log_fn(f" eb_scales: {' '.join(f'{k}={v:.2f}' for k, v in sorted(layer_scales.items()))}") + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) -# ----------------------------- -# TRAINING -# ----------------------------- + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.eval() + + if log_fn: + log_fn(f"ttt_sliding:done val_loss={val_loss:.6f} val_bpb={val_bpb:.6f} " + f"elapsed={time.perf_counter() - t0:.1f}s") + return val_loss, val_bpb + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def collect_hessians( + model: nn.Module, train_loader, args, device: torch.device, + grad_accum_steps: int, num_batches: int = 256, +) -> dict[str, Tensor]: + """Collect H = X^T X for each CastedLinear via forward hooks on calibration data.""" + hessians: dict[str, Tensor] = {} + hooks = [] + for name, module in model.named_modules(): + if isinstance(module, CastedLinear): + pname = name + ".weight" + cols = module.weight.shape[1] + hessians[pname] = torch.zeros(cols, cols, dtype=torch.float32, device="cpu") + def make_hook(pn): + def hook_fn(mod, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + hessians[pn] += (x.T @ x).cpu() + return hook_fn + hooks.append(module.register_forward_hook(make_hook(pname))) + model.eval() + with torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + for _ in range(num_batches): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + model(x, y) + for h in hooks: + h.remove() + for pn in hessians: + H = hessians[pn] + H /= num_batches + damp = 0.01 * torch.diag(H).mean().clamp_min(1e-6) + H += damp * torch.eye(H.shape[0]) + hessians[pn] = H + return hessians +def quantize_int6_gptq( + weight: Tensor, hessian: Tensor, clip_range: int = 31, block_size: int = 128, +) -> tuple[Tensor, Tensor]: + """Full GPTQ: Hessian-aware int6 quantization with Cholesky error compensation.""" + t32 = weight.float() + if t32.ndim != 2: + return quantize_int6_per_row(t32, clip_range) + rows, cols = t32.shape + H = hessian.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * torch.mean(torch.diag(H)) + H[torch.arange(cols, device=H.device), torch.arange(cols, device=H.device)] += damp + # Act-order: quantize most-activated columns first + perm = torch.argsort(torch.diag(H), descending=True) + inv_perm = torch.argsort(perm) + W = t32[:, perm].clone() + W[:, dead[perm]] = 0 + H = H[perm][:, perm] + # Cholesky of H^{-1} + try: + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + except RuntimeError: + # Extra damping fallback + H.diagonal().add_(damp * 10) + Hinv = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(Hinv) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + best_q, best_scale, best_err = None, None, float("inf") + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + sf = s.float() + Q = torch.zeros_like(W, dtype=torch.int8) + W_work = W.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + count = i2 - i1 + W1 = W_work[:, i1:i2].clone() + Q1 = torch.zeros(rows, count, dtype=torch.int8) + Err1 = torch.zeros(rows, count) + Hinv1 = Hinv[i1:i2, i1:i2] + for i in range(count): + w = W1[:, i] + d = Hinv1[i, i] + q = torch.clamp(torch.round(w / sf), -clip_range, clip_range).to(torch.int8) + Q1[:, i] = q + err = (w - q.float() * sf) / d + W1[:, i:] -= err.unsqueeze(1) * Hinv1[i, i:].unsqueeze(0) + Err1[:, i] = err + Q[:, i1:i2] = Q1 + if i2 < cols: + W_work[:, i2:] -= Err1 @ Hinv[i1:i2, i2:] + recon = Q.float() * sf[:, None] + mse = (W - recon).pow(2).mean().item() + if mse < best_err: + best_q, best_scale, best_err = Q, s, mse + best_q = best_q[:, inv_perm] + return best_q, best_scale +def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor] | None = None, + gptq_block_size: int = 128): + num_layers_total = max( + (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")), + default=0, + ) + 1 + late_k_layers = set(range(num_layers_total - 2, num_layers_total)) + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int6_cats and t.ndim >= 1: + H = hessians.get(name) if hessians else None + if H is not None and t.ndim == 2: + q, s = quantize_int6_gptq(t, H, block_size=gptq_block_size) + else: + q, s = quantize_int6_per_row(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + else: + q, s = quantize_float_tensor(t) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out def main() -> None: global zeropower_via_newtonschulz5 - code = Path(__file__).read_text(encoding="utf-8") args = Hyperparameters() zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) - - # ----------------------------- - # DISTRIBUTED + CUDA SETUP - # ----------------------------- - 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 +1284,18 @@ 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 +1304,6 @@ 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) @@ -792,16 +1313,10 @@ 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) @@ -811,18 +1326,16 @@ def log0(msg: str, console: bool = True) -> None: ) 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) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( sp, args.vocab_size, device ) log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") - - # ----------------------------- - # MODEL + OPTIMIZER SETUP - # ----------------------------- - + CastedLinear._qat_enabled = args.qat_enabled base_model = GPT( vocab_size=args.vocab_size, num_layers=args.num_layers, @@ -835,6 +1348,18 @@ 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, + use_vrl=args.vrl, ).to(device).bfloat16() for module in base_model.modules(): if isinstance(module, CastedLinear): @@ -842,18 +1367,14 @@ def log0(msg: str, console: bool = True) -> None: restore_low_dim_params_to_fp32(base_model) compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model - - # 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()) matrix_params = [ p for name, p in block_named_params if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) ] + if base_model.mtp_num_heads > 0: + matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2]) scalar_params = [ p for name, p in block_named_params @@ -861,11 +1382,27 @@ 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: + matrix_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + matrix_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, betas=(args.beta1, args.beta2), eps=args.adam_eps, + weight_decay=args.adam_wd, fused=True, ) optimizer_muon = Muon( @@ -873,13 +1410,15 @@ 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] @@ -891,9 +1430,14 @@ def log0(msg: str, console: bool = True) -> None: fused=True, ) optimizers.insert(1, optimizer_head) - n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + vrl_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_vrl] + log0(f"VRL:{args.vrl} active_layers:{vrl_layers}") log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") @@ -908,19 +1452,11 @@ 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 +1467,58 @@ 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.eval_only: + log0("eval_only:loading saved quantized model, skipping training + GPTQ") + quant_data = torch.load("final_int6_model.pt", map_location="cpu") + quant_result_eo, quant_meta_eo = quant_data["quantized"], quant_data["meta"] + sd_cpu_eo = base_model.state_dict() + sd_cpu_eo = {k: v.detach().cpu() for k, v in sd_cpu_eo.items()} + deq_state = dequantize_mixed_int6(quant_result_eo, quant_meta_eo, sd_cpu_eo) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + CastedLinear._qat_enabled = False + CastedLinear._soft_tau = 1000.0 + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + else: + log0("eval_only:TTT disabled, computing sliding window BPB") + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, eval_seq_len=args.train_seq_len, + ) + log0(f"eval_only_sliding val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if distributed: + dist.destroy_process_group() + return if args.warmup_steps > 0: initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] @@ -959,20 +1544,17 @@ def lr_mul(step: int, elapsed_ms: float) -> float: 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 + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 training_time_ms = 0.0 stop_after_step: int | None = None torch.cuda.synchronize() t0 = time.perf_counter() - step = 0 while True: last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) - should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) if should_validate: torch.cuda.synchronize() @@ -995,7 +1577,6 @@ 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 +1584,58 @@ 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}") + # Anneal soft-rounding temperature: hard for most of QAT, soft at the end + if CastedLinear._qat_enabled: + CastedLinear._soft_tau = 0.1 if scale < 0.02 else 1000.0 zero_grad_all() train_loss = torch.zeros((), device=device) for micro_step in range(grad_accum_steps): if distributed: model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + if args.ttt_burst_enabled and scale < args.ttt_burst_trigger: + if not hasattr(train_loader, '_ttt_buffer'): + train_loader._ttt_buffer = [] + train_loader._ttt_buffer.append((x.detach().clone(), y.detach().clone())) + if len(train_loader._ttt_buffer) > args.ttt_burst_steps: + train_loader._ttt_buffer.pop(0) with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): loss = model(x, y) train_loss += loss.detach() (loss * grad_scale).backward() train_loss /= grad_accum_steps - frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum for group in optimizer_muon.param_groups: group["momentum"] = muon_momentum - for opt in optimizers: for group in opt.param_groups: group["lr"] = group["base_lr"] * scale - if args.grad_clip_norm > 0: torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) for opt in optimizers: opt.step() zero_grad_all() - + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) step += 1 approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 should_log_train = ( args.train_log_every > 0 and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) @@ -1044,8 +1645,6 @@ 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 +1652,252 @@ 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. - + # === TTT BURST: Late-stage sharpening on recent training data === + if args.ttt_burst_enabled and hasattr(train_loader, '_ttt_buffer') and len(train_loader._ttt_buffer) > 0: + ttt_buffer = train_loader._ttt_buffer + log0(f"ttt_burst:start epochs:{args.ttt_burst_epochs} buffer_size:{len(ttt_buffer)} lr_factor:{args.ttt_burst_lr_factor}") + ttt_lr_scale = args.ttt_burst_lr_factor + for ttt_epoch in range(args.ttt_burst_epochs): + ttt_epoch_loss = 0.0 + for ttt_i, (bx, by) in enumerate(ttt_buffer): + zero_grad_all() + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * ttt_lr_scale + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + ttt_loss = model(bx, by) + (ttt_loss * grad_scale).backward() + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + for opt in optimizers: + opt.step() + zero_grad_all() + ttt_epoch_loss += ttt_loss.item() + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + log0(f"ttt_burst:epoch:{ttt_epoch + 1}/{args.ttt_burst_epochs} avg_loss:{ttt_epoch_loss / len(ttt_buffer):.4f}") + log0("ttt_burst:done") + + # Apply averaged weights: blend SWA (if available) with EMA + if swa_state is not None and swa_count > 0: + log0(f"swa:applying {swa_count} snapshots, blending with EMA (0.5/0.5)") + swa_avg = {name: (t / swa_count).to(device) for name, t in swa_state.items()} + current_state = base_model.state_dict() + avg_state = {} + for name in current_state: + ema_w = ema_state[name].to(dtype=current_state[name].dtype) + swa_w = swa_avg[name].to(dtype=current_state[name].dtype) + avg_state[name] = 0.5 * ema_w + 0.5 * swa_w + else: + log0("ema:applying EMA weights (no SWA snapshots)") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") if master_process: - torch.save(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()) + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ: collect Hessians for calibration-based quantization + hessians = None + if args.gptq_enabled: + log0(f"gptq:collecting hessians batches={args.gptq_calib_batches}") + t_hess = time.perf_counter() + calib_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) + hessians = collect_hessians( + base_model, calib_loader, args, device, grad_accum_steps, + num_batches=args.gptq_calib_batches, + ) + log0(f"gptq:hessians collected layers={len(hessians)} time={time.perf_counter() - t_hess:.1f}s") + del calib_loader + torch.cuda.empty_cache() + quant_result, quant_meta = mixed_quantize_int6( + sd_cpu, {"mlp", "attn"}, hessians=hessians, gptq_block_size=args.gptq_block_size, + ) + # Selective +/-1 pruning: zero out least-impactful quantized values to fit target size + target_bytes = 16_000_000 + code_bytes = len(code.encode("utf-8")) + target_model_bytes = target_bytes - code_bytes - 50_000 # headroom + def _serialize_and_compress(qr, qm): + buf = io.BytesIO() + torch.save({"w": qr, "m": qm}, buf) + return lzma.compress(buf.getvalue(), preset=6) + test_blob = _serialize_and_compress(quant_result, quant_meta) + log0(f"gptq:pre_prune artifact={len(test_blob)} target={target_model_bytes}") + if len(test_blob) > target_model_bytes: + # Collect all +/-1 values with Hessian-weighted cost + prune_candidates = [] + for name, info in quant_meta.items(): + if isinstance(info, dict) and info.get("type") == "int6": + qk = name + ".q" + sk = name + ".scale" + q, s = quant_result[qk], quant_result[sk] + H = hessians.get(name) if hessians else None + h_diag = torch.diag(H).float() if H is not None else None + mask = q.abs() == 1 + if mask.any(): + indices = mask.nonzero(as_tuple=False) + for idx in indices: + row = idx[0].item() + col = idx[1].item() if idx.ndim > 0 and len(idx) > 1 else 0 + sc = s[row].float().item() if s.ndim > 0 else s.float().item() + # Hessian-weighted cost: how sensitive is output to this weight? + cost = sc * sc * (h_diag[col].item() if h_diag is not None and col < len(h_diag) else 1.0) + prune_candidates.append((cost, qk, tuple(idx.tolist()))) + prune_candidates.sort(key=lambda x: x[0]) # ascending error = least impactful first + log0(f"gptq:pruning candidates={len(prune_candidates)}") + lo, hi = 0, len(prune_candidates) + best_n = 0 + while lo <= hi: + mid = (lo + hi) // 2 + if mid == 0: + lo = mid + 1 + continue + # Clone and zero + qr_test = {k: v.clone() for k, v in quant_result.items()} + for i in range(mid): + _, qk, idx = prune_candidates[i] + qr_test[qk][idx] = 0 + blob = _serialize_and_compress(qr_test, quant_meta) + if len(blob) <= target_model_bytes: + best_n = mid + hi = mid - 1 + else: + lo = mid + 1 + if best_n > 0: + for i in range(best_n): + _, qk, idx = prune_candidates[i] + quant_result[qk][idx] = 0 + log0(f"gptq:pruned {best_n} values") quant_buf = io.BytesIO() - torch.save(quant_obj, quant_buf) + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + # Save quantized model for fast eval-only iterations + if master_process: + torch.save({"quantized": quant_result, "meta": quant_meta}, "final_int6_model.pt") + log0(f"Saved quantized model to final_int6_model.pt") quant_raw = quant_buf.getvalue() - quant_blob = zlib.compress(quant_raw, level=9) - quant_raw_bytes = len(quant_raw) + quant_blob = lzma.compress(quant_raw, preset=9 | lzma.PRESET_EXTREME) 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") - 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") - + quant_file_bytes = len(quant_blob) + log0(f"Serialized model int6+lzma: {quant_file_bytes} bytes") + log0(f"Total submission size int6+lzma: {quant_file_bytes + code_bytes} bytes") if distributed: dist.barrier() - with open("final_model.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_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, # must match training model + rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + use_vrl=args.vrl, + ).to(device).bfloat16() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + CastedLinear._qat_enabled = False + CastedLinear._soft_tau = 1000.0 + compiled_eval = torch.compile(eval_model, dynamic=False, fullgraph=True) torch.cuda.synchronize() t_qeval = time.perf_counter() q_val_loss, q_val_bpb = eval_val( - args, - 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#461/549 recipe) + if args.ttt_enabled: + if distributed: + dist.barrier() + log0(f"ttt:start lr={args.ttt_lr} epochs={args.ttt_epochs} chunks={args.ttt_chunk_tokens}") + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_sliding_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, batch_seqs=32, log_fn=log0, + ) + log0(f"ttt:elapsed={time.perf_counter() - t_ttt:.1f}s") + log0(f"final_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f}") + log0(f"final_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() if distributed: dist.destroy_process_group() - - if __name__ == "__main__": main()