diff --git a/records/track_10min_16mb/2026-03-19_QAT_SlidingWindowEval/train_gpt.py b/records/track_10min_16mb/2026-03-19_QAT_SlidingWindowEval/train_gpt.py
new file mode 100644
index 000000000..64fd05b9c
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-19_QAT_SlidingWindowEval/train_gpt.py
@@ -0,0 +1,1475 @@
+"""
+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 math
+import os
+import random
+import subprocess
+import sys
+import time
+import uuid
+import zlib
+from pathlib import Path
+
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from torch import Tensor, nn
+from torch.nn.parallel import DistributedDataParallel as DDP
+
+# -----------------------------
+# HYPERPARAMETERS
+# -----------------------------
+# 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
+
+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.
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200))
+    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))
+    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_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_hidden = int(os.environ.get("MLP_HIDDEN", 0))  # 0 = use mlp_mult * model_dim
+    fp16_embed_export = bool(int(os.environ.get("FP16_EMBED_EXPORT", "0")))  # keep tok_emb in fp16 at export
+    num_loops = int(os.environ.get("NUM_LOOPS", 1))
+    lora_rank = int(os.environ.get("LORA_RANK", 0))
+    qat = bool(int(os.environ.get("QAT", "1")))
+    eval_stride = int(os.environ.get("EVAL_STRIDE", 64))
+    eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32))
+    doc_isolated_eval = bool(int(os.environ.get("DOC_ISOLATED_EVAL", "1")))  # eval per-document, no cross-doc context
+    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_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))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    lora_lr = float(os.environ.get("LORA_LR", 0.01))
+    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.
+    a, b, c = (3.4445, -4.7750, 2.0315)
+    X = G.bfloat16()
+    X /= X.norm() + eps
+    transposed = G.size(0) > G.size(1)
+    if transposed:
+        X = X.T
+    for _ in range(steps):
+        A = X @ X.T
+        B = b * A + c * A @ A
+        X = a * X + B @ X
+    return X.T if transposed else X
+
+
+class Muon(torch.optim.Optimizer):
+    def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov),
+        )
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                continue
+            lr = group["lr"]
+            momentum = group["momentum"]
+            backend_steps = group["backend_steps"]
+            nesterov = group["nesterov"]
+
+            total_params = sum(int(p.numel()) for p in params)
+            updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16)
+
+            curr = 0
+            for i, p in enumerate(params):
+                if i % world_size == rank and p.grad is not None:
+                    g = p.grad
+                    state = self.state[p]
+                    if "momentum_buffer" not in state:
+                        state["momentum_buffer"] = torch.zeros_like(g)
+                    buf = state["momentum_buffer"]
+                    buf.mul_(momentum).add_(g)
+                    if nesterov:
+                        g = g.add(buf, alpha=momentum)
+                    g = zeropower_via_newtonschulz5(g, steps=backend_steps)
+                    # 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()
+
+        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]:
+    sp_vocab_size = int(sp.vocab_size())
+    table_size = max(sp_vocab_size, vocab_size)
+    base_bytes_np = np.zeros((table_size,), dtype=np.int16)
+    has_leading_space_np = np.zeros((table_size,), dtype=np.bool_)
+    is_boundary_token_np = np.ones((table_size,), dtype=np.bool_)
+    for token_id in range(sp_vocab_size):
+        if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id):
+            continue
+        is_boundary_token_np[token_id] = False
+        if sp.is_byte(token_id):
+            base_bytes_np[token_id] = 1
+            continue
+        piece = sp.id_to_piece(token_id)
+        if piece.startswith("▁"):
+            has_leading_space_np[token_id] = True
+            piece = piece[1:]
+        base_bytes_np[token_id] = len(piece.encode("utf-8"))
+    return (
+        torch.tensor(base_bytes_np, dtype=torch.int16, device=device),
+        torch.tensor(has_leading_space_np, dtype=torch.bool, device=device),
+        torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device),
+    )
+
+
+def load_validation_tokens(pattern: str, seq_len: int) -> Tensor:
+    files = [Path(p) for p in sorted(glob.glob(pattern))]
+    if not files:
+        raise FileNotFoundError(f"No files found for pattern: {pattern}")
+    # 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,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    grad_accum_steps: int,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+) -> tuple[float, float]:
+    # Validation computes two metrics:
+    # - val_loss: token cross-entropy (natural log)
+    # - val_bpb: tokenizer-agnostic compression metric used by the challenge
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < args.train_seq_len:
+        raise ValueError(
+            "VAL_BATCH_SIZE must provide at least one sequence per rank; "
+            f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, "
+            f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // args.train_seq_len
+    total_seqs = (val_tokens.numel() - 1) // args.train_seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * args.train_seq_len
+            raw_end = batch_seq_end * args.train_seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, args.train_seq_len)
+            y = local[1:].reshape(-1, args.train_seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            val_loss_sum += batch_loss.to(torch.float64) * batch_token_count
+            val_token_count += batch_token_count
+            prev_ids = x.reshape(-1)
+            tgt_ids = y.reshape(-1)
+            token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16)
+            token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16)
+            val_byte_count += token_bytes.to(torch.float64).sum()
+
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM)
+
+    val_loss = val_loss_sum / val_token_count
+    bits_per_token = val_loss.item() / math.log(2.0)
+    tokens_per_byte = val_token_count.item() / val_byte_count.item()
+    model.train()
+    return float(val_loss.item()), float(bits_per_token * tokens_per_byte)
+
+# -----------------------------
+# 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",
+    ).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()
+            else torch.empty((t32.shape[0],), dtype=torch.float32)
+        )
+        clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None])
+        scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0)
+        q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous()
+        return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous()
+
+    # 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
+
+# Names of tensors to keep in fp16 at export instead of quantizing to int8.
+# Populated at runtime when fp16_embed_export=True.
+_FP16_EXPORT_NAMES: set[str] = set()
+
+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
+    # - fp16 passthrough for tensors in _FP16_EXPORT_NAMES (e.g. tied embeddings)
+    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
+
+        # FP16 export bypass: keep specified tensors (e.g. tied embeddings) in fp16
+        # instead of quantizing to int8. Avoids compounding int8 errors through both
+        # the input embedding and output projection paths.
+        if name in _FP16_EXPORT_NAMES:
+            kept = t.to(dtype=torch.float16).contiguous()
+            passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.")
+            passthrough[name] = kept
+            stats["int8_payload_bytes"] += tensor_nbytes(kept)
+            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("<i4").itemsize
+    token_bytes = np.dtype("<u2").itemsize
+    header = np.fromfile(file, dtype="<i4", count=256)
+    # SHARD HEADER INTS & SHARD_MAGIC
+    if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1:
+        raise ValueError(f"Unexpected shard header for {file}")
+    num_tokens = int(header[2])
+    expected_size = header_bytes + num_tokens * token_bytes
+    if file.stat().st_size != expected_size:
+        raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes")
+    tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes)
+    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:
+            raise FileNotFoundError(f"No files found for pattern: {pattern}")
+        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
+        while remaining > 0:
+            avail = self.tokens.numel() - self.pos
+            if avail <= 0:
+                self._advance_file()
+                continue
+            k = min(remaining, avail)
+            chunks.append(self.tokens[self.pos : self.pos + k])
+            self.pos += k
+            remaining -= k
+        return chunks[0] if len(chunks) == 1 else torch.cat(chunks)
+
+
+class DistributedTokenLoader:
+    # 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
+        chunk = self.stream.take(per_rank_span * self.world_size)
+        start = self.rank * per_rank_span
+        local = chunk[start : start + per_rank_span].to(dtype=torch.int64)
+        x = local[:-1].reshape(-1, seq_len)
+        y = local[1:].reshape(-1, seq_len)
+        return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True)
+
+# -----------------------------
+# TRANSFORMER MODULES
+# -----------------------------
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+
+def fake_quantize_int8_per_row(w: Tensor) -> Tensor:
+    """Simulate per-row int8 quantization with straight-through estimator.
+
+    Forward: uses quantized-then-dequantized weights matching the export pipeline exactly
+    (same INT8_CLIP_Q percentile clip, same per-row scale, same rounding as quantize_float_tensor).
+    Backward: gradients pass through as if no quantization happened (STE).
+    """
+    w32 = w.float()
+    clip_abs = torch.quantile(w32.abs(), INT8_CLIP_Q, dim=1)
+    scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0)
+    w_clipped = torch.clamp(w32, -clip_abs[:, None], clip_abs[:, None])
+    w_q = torch.clamp(torch.round(w_clipped / scale[:, None]), -127, 127)
+    w_deq = w_q * scale[:, None]
+    return w + (w_deq.to(w.dtype) - w).detach()
+
+
+class CastedLinear(nn.Linear):
+    # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute.
+    _qat: bool = False
+
+    def forward(self, x: Tensor) -> Tensor:
+        w = self.weight
+        if self._qat and self.training:
+            w = fake_quantize_int8_per_row(w)
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, w.to(x.dtype), bias)
+
+
+class AttentionLoRA(nn.Module):
+    """Per-iteration LoRA adapters for attention Q, K, V, and output projections.
+
+    Initialized so that the LoRA contribution is zero at the start of training
+    (B matrices are zeros). During training, the optimizer learns per-iteration
+    specialization while the base attention weights remain shared across loops.
+    """
+    def __init__(self, dim: int, kv_dim: int, rank: int):
+        super().__init__()
+        self.q_A = nn.Parameter(torch.empty(dim, rank))
+        self.q_B = nn.Parameter(torch.zeros(rank, dim))
+        self.k_A = nn.Parameter(torch.empty(dim, rank))
+        self.k_B = nn.Parameter(torch.zeros(rank, kv_dim))
+        self.v_A = nn.Parameter(torch.empty(dim, rank))
+        self.v_B = nn.Parameter(torch.zeros(rank, kv_dim))
+        self.proj_A = nn.Parameter(torch.empty(dim, rank))
+        self.proj_B = nn.Parameter(torch.zeros(rank, dim))
+        self._init_lora()
+
+    def _init_lora(self) -> None:
+        for name in ("q_A", "k_A", "v_A", "proj_A"):
+            nn.init.kaiming_uniform_(getattr(self, name), a=math.sqrt(5))
+
+
+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):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            or self._sin_cached is None
+            or self._seq_len_cached != seq_len
+            or self._cos_cached.device != device
+        ):
+            t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
+            freqs = torch.outer(t, self.inv_freq.to(device))
+            self._cos_cached = freqs.cos()[None, None, :, :]
+            self._sin_cached = freqs.sin()[None, None, :, :]
+            self._seq_len_cached = seq_len
+        return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype)
+
+
+def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor:
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+
+class CausalSelfAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_base: float,
+        qk_gain_init: float,
+    ):
+        super().__init__()
+        if dim % num_heads != 0:
+            raise ValueError("model_dim must be divisible by num_heads")
+        if num_heads % num_kv_heads != 0:
+            raise ValueError("num_heads must be divisible by num_kv_heads")
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = dim // num_heads
+        if self.head_dim % 2 != 0:
+            raise ValueError("head_dim must be even for RoPE")
+        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
+        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, lora: AttentionLoRA | None = None) -> Tensor:
+        bsz, seqlen, dim = x.shape
+        q = self.c_q(x)
+        k = self.c_k(x)
+        v = self.c_v(x)
+        if lora is not None:
+            # LoRA delta: (bsz, seqlen, dim) @ (dim, rank) @ (rank, out_dim)
+            # autocast handles fp32->bf16 cast of LoRA params automatically
+            q = q + (x @ lora.q_A) @ lora.q_B
+            k = k + (x @ lora.k_A) @ lora.k_B
+            v = v + (x @ lora.v_A) @ lora.v_B
+        q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2)
+        k = k.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        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)
+        out = self.proj(y)
+        if lora is not None:
+            out = out + (y @ lora.proj_A) @ lora.proj_B
+        return out
+
+
+class MLP(nn.Module):
+    # relu^2 MLP from the original modded-nanogpt setup
+    def __init__(self, dim: int, mlp_mult: int, mlp_hidden: int = 0):
+        super().__init__()
+        hidden = mlp_hidden if mlp_hidden > 0 else 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())
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        mlp_mult: int,
+        rope_base: float,
+        qk_gain_init: float,
+        mlp_hidden: int = 0,
+    ):
+        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.mlp = MLP(dim, mlp_mult, mlp_hidden=mlp_hidden)
+        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, lora: AttentionLoRA | None = None) -> 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), lora=lora)
+        x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out
+        x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x))
+        return x
+
+
+class GPT(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        num_layers: int,
+        model_dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        mlp_mult: int,
+        tie_embeddings: bool,
+        tied_embed_init_std: float,
+        logit_softcap: float,
+        rope_base: float,
+        qk_gain_init: float,
+        num_loops: int = 1,
+        lora_rank: int = 0,
+        mlp_hidden: int = 0,
+    ):
+        super().__init__()
+        if logit_softcap <= 0.0:
+            raise ValueError(f"logit_softcap must be positive, got {logit_softcap}")
+        self.tie_embeddings = tie_embeddings
+        self.tied_embed_init_std = tied_embed_init_std
+        self.logit_softcap = logit_softcap
+        self.num_unique_layers = num_layers
+        self.num_loops = num_loops
+        effective_depth = num_layers * num_loops
+        self.tok_emb = nn.Embedding(vocab_size, model_dim)
+        self.num_encoder_layers = effective_depth // 2
+        self.num_decoder_layers = effective_depth - self.num_encoder_layers
+        self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers)
+        self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32))
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    model_dim,
+                    num_heads,
+                    num_kv_heads,
+                    mlp_mult,
+                    rope_base,
+                    qk_gain_init,
+                    mlp_hidden=mlp_hidden,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        # Per-(loop, block) LoRA adapters for attention projections.
+        # Only created when num_loops > 1 and lora_rank > 0.
+        kv_dim = num_kv_heads * (model_dim // num_heads)
+        if lora_rank > 0 and num_loops > 1:
+            self.lora_adapters = nn.ModuleList(
+                [
+                    nn.ModuleList(
+                        [AttentionLoRA(model_dim, kv_dim, lora_rank) for _ in range(num_layers)]
+                    )
+                    for _ in range(num_loops)
+                ]
+            )
+        else:
+            self.lora_adapters = None
+        self.final_norm = RMSNorm()
+        self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False)
+        if self.lm_head is not None:
+            self.lm_head._zero_init = True
+        self._init_weights()
+
+    def _init_weights(self) -> None:
+        if self.tie_embeddings:
+            nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std)
+        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:
+        x = self.tok_emb(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x0 = x
+        skips: list[Tensor] = []
+
+        # Iterate through effective layers: each unique block is reused across loops.
+        # First half (encoder) stores skip connections; second half (decoder) pops them.
+        eff_idx = 0
+        for loop_idx in range(self.num_loops):
+            for block_idx in range(self.num_unique_layers):
+                lora = self.lora_adapters[loop_idx][block_idx] if self.lora_adapters is not None else None
+                if eff_idx < self.num_encoder_layers:
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                    skips.append(x)
+                else:
+                    dec_idx = eff_idx - self.num_encoder_layers
+                    if dec_idx < self.num_skip_weights and skips:
+                        x = x + self.skip_weights[dec_idx].to(dtype=x.dtype)[None, None, :] * skips.pop()
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                eff_idx += 1
+
+        x = self.final_norm(x).reshape(-1, x.size(-1))
+        targets = target_ids.reshape(-1)
+        if self.tie_embeddings:
+            logits_proj = F.linear(x, self.tok_emb.weight)
+        else:
+            if self.lm_head is None:
+                raise RuntimeError("lm_head is required when tie_embeddings=False")
+            logits_proj = self.lm_head(x)
+        logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap)
+        return F.cross_entropy(logits.float(), targets, reduction="mean")
+
+    def forward_logits(self, input_ids: Tensor) -> Tensor:
+        """Return logits (bsz, seq_len, vocab) without computing loss."""
+        x = self.tok_emb(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x0 = x
+        skips: list[Tensor] = []
+        eff_idx = 0
+        for loop_idx in range(self.num_loops):
+            for block_idx in range(self.num_unique_layers):
+                lora = self.lora_adapters[loop_idx][block_idx] if self.lora_adapters is not None else None
+                if eff_idx < self.num_encoder_layers:
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                    skips.append(x)
+                else:
+                    dec_idx = eff_idx - self.num_encoder_layers
+                    if dec_idx < self.num_skip_weights and skips:
+                        x = x + self.skip_weights[dec_idx].to(dtype=x.dtype)[None, None, :] * skips.pop()
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                eff_idx += 1
+        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)
+
+
+BOS_ID = 1  # SentencePiece BOS token ID
+
+def _find_docs(all_tokens: Tensor) -> list[tuple[int, int]]:
+    """Return (start, length) for each document, identified by BOS boundaries.
+
+    Each document starts at a BOS token and extends to just before the next BOS.
+    The last document extends to the end of the token stream.
+    """
+    bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].tolist()
+    if not bos_positions:
+        return [(0, all_tokens.numel())]
+    docs = []
+    for i in range(len(bos_positions)):
+        start = bos_positions[i]
+        end = bos_positions[i + 1] if i + 1 < len(bos_positions) else all_tokens.numel()
+        if end - start >= 2:  # need at least 2 tokens for (x, y) pair
+            docs.append((start, end - start))
+    return docs
+
+
+def _build_sliding_windows(
+    total_tokens: int, seq_len: int, stride: int
+) -> list[tuple[int, int]]:
+    """Return (window_start, score_start) pairs covering every token exactly once.
+
+    Every token in [0, total_tokens) is scored by exactly one window.
+    Full windows score their last `stride` positions (first window scores all seq_len).
+    One tail-aligned window covers any tokens beyond the last full window's end.
+    """
+    if total_tokens <= 0:
+        return []
+    if total_tokens <= seq_len:
+        return [(0, 0)]
+
+    windows: list[tuple[int, int]] = []
+    last_full_end = 0
+    ws = 0
+    while ws + seq_len <= total_tokens:
+        s = 0 if ws == 0 else seq_len - stride
+        windows.append((ws, s))
+        last_full_end = ws + seq_len
+        ws += stride
+
+    # One tail window ending exactly at total_tokens covers any remainder.
+    if last_full_end < total_tokens:
+        tail_ws = total_tokens - seq_len
+        tail_s = last_full_end - tail_ws  # skip already-scored prefix
+        windows.append((tail_ws, tail_s))
+
+    return windows
+
+
+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,
+) -> tuple[float, float]:
+    """Sliding window evaluation: each token scored with maximum context.
+
+    Windows of train_seq_len advance by `stride`. Only newly covered tokens
+    contribute to the score (first window scores all seq_len; non-first full
+    windows score the last `stride` tokens; one tail window covers any remainder).
+    Every validation token is counted exactly once.
+    """
+    seq_len = args.train_seq_len
+    total_tokens = val_tokens.numel() - 1
+
+    if args.doc_isolated_eval:
+        # Build windows per document — context never crosses document boundaries.
+        # Each window's (ws, s) is in absolute token-stream coordinates.
+        docs = _find_docs(val_tokens)
+        all_windows: list[tuple[int, int]] = []
+        for doc_start, doc_len in docs:
+            doc_pred_len = doc_len - 1  # number of prediction positions
+            doc_windows = _build_sliding_windows(doc_pred_len, seq_len, stride)
+            for ws, s in doc_windows:
+                all_windows.append((doc_start + ws, s))
+    else:
+        all_windows = _build_sliding_windows(total_tokens, seq_len, stride)
+    total_windows = len(all_windows)
+
+    # Distribute across ranks
+    my_s = (total_windows * rank) // world_size
+    my_e = (total_windows * (rank + 1)) // world_size
+    my_windows = all_windows[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()
+    with torch.inference_mode():
+        for bi in range(0, len(my_windows), batch_seqs):
+            batch_items = my_windows[bi:bi + batch_seqs]
+            bsz = len(batch_items)
+
+            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] = []
+            score_starts: list[int] = []
+
+            for i, (ws, s) in enumerate(batch_items):
+                end = min(ws + seq_len, total_tokens)
+                wlen = end - ws
+                wlens.append(wlen)
+                score_starts.append(s)
+                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 = 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, s) in enumerate(batch_items):
+                wlen = wlens[i]
+                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()
+
+            # Progress (rank 0 only)
+            if rank == 0 and (bi // batch_seqs) % 50 == 0:
+                done = min(bi + batch_seqs, len(my_windows))
+                pct = done / len(my_windows) * 100
+                running_bpb = 0.0
+                if token_count.item() > 0:
+                    rl = (loss_sum / token_count).item()
+                    running_bpb = rl / math.log(2.0) * (token_count.item() / byte_count.item())
+                print(f"  sliding_eval [{pct:5.1f}%] {done}/{len(my_windows)} windows running_bpb={running_bpb:.6f}", flush=True)
+
+    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
+
+
+# -----------------------------
+# TRAINING
+# -----------------------------
+
+def main() -> None:
+    global zeropower_via_newtonschulz5
+
+    code = Path(__file__).read_text(encoding="utf-8")
+    args = Hyperparameters()
+    zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5)
+
+    # -----------------------------
+    # DISTRIBUTED + 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"))
+    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
+    if world_size <= 0:
+        raise ValueError(f"WORLD_SIZE must be positive, got {world_size}")
+    if 8 % world_size != 0:
+        raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral")
+    grad_accum_steps = 8 // world_size
+    grad_scale = 1.0 / grad_accum_steps
+    if not torch.cuda.is_available():
+        raise RuntimeError("CUDA is required")
+    device = torch.device("cuda", local_rank)
+    torch.cuda.set_device(device)
+    if distributed:
+        dist.init_process_group(backend="nccl", device_id=device)
+        dist.barrier()
+    master_process = rank == 0
+
+    # 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
+        if console:
+            print(msg)
+        if logfile is not None:
+            with open(logfile, "a", encoding="utf-8") as f:
+                print(msg, file=f)
+
+    log0(code, console=False)
+    log0("=" * 100, console=False)
+    log0(f"Running Python {sys.version}", console=False)
+    log0(f"Running PyTorch {torch.__version__}", console=False)
+    log0(
+        subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout,
+        console=False,
+    )
+    log0("=" * 100, console=False)
+
+    # -----------------------------
+    # 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())}"
+        )
+    dataset_dir = Path(args.data_path).resolve()
+    actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin")))
+    val_tokens = load_validation_tokens(args.val_files, args.train_seq_len)
+    base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(
+        sp, args.vocab_size, device
+    )
+    log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}")
+    log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}")
+    log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}")
+
+    # -----------------------------
+    # MODEL + OPTIMIZER SETUP
+    # -----------------------------
+
+    base_model = GPT(
+        vocab_size=args.vocab_size,
+        num_layers=args.num_layers,
+        model_dim=args.model_dim,
+        num_heads=args.num_heads,
+        num_kv_heads=args.num_kv_heads,
+        mlp_mult=args.mlp_mult,
+        tie_embeddings=args.tie_embeddings,
+        tied_embed_init_std=args.tied_embed_init_std,
+        logit_softcap=args.logit_softcap,
+        rope_base=args.rope_base,
+        qk_gain_init=args.qk_gain_init,
+        num_loops=args.num_loops,
+        lora_rank=args.lora_rank,
+        mlp_hidden=args.mlp_hidden,
+    ).to(device).bfloat16()
+
+    # FP16 tied embedding export: skip int8 quantization for tok_emb.weight at export time.
+    # Avoids compounding int8 errors through both input embedding and output projection.
+    if args.fp16_embed_export and args.tie_embeddings:
+        _FP16_EXPORT_NAMES.add("tok_emb.weight")
+        log0(f"fp16_embed_export:enabled (tok_emb.weight kept in fp16, ~{args.vocab_size * args.model_dim * 2 / 1024:.0f}KB)")
+
+    for module in base_model.modules():
+        if isinstance(module, (CastedLinear, AttentionLoRA)):
+            module.float()
+    restore_low_dim_params_to_fp32(base_model)
+    log0(f"qat:{args.qat} (activates at 60% of iterations = step {int(args.iterations * 0.30)})")
+    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)
+    ]
+    scalar_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    if base_model.skip_weights.numel() > 0:
+        scalar_params.append(base_model.skip_weights)
+    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}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.Adam(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    if base_model.lora_adapters is not None:
+        lora_params = list(base_model.lora_adapters.parameters())
+        optimizer_lora = torch.optim.Adam(
+            [{"params": lora_params, "lr": args.lora_lr, "base_lr": args.lora_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.append(optimizer_lora)
+    if base_model.lm_head is not None:
+        optimizer_head = torch.optim.Adam(
+            [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.insert(1, optimizer_head)
+
+    n_params = sum(p.numel() for p in base_model.parameters())
+    n_lora = sum(p.numel() for p in base_model.lora_adapters.parameters()) if base_model.lora_adapters is not None else 0
+    effective_depth = args.num_layers * args.num_loops
+    log0(f"model_params:{n_params} (unique_layers:{args.num_layers} loops:{args.num_loops} effective_depth:{effective_depth} lora_rank:{args.lora_rank} lora_params:{n_lora})")
+    log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}")
+    log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False")
+    log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}")
+    log0(
+        f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} "
+        f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} "
+        f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}"
+    )
+    log0(
+        f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} "
+        f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} "
+        f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}"
+    )
+    log0(f"seed:{args.seed}")
+
+    # -----------------------------
+    # 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
+        if max_wallclock_ms is None:
+            warmdown_start = max(args.iterations - args.warmdown_iters, 0)
+            return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0
+        step_ms = elapsed_ms / max(step, 1)
+        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]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        # NOTE: QAT graph priming removed — it caused torch.compile to use a slower
+        # compilation path for the non-QAT forward pass (step_avg jumped from 44ms to 58ms).
+        # The one-time recompile when QAT activates (~30-90s) is cheaper than the cumulative
+        # overhead of a slower non-QAT path across thousands of steps.
+
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    # -----------------------------
+    # MAIN TRAINING LOOP
+    # -----------------------------
+
+    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()
+            training_time_ms += 1000.0 * (time.perf_counter() - t0)
+            val_loss, val_bpb = eval_val(
+                args,
+                model,
+                rank,
+                world_size,
+                device,
+                grad_accum_steps,
+                val_tokens,
+                base_bytes_lut,
+                has_leading_space_lut,
+                is_boundary_token_lut,
+            )
+            log0(
+                f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} "
+                f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms"
+            )
+            torch.cuda.synchronize()
+            t0 = time.perf_counter()
+
+        if last_step:
+            if stop_after_step is not None and step < args.iterations:
+                log0(
+                    f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms "
+                    f"step:{step}/{args.iterations}"
+                )
+            break
+
+        elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        scale = lr_mul(step, elapsed_ms)
+        zero_grad_all()
+        train_loss = torch.zeros((), device=device)
+        for micro_step in range(grad_accum_steps):
+            if distributed:
+                model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+            x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                loss = model(x, y)
+            train_loss += loss.detach()
+            (loss * grad_scale).backward()
+        train_loss /= grad_accum_steps
+
+        frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0
+        muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum
+        for group in optimizer_muon.param_groups:
+            group["momentum"] = muon_momentum
+
+        for opt in optimizers:
+            for group in opt.param_groups:
+                group["lr"] = group["base_lr"] * scale
+
+        if args.grad_clip_norm > 0:
+            torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm)
+        for opt in optimizers:
+            opt.step()
+        zero_grad_all()
+
+        step += 1
+
+        # Delayed QAT schedule: activate fake-quantize at 60% of training steps to
+        # avoid early-training instability before the loss landscape has settled.
+        if args.qat and step == int(args.iterations * 0.30):
+            for module in base_model.modules():
+                if isinstance(module, CastedLinear):
+                    module._qat = True
+            log0(f"qat_activated step:{step}/{args.iterations}")
+
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        if should_log_train:
+            log0(
+                f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} "
+                f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms"
+            )
+
+        # 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)
+            dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX)
+            reached_cap = bool(reached_cap_tensor.item())
+        if stop_after_step is None and reached_cap:
+            stop_after_step = step
+
+    log0(
+        f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+        f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB"
+    )
+
+    # -----------------------------
+    # 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.
+
+    if master_process:
+        torch.save(base_model.state_dict(), "final_model.pt")
+        model_bytes = os.path.getsize("final_model.pt")
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model: {model_bytes} bytes")
+        log0(f"Code size: {code_bytes} bytes")
+        log0(f"Total submission size: {model_bytes + code_bytes} bytes")
+
+    quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict())
+    quant_buf = io.BytesIO()
+    torch.save(quant_obj, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    quant_blob = zlib.compress(quant_raw, level=9)
+    quant_raw_bytes = len(quant_raw)
+    if master_process:
+        with open("final_model.int8.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")
+
+    if distributed:
+        dist.barrier()
+    with open("final_model.int8.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)
+    torch.cuda.synchronize()
+    t_qeval = time.perf_counter()
+    if args.eval_stride > 0 and args.eval_stride < args.train_seq_len:
+        log0(f"final_eval_mode:sliding_window stride:{args.eval_stride} batch_seqs:{args.eval_batch_seqs} doc_isolated:{args.doc_isolated_eval}")
+        q_val_loss, q_val_bpb = eval_val_sliding(
+            args,
+            base_model,
+            rank,
+            world_size,
+            device,
+            val_tokens,
+            base_bytes_lut,
+            has_leading_space_lut,
+            is_boundary_token_lut,
+            stride=args.eval_stride,
+            batch_seqs=args.eval_batch_seqs,
+        )
+    else:
+        log0("final_eval_mode:standard")
+        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,
+        )
+    torch.cuda.synchronize()
+    log0(
+        f"final_int8_zlib_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}")
+
+    if distributed:
+        dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/README.md b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/README.md
new file mode 100644
index 000000000..54618a216
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/README.md
@@ -0,0 +1,93 @@
+# 2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow
+
+*Non-record: Stacking proven community techniques with doc-isolated sliding window eval*
+
+**val_bpb: 1.2045** (post int8+zlib roundtrip, doc-isolated sliding window eval)
+
+---
+
+## Hardware note
+
+**This result is significantly limited by hardware speed.** Our RunPod 8×H100 SXM node ran at **70ms/step** — roughly 60% slower than the 43-44ms/step reported by the baseline and other entries. This is node-to-node variance within RunPod, not a code issue. The same configuration on a 44ms node would complete ~13,600 steps instead of our 8,528, likely scoring **~1.185-1.190**.
+
+We include this result as-is for transparency. The official evaluation on OpenAI's standardized hardware will produce the canonical score.
+
+---
+
+## Techniques
+
+All techniques are borrowed from published community entries. No novel contributions in this submission — it is a composition of proven ideas.
+
+| Technique | Source | Purpose |
+|-----------|--------|---------|
+| FP16 tied embedding export | WarmdownQuantization (#65), FP16Embed (#42) | Avoid int8 compounding through input+output paths |
+| MLP_HIDDEN=992 | WarmdownQuantization (#65) | Offset FP16 embed artifact cost |
+| WARMDOWN_ITERS=20000 | WarmdownQuantization (#65) | Tighter weight distributions → less quant damage |
+| TRAIN_SEQ_LEN=2048 | LongContextSeq2048 (#49) | Longer context during training |
+| Tuned LRs (0.06/0.06/0.07) | WarmdownQuantization (#65) | Optimal for aggressive warmdown |
+| GRAD_CLIP_NORM=1.0 | WarmdownQuantization (#65) | Training stability |
+| MUON_BACKEND_STEPS=5 | WarmdownQuantization (#65) | Better with aggressive warmdown |
+| MUON_MOMENTUM=0.99 | Int6 MLP3x STE QAT (#128) | Improved Muon convergence |
+| Sliding window eval (stride=64) | SlidingWindowEval (#50) | Maximum context for scoring |
+| Doc-isolated eval | LoRA TTT (#77) | No cross-document context bleed |
+
+---
+
+## Results
+
+| Metric | Naive Baseline | This submission |
+|--------|---------------|-----------------|
+| Steps completed | ~13,800 | 8,528 (hardware-limited) |
+| step_avg | 43.5ms | 70.4ms |
+| Pre-quant val_bpb | 1.2172 | 1.2154 |
+| **Post-quant val_bpb** | **1.2244** | **1.2045** |
+| Improvement | — | -0.0199 BPB |
+| Artifact bytes | 15,863,489 | 15,912,648 |
+| Eval time | ~16s | 43s |
+
+---
+
+## Configuration
+
+```
+VOCAB_SIZE=1024 NUM_LAYERS=9 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4
+MLP_HIDDEN=992 TIE_EMBEDDINGS=1 FP16_EMBED_EXPORT=1
+TRAIN_BATCH_TOKENS=524288 TRAIN_SEQ_LEN=2048
+ITERATIONS=20000 WARMDOWN_ITERS=20000 WARMUP_STEPS=20
+MAX_WALLCLOCK_SECONDS=600
+MATRIX_LR=0.06 SCALAR_LR=0.06 TIED_EMBED_LR=0.07
+GRAD_CLIP_NORM=1.0 MUON_BACKEND_STEPS=5 MUON_MOMENTUM=0.99
+QAT=0 EVAL_STRIDE=64 EVAL_BATCH_SEQS=32 DOC_ISOLATED_EVAL=1
+```
+
+## Command
+
+```bash
+export VOCAB_SIZE=1024 NUM_LAYERS=9 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4
+export MLP_HIDDEN=992 TIE_EMBEDDINGS=1 FP16_EMBED_EXPORT=1
+export TRAIN_BATCH_TOKENS=524288 TRAIN_SEQ_LEN=2048
+export ITERATIONS=20000 WARMDOWN_ITERS=20000 WARMUP_STEPS=20
+export MAX_WALLCLOCK_SECONDS=600
+export MATRIX_LR=0.06 SCALAR_LR=0.06 TIED_EMBED_LR=0.07
+export GRAD_CLIP_NORM=1.0 MUON_BACKEND_STEPS=5 MUON_MOMENTUM=0.99
+export QAT=0 EVAL_STRIDE=64 EVAL_BATCH_SEQS=32 DOC_ISOLATED_EVAL=1
+export SEED=1337 RUN_ID=kitchen_sink_9L_seed1337
+torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
+
+---
+
+## Acknowledgments
+
+This submission composes techniques entirely from community entries:
+- @samuellarson (WarmdownQuantization #65) — warmdown, FP16 embed, tuned LRs
+- @mattqlf (SlidingWindowEval #50) — sliding window evaluation
+- @samacquaviva (LoRA TTT #77) — doc-isolated evaluation concept
+- @spokane-way (LongContextSeq2048 #49) — seq2048 training
+- @rsavitt (Int6 MLP3x STE QAT #128) — MUON_MOMENTUM=0.99
+- Built with [Claude Code](https://claude.com/claude-code)
+
+## Author
+
+GitHub: [@mrdavtan](https://github.com/mrdavtan)
+Date: 2026-03-20
diff --git a/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/logs/kitchen_sink_9L_seed1337.txt b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/logs/kitchen_sink_9L_seed1337.txt
new file mode 100644
index 000000000..fa664f8c6
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/logs/kitchen_sink_9L_seed1337.txt
@@ -0,0 +1,120 @@
+logs/kitchen_sink_9L_seed1337.txt
+val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_1024_bpe.model
+train_loader:dataset:fineweb10B_sp1024 train_shards:80
+val_loader:shards pattern=./data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632
+fp16_embed_export:enabled (tok_emb.weight kept in fp16, ~1024KB)
+qat:False (activates at 60% of iterations = step 6000)
+model_params:16765000 (unique_layers:9 loops:1 effective_depth:9 lora_rank:0 lora_params:0)
+world_size:8 grad_accum_steps:1
+sdp_backends:cudnn=False flash=True mem_efficient=False math=False
+attention_mode:gqa num_heads:8 num_kv_heads:4
+tie_embeddings:True embed_lr:0.07 head_lr:0.0 matrix_lr:0.06 scalar_lr:0.06
+train_batch_tokens:524288 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000
+seed:1337
+warmup_step:1/20
+warmup_step:2/20
+warmup_step:3/20
+warmup_step:4/20
+warmup_step:5/20
+warmup_step:6/20
+warmup_step:7/20
+warmup_step:8/20
+warmup_step:9/20
+warmup_step:10/20
+warmup_step:11/20
+warmup_step:12/20
+warmup_step:13/20
+warmup_step:14/20
+warmup_step:15/20
+warmup_step:16/20
+warmup_step:17/20
+warmup_step:18/20
+warmup_step:19/20
+warmup_step:20/20
+step:1/20000 train_loss:6.9362 train_time:107ms step_avg:106.87ms
+step:2/20000 train_loss:19.5633 train_time:165ms step_avg:82.41ms
+step:3/20000 train_loss:17.0172 train_time:225ms step_avg:75.10ms
+step:4/20000 train_loss:13.0188 train_time:285ms step_avg:71.22ms
+step:5/20000 train_loss:9.0412 train_time:361ms step_avg:72.24ms
+step:6/20000 train_loss:7.9213 train_time:425ms step_avg:70.84ms
+step:7/20000 train_loss:6.3168 train_time:486ms step_avg:69.48ms
+step:8/20000 train_loss:6.1135 train_time:563ms step_avg:70.40ms
+step:9/20000 train_loss:6.0163 train_time:633ms step_avg:70.37ms
+step:10/20000 train_loss:5.9286 train_time:697ms step_avg:69.72ms
+step:200/20000 train_loss:2.7443 train_time:14145ms step_avg:70.73ms
+step:400/20000 train_loss:2.2778 train_time:27827ms step_avg:69.57ms
+step:600/20000 train_loss:2.5130 train_time:41885ms step_avg:69.81ms
+step:800/20000 train_loss:2.2471 train_time:55293ms step_avg:69.12ms
+step:1000/20000 train_loss:2.3379 train_time:69484ms step_avg:69.48ms
+step:1200/20000 train_loss:2.3580 train_time:83991ms step_avg:69.99ms
+step:1400/20000 train_loss:2.3786 train_time:98053ms step_avg:70.04ms
+step:1600/20000 train_loss:2.0441 train_time:111709ms step_avg:69.82ms
+step:1800/20000 train_loss:2.1618 train_time:125866ms step_avg:69.93ms
+step:2000/20000 train_loss:2.2072 train_time:140332ms step_avg:70.17ms
+step:2200/20000 train_loss:2.0254 train_time:154521ms step_avg:70.24ms
+step:2400/20000 train_loss:2.1598 train_time:168816ms step_avg:70.34ms
+step:2600/20000 train_loss:2.3709 train_time:182711ms step_avg:70.27ms
+step:2800/20000 train_loss:2.1925 train_time:198136ms step_avg:70.76ms
+step:3000/20000 train_loss:2.1835 train_time:212511ms step_avg:70.84ms
+step:3200/20000 train_loss:2.1439 train_time:226333ms step_avg:70.73ms
+step:3400/20000 train_loss:2.1130 train_time:239929ms step_avg:70.57ms
+step:3600/20000 train_loss:2.0557 train_time:253341ms step_avg:70.37ms
+step:3800/20000 train_loss:2.1645 train_time:267003ms step_avg:70.26ms
+step:4000/20000 train_loss:2.1249 train_time:281480ms step_avg:70.37ms
+step:4200/20000 train_loss:2.1205 train_time:298900ms step_avg:71.17ms
+step:4400/20000 train_loss:2.0552 train_time:312310ms step_avg:70.98ms
+step:4600/20000 train_loss:1.9263 train_time:326012ms step_avg:70.87ms
+step:4800/20000 train_loss:2.2045 train_time:339779ms step_avg:70.79ms
+step:5000/20000 train_loss:1.9633 train_time:352953ms step_avg:70.59ms
+step:5200/20000 train_loss:2.1193 train_time:366724ms step_avg:70.52ms
+step:5400/20000 train_loss:2.1392 train_time:380995ms step_avg:70.55ms
+step:5600/20000 train_loss:2.1235 train_time:395241ms step_avg:70.58ms
+step:5800/20000 train_loss:2.0798 train_time:408686ms step_avg:70.46ms
+step:6000/20000 train_loss:2.1596 train_time:422639ms step_avg:70.44ms
+step:6200/20000 train_loss:2.0288 train_time:436089ms step_avg:70.34ms
+step:6400/20000 train_loss:2.1074 train_time:449775ms step_avg:70.28ms
+step:6600/20000 train_loss:2.0614 train_time:463526ms step_avg:70.23ms
+step:6800/20000 train_loss:2.1256 train_time:477901ms step_avg:70.28ms
+step:7000/20000 train_loss:2.1672 train_time:492236ms step_avg:70.32ms
+step:7200/20000 train_loss:2.1355 train_time:506416ms step_avg:70.34ms
+step:7400/20000 train_loss:2.0652 train_time:519821ms step_avg:70.25ms
+step:7600/20000 train_loss:1.9445 train_time:533517ms step_avg:70.20ms
+step:7800/20000 train_loss:2.0846 train_time:546851ms step_avg:70.11ms
+step:8000/20000 train_loss:2.0529 train_time:560937ms step_avg:70.12ms
+step:8200/20000 train_loss:2.1287 train_time:574928ms step_avg:70.11ms
+step:8400/20000 train_loss:2.0735 train_time:593421ms step_avg:70.65ms
+step:8528/20000 val_loss:2.0521 val_bpb:1.2154 train_time:599996ms step_avg:70.36ms
+stopping_early: wallclock_cap train_time:599996ms step:8528/20000
+peak memory allocated: 10112 MiB reserved: 10304 MiB
+Serialized model: 66045335 bytes
+Code size: 63581 bytes
+Total submission size: 66108916 bytes
+Serialized model int8+zlib: 15849067 bytes (payload:17405664 raw_torch:17450459 payload_ratio:3.79x)
+Total submission size int8+zlib: 15912648 bytes
+final_eval_mode:sliding_window stride:64 batch_seqs:32 doc_isolated:True
+sliding_eval [  0.1%] 32/37402 windows running_bpb=1.269613
+sliding_eval [  4.4%] 1632/37402 windows running_bpb=1.215079
+sliding_eval [  8.6%] 3232/37402 windows running_bpb=1.221089
+sliding_eval [ 12.9%] 4832/37402 windows running_bpb=1.223454
+sliding_eval [ 17.2%] 6432/37402 windows running_bpb=1.219305
+sliding_eval [ 21.5%] 8032/37402 windows running_bpb=1.214802
+sliding_eval [ 25.8%] 9632/37402 windows running_bpb=1.220540
+sliding_eval [ 30.0%] 11232/37402 windows running_bpb=1.218193
+sliding_eval [ 34.3%] 12832/37402 windows running_bpb=1.217413
+sliding_eval [ 38.6%] 14432/37402 windows running_bpb=1.215383
+sliding_eval [ 42.9%] 16032/37402 windows running_bpb=1.214618
+sliding_eval [ 47.1%] 17632/37402 windows running_bpb=1.215056
+sliding_eval [ 51.4%] 19232/37402 windows running_bpb=1.214596
+sliding_eval [ 55.7%] 20832/37402 windows running_bpb=1.213343
+sliding_eval [ 60.0%] 22432/37402 windows running_bpb=1.211717
+sliding_eval [ 64.3%] 24032/37402 windows running_bpb=1.210711
+sliding_eval [ 68.5%] 25632/37402 windows running_bpb=1.211016
+sliding_eval [ 72.8%] 27232/37402 windows running_bpb=1.211956
+sliding_eval [ 77.1%] 28832/37402 windows running_bpb=1.213183
+sliding_eval [ 81.4%] 30432/37402 windows running_bpb=1.213285
+sliding_eval [ 85.6%] 32032/37402 windows running_bpb=1.214272
+sliding_eval [ 89.9%] 33632/37402 windows running_bpb=1.214467
+sliding_eval [ 94.2%] 35232/37402 windows running_bpb=1.212751
+sliding_eval [ 98.5%] 36832/37402 windows running_bpb=1.213596
+final_int8_zlib_roundtrip val_loss:2.0327 val_bpb:1.2045 eval_time:43360ms
+final_int8_zlib_roundtrip_exact val_loss:2.03272810 val_bpb:1.20453424
diff --git a/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/submission.json b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/submission.json
new file mode 100644
index 000000000..e79fe6dcf
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/submission.json
@@ -0,0 +1,39 @@
+{
+  "author": "mrdavtan",
+  "github_id": "mrdavtan",
+  "name": "Non-record: FP16 Embed + WD20k + Seq2048 + Sliding Window + Doc-Isolated Eval",
+  "blurb": "Stacks proven community techniques on 9L baseline: FP16 tied embedding export, aggressive warmdown (WD=20000), seq2048 training, tuned LRs, sliding window eval (stride=64) with doc-isolated scoring. Score limited by slow RunPod node (70ms/step vs typical 44ms). Expect ~1.185 on standard hardware.",
+  "date": "2026-03-20",
+  "val_loss": 2.03272810,
+  "val_bpb": 1.20453424,
+  "pre_quant_val_loss": 2.0521,
+  "pre_quant_val_bpb": 1.2154,
+  "step_stop": 8528,
+  "wallclock_seconds": 599.996,
+  "eval_time_seconds": 43.360,
+  "bytes_total": 15912648,
+  "bytes_model_int8_zlib": 15849067,
+  "bytes_code": 63581,
+  "hardware": "8xH100 SXM (RunPod), PyTorch 2.9.1+cu128, step_avg 70ms (slow node)",
+  "seed": 1337,
+  "track": "track_10min_16mb",
+  "model": {
+    "num_layers": 9,
+    "model_dim": 512,
+    "num_heads": 8,
+    "num_kv_heads": 4,
+    "mlp_hidden": 992,
+    "vocab_size": 1024,
+    "tie_embeddings": true
+  },
+  "changes_from_baseline": [
+    "FP16 tied embedding export (tok_emb.weight kept in fp16, avoids int8 compounding)",
+    "MLP_HIDDEN=992 (offset FP16 embed artifact cost)",
+    "WARMDOWN_ITERS=20000 (aggressive LR decay for tighter weight distributions)",
+    "TRAIN_SEQ_LEN=2048 (longer context training)",
+    "MATRIX_LR=0.06, SCALAR_LR=0.06, TIED_EMBED_LR=0.07 (tuned for WD20k)",
+    "GRAD_CLIP_NORM=1.0, MUON_BACKEND_STEPS=5, MUON_MOMENTUM=0.99",
+    "Sliding window eval stride=64 with doc-isolated scoring"
+  ],
+  "notes": "Score limited by hardware: 70ms/step on this RunPod node vs 44ms typical. Only 8,528 steps vs ~13,000 on standard hardware."
+}
diff --git a/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/train_gpt.py b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/train_gpt.py
new file mode 100644
index 000000000..64fd05b9c
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-20_FP16Embed_WD20k_Seq2048_SlidingWindow/train_gpt.py
@@ -0,0 +1,1475 @@
+"""
+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 math
+import os
+import random
+import subprocess
+import sys
+import time
+import uuid
+import zlib
+from pathlib import Path
+
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from torch import Tensor, nn
+from torch.nn.parallel import DistributedDataParallel as DDP
+
+# -----------------------------
+# HYPERPARAMETERS
+# -----------------------------
+# 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
+
+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.
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200))
+    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))
+    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_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_hidden = int(os.environ.get("MLP_HIDDEN", 0))  # 0 = use mlp_mult * model_dim
+    fp16_embed_export = bool(int(os.environ.get("FP16_EMBED_EXPORT", "0")))  # keep tok_emb in fp16 at export
+    num_loops = int(os.environ.get("NUM_LOOPS", 1))
+    lora_rank = int(os.environ.get("LORA_RANK", 0))
+    qat = bool(int(os.environ.get("QAT", "1")))
+    eval_stride = int(os.environ.get("EVAL_STRIDE", 64))
+    eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32))
+    doc_isolated_eval = bool(int(os.environ.get("DOC_ISOLATED_EVAL", "1")))  # eval per-document, no cross-doc context
+    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_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))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    lora_lr = float(os.environ.get("LORA_LR", 0.01))
+    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.
+    a, b, c = (3.4445, -4.7750, 2.0315)
+    X = G.bfloat16()
+    X /= X.norm() + eps
+    transposed = G.size(0) > G.size(1)
+    if transposed:
+        X = X.T
+    for _ in range(steps):
+        A = X @ X.T
+        B = b * A + c * A @ A
+        X = a * X + B @ X
+    return X.T if transposed else X
+
+
+class Muon(torch.optim.Optimizer):
+    def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov),
+        )
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                continue
+            lr = group["lr"]
+            momentum = group["momentum"]
+            backend_steps = group["backend_steps"]
+            nesterov = group["nesterov"]
+
+            total_params = sum(int(p.numel()) for p in params)
+            updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16)
+
+            curr = 0
+            for i, p in enumerate(params):
+                if i % world_size == rank and p.grad is not None:
+                    g = p.grad
+                    state = self.state[p]
+                    if "momentum_buffer" not in state:
+                        state["momentum_buffer"] = torch.zeros_like(g)
+                    buf = state["momentum_buffer"]
+                    buf.mul_(momentum).add_(g)
+                    if nesterov:
+                        g = g.add(buf, alpha=momentum)
+                    g = zeropower_via_newtonschulz5(g, steps=backend_steps)
+                    # 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()
+
+        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]:
+    sp_vocab_size = int(sp.vocab_size())
+    table_size = max(sp_vocab_size, vocab_size)
+    base_bytes_np = np.zeros((table_size,), dtype=np.int16)
+    has_leading_space_np = np.zeros((table_size,), dtype=np.bool_)
+    is_boundary_token_np = np.ones((table_size,), dtype=np.bool_)
+    for token_id in range(sp_vocab_size):
+        if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id):
+            continue
+        is_boundary_token_np[token_id] = False
+        if sp.is_byte(token_id):
+            base_bytes_np[token_id] = 1
+            continue
+        piece = sp.id_to_piece(token_id)
+        if piece.startswith("▁"):
+            has_leading_space_np[token_id] = True
+            piece = piece[1:]
+        base_bytes_np[token_id] = len(piece.encode("utf-8"))
+    return (
+        torch.tensor(base_bytes_np, dtype=torch.int16, device=device),
+        torch.tensor(has_leading_space_np, dtype=torch.bool, device=device),
+        torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device),
+    )
+
+
+def load_validation_tokens(pattern: str, seq_len: int) -> Tensor:
+    files = [Path(p) for p in sorted(glob.glob(pattern))]
+    if not files:
+        raise FileNotFoundError(f"No files found for pattern: {pattern}")
+    # 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,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    grad_accum_steps: int,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+) -> tuple[float, float]:
+    # Validation computes two metrics:
+    # - val_loss: token cross-entropy (natural log)
+    # - val_bpb: tokenizer-agnostic compression metric used by the challenge
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < args.train_seq_len:
+        raise ValueError(
+            "VAL_BATCH_SIZE must provide at least one sequence per rank; "
+            f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, "
+            f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // args.train_seq_len
+    total_seqs = (val_tokens.numel() - 1) // args.train_seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * args.train_seq_len
+            raw_end = batch_seq_end * args.train_seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, args.train_seq_len)
+            y = local[1:].reshape(-1, args.train_seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            val_loss_sum += batch_loss.to(torch.float64) * batch_token_count
+            val_token_count += batch_token_count
+            prev_ids = x.reshape(-1)
+            tgt_ids = y.reshape(-1)
+            token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16)
+            token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16)
+            val_byte_count += token_bytes.to(torch.float64).sum()
+
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM)
+
+    val_loss = val_loss_sum / val_token_count
+    bits_per_token = val_loss.item() / math.log(2.0)
+    tokens_per_byte = val_token_count.item() / val_byte_count.item()
+    model.train()
+    return float(val_loss.item()), float(bits_per_token * tokens_per_byte)
+
+# -----------------------------
+# 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",
+    ).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()
+            else torch.empty((t32.shape[0],), dtype=torch.float32)
+        )
+        clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None])
+        scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0)
+        q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous()
+        return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous()
+
+    # 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
+
+# Names of tensors to keep in fp16 at export instead of quantizing to int8.
+# Populated at runtime when fp16_embed_export=True.
+_FP16_EXPORT_NAMES: set[str] = set()
+
+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
+    # - fp16 passthrough for tensors in _FP16_EXPORT_NAMES (e.g. tied embeddings)
+    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
+
+        # FP16 export bypass: keep specified tensors (e.g. tied embeddings) in fp16
+        # instead of quantizing to int8. Avoids compounding int8 errors through both
+        # the input embedding and output projection paths.
+        if name in _FP16_EXPORT_NAMES:
+            kept = t.to(dtype=torch.float16).contiguous()
+            passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.")
+            passthrough[name] = kept
+            stats["int8_payload_bytes"] += tensor_nbytes(kept)
+            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("<i4").itemsize
+    token_bytes = np.dtype("<u2").itemsize
+    header = np.fromfile(file, dtype="<i4", count=256)
+    # SHARD HEADER INTS & SHARD_MAGIC
+    if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1:
+        raise ValueError(f"Unexpected shard header for {file}")
+    num_tokens = int(header[2])
+    expected_size = header_bytes + num_tokens * token_bytes
+    if file.stat().st_size != expected_size:
+        raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes")
+    tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes)
+    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:
+            raise FileNotFoundError(f"No files found for pattern: {pattern}")
+        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
+        while remaining > 0:
+            avail = self.tokens.numel() - self.pos
+            if avail <= 0:
+                self._advance_file()
+                continue
+            k = min(remaining, avail)
+            chunks.append(self.tokens[self.pos : self.pos + k])
+            self.pos += k
+            remaining -= k
+        return chunks[0] if len(chunks) == 1 else torch.cat(chunks)
+
+
+class DistributedTokenLoader:
+    # 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
+        chunk = self.stream.take(per_rank_span * self.world_size)
+        start = self.rank * per_rank_span
+        local = chunk[start : start + per_rank_span].to(dtype=torch.int64)
+        x = local[:-1].reshape(-1, seq_len)
+        y = local[1:].reshape(-1, seq_len)
+        return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True)
+
+# -----------------------------
+# TRANSFORMER MODULES
+# -----------------------------
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+
+def fake_quantize_int8_per_row(w: Tensor) -> Tensor:
+    """Simulate per-row int8 quantization with straight-through estimator.
+
+    Forward: uses quantized-then-dequantized weights matching the export pipeline exactly
+    (same INT8_CLIP_Q percentile clip, same per-row scale, same rounding as quantize_float_tensor).
+    Backward: gradients pass through as if no quantization happened (STE).
+    """
+    w32 = w.float()
+    clip_abs = torch.quantile(w32.abs(), INT8_CLIP_Q, dim=1)
+    scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0)
+    w_clipped = torch.clamp(w32, -clip_abs[:, None], clip_abs[:, None])
+    w_q = torch.clamp(torch.round(w_clipped / scale[:, None]), -127, 127)
+    w_deq = w_q * scale[:, None]
+    return w + (w_deq.to(w.dtype) - w).detach()
+
+
+class CastedLinear(nn.Linear):
+    # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute.
+    _qat: bool = False
+
+    def forward(self, x: Tensor) -> Tensor:
+        w = self.weight
+        if self._qat and self.training:
+            w = fake_quantize_int8_per_row(w)
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, w.to(x.dtype), bias)
+
+
+class AttentionLoRA(nn.Module):
+    """Per-iteration LoRA adapters for attention Q, K, V, and output projections.
+
+    Initialized so that the LoRA contribution is zero at the start of training
+    (B matrices are zeros). During training, the optimizer learns per-iteration
+    specialization while the base attention weights remain shared across loops.
+    """
+    def __init__(self, dim: int, kv_dim: int, rank: int):
+        super().__init__()
+        self.q_A = nn.Parameter(torch.empty(dim, rank))
+        self.q_B = nn.Parameter(torch.zeros(rank, dim))
+        self.k_A = nn.Parameter(torch.empty(dim, rank))
+        self.k_B = nn.Parameter(torch.zeros(rank, kv_dim))
+        self.v_A = nn.Parameter(torch.empty(dim, rank))
+        self.v_B = nn.Parameter(torch.zeros(rank, kv_dim))
+        self.proj_A = nn.Parameter(torch.empty(dim, rank))
+        self.proj_B = nn.Parameter(torch.zeros(rank, dim))
+        self._init_lora()
+
+    def _init_lora(self) -> None:
+        for name in ("q_A", "k_A", "v_A", "proj_A"):
+            nn.init.kaiming_uniform_(getattr(self, name), a=math.sqrt(5))
+
+
+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):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            or self._sin_cached is None
+            or self._seq_len_cached != seq_len
+            or self._cos_cached.device != device
+        ):
+            t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
+            freqs = torch.outer(t, self.inv_freq.to(device))
+            self._cos_cached = freqs.cos()[None, None, :, :]
+            self._sin_cached = freqs.sin()[None, None, :, :]
+            self._seq_len_cached = seq_len
+        return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype)
+
+
+def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor:
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+
+class CausalSelfAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_base: float,
+        qk_gain_init: float,
+    ):
+        super().__init__()
+        if dim % num_heads != 0:
+            raise ValueError("model_dim must be divisible by num_heads")
+        if num_heads % num_kv_heads != 0:
+            raise ValueError("num_heads must be divisible by num_kv_heads")
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = dim // num_heads
+        if self.head_dim % 2 != 0:
+            raise ValueError("head_dim must be even for RoPE")
+        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
+        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, lora: AttentionLoRA | None = None) -> Tensor:
+        bsz, seqlen, dim = x.shape
+        q = self.c_q(x)
+        k = self.c_k(x)
+        v = self.c_v(x)
+        if lora is not None:
+            # LoRA delta: (bsz, seqlen, dim) @ (dim, rank) @ (rank, out_dim)
+            # autocast handles fp32->bf16 cast of LoRA params automatically
+            q = q + (x @ lora.q_A) @ lora.q_B
+            k = k + (x @ lora.k_A) @ lora.k_B
+            v = v + (x @ lora.v_A) @ lora.v_B
+        q = q.reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2)
+        k = k.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        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)
+        out = self.proj(y)
+        if lora is not None:
+            out = out + (y @ lora.proj_A) @ lora.proj_B
+        return out
+
+
+class MLP(nn.Module):
+    # relu^2 MLP from the original modded-nanogpt setup
+    def __init__(self, dim: int, mlp_mult: int, mlp_hidden: int = 0):
+        super().__init__()
+        hidden = mlp_hidden if mlp_hidden > 0 else 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())
+
+
+class Block(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        mlp_mult: int,
+        rope_base: float,
+        qk_gain_init: float,
+        mlp_hidden: int = 0,
+    ):
+        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.mlp = MLP(dim, mlp_mult, mlp_hidden=mlp_hidden)
+        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, lora: AttentionLoRA | None = None) -> 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), lora=lora)
+        x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out
+        x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x))
+        return x
+
+
+class GPT(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        num_layers: int,
+        model_dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        mlp_mult: int,
+        tie_embeddings: bool,
+        tied_embed_init_std: float,
+        logit_softcap: float,
+        rope_base: float,
+        qk_gain_init: float,
+        num_loops: int = 1,
+        lora_rank: int = 0,
+        mlp_hidden: int = 0,
+    ):
+        super().__init__()
+        if logit_softcap <= 0.0:
+            raise ValueError(f"logit_softcap must be positive, got {logit_softcap}")
+        self.tie_embeddings = tie_embeddings
+        self.tied_embed_init_std = tied_embed_init_std
+        self.logit_softcap = logit_softcap
+        self.num_unique_layers = num_layers
+        self.num_loops = num_loops
+        effective_depth = num_layers * num_loops
+        self.tok_emb = nn.Embedding(vocab_size, model_dim)
+        self.num_encoder_layers = effective_depth // 2
+        self.num_decoder_layers = effective_depth - self.num_encoder_layers
+        self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers)
+        self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32))
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    model_dim,
+                    num_heads,
+                    num_kv_heads,
+                    mlp_mult,
+                    rope_base,
+                    qk_gain_init,
+                    mlp_hidden=mlp_hidden,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        # Per-(loop, block) LoRA adapters for attention projections.
+        # Only created when num_loops > 1 and lora_rank > 0.
+        kv_dim = num_kv_heads * (model_dim // num_heads)
+        if lora_rank > 0 and num_loops > 1:
+            self.lora_adapters = nn.ModuleList(
+                [
+                    nn.ModuleList(
+                        [AttentionLoRA(model_dim, kv_dim, lora_rank) for _ in range(num_layers)]
+                    )
+                    for _ in range(num_loops)
+                ]
+            )
+        else:
+            self.lora_adapters = None
+        self.final_norm = RMSNorm()
+        self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False)
+        if self.lm_head is not None:
+            self.lm_head._zero_init = True
+        self._init_weights()
+
+    def _init_weights(self) -> None:
+        if self.tie_embeddings:
+            nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std)
+        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:
+        x = self.tok_emb(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x0 = x
+        skips: list[Tensor] = []
+
+        # Iterate through effective layers: each unique block is reused across loops.
+        # First half (encoder) stores skip connections; second half (decoder) pops them.
+        eff_idx = 0
+        for loop_idx in range(self.num_loops):
+            for block_idx in range(self.num_unique_layers):
+                lora = self.lora_adapters[loop_idx][block_idx] if self.lora_adapters is not None else None
+                if eff_idx < self.num_encoder_layers:
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                    skips.append(x)
+                else:
+                    dec_idx = eff_idx - self.num_encoder_layers
+                    if dec_idx < self.num_skip_weights and skips:
+                        x = x + self.skip_weights[dec_idx].to(dtype=x.dtype)[None, None, :] * skips.pop()
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                eff_idx += 1
+
+        x = self.final_norm(x).reshape(-1, x.size(-1))
+        targets = target_ids.reshape(-1)
+        if self.tie_embeddings:
+            logits_proj = F.linear(x, self.tok_emb.weight)
+        else:
+            if self.lm_head is None:
+                raise RuntimeError("lm_head is required when tie_embeddings=False")
+            logits_proj = self.lm_head(x)
+        logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap)
+        return F.cross_entropy(logits.float(), targets, reduction="mean")
+
+    def forward_logits(self, input_ids: Tensor) -> Tensor:
+        """Return logits (bsz, seq_len, vocab) without computing loss."""
+        x = self.tok_emb(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x0 = x
+        skips: list[Tensor] = []
+        eff_idx = 0
+        for loop_idx in range(self.num_loops):
+            for block_idx in range(self.num_unique_layers):
+                lora = self.lora_adapters[loop_idx][block_idx] if self.lora_adapters is not None else None
+                if eff_idx < self.num_encoder_layers:
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                    skips.append(x)
+                else:
+                    dec_idx = eff_idx - self.num_encoder_layers
+                    if dec_idx < self.num_skip_weights and skips:
+                        x = x + self.skip_weights[dec_idx].to(dtype=x.dtype)[None, None, :] * skips.pop()
+                    x = self.blocks[block_idx](x, x0, lora=lora)
+                eff_idx += 1
+        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)
+
+
+BOS_ID = 1  # SentencePiece BOS token ID
+
+def _find_docs(all_tokens: Tensor) -> list[tuple[int, int]]:
+    """Return (start, length) for each document, identified by BOS boundaries.
+
+    Each document starts at a BOS token and extends to just before the next BOS.
+    The last document extends to the end of the token stream.
+    """
+    bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].tolist()
+    if not bos_positions:
+        return [(0, all_tokens.numel())]
+    docs = []
+    for i in range(len(bos_positions)):
+        start = bos_positions[i]
+        end = bos_positions[i + 1] if i + 1 < len(bos_positions) else all_tokens.numel()
+        if end - start >= 2:  # need at least 2 tokens for (x, y) pair
+            docs.append((start, end - start))
+    return docs
+
+
+def _build_sliding_windows(
+    total_tokens: int, seq_len: int, stride: int
+) -> list[tuple[int, int]]:
+    """Return (window_start, score_start) pairs covering every token exactly once.
+
+    Every token in [0, total_tokens) is scored by exactly one window.
+    Full windows score their last `stride` positions (first window scores all seq_len).
+    One tail-aligned window covers any tokens beyond the last full window's end.
+    """
+    if total_tokens <= 0:
+        return []
+    if total_tokens <= seq_len:
+        return [(0, 0)]
+
+    windows: list[tuple[int, int]] = []
+    last_full_end = 0
+    ws = 0
+    while ws + seq_len <= total_tokens:
+        s = 0 if ws == 0 else seq_len - stride
+        windows.append((ws, s))
+        last_full_end = ws + seq_len
+        ws += stride
+
+    # One tail window ending exactly at total_tokens covers any remainder.
+    if last_full_end < total_tokens:
+        tail_ws = total_tokens - seq_len
+        tail_s = last_full_end - tail_ws  # skip already-scored prefix
+        windows.append((tail_ws, tail_s))
+
+    return windows
+
+
+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,
+) -> tuple[float, float]:
+    """Sliding window evaluation: each token scored with maximum context.
+
+    Windows of train_seq_len advance by `stride`. Only newly covered tokens
+    contribute to the score (first window scores all seq_len; non-first full
+    windows score the last `stride` tokens; one tail window covers any remainder).
+    Every validation token is counted exactly once.
+    """
+    seq_len = args.train_seq_len
+    total_tokens = val_tokens.numel() - 1
+
+    if args.doc_isolated_eval:
+        # Build windows per document — context never crosses document boundaries.
+        # Each window's (ws, s) is in absolute token-stream coordinates.
+        docs = _find_docs(val_tokens)
+        all_windows: list[tuple[int, int]] = []
+        for doc_start, doc_len in docs:
+            doc_pred_len = doc_len - 1  # number of prediction positions
+            doc_windows = _build_sliding_windows(doc_pred_len, seq_len, stride)
+            for ws, s in doc_windows:
+                all_windows.append((doc_start + ws, s))
+    else:
+        all_windows = _build_sliding_windows(total_tokens, seq_len, stride)
+    total_windows = len(all_windows)
+
+    # Distribute across ranks
+    my_s = (total_windows * rank) // world_size
+    my_e = (total_windows * (rank + 1)) // world_size
+    my_windows = all_windows[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()
+    with torch.inference_mode():
+        for bi in range(0, len(my_windows), batch_seqs):
+            batch_items = my_windows[bi:bi + batch_seqs]
+            bsz = len(batch_items)
+
+            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] = []
+            score_starts: list[int] = []
+
+            for i, (ws, s) in enumerate(batch_items):
+                end = min(ws + seq_len, total_tokens)
+                wlen = end - ws
+                wlens.append(wlen)
+                score_starts.append(s)
+                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 = 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, s) in enumerate(batch_items):
+                wlen = wlens[i]
+                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()
+
+            # Progress (rank 0 only)
+            if rank == 0 and (bi // batch_seqs) % 50 == 0:
+                done = min(bi + batch_seqs, len(my_windows))
+                pct = done / len(my_windows) * 100
+                running_bpb = 0.0
+                if token_count.item() > 0:
+                    rl = (loss_sum / token_count).item()
+                    running_bpb = rl / math.log(2.0) * (token_count.item() / byte_count.item())
+                print(f"  sliding_eval [{pct:5.1f}%] {done}/{len(my_windows)} windows running_bpb={running_bpb:.6f}", flush=True)
+
+    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
+
+
+# -----------------------------
+# TRAINING
+# -----------------------------
+
+def main() -> None:
+    global zeropower_via_newtonschulz5
+
+    code = Path(__file__).read_text(encoding="utf-8")
+    args = Hyperparameters()
+    zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5)
+
+    # -----------------------------
+    # DISTRIBUTED + 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"))
+    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
+    if world_size <= 0:
+        raise ValueError(f"WORLD_SIZE must be positive, got {world_size}")
+    if 8 % world_size != 0:
+        raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral")
+    grad_accum_steps = 8 // world_size
+    grad_scale = 1.0 / grad_accum_steps
+    if not torch.cuda.is_available():
+        raise RuntimeError("CUDA is required")
+    device = torch.device("cuda", local_rank)
+    torch.cuda.set_device(device)
+    if distributed:
+        dist.init_process_group(backend="nccl", device_id=device)
+        dist.barrier()
+    master_process = rank == 0
+
+    # 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
+        if console:
+            print(msg)
+        if logfile is not None:
+            with open(logfile, "a", encoding="utf-8") as f:
+                print(msg, file=f)
+
+    log0(code, console=False)
+    log0("=" * 100, console=False)
+    log0(f"Running Python {sys.version}", console=False)
+    log0(f"Running PyTorch {torch.__version__}", console=False)
+    log0(
+        subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout,
+        console=False,
+    )
+    log0("=" * 100, console=False)
+
+    # -----------------------------
+    # 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())}"
+        )
+    dataset_dir = Path(args.data_path).resolve()
+    actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin")))
+    val_tokens = load_validation_tokens(args.val_files, args.train_seq_len)
+    base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(
+        sp, args.vocab_size, device
+    )
+    log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}")
+    log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}")
+    log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}")
+
+    # -----------------------------
+    # MODEL + OPTIMIZER SETUP
+    # -----------------------------
+
+    base_model = GPT(
+        vocab_size=args.vocab_size,
+        num_layers=args.num_layers,
+        model_dim=args.model_dim,
+        num_heads=args.num_heads,
+        num_kv_heads=args.num_kv_heads,
+        mlp_mult=args.mlp_mult,
+        tie_embeddings=args.tie_embeddings,
+        tied_embed_init_std=args.tied_embed_init_std,
+        logit_softcap=args.logit_softcap,
+        rope_base=args.rope_base,
+        qk_gain_init=args.qk_gain_init,
+        num_loops=args.num_loops,
+        lora_rank=args.lora_rank,
+        mlp_hidden=args.mlp_hidden,
+    ).to(device).bfloat16()
+
+    # FP16 tied embedding export: skip int8 quantization for tok_emb.weight at export time.
+    # Avoids compounding int8 errors through both input embedding and output projection.
+    if args.fp16_embed_export and args.tie_embeddings:
+        _FP16_EXPORT_NAMES.add("tok_emb.weight")
+        log0(f"fp16_embed_export:enabled (tok_emb.weight kept in fp16, ~{args.vocab_size * args.model_dim * 2 / 1024:.0f}KB)")
+
+    for module in base_model.modules():
+        if isinstance(module, (CastedLinear, AttentionLoRA)):
+            module.float()
+    restore_low_dim_params_to_fp32(base_model)
+    log0(f"qat:{args.qat} (activates at 60% of iterations = step {int(args.iterations * 0.30)})")
+    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)
+    ]
+    scalar_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    if base_model.skip_weights.numel() > 0:
+        scalar_params.append(base_model.skip_weights)
+    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}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.Adam(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    if base_model.lora_adapters is not None:
+        lora_params = list(base_model.lora_adapters.parameters())
+        optimizer_lora = torch.optim.Adam(
+            [{"params": lora_params, "lr": args.lora_lr, "base_lr": args.lora_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.append(optimizer_lora)
+    if base_model.lm_head is not None:
+        optimizer_head = torch.optim.Adam(
+            [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.insert(1, optimizer_head)
+
+    n_params = sum(p.numel() for p in base_model.parameters())
+    n_lora = sum(p.numel() for p in base_model.lora_adapters.parameters()) if base_model.lora_adapters is not None else 0
+    effective_depth = args.num_layers * args.num_loops
+    log0(f"model_params:{n_params} (unique_layers:{args.num_layers} loops:{args.num_loops} effective_depth:{effective_depth} lora_rank:{args.lora_rank} lora_params:{n_lora})")
+    log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}")
+    log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False")
+    log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}")
+    log0(
+        f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} "
+        f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} "
+        f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}"
+    )
+    log0(
+        f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} "
+        f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} "
+        f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}"
+    )
+    log0(f"seed:{args.seed}")
+
+    # -----------------------------
+    # 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
+        if max_wallclock_ms is None:
+            warmdown_start = max(args.iterations - args.warmdown_iters, 0)
+            return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0
+        step_ms = elapsed_ms / max(step, 1)
+        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]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        # NOTE: QAT graph priming removed — it caused torch.compile to use a slower
+        # compilation path for the non-QAT forward pass (step_avg jumped from 44ms to 58ms).
+        # The one-time recompile when QAT activates (~30-90s) is cheaper than the cumulative
+        # overhead of a slower non-QAT path across thousands of steps.
+
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    # -----------------------------
+    # MAIN TRAINING LOOP
+    # -----------------------------
+
+    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()
+            training_time_ms += 1000.0 * (time.perf_counter() - t0)
+            val_loss, val_bpb = eval_val(
+                args,
+                model,
+                rank,
+                world_size,
+                device,
+                grad_accum_steps,
+                val_tokens,
+                base_bytes_lut,
+                has_leading_space_lut,
+                is_boundary_token_lut,
+            )
+            log0(
+                f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} "
+                f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms"
+            )
+            torch.cuda.synchronize()
+            t0 = time.perf_counter()
+
+        if last_step:
+            if stop_after_step is not None and step < args.iterations:
+                log0(
+                    f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms "
+                    f"step:{step}/{args.iterations}"
+                )
+            break
+
+        elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        scale = lr_mul(step, elapsed_ms)
+        zero_grad_all()
+        train_loss = torch.zeros((), device=device)
+        for micro_step in range(grad_accum_steps):
+            if distributed:
+                model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+            x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                loss = model(x, y)
+            train_loss += loss.detach()
+            (loss * grad_scale).backward()
+        train_loss /= grad_accum_steps
+
+        frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0
+        muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum
+        for group in optimizer_muon.param_groups:
+            group["momentum"] = muon_momentum
+
+        for opt in optimizers:
+            for group in opt.param_groups:
+                group["lr"] = group["base_lr"] * scale
+
+        if args.grad_clip_norm > 0:
+            torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm)
+        for opt in optimizers:
+            opt.step()
+        zero_grad_all()
+
+        step += 1
+
+        # Delayed QAT schedule: activate fake-quantize at 60% of training steps to
+        # avoid early-training instability before the loss landscape has settled.
+        if args.qat and step == int(args.iterations * 0.30):
+            for module in base_model.modules():
+                if isinstance(module, CastedLinear):
+                    module._qat = True
+            log0(f"qat_activated step:{step}/{args.iterations}")
+
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        if should_log_train:
+            log0(
+                f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} "
+                f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms"
+            )
+
+        # 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)
+            dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX)
+            reached_cap = bool(reached_cap_tensor.item())
+        if stop_after_step is None and reached_cap:
+            stop_after_step = step
+
+    log0(
+        f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+        f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB"
+    )
+
+    # -----------------------------
+    # 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.
+
+    if master_process:
+        torch.save(base_model.state_dict(), "final_model.pt")
+        model_bytes = os.path.getsize("final_model.pt")
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model: {model_bytes} bytes")
+        log0(f"Code size: {code_bytes} bytes")
+        log0(f"Total submission size: {model_bytes + code_bytes} bytes")
+
+    quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict())
+    quant_buf = io.BytesIO()
+    torch.save(quant_obj, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    quant_blob = zlib.compress(quant_raw, level=9)
+    quant_raw_bytes = len(quant_raw)
+    if master_process:
+        with open("final_model.int8.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")
+
+    if distributed:
+        dist.barrier()
+    with open("final_model.int8.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)
+    torch.cuda.synchronize()
+    t_qeval = time.perf_counter()
+    if args.eval_stride > 0 and args.eval_stride < args.train_seq_len:
+        log0(f"final_eval_mode:sliding_window stride:{args.eval_stride} batch_seqs:{args.eval_batch_seqs} doc_isolated:{args.doc_isolated_eval}")
+        q_val_loss, q_val_bpb = eval_val_sliding(
+            args,
+            base_model,
+            rank,
+            world_size,
+            device,
+            val_tokens,
+            base_bytes_lut,
+            has_leading_space_lut,
+            is_boundary_token_lut,
+            stride=args.eval_stride,
+            batch_seqs=args.eval_batch_seqs,
+        )
+    else:
+        log0("final_eval_mode:standard")
+        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,
+        )
+    torch.cuda.synchronize()
+    log0(
+        f"final_int8_zlib_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}")
+
+    if distributed:
+        dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()