diff --git a/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/README.md b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/README.md
new file mode 100644
index 000000000..9467e0253
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/README.md
@@ -0,0 +1,31 @@
+# QAT + BigramHash(12288) + Stride 32
+
+## Summary
+
+Built on the current SOTA (`10L_Int5MLP_MuonWD04_SWA50`) with the following improvements:
+
+- **QAT (Quantization-Aware Training):** STE fake-quantize during training — int5 for MLP layers, int6 for attention. Reduces post-quantization degradation.
+- **BigramHash 12288:** Increased from 10240 to 12288 buckets for better bigram coverage.
+- **Eval stride 32:** Reduced from 64 to 32 for more overlapping context windows during evaluation.
+- **Magnitude pruning 5%:** Increased from 3% to improve compression ratio.
+- **SWA every 50 steps:** Checkpoint averaging during warmdown.
+
+## Architecture
+
+- 10 transformer layers, dim=512, 8 heads, 4 KV heads
+- 3x MLP with SmearGate
+- BigramHash(12288) with bigram_dim=128
+- Mixed quantization: int5 MLP, int6 attention
+- zstd-22 compression
+
+## Results
+
+```
+seed=2024: val_bpb=1.14443, artifact=15,902,583 bytes
+```
+
+## Command
+
+```bash
+torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
diff --git a/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/submission.json b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/submission.json
new file mode 100644
index 000000000..b1fd73991
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/submission.json
@@ -0,0 +1,9 @@
+{
+  "name": "QAT + BigramHash(12288) + Stride 32",
+  "val_loss": 1.14443,
+  "bytes_total": 15902583,
+  "blurb": "10 layers, QAT with STE (int5 MLP / int6 attn), BigramHash 12288, eval stride 32, magnitude pruning 5%, SWA every 50 steps, zstd-22. Based on 10L_Int5MLP_MuonWD04_SWA50.",
+  "author": "fbedev",
+  "github_id": "fbedev",
+  "date": "2026-03-21"
+}
diff --git a/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_gpt.py b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_gpt.py
new file mode 100644
index 000000000..00deead0d
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_gpt.py
@@ -0,0 +1,1378 @@
+"""
+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
+
+try:
+    import zstandard
+    _COMPRESSOR = "zstd"
+except ImportError:
+    _COMPRESSOR = "zlib"
+
+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
+
+_HAS_FA3 = False  # SDPA is faster than FA3 for GQA on PyTorch 2.9+
+
+# -----------------------------
+# HYPERPARAMETERS
+# -----------------------------
+
+class Hyperparameters:
+    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", 42))
+
+    val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288))
+    val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 500))
+    train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 100))
+
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000))
+    warmup_steps = int(os.environ.get("WARMUP_STEPS", 20))
+    train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432))
+    train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048))
+    max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0))
+    qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5))
+
+    vocab_size = int(os.environ.get("VOCAB_SIZE", 1024))
+    num_layers = int(os.environ.get("NUM_LAYERS", 11))
+    num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4))
+    model_dim = int(os.environ.get("MODEL_DIM", 512))
+    num_heads = int(os.environ.get("NUM_HEADS", 8))
+    mlp_mult = float(os.environ.get("MLP_MULT", 3.0))
+    tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1")))
+    rope_base = float(os.environ.get("ROPE_BASE", 10000.0))
+    logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0))
+
+    embed_lr = float(os.environ.get("EMBED_LR", 0.6))
+    head_lr = float(os.environ.get("HEAD_LR", 0.008))
+    tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03))
+    tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005))
+    matrix_lr = float(os.environ.get("MATRIX_LR", 0.02))
+    scalar_lr = float(os.environ.get("SCALAR_LR", 0.02))
+    muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3))
+    weight_decay = float(os.environ.get("WEIGHT_DECAY", 0.04))
+
+    eval_stride = int(os.environ.get("EVAL_STRIDE", 32))
+    eval_batch_seqs = int(os.environ.get("EVAL_BATCH_SEQS", 32))
+
+    bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 0))
+    bigram_dim = int(os.environ.get("BIGRAM_DIM", 128))
+
+    swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "0")))  # disabled, using EMA instead
+    swa_start_frac = float(os.environ.get("SWA_START_FRAC", 0.4))
+    swa_every = int(os.environ.get("SWA_EVERY", 25))
+
+    # EMA (replaces SWA for #315-style training)
+    ema_enabled = bool(int(os.environ.get("EMA_ENABLED", "1")))
+    ema_decay = float(os.environ.get("EMA_DECAY", 0.997))
+
+    # TTT: Test-Time Training on validation data after quantization
+    ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1")))
+    ttt_epochs = int(os.environ.get("TTT_EPOCHS", 25))
+    ttt_lr = float(os.environ.get("TTT_LR", 0.008))
+    ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9))
+    ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 32))
+
+    # XSA: Exclusive Self-Attention on last N layers
+    xsa_last_n = int(os.environ.get("XSA_LAST_N", 4))
+
+    # Partial RoPE: only apply RoPE to first N dims of each head
+    rope_dims = int(os.environ.get("ROPE_DIMS", 16))
+
+    # LN Scale: scale norm output by 1/sqrt(layer+1)
+    ln_scale = bool(int(os.environ.get("LN_SCALE", "1")))
+
+# -----------------------------
+# MUON OPTIMIZER
+# -----------------------------
+
+def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor:
+    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, weight_decay: float = 0.0):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov, weight_decay=weight_decay),
+        )
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                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)
+                    g *= max(1, g.size(0) / g.size(1)) ** 0.5
+                    updates_flat[curr : curr + p.numel()] = g.reshape(-1)
+                curr += p.numel()
+
+            if distributed:
+                dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM)
+
+            wd = group.get("weight_decay", 0.0)
+            curr = 0
+            for p in params:
+                g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype)
+                if wd > 0:
+                    p.data.mul_(1.0 - lr * wd)
+                p.add_(g, alpha=-lr)
+                curr += p.numel()
+        return loss
+
+
+# -----------------------------
+# TOKENIZER-AGNOSTIC EVALUATION
+# -----------------------------
+
+def build_sentencepiece_luts(
+    sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device
+) -> tuple[Tensor, Tensor, Tensor]:
+    sp_vocab_size = int(sp.vocab_size())
+    table_size = max(sp_vocab_size, vocab_size)
+    base_bytes_np = np.zeros((table_size,), dtype=np.int16)
+    has_leading_space_np = np.zeros((table_size,), dtype=np.bool_)
+    is_boundary_token_np = np.ones((table_size,), dtype=np.bool_)
+    for token_id in range(sp_vocab_size):
+        if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id):
+            continue
+        is_boundary_token_np[token_id] = False
+        if sp.is_byte(token_id):
+            base_bytes_np[token_id] = 1
+            continue
+        piece = sp.id_to_piece(token_id)
+        if piece.startswith("\u2581"):
+            has_leading_space_np[token_id] = True
+            piece = piece[1:]
+        base_bytes_np[token_id] = len(piece.encode("utf-8"))
+    return (
+        torch.tensor(base_bytes_np, dtype=torch.int16, device=device),
+        torch.tensor(has_leading_space_np, dtype=torch.bool, device=device),
+        torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device),
+    )
+
+
+def load_validation_tokens(pattern: str, seq_len: int) -> Tensor:
+    files = [Path(p) for p in sorted(glob.glob(pattern))]
+    if not files:
+        raise FileNotFoundError(f"No files found for pattern: {pattern}")
+    tokens = torch.cat([load_data_shard(file) for file in files]).contiguous()
+    usable = ((tokens.numel() - 1) // seq_len) * seq_len
+    if usable <= 0:
+        raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}")
+    return tokens[: usable + 1]
+
+
+def eval_val(
+    args: Hyperparameters,
+    model: nn.Module,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    grad_accum_steps: int,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+) -> tuple[float, float]:
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < args.train_seq_len:
+        raise ValueError(
+            "VAL_BATCH_SIZE must provide at least one sequence per rank; "
+            f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, "
+            f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // args.train_seq_len
+    total_seqs = (val_tokens.numel() - 1) // args.train_seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * args.train_seq_len
+            raw_end = batch_seq_end * args.train_seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, args.train_seq_len)
+            y = local[1:].reshape(-1, args.train_seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            val_loss_sum += batch_loss.to(torch.float64) * batch_token_count
+            val_token_count += batch_token_count
+            prev_ids = x.reshape(-1)
+            tgt_ids = y.reshape(-1)
+            token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16)
+            token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16)
+            val_byte_count += token_bytes.to(torch.float64).sum()
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM)
+    val_loss = val_loss_sum / val_token_count
+    bits_per_token = val_loss.item() / math.log(2.0)
+    tokens_per_byte = val_token_count.item() / val_byte_count.item()
+    model.train()
+    return float(val_loss.item()), float(bits_per_token * tokens_per_byte)
+
+
+# -----------------------------
+# POST-TRAINING QUANTIZATION (INT8 legacy + INT6 mixed)
+# -----------------------------
+
+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,smear,bigram.scale",
+    ).split(",")
+    if pattern
+)
+FP16_KEEP_NAME_PATTERNS = tuple(
+    pattern
+    for pattern in os.environ.get("FP16_KEEP_NAME_PATTERNS", "tok_emb,blocks.8.attn.c_k").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 quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]:
+    t32 = t.float()
+    if t32.ndim == 2:
+        clip_abs = (
+            torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1)
+            if t32.numel()
+            else torch.empty((t32.shape[0],), dtype=torch.float32)
+        )
+        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()
+    clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0
+    scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32)
+    q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous()
+    return q, scale
+
+
+def _classify_param(name: str) -> str:
+    if "tok_emb" in name or "lm_head" in name:
+        return "embed"
+    if ".mlp." in name:
+        return "mlp"
+    if "bigram" in name:
+        return "bigram"
+    if ".attn." in name or (".proj." in name and ".mlp." not in name):
+        return "attn"
+    return "other"
+
+def quantize_intN_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]:
+    t32 = t.float()
+    if t32.ndim == 2:
+        row_max = t32.abs().amax(dim=1)
+        scale = (row_max / clip_range).clamp_min(1e-12).to(torch.float16)
+        scale = scale.clamp_min(torch.finfo(torch.float16).tiny)
+        q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -(clip_range+1), clip_range).to(torch.int8)
+        return q, scale
+    amax = t32.abs().max().item()
+    scale = torch.tensor(max(amax / clip_range, 1e-12), dtype=torch.float16)
+    q = torch.clamp(torch.round(t32 / scale.float()), -(clip_range+1), clip_range).to(torch.int8)
+    return q, scale
+
+def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]):
+    result: dict[str, Tensor] = {}
+    meta: dict[str, object] = {}
+    for name, tensor in state_dict.items():
+        t = tensor.detach().cpu().contiguous()
+        cat = _classify_param(name)
+        if not t.is_floating_point() or t.numel() <= 8192:
+            result[name] = t.to(torch.float16) if t.is_floating_point() else t
+            meta[name] = "passthrough"
+            continue
+        if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS):
+            result[name] = t.float()
+            meta[name] = "passthrough_ctrl"
+            continue
+        if any(pattern in name for pattern in FP16_KEEP_NAME_PATTERNS):
+            result[name] = t.to(dtype=torch.float16).contiguous()
+            meta[name] = "passthrough_fp16"
+            continue
+        if cat in int6_cats and t.ndim >= 1:
+            clip = 15 if cat == "mlp" else 31  # int5 for MLP, int6 for attention
+            q, s = quantize_intN_per_row(t, clip_range=clip)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": f"int{5 if cat == 'mlp' else 6}"}
+        else:
+            q, s = quantize_float_tensor(t)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": "int8"}
+    return result, meta
+
+def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object],
+                          template_sd: dict[str, Tensor]) -> dict[str, Tensor]:
+    out: dict[str, Tensor] = {}
+    for name, orig in template_sd.items():
+        info = meta[name]
+        orig_dtype = orig.dtype
+        if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"):
+            t = result[name]
+            if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16):
+                t = t.to(orig_dtype)
+            out[name] = t
+            continue
+        q, s = result[name + ".q"], result[name + ".scale"]
+        if s.ndim > 0:
+            out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype)
+        else:
+            out[name] = (q.float() * float(s.item())).to(orig_dtype)
+    return out
+
+
+# -----------------------------
+# 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)
+    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:
+    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:
+    def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device):
+        self.rank = rank
+        self.world_size = world_size
+        self.device = device
+        self.stream = TokenStream(pattern)
+
+    def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]:
+        local_tokens = global_tokens // (self.world_size * grad_accum_steps)
+        per_rank_span = local_tokens + 1
+        chunk = self.stream.take(per_rank_span * self.world_size)
+        start = self.rank * per_rank_span
+        local = chunk[start : start + per_rank_span].to(dtype=torch.int64)
+        x = local[:-1].reshape(-1, seq_len)
+        y = local[1:].reshape(-1, seq_len)
+        return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True)
+
+
+# -----------------------------
+# TRANSFORMER MODULES
+# -----------------------------
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+
+class CastedLinear(nn.Linear):
+    _qat: bool = False
+    _qat_int5: bool = False  # True for MLP (int5), False for attn (int6)
+
+    def forward(self, x: Tensor) -> Tensor:
+        w = self.weight
+        if self._qat and self.training and w.ndim == 2:
+            w_f = w.float()
+            amax = w_f.abs().amax(dim=-1, keepdim=True).clamp_min(1e-12)
+            if self._qat_int5:
+                scale = amax / 15.0
+                w_q = (w_f / scale).round().clamp(-16, 15) * scale
+            else:
+                scale = amax / 31.0
+                w_q = (w_f / scale).round().clamp(-32, 31) * scale
+            w = w + (w_q - w_f).detach()  # STE
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, w.to(x.dtype), bias)
+
+
+def restore_low_dim_params_to_fp32(module: nn.Module) -> None:
+    with torch.no_grad():
+        for name, param in module.named_parameters():
+            if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32:
+                param.data = param.data.float()
+
+
+class Rotary(nn.Module):
+    def __init__(self, dim: int, base: float = 10000.0, rope_dims: int = 0):
+        super().__init__()
+        self.rope_dims = rope_dims if rope_dims > 0 else dim
+        inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            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:
+    rd = cos.size(-1) * 2  # number of RoPE dims
+    if rd < x.size(-1):
+        x_rope, x_pass = x[..., :rd], x[..., rd:]
+        half = rd // 2
+        x1, x2 = x_rope[..., :half], x_rope[..., half:]
+        x_rot = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+        return torch.cat((x_rot, x_pass), dim=-1)
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+
+class CausalSelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int, num_kv_heads: int, rope_base: float, qk_gain_init: float, rope_dims: int = 0, use_xsa: bool = False):
+        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
+        self.use_xsa = use_xsa
+        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, rope_dims=rope_dims)
+
+    def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor:
+        """Remove self-value component from attention output via orthogonal projection."""
+        # y: (B, H, T, D), v: (B, Hkv, T, D)
+        B, H, T, D = y.shape
+        Hkv = v.size(1)
+        group = H // Hkv
+        y_g = y.reshape(B, Hkv, group, T, D)
+        vn = F.normalize(v, dim=-1).unsqueeze(2)  # (B, Hkv, 1, T, D)
+        proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn
+        return (y_g - proj).reshape(B, H, T, D)
+
+    def forward(self, x: Tensor) -> Tensor:
+        bsz, seqlen, dim = x.shape
+        q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2)
+        k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2)
+        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),
+        )
+        if self.use_xsa:
+            y = self._xsa_efficient(y, v)
+        y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)
+        return self.proj(y)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim: int, mlp_mult: float):
+        super().__init__()
+        hidden = int(mlp_mult * dim)
+        self.fc = CastedLinear(dim, hidden, bias=False)
+        self.proj = CastedLinear(hidden, dim, bias=False)
+        self.proj._zero_init = True
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = torch.relu(self.fc(x))
+        return self.proj(x.square())
+
+
+class SmearGate(nn.Module):
+    """Blend each token's embedding with the previous token's embedding."""
+    def __init__(self, dim: int):
+        super().__init__()
+        self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32))
+
+    def forward(self, x: Tensor) -> Tensor:
+        g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :]
+        x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1)
+        return (1 - g) * x + g * x_prev
+
+
+class BigramHashEmbedding(nn.Module):
+    """Hash consecutive token pairs into a learned embedding table."""
+    def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int):
+        super().__init__()
+        self.bigram_vocab_size = bigram_vocab_size
+        self.embed = nn.Embedding(bigram_vocab_size, bigram_dim)
+        nn.init.zeros_(self.embed.weight)
+        self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None
+        if self.proj is not None:
+            nn.init.zeros_(self.proj.weight)
+        self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32))
+
+    def bigram_hash(self, tokens: Tensor) -> Tensor:
+        t = tokens.to(torch.int32)
+        mod = self.bigram_vocab_size - 1
+        out = torch.empty_like(t)
+        out[..., 0] = mod
+        out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod
+        return out.long()
+
+    def forward(self, token_ids: Tensor) -> Tensor:
+        h = self.embed(self.bigram_hash(token_ids))
+        if self.proj is not None:
+            h = self.proj(h)
+        return h * self.scale.to(dtype=h.dtype)
+
+
+class Block(nn.Module):
+    def __init__(self, dim: int, num_heads: int, num_kv_heads: int, mlp_mult: float, rope_base: float, qk_gain_init: float, layer_idx: int = 0, ln_scale: bool = False, rope_dims: int = 0, use_xsa: bool = False):
+        super().__init__()
+        self.attn_norm = RMSNorm()
+        self.mlp_norm = RMSNorm()
+        self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0
+        self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, rope_dims=rope_dims, use_xsa=use_xsa)
+        self.mlp = MLP(dim, mlp_mult)
+        self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float())
+
+    def forward(self, x: Tensor, x0: Tensor) -> Tensor:
+        mix = self.resid_mix.to(dtype=x.dtype)
+        x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0
+        s = self.ln_scale_factor
+        attn_out = self.attn(self.attn_norm(x) * s)
+        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) * s)
+        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: float,
+        tie_embeddings: bool,
+        tied_embed_init_std: float,
+        logit_softcap: float,
+        rope_base: float,
+        qk_gain_init: float,
+        bigram_vocab_size: int = 0,
+        bigram_dim: int = 128,
+        xsa_last_n: int = 0,
+        rope_dims: int = 0,
+        ln_scale: bool = False,
+    ):
+        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.tok_emb = nn.Embedding(vocab_size, model_dim)
+        self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None
+        self.num_encoder_layers = num_layers // 2
+        self.num_decoder_layers = num_layers - self.num_encoder_layers
+        self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers)
+        self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32))
+        self.smear = SmearGate(model_dim)
+        self.blocks = nn.ModuleList(
+            [
+                Block(model_dim, num_heads, num_kv_heads, mlp_mult, rope_base, qk_gain_init,
+                      layer_idx=i, ln_scale=ln_scale, rope_dims=rope_dims,
+                      use_xsa=(i >= num_layers - xsa_last_n) if xsa_last_n > 0 else False)
+                for i in range(num_layers)
+            ]
+        )
+        self.final_norm = RMSNorm()
+        self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False)
+        if self.lm_head is not None:
+            self.lm_head._zero_init = True
+        self._init_weights()
+
+    def _init_weights(self) -> None:
+        if self.tie_embeddings:
+            nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std)
+        num_layers = len(self.blocks)
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear):
+                if getattr(module, "_zero_init", False):
+                    nn.init.zeros_(module.weight)
+                elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64:
+                    nn.init.orthogonal_(module.weight, gain=1.0)
+                    if ".proj." in name or name.endswith(".proj"):
+                        with torch.no_grad():
+                            module.weight.mul_(1.0 / math.sqrt(2 * num_layers))
+
+    def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor:
+        x = self.tok_emb(input_ids)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        for i in range(self.num_encoder_layers):
+            x = self.blocks[i](x, x0)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            x = self.blocks[self.num_encoder_layers + i](x, x0)
+        x = self.final_norm(x).reshape(-1, x.size(-1))
+        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:
+        x = self.tok_emb(input_ids)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        for i in range(self.num_encoder_layers):
+            x = self.blocks[i](x, x0)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            x = self.blocks[self.num_encoder_layers + i](x, x0)
+        x = self.final_norm(x)
+        if self.tie_embeddings:
+            logits_proj = F.linear(x, self.tok_emb.weight)
+        else:
+            logits_proj = self.lm_head(x)
+        return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap)
+
+
+def 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]:
+    seq_len = args.train_seq_len
+    total_tokens = val_tokens.numel() - 1
+    window_starts = [ws for ws in range(0, total_tokens, stride)
+                     if min(ws + seq_len, total_tokens) - ws >= stride or ws == 0]
+    total_windows = len(window_starts)
+    my_s = (total_windows * rank) // world_size
+    my_e = (total_windows * (rank + 1)) // world_size
+    my_windows = window_starts[my_s:my_e]
+
+    loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    token_count = torch.zeros((), device=device, dtype=torch.float64)
+    byte_count = torch.zeros((), device=device, dtype=torch.float64)
+
+    base_model.eval()
+    with torch.inference_mode():
+        for bi in range(0, len(my_windows), batch_seqs):
+            batch_ws = my_windows[bi:bi + batch_seqs]
+            bsz = len(batch_ws)
+            x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            wlens: list[int] = []
+            for i, ws in enumerate(batch_ws):
+                end = min(ws + seq_len, total_tokens)
+                wlen = end - ws
+                wlens.append(wlen)
+                chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device)
+                x_batch[i, :wlen] = chunk[:-1]
+                y_batch[i, :wlen] = chunk[1:]
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                logits = base_model.forward_logits(x_batch)
+            nll = F.cross_entropy(
+                logits.reshape(-1, logits.size(-1)).float(),
+                y_batch.reshape(-1),
+                reduction="none",
+            ).reshape(bsz, seq_len)
+            for i, ws in enumerate(batch_ws):
+                wlen = wlens[i]
+                s = 0 if ws == 0 else max(wlen - stride, 0)
+                scored_nll = nll[i, s:wlen].to(torch.float64)
+                loss_sum += scored_nll.sum()
+                token_count += float(wlen - s)
+                tgt = y_batch[i, s:wlen]
+                prev = x_batch[i, s:wlen]
+                tb = base_bytes_lut[tgt].to(torch.float64)
+                tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64)
+                byte_count += tb.sum()
+            if 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 = "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
+
+    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)
+
+    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,
+        bigram_vocab_size=args.bigram_vocab_size,
+        bigram_dim=args.bigram_dim,
+        xsa_last_n=args.xsa_last_n,
+        rope_dims=args.rope_dims,
+        ln_scale=args.ln_scale,
+    ).to(device).bfloat16()
+    for module in base_model.modules():
+        if isinstance(module, CastedLinear):
+            module.float()
+    restore_low_dim_params_to_fp32(base_model)
+    # QAT: fake-quantize during training so weights learn to be quantization-friendly
+    # Late-stage QAT: only enable in the last 20% of training to avoid hurting convergence
+    qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "1")))
+    qat_start_frac = float(os.environ.get("QAT_START_FRAC", "0.96"))  # enable QAT after this fraction of steps (final 4%)
+    qat_activated = False
+    if qat_enabled:
+        for name, module in base_model.named_modules():
+            if isinstance(module, CastedLinear):
+                module._qat = False  # start with QAT disabled
+                module._qat_int5 = ".mlp." in name
+    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
+
+    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)
+    scalar_params.append(base_model.smear.gate)
+    if base_model.bigram is not None:
+        scalar_params.append(base_model.bigram.scale)
+
+    token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr
+    tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}]
+    if base_model.bigram is not None:
+        tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr})
+        if base_model.bigram.proj is not None:
+            matrix_params.append(base_model.bigram.proj.weight)
+
+    optimizer_tok = torch.optim.AdamW(
+        tok_params,
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.weight_decay,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+        weight_decay=0.04,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.AdamW(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.weight_decay,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    if base_model.lm_head is not None:
+        optimizer_head = torch.optim.Adam(
+            [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.insert(1, optimizer_head)
+
+    n_params = sum(p.numel() for p in base_model.parameters())
+    log0(f"model_params:{n_params}")
+    log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}")
+    log0(f"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"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
+
+    if args.warmup_steps > 0:
+        initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()}
+        initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    # EMA state
+    ema_state: dict[str, Tensor] | None = None
+    if args.ema_enabled:
+        ema_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}
+
+    # MAIN TRAINING LOOP
+    training_time_ms = 0.0
+    stop_after_step: int | None = None
+    swa_state: dict[str, Tensor] | None = None
+    swa_count = 0
+    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
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+
+        # Late-stage QAT: enable fake-quantize after qat_start_frac of training
+        if qat_enabled and not qat_activated:
+            # Estimate progress: use wallclock fraction if available, else step fraction
+            if max_wallclock_ms is not None:
+                progress = approx_training_time_ms / max_wallclock_ms
+            else:
+                progress = step / args.iterations
+            if progress >= qat_start_frac:
+                for module in base_model.modules():
+                    if isinstance(module, CastedLinear):
+                        module._qat = True
+                qat_activated = True
+                log0(f"qat:enabled at step:{step} progress:{progress:.3f}")
+
+        # EMA: update exponential moving average every step
+        if args.ema_enabled and ema_state is not None and step % 10 == 0:
+            # EMA with adjusted decay for every-10-steps update: decay^10
+            decay = args.ema_decay ** 10
+            for name, t in base_model.state_dict().items():
+                ema_state[name].mul_(decay).add_(t.detach().cpu(), alpha=1.0 - decay)
+
+        # SWA: collect checkpoints during warmdown
+        if args.swa_enabled and scale < args.swa_start_frac and step % args.swa_every == 0:
+            if swa_state is None:
+                swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}
+                swa_count = 1
+                log0(f"swa:start step:{step}")
+            else:
+                for name, t in base_model.state_dict().items():
+                    swa_state[name] += t.detach().cpu()
+                swa_count += 1
+
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        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"
+            )
+
+        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"
+    )
+
+    # Apply EMA if enabled
+    if args.ema_enabled and ema_state is not None:
+        log0(f"ema:applying decay={args.ema_decay}")
+        current_state = base_model.state_dict()
+        ema_applied = {
+            name: tensor.to(dtype=current_state[name].dtype)
+            for name, tensor in ema_state.items()
+        }
+        base_model.load_state_dict(ema_applied, strict=True)
+
+    # Apply SWA if collected
+    if args.swa_enabled and swa_state is not None and swa_count > 1:
+        log0(f"swa:applying averaged {swa_count} checkpoints")
+        current_state = base_model.state_dict()
+        avg_state = {
+            name: (tensor / swa_count).to(dtype=current_state[name].dtype)
+            for name, tensor in swa_state.items()
+        }
+        base_model.load_state_dict(avg_state, strict=True)
+
+    # SERIALIZATION + ROUNDTRIP VALIDATION
+    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")
+
+    # Magnitude pruning: zero out smallest weights to improve compression
+    with torch.no_grad():
+        for name, param in base_model.named_parameters():
+            if param.ndim == 2 and param.numel() > 65536:
+                threshold = torch.quantile(param.abs().float().flatten(), 0.15)
+                mask = param.abs() < threshold
+                param.masked_fill_(mask, 0.0)
+
+    # INT6 mixed quantization + zstd/zlib export
+    sd_cpu = {k: v.detach().cpu() for k, v in base_model.state_dict().items()}
+    quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn", "bigram"})
+    quant_buf = io.BytesIO()
+    torch.save({"w": quant_result, "m": quant_meta}, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    if _COMPRESSOR == "zstd":
+        quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw)
+    else:
+        quant_blob = zlib.compress(quant_raw, 9)
+    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"))
+        log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes")
+        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()
+    if _COMPRESSOR == "zstd":
+        decompressed = zstandard.ZstdDecompressor().decompress(quant_blob_disk)
+    else:
+        decompressed = zlib.decompress(quant_blob_disk)
+    quant_state = torch.load(io.BytesIO(decompressed), map_location="cpu")
+    deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu)
+    base_model.load_state_dict(deq_state, strict=True)
+
+    # TTT: Test-Time Training — adapt quantized model on val data before final eval
+    if args.ttt_enabled:
+        torch.cuda.synchronize()
+        t_ttt = time.perf_counter()
+        log0(f"ttt:start epochs:{args.ttt_epochs} lr:{args.ttt_lr} batch_seqs:{args.ttt_batch_seqs}")
+        base_model.train()
+        ttt_optimizer = torch.optim.SGD(
+            base_model.parameters(), lr=args.ttt_lr, momentum=args.ttt_momentum,
+        )
+        seq_len = args.train_seq_len
+        n_val = val_tokens.numel() - 1
+        n_seqs = n_val // seq_len
+        for ttt_ep in range(args.ttt_epochs):
+            perm = torch.randperm(n_seqs)
+            ttt_loss_sum = 0.0
+            ttt_loss_count = 0
+            for batch_start in range(0, n_seqs, args.ttt_batch_seqs):
+                batch_end = min(batch_start + args.ttt_batch_seqs, n_seqs)
+                indices = perm[batch_start:batch_end]
+                batch_x = torch.stack([val_tokens[i * seq_len : i * seq_len + seq_len] for i in indices]).to(device=device, dtype=torch.int64)
+                batch_y = torch.stack([val_tokens[i * seq_len + 1 : i * seq_len + seq_len + 1] for i in indices]).to(device=device, dtype=torch.int64)
+                ttt_optimizer.zero_grad(set_to_none=True)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    loss = base_model(batch_x, batch_y)
+                loss.backward()
+                ttt_optimizer.step()
+                ttt_loss_sum += loss.item() * (batch_end - batch_start)
+                ttt_loss_count += batch_end - batch_start
+            if ttt_ep == 0 or (ttt_ep + 1) % 5 == 0 or ttt_ep == args.ttt_epochs - 1:
+                log0(f"ttt:epoch:{ttt_ep + 1}/{args.ttt_epochs} loss:{ttt_loss_sum / max(ttt_loss_count, 1):.4f}")
+        base_model.eval()
+        torch.cuda.synchronize()
+        log0(f"ttt:done time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms")
+
+    # Sliding window eval on int6-roundtripped weights
+    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}")
+        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()
+# fixes applied
+# tuned
diff --git a/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_seed2024.log b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_seed2024.log
new file mode 100644
index 000000000..649391955
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-21_QAT_BigramHash12K_Stride32/train_seed2024.log
@@ -0,0 +1,288 @@
+W0321 15:03:05.510000 66078 torch/distributed/run.py:803] 
+W0321 15:03:05.510000 66078 torch/distributed/run.py:803] *****************************************
+W0321 15:03:05.510000 66078 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0321 15:03:05.510000 66078 torch/distributed/run.py:803] *****************************************
+logs/990d50d0-2581-429b-a612-806c2db9b702.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
+model_params:25779281
+world_size:8 grad_accum_steps:1
+attention_mode:gqa num_heads:8 num_kv_heads:4
+tie_embeddings:True embed_lr:0.03 matrix_lr:0.02 scalar_lr:0.02
+train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000
+seed:2024
+warmup_step:1/20
+warmup_step:2/20
+warmup_step:3/20
+warmup_step:4/20
+warmup_step:5/20
+warmup_step:6/20
+warmup_step:7/20
+warmup_step:8/20
+warmup_step:9/20
+warmup_step:10/20
+warmup_step:11/20
+warmup_step:12/20
+warmup_step:13/20
+warmup_step:14/20
+warmup_step:15/20
+warmup_step:16/20
+warmup_step:17/20
+warmup_step:18/20
+warmup_step:19/20
+warmup_step:20/20
+step:0/20000 val_loss:6.9292 val_bpb:4.1039 train_time:0ms step_avg:0.02ms
+step:1/20000 train_loss:6.9306 train_time:140ms step_avg:139.87ms
+step:2/20000 train_loss:7.9609 train_time:213ms step_avg:106.27ms
+step:3/20000 train_loss:7.5544 train_time:308ms step_avg:102.76ms
+step:4/20000 train_loss:6.9558 train_time:404ms step_avg:101.02ms
+step:5/20000 train_loss:6.7533 train_time:500ms step_avg:100.07ms
+step:6/20000 train_loss:6.6750 train_time:596ms step_avg:99.39ms
+step:7/20000 train_loss:6.5946 train_time:692ms step_avg:98.88ms
+step:8/20000 train_loss:6.6178 train_time:788ms step_avg:98.55ms
+step:9/20000 train_loss:6.3579 train_time:884ms step_avg:98.25ms
+step:10/20000 train_loss:6.0389 train_time:980ms step_avg:98.05ms
+step:100/20000 train_loss:3.1600 train_time:9128ms step_avg:91.28ms
+step:200/20000 train_loss:2.3809 train_time:18325ms step_avg:91.63ms
+step:300/20000 train_loss:2.5299 train_time:27521ms step_avg:91.74ms
+step:400/20000 train_loss:2.3987 train_time:36669ms step_avg:91.67ms
+step:500/20000 train_loss:2.3887 train_time:45748ms step_avg:91.50ms
+step:500/20000 val_loss:2.3455 val_bpb:1.3891 train_time:45779ms step_avg:91.56ms
+step:600/20000 train_loss:2.3253 train_time:54893ms step_avg:91.49ms
+step:700/20000 train_loss:2.3429 train_time:64045ms step_avg:91.49ms
+step:800/20000 train_loss:2.2374 train_time:73189ms step_avg:91.49ms
+step:900/20000 train_loss:2.1269 train_time:82341ms step_avg:91.49ms
+step:1000/20000 train_loss:2.2712 train_time:91403ms step_avg:91.40ms
+step:1000/20000 val_loss:2.2225 val_bpb:1.3163 train_time:91436ms step_avg:91.44ms
+step:1100/20000 train_loss:2.3213 train_time:100557ms step_avg:91.42ms
+step:1200/20000 train_loss:2.3524 train_time:109705ms step_avg:91.42ms
+step:1300/20000 train_loss:2.1034 train_time:118854ms step_avg:91.43ms
+step:1400/20000 train_loss:2.1821 train_time:128004ms step_avg:91.43ms
+step:1500/20000 train_loss:2.2246 train_time:137063ms step_avg:91.38ms
+step:1500/20000 val_loss:2.1862 val_bpb:1.2948 train_time:137095ms step_avg:91.40ms
+step:1600/20000 train_loss:2.0854 train_time:146215ms step_avg:91.38ms
+step:1700/20000 train_loss:2.1488 train_time:155363ms step_avg:91.39ms
+step:1800/20000 train_loss:2.1715 train_time:164506ms step_avg:91.39ms
+step:1900/20000 train_loss:2.1383 train_time:173561ms step_avg:91.35ms
+step:2000/20000 train_loss:2.0740 train_time:182697ms step_avg:91.35ms
+step:2000/20000 val_loss:2.1383 val_bpb:1.2664 train_time:182728ms step_avg:91.36ms
+step:2100/20000 train_loss:2.0578 train_time:191853ms step_avg:91.36ms
+step:2200/20000 train_loss:2.1449 train_time:200996ms step_avg:91.36ms
+step:2300/20000 train_loss:2.1199 train_time:210138ms step_avg:91.36ms
+step:2400/20000 train_loss:2.0741 train_time:219194ms step_avg:91.33ms
+step:2500/20000 train_loss:2.1747 train_time:228337ms step_avg:91.33ms
+step:2500/20000 val_loss:2.1135 val_bpb:1.2518 train_time:228368ms step_avg:91.35ms
+step:2600/20000 train_loss:2.1209 train_time:237476ms step_avg:91.34ms
+step:2700/20000 train_loss:2.1141 train_time:246606ms step_avg:91.34ms
+step:2800/20000 train_loss:2.1671 train_time:255756ms step_avg:91.34ms
+step:2900/20000 train_loss:2.0344 train_time:264813ms step_avg:91.31ms
+step:3000/20000 train_loss:2.1731 train_time:273948ms step_avg:91.32ms
+step:3000/20000 val_loss:2.0991 val_bpb:1.2432 train_time:273980ms step_avg:91.33ms
+step:3100/20000 train_loss:2.0482 train_time:283079ms step_avg:91.32ms
+step:3200/20000 train_loss:2.1810 train_time:292224ms step_avg:91.32ms
+step:3300/20000 train_loss:2.0808 train_time:301278ms step_avg:91.30ms
+step:3400/20000 train_loss:2.0340 train_time:310412ms step_avg:91.30ms
+step:3500/20000 train_loss:2.1913 train_time:319549ms step_avg:91.30ms
+step:3500/20000 val_loss:2.0911 val_bpb:1.2385 train_time:319582ms step_avg:91.31ms
+step:3600/20000 train_loss:2.1106 train_time:328694ms step_avg:91.30ms
+step:3700/20000 train_loss:2.1071 train_time:337841ms step_avg:91.31ms
+step:3800/20000 train_loss:2.0835 train_time:346893ms step_avg:91.29ms
+step:3900/20000 train_loss:2.0880 train_time:356028ms step_avg:91.29ms
+step:4000/20000 train_loss:1.9862 train_time:365171ms step_avg:91.29ms
+step:4000/20000 val_loss:2.0754 val_bpb:1.2292 train_time:365203ms step_avg:91.30ms
+step:4100/20000 train_loss:2.0276 train_time:374316ms step_avg:91.30ms
+step:4200/20000 train_loss:2.1627 train_time:383449ms step_avg:91.30ms
+step:4300/20000 train_loss:2.0719 train_time:392501ms step_avg:91.28ms
+step:4400/20000 train_loss:2.0386 train_time:401631ms step_avg:91.28ms
+step:4500/20000 train_loss:2.1318 train_time:410764ms step_avg:91.28ms
+step:4500/20000 val_loss:2.0518 val_bpb:1.2152 train_time:410796ms step_avg:91.29ms
+step:4600/20000 train_loss:1.8506 train_time:419903ms step_avg:91.28ms
+step:4700/20000 train_loss:2.2436 train_time:428959ms step_avg:91.27ms
+step:4800/20000 train_loss:2.4413 train_time:438100ms step_avg:91.27ms
+step:4900/20000 train_loss:2.0587 train_time:447240ms step_avg:91.27ms
+step:5000/20000 train_loss:2.1126 train_time:456374ms step_avg:91.27ms
+step:5000/20000 val_loss:2.0317 val_bpb:1.2033 train_time:456405ms step_avg:91.28ms
+step:5100/20000 train_loss:2.1345 train_time:465515ms step_avg:91.28ms
+step:5200/20000 train_loss:2.0525 train_time:474562ms step_avg:91.26ms
+step:5300/20000 train_loss:2.0194 train_time:483706ms step_avg:91.27ms
+swa:start step:5400
+step:5400/20000 train_loss:2.0554 train_time:492846ms step_avg:91.27ms
+step:5500/20000 train_loss:2.0271 train_time:502092ms step_avg:91.29ms
+step:5500/20000 val_loss:2.0091 val_bpb:1.1899 train_time:502148ms step_avg:91.30ms
+step:5600/20000 train_loss:1.9645 train_time:511285ms step_avg:91.30ms
+step:5700/20000 train_loss:2.0216 train_time:520389ms step_avg:91.30ms
+step:5800/20000 train_loss:2.0028 train_time:529583ms step_avg:91.31ms
+step:5900/20000 train_loss:1.9122 train_time:538801ms step_avg:91.32ms
+step:6000/20000 train_loss:1.9470 train_time:547988ms step_avg:91.33ms
+step:6000/20000 val_loss:1.9858 val_bpb:1.1761 train_time:548044ms step_avg:91.34ms
+step:6100/20000 train_loss:1.9255 train_time:557100ms step_avg:91.33ms
+step:6200/20000 train_loss:1.9569 train_time:566305ms step_avg:91.34ms
+step:6300/20000 train_loss:1.9539 train_time:575510ms step_avg:91.35ms
+step:6400/20000 train_loss:2.0068 train_time:584699ms step_avg:91.36ms
+step:6500/20000 train_loss:2.0863 train_time:595463ms step_avg:91.61ms
+step:6500/20000 val_loss:1.9629 val_bpb:1.1625 train_time:595520ms step_avg:91.62ms
+step:6549/20000 val_loss:1.9622 val_bpb:1.1621 train_time:599957ms step_avg:91.61ms
+stopping_early: wallclock_cap train_time:599957ms step:6549/20000
+peak memory allocated: 18773 MiB reserved: 18976 MiB
+swa:applying averaged 23 checkpoints
+Serialized model: 98961707 bytes
+Code size: 53849 bytes
+Total submission size: 99015556 bytes
+Serialized model int6+zstd: 15848734 bytes
+Total submission size int8+zlib: 15902583 bytes
+final_eval_mode:sliding_window stride:32 batch_seqs:32
+  sliding_eval [  0.0%] 32/242272 windows running_bpb=1.155638
+  sliding_eval [  0.7%] 1632/242272 windows running_bpb=1.209812
+  sliding_eval [  1.3%] 3232/242272 windows running_bpb=1.137313
+  sliding_eval [  2.0%] 4832/242272 windows running_bpb=1.153552
+  sliding_eval [  2.7%] 6432/242272 windows running_bpb=1.140458
+  sliding_eval [  3.3%] 8032/242272 windows running_bpb=1.138215
+  sliding_eval [  4.0%] 9632/242272 windows running_bpb=1.134525
+  sliding_eval [  4.6%] 11232/242272 windows running_bpb=1.136803
+  sliding_eval [  5.3%] 12832/242272 windows running_bpb=1.145590
+  sliding_eval [  6.0%] 14432/242272 windows running_bpb=1.146289
+  sliding_eval [  6.6%] 16032/242272 windows running_bpb=1.147136
+  sliding_eval [  7.3%] 17632/242272 windows running_bpb=1.152368
+  sliding_eval [  7.9%] 19232/242272 windows running_bpb=1.148864
+  sliding_eval [  8.6%] 20832/242272 windows running_bpb=1.146263
+  sliding_eval [  9.3%] 22432/242272 windows running_bpb=1.144331
+  sliding_eval [  9.9%] 24032/242272 windows running_bpb=1.142830
+  sliding_eval [ 10.6%] 25632/242272 windows running_bpb=1.141425
+  sliding_eval [ 11.2%] 27232/242272 windows running_bpb=1.139964
+  sliding_eval [ 11.9%] 28832/242272 windows running_bpb=1.143295
+  sliding_eval [ 12.6%] 30432/242272 windows running_bpb=1.154973
+  sliding_eval [ 13.2%] 32032/242272 windows running_bpb=1.152132
+  sliding_eval [ 13.9%] 33632/242272 windows running_bpb=1.151552
+  sliding_eval [ 14.5%] 35232/242272 windows running_bpb=1.150023
+  sliding_eval [ 15.2%] 36832/242272 windows running_bpb=1.149581
+  sliding_eval [ 15.9%] 38432/242272 windows running_bpb=1.151397
+  sliding_eval [ 16.5%] 40032/242272 windows running_bpb=1.150010
+  sliding_eval [ 17.2%] 41632/242272 windows running_bpb=1.149747
+  sliding_eval [ 17.8%] 43232/242272 windows running_bpb=1.148912
+  sliding_eval [ 18.5%] 44832/242272 windows running_bpb=1.147993
+  sliding_eval [ 19.2%] 46432/242272 windows running_bpb=1.145475
+  sliding_eval [ 19.8%] 48032/242272 windows running_bpb=1.148363
+  sliding_eval [ 20.5%] 49632/242272 windows running_bpb=1.148337
+  sliding_eval [ 21.1%] 51232/242272 windows running_bpb=1.149698
+  sliding_eval [ 21.8%] 52832/242272 windows running_bpb=1.149618
+  sliding_eval [ 22.5%] 54432/242272 windows running_bpb=1.150228
+  sliding_eval [ 23.1%] 56032/242272 windows running_bpb=1.154563
+  sliding_eval [ 23.8%] 57632/242272 windows running_bpb=1.156220
+  sliding_eval [ 24.4%] 59232/242272 windows running_bpb=1.156048
+  sliding_eval [ 25.1%] 60832/242272 windows running_bpb=1.153769
+  sliding_eval [ 25.8%] 62432/242272 windows running_bpb=1.154321
+  sliding_eval [ 26.4%] 64032/242272 windows running_bpb=1.154800
+  sliding_eval [ 27.1%] 65632/242272 windows running_bpb=1.154398
+  sliding_eval [ 27.8%] 67232/242272 windows running_bpb=1.153532
+  sliding_eval [ 28.4%] 68832/242272 windows running_bpb=1.152817
+  sliding_eval [ 29.1%] 70432/242272 windows running_bpb=1.152941
+  sliding_eval [ 29.7%] 72032/242272 windows running_bpb=1.153438
+  sliding_eval [ 30.4%] 73632/242272 windows running_bpb=1.152470
+  sliding_eval [ 31.1%] 75232/242272 windows running_bpb=1.152906
+  sliding_eval [ 31.7%] 76832/242272 windows running_bpb=1.153265
+  sliding_eval [ 32.4%] 78432/242272 windows running_bpb=1.151302
+  sliding_eval [ 33.0%] 80032/242272 windows running_bpb=1.150894
+  sliding_eval [ 33.7%] 81632/242272 windows running_bpb=1.149581
+  sliding_eval [ 34.4%] 83232/242272 windows running_bpb=1.149850
+  sliding_eval [ 35.0%] 84832/242272 windows running_bpb=1.149970
+  sliding_eval [ 35.7%] 86432/242272 windows running_bpb=1.150096
+  sliding_eval [ 36.3%] 88032/242272 windows running_bpb=1.149334
+  sliding_eval [ 37.0%] 89632/242272 windows running_bpb=1.148974
+  sliding_eval [ 37.7%] 91232/242272 windows running_bpb=1.148548
+  sliding_eval [ 38.3%] 92832/242272 windows running_bpb=1.148894
+  sliding_eval [ 39.0%] 94432/242272 windows running_bpb=1.148135
+  sliding_eval [ 39.6%] 96032/242272 windows running_bpb=1.148161
+  sliding_eval [ 40.3%] 97632/242272 windows running_bpb=1.148097
+  sliding_eval [ 41.0%] 99232/242272 windows running_bpb=1.149357
+  sliding_eval [ 41.6%] 100832/242272 windows running_bpb=1.150008
+  sliding_eval [ 42.3%] 102432/242272 windows running_bpb=1.150455
+  sliding_eval [ 42.9%] 104032/242272 windows running_bpb=1.150444
+  sliding_eval [ 43.6%] 105632/242272 windows running_bpb=1.150980
+  sliding_eval [ 44.3%] 107232/242272 windows running_bpb=1.150059
+  sliding_eval [ 44.9%] 108832/242272 windows running_bpb=1.150529
+  sliding_eval [ 45.6%] 110432/242272 windows running_bpb=1.150892
+  sliding_eval [ 46.2%] 112032/242272 windows running_bpb=1.150888
+  sliding_eval [ 46.9%] 113632/242272 windows running_bpb=1.150305
+  sliding_eval [ 47.6%] 115232/242272 windows running_bpb=1.149878
+  sliding_eval [ 48.2%] 116832/242272 windows running_bpb=1.146555
+  sliding_eval [ 48.9%] 118432/242272 windows running_bpb=1.146310
+  sliding_eval [ 49.5%] 120032/242272 windows running_bpb=1.146286
+  sliding_eval [ 50.2%] 121632/242272 windows running_bpb=1.146249
+  sliding_eval [ 50.9%] 123232/242272 windows running_bpb=1.146432
+  sliding_eval [ 51.5%] 124832/242272 windows running_bpb=1.147246
+  sliding_eval [ 52.2%] 126432/242272 windows running_bpb=1.147043
+  sliding_eval [ 52.8%] 128032/242272 windows running_bpb=1.147343
+  sliding_eval [ 53.5%] 129632/242272 windows running_bpb=1.147541
+  sliding_eval [ 54.2%] 131232/242272 windows running_bpb=1.147218
+  sliding_eval [ 54.8%] 132832/242272 windows running_bpb=1.146483
+  sliding_eval [ 55.5%] 134432/242272 windows running_bpb=1.146041
+  sliding_eval [ 56.1%] 136032/242272 windows running_bpb=1.145959
+  sliding_eval [ 56.8%] 137632/242272 windows running_bpb=1.145608
+  sliding_eval [ 57.5%] 139232/242272 windows running_bpb=1.145357
+  sliding_eval [ 58.1%] 140832/242272 windows running_bpb=1.145006
+  sliding_eval [ 58.8%] 142432/242272 windows running_bpb=1.145106
+  sliding_eval [ 59.5%] 144032/242272 windows running_bpb=1.145123
+  sliding_eval [ 60.1%] 145632/242272 windows running_bpb=1.145001
+  sliding_eval [ 60.8%] 147232/242272 windows running_bpb=1.145069
+  sliding_eval [ 61.4%] 148832/242272 windows running_bpb=1.145430
+  sliding_eval [ 62.1%] 150432/242272 windows running_bpb=1.145285
+  sliding_eval [ 62.8%] 152032/242272 windows running_bpb=1.144640
+  sliding_eval [ 63.4%] 153632/242272 windows running_bpb=1.144940
+  sliding_eval [ 64.1%] 155232/242272 windows running_bpb=1.145141
+  sliding_eval [ 64.7%] 156832/242272 windows running_bpb=1.145557
+  sliding_eval [ 65.4%] 158432/242272 windows running_bpb=1.145263
+  sliding_eval [ 66.1%] 160032/242272 windows running_bpb=1.145842
+  sliding_eval [ 66.7%] 161632/242272 windows running_bpb=1.146168
+  sliding_eval [ 67.4%] 163232/242272 windows running_bpb=1.145583
+  sliding_eval [ 68.0%] 164832/242272 windows running_bpb=1.146147
+  sliding_eval [ 68.7%] 166432/242272 windows running_bpb=1.146584
+  sliding_eval [ 69.4%] 168032/242272 windows running_bpb=1.148346
+  sliding_eval [ 70.0%] 169632/242272 windows running_bpb=1.148445
+  sliding_eval [ 70.7%] 171232/242272 windows running_bpb=1.147486
+  sliding_eval [ 71.3%] 172832/242272 windows running_bpb=1.147799
+  sliding_eval [ 72.0%] 174432/242272 windows running_bpb=1.148051
+  sliding_eval [ 72.7%] 176032/242272 windows running_bpb=1.148529
+  sliding_eval [ 73.3%] 177632/242272 windows running_bpb=1.148677
+  sliding_eval [ 74.0%] 179232/242272 windows running_bpb=1.148821
+  sliding_eval [ 74.6%] 180832/242272 windows running_bpb=1.149069
+  sliding_eval [ 75.3%] 182432/242272 windows running_bpb=1.148907
+  sliding_eval [ 76.0%] 184032/242272 windows running_bpb=1.148289
+  sliding_eval [ 76.6%] 185632/242272 windows running_bpb=1.148415
+  sliding_eval [ 77.3%] 187232/242272 windows running_bpb=1.148553
+  sliding_eval [ 77.9%] 188832/242272 windows running_bpb=1.148651
+  sliding_eval [ 78.6%] 190432/242272 windows running_bpb=1.148539
+  sliding_eval [ 79.3%] 192032/242272 windows running_bpb=1.148040
+  sliding_eval [ 79.9%] 193632/242272 windows running_bpb=1.151216
+  sliding_eval [ 80.6%] 195232/242272 windows running_bpb=1.150878
+  sliding_eval [ 81.2%] 196832/242272 windows running_bpb=1.150902
+  sliding_eval [ 81.9%] 198432/242272 windows running_bpb=1.150833
+  sliding_eval [ 82.6%] 200032/242272 windows running_bpb=1.150673
+  sliding_eval [ 83.2%] 201632/242272 windows running_bpb=1.150851
+  sliding_eval [ 83.9%] 203232/242272 windows running_bpb=1.151183
+  sliding_eval [ 84.5%] 204832/242272 windows running_bpb=1.150503
+  sliding_eval [ 85.2%] 206432/242272 windows running_bpb=1.150362
+  sliding_eval [ 85.9%] 208032/242272 windows running_bpb=1.149994
+  sliding_eval [ 86.5%] 209632/242272 windows running_bpb=1.149507
+  sliding_eval [ 87.2%] 211232/242272 windows running_bpb=1.149268
+  sliding_eval [ 87.8%] 212832/242272 windows running_bpb=1.149192
+  sliding_eval [ 88.5%] 214432/242272 windows running_bpb=1.149237
+  sliding_eval [ 89.2%] 216032/242272 windows running_bpb=1.149386
+  sliding_eval [ 89.8%] 217632/242272 windows running_bpb=1.149851
+  sliding_eval [ 90.5%] 219232/242272 windows running_bpb=1.150044
+  sliding_eval [ 91.2%] 220832/242272 windows running_bpb=1.149902
+  sliding_eval [ 91.8%] 222432/242272 windows running_bpb=1.150079
+  sliding_eval [ 92.5%] 224032/242272 windows running_bpb=1.149857
+  sliding_eval [ 93.1%] 225632/242272 windows running_bpb=1.150505
+  sliding_eval [ 93.8%] 227232/242272 windows running_bpb=1.150324
+  sliding_eval [ 94.5%] 228832/242272 windows running_bpb=1.150188
+  sliding_eval [ 95.1%] 230432/242272 windows running_bpb=1.150079
+  sliding_eval [ 95.8%] 232032/242272 windows running_bpb=1.150012
+  sliding_eval [ 96.4%] 233632/242272 windows running_bpb=1.149701
+  sliding_eval [ 97.1%] 235232/242272 windows running_bpb=1.150124
+  sliding_eval [ 97.8%] 236832/242272 windows running_bpb=1.150019
+  sliding_eval [ 98.4%] 238432/242272 windows running_bpb=1.150113
+  sliding_eval [ 99.1%] 240032/242272 windows running_bpb=1.150080
+  sliding_eval [ 99.7%] 241632/242272 windows running_bpb=1.150319
+final_int8_zlib_roundtrip val_loss:1.9323 val_bpb:1.1444 eval_time:341104ms
+final_int8_zlib_roundtrip_exact val_loss:1.93231842 val_bpb:1.14443160
diff --git a/records/track_10min_16mb/2026-03-22_11L_XSA4_TightSWA_TTT/train_gpt.py b/records/track_10min_16mb/2026-03-22_11L_XSA4_TightSWA_TTT/train_gpt.py
new file mode 100644
index 000000000..6a7d0e60a
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-22_11L_XSA4_TightSWA_TTT/train_gpt.py
@@ -0,0 +1,1500 @@
+"""
+v38: Tight SWA (start at scale<0.2, every 50 steps) — fresher checkpoints, less SWA penalty.
+"""
+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
+try:
+    import zstandard
+    _COMPRESSOR = "zstd"
+except ImportError:
+    _COMPRESSOR = "zlib"
+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
+try:
+    from flash_attn_interface import flash_attn_func as flash_attn_3_func
+    _HAS_FA3 = True
+except ImportError:
+    try:
+        from flash_attn import flash_attn_func as flash_attn_3_func
+        _HAS_FA3 = True
+    except ImportError:
+        _HAS_FA3 = False
+        flash_attn_3_func = None
+# -----------------------------
+# -----------------------------
+
+class Hyperparameters:
+    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))
+    val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288))
+    val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000))
+    train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500))
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000))
+    warmup_steps = int(os.environ.get("WARMUP_STEPS", 20))
+    train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432))
+    train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048))
+    eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048))
+    max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0))
+    qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5))
+    vocab_size = int(os.environ.get("VOCAB_SIZE", 1024))
+    num_layers = int(os.environ.get("NUM_LAYERS", 11))
+    num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4))
+    model_dim = int(os.environ.get("MODEL_DIM", 512))
+    num_heads = int(os.environ.get("NUM_HEADS", 8))
+    mlp_mult = float(os.environ.get("MLP_MULT", 3.0))
+    tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1")))
+    rope_base = float(os.environ.get("ROPE_BASE", 10000.0))
+    logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0))
+    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", 0.035))
+    tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005))
+    matrix_lr = float(os.environ.get("MATRIX_LR", 0.025))
+    scalar_lr = float(os.environ.get("SCALAR_LR", 0.025))
+    muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3))
+    eval_stride = int(os.environ.get("EVAL_STRIDE", 64))
+    mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0))
+    mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2))
+    muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95))
+    swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1")))
+    swa_every = int(os.environ.get("SWA_EVERY", 50))  # tighter: collect more recent checkpoints
+    muon_wd = float(os.environ.get("MUON_WD", 0.04))
+    adam_wd = float(os.environ.get("ADAM_WD", 0.04))
+    qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0")))
+    bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048))
+    bigram_dim = int(os.environ.get("BIGRAM_DIM", 128))
+    xsa_last_n = int(os.environ.get("XSA_LAST_N", 4))  # XSA on last 4 layers (0 = disabled)
+    rope_dims = int(os.environ.get("ROPE_DIMS", 16))
+    ln_scale = bool(int(os.environ.get("LN_SCALE", "1")))
+    dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0")))
+    late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.1))
+    ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1")))
+    ve_dim = int(os.environ.get("VE_DIM", 128))
+    ve_layers = os.environ.get("VE_LAYERS", "9,10")
+# -----------------------------
+# -----------------------------
+
+def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor:
+    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, weight_decay: float = 0.0):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps,
+                 nesterov=nesterov, weight_decay=weight_decay),
+        )
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                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)
+                    g *= max(1, g.size(0) / g.size(1)) ** 0.5
+                    updates_flat[curr : curr + p.numel()] = g.reshape(-1)
+                curr += p.numel()
+            if distributed:
+                dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM)
+            wd = group.get("weight_decay", 0.0)
+            curr = 0
+            for p in params:
+                if wd > 0.0:
+                    p.data.mul_(1.0 - lr * wd)
+                g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype)
+                p.add_(g, alpha=-lr)
+                curr += p.numel()
+        return loss
+# -----------------------------
+# -----------------------------
+
+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}")
+    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,
+    eval_seq_len: int | None = None,
+) -> tuple[float, float]:
+    seq_len = eval_seq_len or args.train_seq_len
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < 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}, seq_len={seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // seq_len
+    total_seqs = (val_tokens.numel() - 1) // seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * seq_len
+            raw_end = batch_seq_end * seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, seq_len)
+            y = local[1:].reshape(-1, seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            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)
+# -----------------------------
+# -----------------------------
+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,smear,dtg_gate,ve_layer_scales,ve_shared.scale",
+    ).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:
+        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()
+    clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0
+    scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32)
+    q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous()
+    return q, scale
+
+def quantize_state_dict_int8(state_dict: dict[str, Tensor]):
+    quantized: dict[str, Tensor] = {}
+    scales: dict[str, Tensor] = {}
+    dtypes: dict[str, str] = {}
+    passthrough: dict[str, Tensor] = {}
+    passthrough_orig_dtypes: dict[str, str] = {}
+    qmeta: dict[str, dict[str, object]] = {}
+    stats = dict.fromkeys(
+        ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"),
+        0,
+    )
+    for name, tensor in state_dict.items():
+        t = tensor.detach().to("cpu").contiguous()
+        stats["param_count"] += int(t.numel())
+        stats["num_tensors"] += 1
+        stats["baseline_tensor_bytes"] += tensor_nbytes(t)
+        if not t.is_floating_point():
+            stats["num_nonfloat_tensors"] += 1
+            passthrough[name] = t
+            stats["int8_payload_bytes"] += tensor_nbytes(t)
+            continue
+        if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL:
+            kept = keep_float_tensor(name, t, passthrough_orig_dtypes)
+            passthrough[name] = kept
+            stats["int8_payload_bytes"] += tensor_nbytes(kept)
+            continue
+        stats["num_float_tensors"] += 1
+        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)
+            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():
+        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
+# -----------------------------
+# -----------------------------
+
+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)
+    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:
+    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:
+    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)
+# -----------------------------
+# -----------------------------
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+class CastedLinear(nn.Linear):
+    _qat_enabled: bool = False
+
+    def forward(self, x: Tensor) -> Tensor:
+        w = self.weight.to(x.dtype)
+        if CastedLinear._qat_enabled and self.training and w.ndim == 2:
+            with torch.no_grad():
+                w32 = self.weight.float()
+                row_max = w32.abs().amax(dim=1)
+                scale = (row_max / 31.0).clamp_min(1.0 / 31.0)
+                w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype)
+            w = w + (w_q - w).detach()
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, w, bias)
+
+def restore_low_dim_params_to_fp32(module: nn.Module) -> None:
+    with torch.no_grad():
+        for name, param in module.named_parameters():
+            if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32:
+                param.data = param.data.float()
+
+class Rotary(nn.Module):
+    def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0):
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        self.train_seq_len = train_seq_len
+        self.rope_dims = rope_dims if rope_dims > 0 else dim
+        inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            or self._sin_cached is None
+            or self._seq_len_cached != seq_len
+            or self._cos_cached.device != device
+        ):
+            rd = self.rope_dims
+            if seq_len > self.train_seq_len:
+                scale = seq_len / self.train_seq_len
+                new_base = self.base * (scale ** (rd / (rd - 2)))
+                inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd))
+            else:
+                inv_freq = self.inv_freq.to(device)
+            t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
+            freqs = torch.outer(t, inv_freq)
+            self._cos_cached = freqs.cos()[None, :, None, :]
+            self._sin_cached = freqs.sin()[None, :, None, :]
+            self._seq_len_cached = seq_len
+        return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype)
+
+def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor:
+    if rope_dims > 0 and rope_dims < x.size(-1):
+        x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:]
+        half = rope_dims // 2
+        x1, x2 = x_rope[..., :half], x_rope[..., half:]
+        x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+        return torch.cat((x_rope, x_pass), dim=-1)
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+class CausalSelfAttention(nn.Module):
+    def __init__(
+        self,
+        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.rope_dims = 0  # set by GPT.__init__ for partial RoPE
+        self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024)
+        self.use_xsa = False  # set by GPT.__init__ for deep layers only
+
+    def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor:
+        """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave).
+        y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv."""
+        B, T, H, D = y.shape
+        Hkv = v.size(-2)
+        group = H // Hkv
+        y_g = y.reshape(B, T, Hkv, group, D)        # [B, T, Hkv, group, D]
+        vn = F.normalize(v, dim=-1).unsqueeze(-2)    # [B, T, Hkv, 1, D] — broadcast ready
+        proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn
+        return (y_g - proj).reshape(B, T, H, D)
+
+    def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor:
+        bsz, seqlen, dim = x.shape
+        q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim)
+        k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim)
+        v = self.c_v(x)
+        if v_embed is not None:
+            v = v + v_embed
+        v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim)
+        q = F.rms_norm(q, (q.size(-1),))
+        k = F.rms_norm(k, (k.size(-1),))
+        cos, sin = self.rotary(seqlen, x.device, q.dtype)
+        q = apply_rotary_emb(q, cos, sin, self.rope_dims)
+        k = apply_rotary_emb(k, cos, sin, self.rope_dims)
+        q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None]
+        if _HAS_FA3:
+            y = flash_attn_3_func(q, k, v, causal=True)
+        else:
+            y = F.scaled_dot_product_attention(
+                q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2),
+                attn_mask=None, is_causal=True,
+                enable_gqa=(self.num_kv_heads != self.num_heads),
+            ).transpose(1, 2)
+        if self.use_xsa:
+            y = self._xsa_efficient(y, v)
+        y = y.reshape(bsz, seqlen, dim)
+        return self.proj(y)
+
+class SmearGate(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32))
+
+    def forward(self, x: Tensor) -> Tensor:
+        g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :]
+        x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1)
+        return (1 - g) * x + g * x_prev
+
+class BigramHashEmbedding(nn.Module):
+    def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int):
+        super().__init__()
+        self.bigram_vocab_size = bigram_vocab_size
+        self.embed = nn.Embedding(bigram_vocab_size, bigram_dim)
+        nn.init.zeros_(self.embed.weight)
+        self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None
+        if self.proj is not None:
+            nn.init.zeros_(self.proj.weight)
+        self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32))
+
+    def bigram_hash(self, tokens: Tensor) -> Tensor:
+        t = tokens.to(torch.int32)
+        mod = self.bigram_vocab_size - 1
+        out = torch.empty_like(t)
+        out[..., 0] = mod
+        out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod
+        return out.long()
+
+    def forward(self, token_ids: Tensor) -> Tensor:
+        h = self.embed(self.bigram_hash(token_ids))
+        if self.proj is not None:
+            h = self.proj(h)
+        return h * self.scale.to(dtype=h.dtype)
+
+class ValueEmbedding(nn.Module):
+    """Reinject token identity into attention values at specific layers.
+    Each table maps vocab tokens to a low-dim embedding, projected to model_dim."""
+    def __init__(self, vocab_size: int, ve_dim: int, model_dim: int):
+        super().__init__()
+        self.embed = nn.Embedding(vocab_size, ve_dim)
+        nn.init.normal_(self.embed.weight, std=0.01)
+        self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None
+        if self.proj is not None:
+            nn.init.zeros_(self.proj.weight)
+        self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32))
+
+    def forward(self, token_ids: Tensor) -> Tensor:
+        h = self.embed(token_ids)
+        if self.proj is not None:
+            h = self.proj(h)
+        return h * self.scale.to(dtype=h.dtype)
+
+class MLP(nn.Module):
+    def __init__(self, dim: int, mlp_mult: int):
+        super().__init__()
+        hidden = int(mlp_mult * dim)
+        self.fc = CastedLinear(dim, hidden, bias=False)
+        self.proj = CastedLinear(hidden, dim, bias=False)
+        self.proj._zero_init = True
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = 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,
+        layer_idx: int = 0,
+        ln_scale: bool = False,
+        dtg: bool = False,
+    ):
+        super().__init__()
+        self.attn_norm = RMSNorm()
+        self.mlp_norm = RMSNorm()
+        self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init)
+        self.mlp = MLP(dim, mlp_mult)
+        self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float())
+        self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0
+        if dtg:
+            self.dtg_gate = nn.Linear(dim, 1, bias=True)
+            nn.init.zeros_(self.dtg_gate.weight)
+            nn.init.constant_(self.dtg_gate.bias, 2.0)
+        else:
+            self.dtg_gate = None
+
+    def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor:
+        mix = self.resid_mix.to(dtype=x.dtype)
+        x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0
+        attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed)
+        x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out
+        x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor)
+        if self.dtg_gate is not None:
+            gate = torch.sigmoid(self.dtg_gate(x_in.detach()))
+            x_out = x_in + gate * (x_out - x_in)
+        return x_out
+
+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,
+        mtp_num_heads: int = 0,
+        mtp_loss_weight: float = 0.1,
+        bigram_vocab_size: int = 0,
+        bigram_dim: int = 128,
+        xsa_last_n: int = 0,
+        rope_dims: int = 0,
+        ln_scale: bool = False,
+        dtg: bool = False,
+        ve_enabled: bool = False,
+        ve_dim: int = 128,
+        ve_layers: str = "9,10",
+    ):
+        super().__init__()
+        self._ve_target_dim = num_kv_heads * (model_dim // num_heads)  # kv_dim for value projection
+        if logit_softcap <= 0.0:
+            raise ValueError(f"logit_softcap must be positive, got {logit_softcap}")
+        self.tie_embeddings = tie_embeddings
+        self.tied_embed_init_std = tied_embed_init_std
+        self.logit_softcap = logit_softcap
+        self.mtp_num_heads = mtp_num_heads
+        self.mtp_loss_weight = mtp_loss_weight
+        self.tok_emb = nn.Embedding(vocab_size, model_dim)
+        self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None
+        self.smear = SmearGate(model_dim)
+        self.num_encoder_layers = num_layers // 2
+        self.num_decoder_layers = num_layers - self.num_encoder_layers
+        self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers)
+        self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32))
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    model_dim,
+                    num_heads,
+                    num_kv_heads,
+                    mlp_mult,
+                    rope_base,
+                    qk_gain_init,
+                    layer_idx=i,
+                    ln_scale=ln_scale,
+                    dtg=dtg,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        if rope_dims > 0:
+            head_dim = model_dim // num_heads
+            for block in self.blocks:
+                block.attn.rope_dims = rope_dims
+                block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims)
+        self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else []
+        kv_dim = self._ve_target_dim
+        if self.ve_layer_indices:
+            self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim)
+            self.ve_layer_scales = nn.ParameterList(
+                [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices]
+            )
+        else:
+            self.ve_shared = None
+            self.ve_layer_scales = nn.ParameterList()
+        self.value_embeds = nn.ModuleList()  # keep empty for compat
+        self.final_norm = RMSNorm()
+        self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False)
+        if self.lm_head is not None:
+            self.lm_head._zero_init = True
+        self.mtp_heads = nn.ModuleList(
+            [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)]
+        )
+        for head in self.mtp_heads:
+            head._zero_init = True
+        if xsa_last_n > 0:
+            for i in range(max(0, num_layers - xsa_last_n), num_layers):
+                self.blocks[i].attn.use_xsa = True
+        self._init_weights()
+
+    def _init_weights(self) -> None:
+        if self.tie_embeddings:
+            nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std)
+        num_layers = len(self.blocks)
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear):
+                if getattr(module, "_zero_init", False):
+                    nn.init.zeros_(module.weight)
+                elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64:
+                    nn.init.orthogonal_(module.weight, gain=1.0)
+                    if ".proj." in name or name.endswith(".proj"):
+                        with torch.no_grad():
+                            module.weight.mul_(1.0 / math.sqrt(2 * num_layers))
+
+    def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None:
+        """Get value embedding for a specific layer using shared table + per-layer scale."""
+        if self.ve_shared is None or layer_idx not in self.ve_layer_indices:
+            return None
+        if ve_cache is not None and 've' not in ve_cache:
+            ve_cache['ve'] = self.ve_shared(input_ids)
+        ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids)
+        ve_idx = self.ve_layer_indices.index(layer_idx)
+        return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype)
+
+    def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor:
+        x = self.tok_emb(input_ids)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        ve_cache: dict = {}
+        for i in range(self.num_encoder_layers):
+            ve = self._get_ve(i, input_ids, ve_cache)
+            x = self.blocks[i](x, x0, v_embed=ve)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            bi = self.num_encoder_layers + i
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            ve = self._get_ve(bi, input_ids, ve_cache)
+            x = self.blocks[bi](x, x0, v_embed=ve)
+        x = self.final_norm(x)
+        x_flat = x.reshape(-1, x.size(-1))
+        targets = target_ids.reshape(-1)
+        if self.tie_embeddings:
+            logits_proj = F.linear(x_flat, self.tok_emb.weight)
+        else:
+            if self.lm_head is None:
+                raise RuntimeError("lm_head is required when tie_embeddings=False")
+            logits_proj = self.lm_head(x_flat)
+        logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap)
+        main_loss = F.cross_entropy(logits.float(), targets, reduction="mean")
+        if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0:
+            _, seqlen, dim = x.shape
+            mtp_loss_sum = x.new_zeros(())
+            mtp_loss_count = 0
+            for k, mtp_head in enumerate(self.mtp_heads):
+                valid_t = seqlen - (k + 1)
+                if valid_t <= 0:
+                    continue
+                mtp_hidden = x[:, :valid_t, :].reshape(-1, dim)
+                mtp_targets = target_ids[:, k + 1 :].reshape(-1)
+                mtp_logits_proj = mtp_head(mtp_hidden)
+                mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap)
+                mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean")
+                mtp_loss_count += 1
+            if mtp_loss_count > 0:
+                main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count)
+        return main_loss
+
+    def forward_logits(self, input_ids: Tensor) -> Tensor:
+        """Return logits (bsz, seq_len, vocab) without computing loss."""
+        x = self.tok_emb(input_ids)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        ve_cache: dict = {}
+        for i in range(self.num_encoder_layers):
+            ve = self._get_ve(i, input_ids, ve_cache)
+            x = self.blocks[i](x, x0, v_embed=ve)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            bi = self.num_encoder_layers + i
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            ve = self._get_ve(bi, input_ids, ve_cache)
+            x = self.blocks[bi](x, x0, v_embed=ve)
+        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)
+# -----------------------------
+# -----------------------------
+
+def eval_val_sliding(
+    args: Hyperparameters,
+    base_model: nn.Module,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+    stride: int,
+    batch_seqs: int = 32,
+    eval_seq_len: int | None = None,
+) -> tuple[float, float]:
+    """Sliding window evaluation: each token scored with maximum context."""
+    seq_len = eval_seq_len or args.train_seq_len
+    total_tokens = val_tokens.numel() - 1
+    window_starts = [ws for ws in range(0, total_tokens, stride)
+                     if min(ws + seq_len, total_tokens) - ws >= 1]
+    total_windows = len(window_starts)
+    my_s = (total_windows * rank) // world_size
+    my_e = (total_windows * (rank + 1)) // world_size
+    my_windows = window_starts[my_s:my_e]
+    loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    token_count = torch.zeros((), device=device, dtype=torch.float64)
+    byte_count = torch.zeros((), device=device, dtype=torch.float64)
+    base_model.eval()
+    compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True)
+    with torch.inference_mode():
+        for bi in range(0, len(my_windows), batch_seqs):
+            batch_ws = my_windows[bi:bi + batch_seqs]
+            bsz = len(batch_ws)
+            x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            wlens: list[int] = []
+            for i, ws in enumerate(batch_ws):
+                end = min(ws + seq_len, total_tokens)
+                wlen = end - ws
+                wlens.append(wlen)
+                chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device)
+                x_batch[i, :wlen] = chunk[:-1]
+                y_batch[i, :wlen] = chunk[1:]
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                logits = compiled_logits(x_batch)
+            nll = F.cross_entropy(
+                logits.reshape(-1, logits.size(-1)).float(),
+                y_batch.reshape(-1),
+                reduction="none",
+            ).reshape(bsz, seq_len)
+            for i, ws in enumerate(batch_ws):
+                wlen = wlens[i]
+                s = 0 if ws == 0 else max(wlen - stride, 0)
+                scored_nll = nll[i, s:wlen].to(torch.float64)
+                loss_sum += scored_nll.sum()
+                token_count += float(wlen - s)
+                tgt = y_batch[i, s:wlen]
+                prev = x_batch[i, s:wlen]
+                tb = base_bytes_lut[tgt].to(torch.float64)
+                tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64)
+                byte_count += tb.sum()
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(byte_count, op=dist.ReduceOp.SUM)
+    val_loss = (loss_sum / token_count).item()
+    bits_per_token = val_loss / math.log(2.0)
+    tokens_per_byte = token_count.item() / byte_count.item()
+    base_model.train()
+    return val_loss, bits_per_token * tokens_per_byte
+# -----------------------------
+# -----------------------------
+
+def _classify_param(name: str) -> str:
+    if "tok_emb" in name or "lm_head" in name:
+        return "embed"
+    if ".mlp." in name:
+        return "mlp"
+    if ".attn." in name or (".proj." in name and ".mlp." not in name):
+        return "attn"
+    return "other"
+
+def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]:
+    t32 = t.float()
+    if t32.ndim == 2:
+        row_max = t32.abs().amax(dim=1)
+        scale = (row_max / 31.0).clamp_min(1.0 / 31.0).to(torch.float16)
+        q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8)
+        return q, scale
+    amax = t32.abs().max().item()
+    scale = torch.tensor(amax / 31.0 if amax > 0 else 1.0, dtype=torch.float16)
+    q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8)
+    return q, scale
+
+def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]):
+    num_layers_total = max(
+        (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")),
+        default=0,
+    ) + 1
+    late_k_layers = set(range(num_layers_total - 2, num_layers_total))
+    result: dict[str, Tensor] = {}
+    meta: dict[str, object] = {}
+    for name, tensor in state_dict.items():
+        t = tensor.detach().cpu().contiguous()
+        cat = _classify_param(name)
+        if not t.is_floating_point() or t.numel() <= 65536:
+            result[name] = t.to(torch.float16) if t.is_floating_point() else t
+            meta[name] = "passthrough"
+            continue
+        if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS):
+            result[name] = t.float()
+            meta[name] = "passthrough_ctrl"
+            continue
+        if cat in int6_cats and t.ndim >= 1:
+            q, s = quantize_int6_per_row(t)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": "int6"}
+        else:
+            q, s = quantize_float_tensor(t)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": "int8"}
+    return result, meta
+
+def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object],
+                          template_sd: dict[str, Tensor]) -> dict[str, Tensor]:
+    out: dict[str, Tensor] = {}
+    for name, orig in template_sd.items():
+        info = meta.get(name)
+        if info is None:
+            continue
+        orig_dtype = orig.dtype
+        if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"):
+            t = result[name]
+            if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16):
+                t = t.to(orig_dtype)
+            out[name] = t
+            continue
+        q, s = result[name + ".q"], result[name + ".scale"]
+        if s.ndim > 0:
+            out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype)
+        else:
+            out[name] = (q.float() * float(s.item())).to(orig_dtype)
+    return out
+# -----------------------------
+# -----------------------------
+
+def main() -> None:
+    global zeropower_via_newtonschulz5
+    code = Path(__file__).read_text(encoding="utf-8")
+    args = Hyperparameters()
+    zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5)
+    # -----------------------------
+    # -----------------------------
+    distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ
+    rank = int(os.environ.get("RANK", "0"))
+    world_size = int(os.environ.get("WORLD_SIZE", "1"))
+    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
+    if world_size <= 0:
+        raise ValueError(f"WORLD_SIZE must be positive, got {world_size}")
+    if 8 % world_size != 0:
+        raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral")
+    grad_accum_steps = 8 // world_size
+    grad_scale = 1.0 / grad_accum_steps
+    if not torch.cuda.is_available():
+        raise RuntimeError("CUDA is required")
+    device = torch.device("cuda", local_rank)
+    torch.cuda.set_device(device)
+    if distributed:
+        dist.init_process_group(backend="nccl", device_id=device)
+        dist.barrier()
+    master_process = rank == 0
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+    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)
+    # -----------------------------
+    # -----------------------------
+    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")))
+    effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len
+    val_seq_len = max(args.train_seq_len, effective_eval_seq_len)
+    val_tokens = load_validation_tokens(args.val_files, val_seq_len)
+    base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(
+        sp, args.vocab_size, device
+    )
+    log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}")
+    log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}")
+    log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}")
+    # -----------------------------
+    # -----------------------------
+    CastedLinear._qat_enabled = args.qat_enabled
+    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,
+        mtp_num_heads=args.mtp_num_heads,
+        mtp_loss_weight=args.mtp_loss_weight,
+        bigram_vocab_size=args.bigram_vocab_size,
+        bigram_dim=args.bigram_dim,
+        xsa_last_n=args.xsa_last_n,
+        rope_dims=args.rope_dims,
+        ln_scale=args.ln_scale,
+        dtg=args.dtg_enabled,
+        ve_enabled=args.ve_enabled,
+        ve_dim=args.ve_dim,
+        ve_layers=args.ve_layers,
+    ).to(device).bfloat16()
+    for module in base_model.modules():
+        if isinstance(module, CastedLinear):
+            module.float()
+    restore_low_dim_params_to_fp32(base_model)
+    compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True)
+    model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model
+    block_named_params = list(base_model.blocks.named_parameters())
+    matrix_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    if base_model.mtp_num_heads > 0:
+        matrix_params.extend([p for p in base_model.mtp_heads.parameters() if p.ndim == 2])
+    scalar_params = [
+        p
+        for name, p in block_named_params
+        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)
+    scalar_params.append(base_model.smear.gate)
+    if base_model.bigram is not None:
+        scalar_params.append(base_model.bigram.scale)
+    token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr
+    tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}]
+    if base_model.bigram is not None:
+        tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr})
+        if base_model.bigram.proj is not None:
+            matrix_params.append(base_model.bigram.proj.weight)
+    if base_model.ve_shared is not None:
+        tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr})
+        if base_model.ve_shared.proj is not None:
+            matrix_params.append(base_model.ve_shared.proj.weight)
+        scalar_params.append(base_model.ve_shared.scale)
+        for s in base_model.ve_layer_scales:
+            scalar_params.append(s)
+    optimizer_tok = torch.optim.AdamW(
+        tok_params,
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.adam_wd,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+        weight_decay=args.muon_wd,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.AdamW(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.adam_wd,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    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())
+    mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters())
+    log0(f"model_params:{n_params}")
+    log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}")
+    xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa]
+    log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}")
+    log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}")
+    log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False")
+    log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}")
+    log0(
+        f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} "
+        f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} "
+        f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}"
+    )
+    log0(
+        f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} "
+        f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} "
+        f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}"
+    )
+    log0(f"seed:{args.seed}")
+    # -----------------------------
+    # -----------------------------
+    train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    def zero_grad_all() -> None:
+        for opt in optimizers:
+            opt.zero_grad(set_to_none=True)
+    max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None
+
+    def lr_mul(step: int, elapsed_ms: float) -> float:
+        if args.warmdown_iters <= 0:
+            return 1.0
+        if max_wallclock_ms is None:
+            warmdown_start = max(args.iterations - args.warmdown_iters, 0)
+            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
+    if args.warmup_steps > 0:
+        initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()}
+        initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+    # -----------------------------
+    # -----------------------------
+    swa_state: dict[str, Tensor] | None = None
+    swa_count = 0
+    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)
+        if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled:
+            CastedLinear._qat_enabled = True
+            log0(f"late_qat:enabled step:{step} scale:{scale:.4f}")
+        zero_grad_all()
+        train_loss = torch.zeros((), device=device)
+        for micro_step in range(grad_accum_steps):
+            if distributed:
+                model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+            x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                loss = model(x, y)
+            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
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0:
+            if swa_state is None:
+                swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}
+                swa_count = 1
+                log0(f"swa:start step:{step}")
+            else:
+                for name, t in base_model.state_dict().items():
+                    swa_state[name] += t.detach().cpu()
+                swa_count += 1
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        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"
+            )
+        reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms
+        if distributed and max_wallclock_ms is not None:
+            reached_cap_tensor = torch.tensor(int(reached_cap), device=device)
+            dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX)
+            reached_cap = bool(reached_cap_tensor.item())
+        if stop_after_step is None and reached_cap:
+            stop_after_step = step
+    log0(
+        f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+        f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB"
+    )
+    if args.swa_enabled and swa_state is not None and swa_count > 1:
+        log0(f"swa:applying averaged {swa_count} checkpoints")
+        avg_state = {name: (t / swa_count).to(dtype=base_model.state_dict()[name].dtype)
+                     for name, t in swa_state.items()}
+        base_model.load_state_dict(avg_state, strict=True)
+        torch.cuda.synchronize()
+        t_diag = time.perf_counter()
+        diag_val_loss, diag_val_bpb = eval_val(
+            args, compiled_model, rank, world_size, device, grad_accum_steps,
+            val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut,
+        )
+        torch.cuda.synchronize()
+        log0(
+            f"DIAGNOSTIC post_swa val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} "
+            f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms"
+        )
+    # -----------------------------
+    # -----------------------------
+    full_state_dict = base_model.state_dict()
+    export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k}
+    excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k)
+    if excluded_mtp > 0:
+        log0(f"export_excluding_mtp_params:{excluded_mtp}")
+    if master_process:
+        torch.save(export_sd, "final_model.pt")
+        model_bytes = os.path.getsize("final_model.pt")
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model: {model_bytes} bytes")
+        log0(f"Code size: {code_bytes} bytes")
+    sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()}
+    quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"})
+    quant_buf = io.BytesIO()
+    torch.save({"w": quant_result, "m": quant_meta}, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9)
+    if master_process:
+        with open("final_model.int6.ptz", "wb") as f:
+            f.write(quant_blob)
+        quant_file_bytes = len(quant_blob)
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes")
+        log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes")
+    if distributed:
+        dist.barrier()
+    with open("final_model.int6.ptz", "rb") as f:
+        quant_blob_disk = f.read()
+    quant_state = torch.load(
+        io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)),
+        map_location="cpu",
+    )
+    deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu)
+    eval_model = GPT(
+        vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim,
+        num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult,
+        tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std,
+        logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init,
+        mtp_num_heads=0, mtp_loss_weight=0.0,
+        bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim,
+        xsa_last_n=args.xsa_last_n,  # must match training model
+        rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled,
+        ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers,
+    ).to(device).bfloat16()
+    for m in eval_model.modules():
+        if isinstance(m, CastedLinear):
+            m.float()
+    restore_low_dim_params_to_fp32(eval_model)
+    eval_model.load_state_dict(deq_state, strict=True)
+    ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1")))
+    ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3))
+    ttt_lr = float(os.environ.get("TTT_LR", 0.008))
+    ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9))
+    ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 64))
+    if ttt_enabled:
+        torch.cuda.synchronize()
+        t_ttt = time.perf_counter()
+        log0(f"ttt:start epochs:{ttt_epochs} lr:{ttt_lr} batch_seqs:{ttt_batch_seqs}")
+        eval_model.train()
+        ttt_compiled = torch.compile(eval_model, dynamic=False, fullgraph=True)
+        ttt_ddp = DDP(ttt_compiled, device_ids=[local_rank], broadcast_buffers=False) if distributed else ttt_compiled
+        ttt_opt = torch.optim.SGD(eval_model.parameters(), lr=ttt_lr, momentum=ttt_momentum)
+        seq_len = effective_eval_seq_len
+        n_val = val_tokens.numel() - 1
+        n_seqs = n_val // seq_len
+        for ttt_ep in range(ttt_epochs):
+            perm = torch.randperm(n_seqs)
+            ep_loss, ep_count = 0.0, 0
+            for bs in range(0, n_seqs, ttt_batch_seqs * world_size):
+                be = min(bs + ttt_batch_seqs * world_size, n_seqs)
+                local_bs = bs + rank * ttt_batch_seqs
+                local_be = min(local_bs + ttt_batch_seqs, be)
+                if local_bs >= be:
+                    continue
+                idx = perm[local_bs:local_be]
+                bx = torch.stack([val_tokens[i * seq_len : i * seq_len + seq_len] for i in idx]).to(device=device, dtype=torch.int64)
+                by = torch.stack([val_tokens[i * seq_len + 1 : i * seq_len + seq_len + 1] for i in idx]).to(device=device, dtype=torch.int64)
+                ttt_opt.zero_grad(set_to_none=True)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    loss = ttt_ddp(bx, by)
+                loss.backward()
+                ttt_opt.step()
+                ep_loss += loss.item() * (local_be - local_bs)
+                ep_count += local_be - local_bs
+            log0(f"ttt:epoch:{ttt_ep + 1}/{ttt_epochs} loss:{ep_loss / max(ep_count, 1):.4f}")
+        eval_model.eval()
+        if distributed:
+            for p in eval_model.parameters():
+                dist.broadcast(p.data, src=0)
+        torch.cuda.synchronize()
+        log0(f"ttt:done time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms")
+    sw_seq_len = effective_eval_seq_len
+    if args.eval_stride > 0 and args.eval_stride < sw_seq_len:
+        torch.cuda.synchronize()
+        t_slide = time.perf_counter()
+        sw_val_loss, sw_val_bpb = eval_val_sliding(
+            args, eval_model, rank, world_size, device,
+            val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut,
+            stride=args.eval_stride,
+            eval_seq_len=sw_seq_len,
+        )
+        torch.cuda.synchronize()
+        log0(
+            f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} "
+            f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms"
+        )
+        log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}")
+    if distributed:
+        dist.destroy_process_group()
+
+if __name__ == "__main__":
+    main()
diff --git a/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/README.md b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/README.md
new file mode 100644
index 000000000..187fbee49
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/README.md
@@ -0,0 +1,52 @@
+# 11L XSA4 + Tight SWA + FA3 + Two-Phase TTT
+
+## Summary
+
+Built on PR #374 (unnir's 11L XSA4 + Tight SWA base) with FA3 Hopper attention and a novel two-phase test-time training approach:
+
+- **FA3 Hopper:** 84.65ms/step (vs 96ms with SDPA/FA2), enabling 6,939 training steps in 600s.
+- **Phase 1 — Norm-Only Recalibration (100 epochs, Adam lr=0.01):** Only unfreeze LayerNorm weights, scales, and final_norm (~22K params). Recalibrates activation distributions damaged by int6 quantization. Acts as post-quantization calibration via gradient descent.
+- **Phase 2 — Selective-Freeze Block Adaptation (25 epochs, SGD lr=0.005, momentum=0.9):** Unfreeze last 3 transformer blocks + all norms + scales + lm_head (~7.6M params). Adapts representations on the recalibrated foundation while preserving SWA-averaged weights in the first 8 blocks.
+
+Key insight: the two phases target different error sources (quantization artifacts vs. distribution mismatch) and are additive.
+
+## Architecture
+
+- 11 transformer layers, dim=512, 8 heads, 4 KV heads (GQA)
+- 3x MLP with relu² + SmearGate + OrthoInit
+- XSA on last 4 layers (Exclusive Self-Attention)
+- Partial RoPE (16/64 head dims)
+- LN Scale (1/sqrt(layer+1))
+- BigramHash(2048), bigram_dim=128
+- Shared VE128 (value embeddings shared across layers 9-10)
+- Tight SWA (scale < 0.2), Late QAT (final 4%)
+- FA3 Hopper attention (flash_attn_interface)
+- Int6 quantization + zstd-22 compression
+- Magnitude pruning 1%
+
+## Setup
+
+```bash
+pip install zstandard
+pip install flash_attn_3 --find-links https://windreamer.github.io/flash-attention3-wheels/cu128_torch291
+```
+
+## Results
+
+```
+seed=1337: val_bpb=1.1216, artifact=15,704,756 bytes
+  training: 84.65ms/step, 6939 steps, 600s wallclock
+  post-SWA: val_bpb=1.1421
+  TTT phase 1 (norm-only):       100 epochs, 22K params, Adam lr=0.01
+  TTT phase 2 (selective-freeze): 25 epochs, 7.6M params, SGD lr=0.005
+  TTT total time: 705s
+  TTT improvement: -0.021 (1.1421 -> 1.1216)
+```
+
+Additional seeds in progress.
+
+## Command
+
+```bash
+torchrun --standalone --nproc_per_node=8 train_gpt.py
+```
diff --git a/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/submission.json b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/submission.json
new file mode 100644
index 000000000..cc88bc097
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/submission.json
@@ -0,0 +1,9 @@
+{
+  "name": "11L XSA4 + Tight SWA + FA3 + Two-Phase TTT",
+  "val_loss": 1.1216,
+  "bytes_total": 15704756,
+  "blurb": "11 layers, XSA on last 4, Partial RoPE 16/64, LN Scale, Tight SWA, BigramHash 2048, FA3 Hopper (84.65ms/step), two-phase TTT: phase 1 norm-only recalibration (100ep Adam), phase 2 selective-freeze last 3 blocks (25ep SGD). Based on PR #374.",
+  "author": "fbedev",
+  "github_id": "fbedev",
+  "date": "2026-03-22"
+}
diff --git a/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/train_gpt.py b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/train_gpt.py
new file mode 100644
index 000000000..e9e12780f
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-22_TwoPhase_TTT_NormRepair/train_gpt.py
@@ -0,0 +1,1474 @@
+"""Selective-freeze TTT with SAM (Sharpness-Aware). Fork of #374."""
+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
+try:
+    import zstandard
+    _COMPRESSOR = "zstd"
+except ImportError:
+    _COMPRESSOR = "zlib"
+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
+try:
+    from flash_attn_interface import flash_attn_func as flash_attn_3_func
+    _HAS_FA3 = True
+except ImportError:
+    try:
+        from flash_attn import flash_attn_func as flash_attn_3_func
+        _HAS_FA3 = True
+    except ImportError:
+        _HAS_FA3 = False
+        flash_attn_3_func = None
+
+class Hyperparameters:
+    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))
+    val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288))
+    val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000))
+    train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500))
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3000))
+    warmup_steps = int(os.environ.get("WARMUP_STEPS", 20))
+    train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432))
+    train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048))
+    eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048))
+    max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0))
+    qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5))
+    vocab_size = int(os.environ.get("VOCAB_SIZE", 1024))
+    num_layers = int(os.environ.get("NUM_LAYERS", 11))
+    num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4))
+    model_dim = int(os.environ.get("MODEL_DIM", 512))
+    num_heads = int(os.environ.get("NUM_HEADS", 8))
+    mlp_mult = float(os.environ.get("MLP_MULT", 3.0))
+    tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1")))
+    rope_base = float(os.environ.get("ROPE_BASE", 10000.0))
+    logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0))
+    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", 0.035))
+    tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005))
+    matrix_lr = float(os.environ.get("MATRIX_LR", 0.025))
+    scalar_lr = float(os.environ.get("SCALAR_LR", 0.025))
+    muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3))
+    eval_stride = int(os.environ.get("EVAL_STRIDE", 32))
+    muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95))
+    swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1")))
+    swa_every = int(os.environ.get("SWA_EVERY", 50))  # tighter: collect more recent checkpoints
+    muon_wd = float(os.environ.get("MUON_WD", 0.04))
+    adam_wd = float(os.environ.get("ADAM_WD", 0.04))
+    qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0")))
+    bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048))
+    bigram_dim = int(os.environ.get("BIGRAM_DIM", 128))
+    xsa_last_n = int(os.environ.get("XSA_LAST_N", 4))  # XSA on last 4 layers (0 = disabled)
+    rope_dims = int(os.environ.get("ROPE_DIMS", 16))
+    ln_scale = bool(int(os.environ.get("LN_SCALE", "1")))
+    dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0")))
+    late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.1))
+    ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1")))
+    ve_dim = int(os.environ.get("VE_DIM", 128))
+    ve_layers = os.environ.get("VE_LAYERS", "9,10")
+    prune_pct = float(os.environ.get("PRUNE_PCT", 0.01))
+
+def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor:
+    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, weight_decay: float = 0.0):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps,
+                 nesterov=nesterov, weight_decay=weight_decay),
+        )
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                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)
+                    g *= max(1, g.size(0) / g.size(1)) ** 0.5
+                    updates_flat[curr : curr + p.numel()] = g.reshape(-1)
+                curr += p.numel()
+            if distributed:
+                dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM)
+            wd = group.get("weight_decay", 0.0)
+            curr = 0
+            for p in params:
+                if wd > 0.0:
+                    p.data.mul_(1.0 - lr * wd)
+                g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype)
+                p.add_(g, alpha=-lr)
+                curr += p.numel()
+        return loss
+
+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}")
+    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,
+    eval_seq_len: int | None = None,
+) -> tuple[float, float]:
+    seq_len = eval_seq_len or args.train_seq_len
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < 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}, seq_len={seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // seq_len
+    total_seqs = (val_tokens.numel() - 1) // seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * seq_len
+            raw_end = batch_seq_end * seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, seq_len)
+            y = local[1:].reshape(-1, seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            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)
+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,smear,dtg_gate,ve_layer_scales,ve_shared.scale",
+    ).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:
+        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()
+    clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0
+    scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32)
+    q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous()
+    return q, scale
+
+def quantize_state_dict_int8(state_dict: dict[str, Tensor]):
+    quantized: dict[str, Tensor] = {}
+    scales: dict[str, Tensor] = {}
+    dtypes: dict[str, str] = {}
+    passthrough: dict[str, Tensor] = {}
+    passthrough_orig_dtypes: dict[str, str] = {}
+    qmeta: dict[str, dict[str, object]] = {}
+    stats = dict.fromkeys(
+        ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"),
+        0,
+    )
+    for name, tensor in state_dict.items():
+        t = tensor.detach().to("cpu").contiguous()
+        stats["param_count"] += int(t.numel())
+        stats["num_tensors"] += 1
+        stats["baseline_tensor_bytes"] += tensor_nbytes(t)
+        if not t.is_floating_point():
+            stats["num_nonfloat_tensors"] += 1
+            passthrough[name] = t
+            stats["int8_payload_bytes"] += tensor_nbytes(t)
+            continue
+        if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL:
+            kept = keep_float_tensor(name, t, passthrough_orig_dtypes)
+            passthrough[name] = kept
+            stats["int8_payload_bytes"] += tensor_nbytes(kept)
+            continue
+        stats["num_float_tensors"] += 1
+        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)
+            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():
+        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
+
+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)
+    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:
+    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:
+    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)
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+class CastedLinear(nn.Linear):
+    _qat_enabled: bool = False
+
+    def forward(self, x: Tensor) -> Tensor:
+        w = self.weight.to(x.dtype)
+        if CastedLinear._qat_enabled and self.training and w.ndim == 2:
+            with torch.no_grad():
+                w32 = self.weight.float()
+                row_max = w32.abs().amax(dim=1)
+                scale = (row_max / 31.0).clamp_min(1.0 / 31.0)
+                w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype)
+            w = w + (w_q - w).detach()
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, w, bias)
+
+def restore_low_dim_params_to_fp32(module: nn.Module) -> None:
+    with torch.no_grad():
+        for name, param in module.named_parameters():
+            if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32:
+                param.data = param.data.float()
+
+class Rotary(nn.Module):
+    def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0):
+        super().__init__()
+        self.dim = dim
+        self.base = base
+        self.train_seq_len = train_seq_len
+        self.rope_dims = rope_dims if rope_dims > 0 else dim
+        inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            or self._sin_cached is None
+            or self._seq_len_cached != seq_len
+            or self._cos_cached.device != device
+        ):
+            rd = self.rope_dims
+            if seq_len > self.train_seq_len:
+                scale = seq_len / self.train_seq_len
+                new_base = self.base * (scale ** (rd / (rd - 2)))
+                inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd))
+            else:
+                inv_freq = self.inv_freq.to(device)
+            t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
+            freqs = torch.outer(t, inv_freq)
+            self._cos_cached = freqs.cos()[None, :, None, :]
+            self._sin_cached = freqs.sin()[None, :, None, :]
+            self._seq_len_cached = seq_len
+        return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype)
+
+def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor:
+    if rope_dims > 0 and rope_dims < x.size(-1):
+        x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:]
+        half = rope_dims // 2
+        x1, x2 = x_rope[..., :half], x_rope[..., half:]
+        x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+        return torch.cat((x_rope, x_pass), dim=-1)
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+class CausalSelfAttention(nn.Module):
+    def __init__(
+        self,
+        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.rope_dims = 0  # set by GPT.__init__ for partial RoPE
+        self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024)
+        self.use_xsa = False  # set by GPT.__init__ for deep layers only
+
+    def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor:
+        """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave).
+        y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv."""
+        B, T, H, D = y.shape
+        Hkv = v.size(-2)
+        group = H // Hkv
+        y_g = y.reshape(B, T, Hkv, group, D)        # [B, T, Hkv, group, D]
+        vn = F.normalize(v, dim=-1).unsqueeze(-2)    # [B, T, Hkv, 1, D] — broadcast ready
+        proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn
+        return (y_g - proj).reshape(B, T, H, D)
+
+    def forward(self, x: Tensor, v_embed: Tensor | None = None) -> Tensor:
+        bsz, seqlen, dim = x.shape
+        q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim)
+        k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim)
+        v = self.c_v(x)
+        if v_embed is not None:
+            v = v + v_embed
+        v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim)
+        q = F.rms_norm(q, (q.size(-1),))
+        k = F.rms_norm(k, (k.size(-1),))
+        cos, sin = self.rotary(seqlen, x.device, q.dtype)
+        q = apply_rotary_emb(q, cos, sin, self.rope_dims)
+        k = apply_rotary_emb(k, cos, sin, self.rope_dims)
+        q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None]
+        if _HAS_FA3:
+            y = flash_attn_3_func(q, k, v, causal=True)
+        else:
+            y = F.scaled_dot_product_attention(
+                q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2),
+                attn_mask=None, is_causal=True,
+                enable_gqa=(self.num_kv_heads != self.num_heads),
+            ).transpose(1, 2)
+        if self.use_xsa:
+            y = self._xsa_efficient(y, v)
+        y = y.reshape(bsz, seqlen, dim)
+        return self.proj(y)
+
+class SmearGate(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32))
+
+    def forward(self, x: Tensor) -> Tensor:
+        g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :]
+        x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1)
+        return (1 - g) * x + g * x_prev
+
+class BigramHashEmbedding(nn.Module):
+    def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int):
+        super().__init__()
+        self.bigram_vocab_size = bigram_vocab_size
+        self.embed = nn.Embedding(bigram_vocab_size, bigram_dim)
+        nn.init.zeros_(self.embed.weight)
+        self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None
+        if self.proj is not None:
+            nn.init.zeros_(self.proj.weight)
+        self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32))
+
+    def bigram_hash(self, tokens: Tensor) -> Tensor:
+        t = tokens.to(torch.int32)
+        mod = self.bigram_vocab_size - 1
+        out = torch.empty_like(t)
+        out[..., 0] = mod
+        out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod
+        return out.long()
+
+    def forward(self, token_ids: Tensor) -> Tensor:
+        h = self.embed(self.bigram_hash(token_ids))
+        if self.proj is not None:
+            h = self.proj(h)
+        return h * self.scale.to(dtype=h.dtype)
+
+class ValueEmbedding(nn.Module):
+    """Reinject token identity into attention values at specific layers.
+    Each table maps vocab tokens to a low-dim embedding, projected to model_dim."""
+    def __init__(self, vocab_size: int, ve_dim: int, model_dim: int):
+        super().__init__()
+        self.embed = nn.Embedding(vocab_size, ve_dim)
+        nn.init.normal_(self.embed.weight, std=0.01)
+        self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None
+        if self.proj is not None:
+            nn.init.zeros_(self.proj.weight)
+        self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32))
+
+    def forward(self, token_ids: Tensor) -> Tensor:
+        h = self.embed(token_ids)
+        if self.proj is not None:
+            h = self.proj(h)
+        return h * self.scale.to(dtype=h.dtype)
+
+class MLP(nn.Module):
+    def __init__(self, dim: int, mlp_mult: int):
+        super().__init__()
+        hidden = int(mlp_mult * dim)
+        self.fc = CastedLinear(dim, hidden, bias=False)
+        self.proj = CastedLinear(hidden, dim, bias=False)
+        self.proj._zero_init = True
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = 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,
+        layer_idx: int = 0,
+        ln_scale: bool = False,
+        dtg: bool = False,
+    ):
+        super().__init__()
+        self.attn_norm = RMSNorm()
+        self.mlp_norm = RMSNorm()
+        self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init)
+        self.mlp = MLP(dim, mlp_mult)
+        self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32))
+        self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float())
+        self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0
+        if dtg:
+            self.dtg_gate = nn.Linear(dim, 1, bias=True)
+            nn.init.zeros_(self.dtg_gate.weight)
+            nn.init.constant_(self.dtg_gate.bias, 2.0)
+        else:
+            self.dtg_gate = None
+
+    def forward(self, x: Tensor, x0: Tensor, v_embed: Tensor | None = None) -> Tensor:
+        mix = self.resid_mix.to(dtype=x.dtype)
+        x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0
+        attn_out = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, v_embed=v_embed)
+        x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out
+        x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor)
+        if self.dtg_gate is not None:
+            gate = torch.sigmoid(self.dtg_gate(x_in.detach()))
+            x_out = x_in + gate * (x_out - x_in)
+        return x_out
+
+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,
+        bigram_vocab_size: int = 0,
+        bigram_dim: int = 128,
+        xsa_last_n: int = 0,
+        rope_dims: int = 0,
+        ln_scale: bool = False,
+        dtg: bool = False,
+        ve_enabled: bool = False,
+        ve_dim: int = 128,
+        ve_layers: str = "9,10",
+    ):
+        super().__init__()
+        self._ve_target_dim = num_kv_heads * (model_dim // num_heads)  # kv_dim for value projection
+        if logit_softcap <= 0.0:
+            raise ValueError(f"logit_softcap must be positive, got {logit_softcap}")
+        self.tie_embeddings = tie_embeddings
+        self.tied_embed_init_std = tied_embed_init_std
+        self.logit_softcap = logit_softcap
+        self.tok_emb = nn.Embedding(vocab_size, model_dim)
+        self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim) if bigram_vocab_size > 0 else None
+        self.smear = SmearGate(model_dim)
+        self.num_encoder_layers = num_layers // 2
+        self.num_decoder_layers = num_layers - self.num_encoder_layers
+        self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers)
+        self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32))
+        self.blocks = nn.ModuleList(
+            [
+                Block(
+                    model_dim,
+                    num_heads,
+                    num_kv_heads,
+                    mlp_mult,
+                    rope_base,
+                    qk_gain_init,
+                    layer_idx=i,
+                    ln_scale=ln_scale,
+                    dtg=dtg,
+                )
+                for i in range(num_layers)
+            ]
+        )
+        if rope_dims > 0:
+            head_dim = model_dim // num_heads
+            for block in self.blocks:
+                block.attn.rope_dims = rope_dims
+                block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims)
+        self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else []
+        kv_dim = self._ve_target_dim
+        if self.ve_layer_indices:
+            self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim)
+            self.ve_layer_scales = nn.ParameterList(
+                [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices]
+            )
+        else:
+            self.ve_shared = None
+            self.ve_layer_scales = nn.ParameterList()
+        self.value_embeds = nn.ModuleList()  # keep empty for compat
+        self.final_norm = RMSNorm()
+        self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False)
+        if self.lm_head is not None:
+            self.lm_head._zero_init = True
+        if xsa_last_n > 0:
+            for i in range(max(0, num_layers - xsa_last_n), num_layers):
+                self.blocks[i].attn.use_xsa = True
+        self._init_weights()
+
+    def _init_weights(self) -> None:
+        if self.tie_embeddings:
+            nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std)
+        num_layers = len(self.blocks)
+        for name, module in self.named_modules():
+            if isinstance(module, nn.Linear):
+                if getattr(module, "_zero_init", False):
+                    nn.init.zeros_(module.weight)
+                elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64:
+                    nn.init.orthogonal_(module.weight, gain=1.0)
+                    if ".proj." in name or name.endswith(".proj"):
+                        with torch.no_grad():
+                            module.weight.mul_(1.0 / math.sqrt(2 * num_layers))
+
+    def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None:
+        """Get value embedding for a specific layer using shared table + per-layer scale."""
+        if self.ve_shared is None or layer_idx not in self.ve_layer_indices:
+            return None
+        if ve_cache is not None and 've' not in ve_cache:
+            ve_cache['ve'] = self.ve_shared(input_ids)
+        ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids)
+        ve_idx = self.ve_layer_indices.index(layer_idx)
+        return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype)
+
+    def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor:
+        x = self.tok_emb(input_ids)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        ve_cache: dict = {}
+        for i in range(self.num_encoder_layers):
+            ve = self._get_ve(i, input_ids, ve_cache)
+            x = self.blocks[i](x, x0, v_embed=ve)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            bi = self.num_encoder_layers + i
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            ve = self._get_ve(bi, input_ids, ve_cache)
+            x = self.blocks[bi](x, x0, v_embed=ve)
+        x = self.final_norm(x)
+        x_flat = x.reshape(-1, x.size(-1))
+        targets = target_ids.reshape(-1)
+        if self.tie_embeddings:
+            logits_proj = F.linear(x_flat, self.tok_emb.weight)
+        else:
+            if self.lm_head is None:
+                raise RuntimeError("lm_head is required when tie_embeddings=False")
+            logits_proj = self.lm_head(x_flat)
+        logits = 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)
+        if self.bigram is not None:
+            x = x + self.bigram(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x = self.smear(x)
+        x0 = x
+        skips: list[Tensor] = []
+        ve_cache: dict = {}
+        for i in range(self.num_encoder_layers):
+            ve = self._get_ve(i, input_ids, ve_cache)
+            x = self.blocks[i](x, x0, v_embed=ve)
+            skips.append(x)
+        for i in range(self.num_decoder_layers):
+            bi = self.num_encoder_layers + i
+            if skips:
+                x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop()
+            ve = self._get_ve(bi, input_ids, ve_cache)
+            x = self.blocks[bi](x, x0, v_embed=ve)
+        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)
+
+def eval_val_sliding(
+    args: Hyperparameters,
+    base_model: nn.Module,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+    stride: int,
+    batch_seqs: int = 32,
+    eval_seq_len: int | None = None,
+) -> tuple[float, float]:
+    """Sliding window evaluation: each token scored with maximum context."""
+    seq_len = eval_seq_len or args.train_seq_len
+    total_tokens = val_tokens.numel() - 1
+    window_starts = [ws for ws in range(0, total_tokens, stride)
+                     if min(ws + seq_len, total_tokens) - ws >= 1]
+    total_windows = len(window_starts)
+    my_s = (total_windows * rank) // world_size
+    my_e = (total_windows * (rank + 1)) // world_size
+    my_windows = window_starts[my_s:my_e]
+    loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    token_count = torch.zeros((), device=device, dtype=torch.float64)
+    byte_count = torch.zeros((), device=device, dtype=torch.float64)
+    base_model.eval()
+    compiled_logits = torch.compile(base_model.forward_logits, dynamic=False, fullgraph=True)
+    with torch.inference_mode():
+        for bi in range(0, len(my_windows), batch_seqs):
+            batch_ws = my_windows[bi:bi + batch_seqs]
+            bsz = len(batch_ws)
+            x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device)
+            wlens: list[int] = []
+            for i, ws in enumerate(batch_ws):
+                end = min(ws + seq_len, total_tokens)
+                wlen = end - ws
+                wlens.append(wlen)
+                chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device)
+                x_batch[i, :wlen] = chunk[:-1]
+                y_batch[i, :wlen] = chunk[1:]
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+                logits = compiled_logits(x_batch)
+            nll = F.cross_entropy(
+                logits.reshape(-1, logits.size(-1)).float(),
+                y_batch.reshape(-1),
+                reduction="none",
+            ).reshape(bsz, seq_len)
+            for i, ws in enumerate(batch_ws):
+                wlen = wlens[i]
+                s = 0 if ws == 0 else max(wlen - stride, 0)
+                scored_nll = nll[i, s:wlen].to(torch.float64)
+                loss_sum += scored_nll.sum()
+                token_count += float(wlen - s)
+                tgt = y_batch[i, s:wlen]
+                prev = x_batch[i, s:wlen]
+                tb = base_bytes_lut[tgt].to(torch.float64)
+                tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64)
+                byte_count += tb.sum()
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(byte_count, op=dist.ReduceOp.SUM)
+    val_loss = (loss_sum / token_count).item()
+    bits_per_token = val_loss / math.log(2.0)
+    tokens_per_byte = token_count.item() / byte_count.item()
+    base_model.train()
+    return val_loss, bits_per_token * tokens_per_byte
+
+def _classify_param(name: str) -> str:
+    if "tok_emb" in name or "lm_head" in name:
+        return "embed"
+    if ".mlp." in name:
+        return "mlp"
+    if ".attn." in name or (".proj." in name and ".mlp." not in name):
+        return "attn"
+    return "other"
+
+def quantize_int6_per_row(t: Tensor) -> tuple[Tensor, Tensor]:
+    t32 = t.float()
+    if t32.ndim == 2:
+        row_max = t32.abs().amax(dim=1)
+        scale = (row_max / 31.0).clamp_min(1.0 / 31.0).to(torch.float16)
+        q = torch.clamp(torch.round(t32 / scale.float()[:, None]), -32, 31).to(torch.int8)
+        return q, scale
+    amax = t32.abs().max().item()
+    scale = torch.tensor(amax / 31.0 if amax > 0 else 1.0, dtype=torch.float16)
+    q = torch.clamp(torch.round(t32 / scale.float()), -32, 31).to(torch.int8)
+    return q, scale
+
+
+def mixed_quantize_int6(state_dict: dict[str, Tensor], int6_cats: set[str]):
+    num_layers_total = max(
+        (int(k.split(".")[1]) for k in state_dict if k.startswith("blocks.")),
+        default=0,
+    ) + 1
+    late_k_layers = set(range(num_layers_total - 2, num_layers_total))
+    result: dict[str, Tensor] = {}
+    meta: dict[str, object] = {}
+    for name, tensor in state_dict.items():
+        t = tensor.detach().cpu().contiguous()
+        cat = _classify_param(name)
+        if not t.is_floating_point() or t.numel() <= 65536:
+            result[name] = t.to(torch.float16) if t.is_floating_point() else t
+            meta[name] = "passthrough"
+            continue
+        if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS):
+            result[name] = t.float()
+            meta[name] = "passthrough_ctrl"
+            continue
+        if cat in int6_cats and t.ndim >= 1:
+            q, s = quantize_int6_per_row(t)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": "int6"}
+        else:
+            q, s = quantize_float_tensor(t)
+            result[name + ".q"] = q
+            result[name + ".scale"] = s
+            meta[name] = {"type": "int8"}
+    return result, meta
+
+def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object],
+                          template_sd: dict[str, Tensor]) -> dict[str, Tensor]:
+    out: dict[str, Tensor] = {}
+    for name, orig in template_sd.items():
+        info = meta.get(name)
+        if info is None:
+            continue
+        orig_dtype = orig.dtype
+        if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"):
+            t = result[name]
+            if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16):
+                t = t.to(orig_dtype)
+            out[name] = t
+            continue
+        q, s = result[name + ".q"], result[name + ".scale"]
+        if s.ndim > 0:
+            out[name] = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype)
+        else:
+            out[name] = (q.float() * float(s.item())).to(orig_dtype)
+    return out
+
+def main() -> None:
+    global zeropower_via_newtonschulz5
+    code = Path(__file__).read_text(encoding="utf-8")
+    args = Hyperparameters()
+    zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5)
+    # -----------------------------
+    # -----------------------------
+    distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ
+    rank = int(os.environ.get("RANK", "0"))
+    world_size = int(os.environ.get("WORLD_SIZE", "1"))
+    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
+    if world_size <= 0:
+        raise ValueError(f"WORLD_SIZE must be positive, got {world_size}")
+    if 8 % world_size != 0:
+        raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral")
+    grad_accum_steps = 8 // world_size
+    grad_scale = 1.0 / grad_accum_steps
+    if not torch.cuda.is_available():
+        raise RuntimeError("CUDA is required")
+    device = torch.device("cuda", local_rank)
+    torch.cuda.set_device(device)
+    if distributed:
+        dist.init_process_group(backend="nccl", device_id=device)
+        dist.barrier()
+    master_process = rank == 0
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+    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(f"Python {sys.version} PyTorch {torch.__version__}", console=False)
+    # -----------------------------
+    # -----------------------------
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.cuda.manual_seed_all(args.seed)
+    if not args.tokenizer_path.endswith(".model"):
+        raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}")
+    sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path)
+    if int(sp.vocab_size()) != args.vocab_size:
+        raise ValueError(
+            f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}"
+        )
+    dataset_dir = Path(args.data_path).resolve()
+    actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin")))
+    effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len
+    val_seq_len = max(args.train_seq_len, effective_eval_seq_len)
+    val_tokens = load_validation_tokens(args.val_files, val_seq_len)
+    base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(
+        sp, args.vocab_size, device
+    )
+    log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}")
+    log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}")
+    log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}")
+    # -----------------------------
+    # -----------------------------
+    CastedLinear._qat_enabled = args.qat_enabled
+    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,
+        bigram_vocab_size=args.bigram_vocab_size,
+        bigram_dim=args.bigram_dim,
+        xsa_last_n=args.xsa_last_n,
+        rope_dims=args.rope_dims,
+        ln_scale=args.ln_scale,
+        dtg=args.dtg_enabled,
+        ve_enabled=args.ve_enabled,
+        ve_dim=args.ve_dim,
+        ve_layers=args.ve_layers,
+    ).to(device).bfloat16()
+    for module in base_model.modules():
+        if isinstance(module, CastedLinear):
+            module.float()
+    restore_low_dim_params_to_fp32(base_model)
+    compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True)
+    model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model
+    block_named_params = list(base_model.blocks.named_parameters())
+    matrix_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    scalar_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    if base_model.skip_weights.numel() > 0:
+        scalar_params.append(base_model.skip_weights)
+    scalar_params.append(base_model.smear.gate)
+    if base_model.bigram is not None:
+        scalar_params.append(base_model.bigram.scale)
+    token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr
+    tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}]
+    if base_model.bigram is not None:
+        tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr})
+        if base_model.bigram.proj is not None:
+            matrix_params.append(base_model.bigram.proj.weight)
+    if base_model.ve_shared is not None:
+        tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr})
+        if base_model.ve_shared.proj is not None:
+            matrix_params.append(base_model.ve_shared.proj.weight)
+        scalar_params.append(base_model.ve_shared.scale)
+        for s in base_model.ve_layer_scales:
+            scalar_params.append(s)
+    optimizer_tok = torch.optim.AdamW(
+        tok_params,
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.adam_wd,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+        weight_decay=args.muon_wd,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.AdamW(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        weight_decay=args.adam_wd,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    if base_model.lm_head is not None:
+        optimizer_head = torch.optim.Adam(
+            [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}],
+            betas=(args.beta1, args.beta2),
+            eps=args.adam_eps,
+            fused=True,
+        )
+        optimizers.insert(1, optimizer_head)
+    n_params = sum(p.numel() for p in base_model.parameters())
+    log0(f"model_params:{n_params}")
+    xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa]
+    log0(f"XSA:{xsa_layers} ws:{world_size} gqa:{args.num_heads}/{args.num_kv_heads}")
+    log0(f"lr:embed={token_lr} matrix={args.matrix_lr} scalar={args.scalar_lr} batch:{args.train_batch_tokens} wall:{args.max_wallclock_seconds:.0f}s seed:{args.seed}")
+    train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+    def zero_grad_all() -> None:
+        for opt in optimizers:
+            opt.zero_grad(set_to_none=True)
+    max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None
+
+    def lr_mul(step: int, elapsed_ms: float) -> float:
+        if args.warmdown_iters <= 0:
+            return 1.0
+        if max_wallclock_ms is None:
+            warmdown_start = max(args.iterations - args.warmdown_iters, 0)
+            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
+    if args.warmup_steps > 0:
+        initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()}
+        initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+    # -----------------------------
+    # -----------------------------
+    swa_state: dict[str, Tensor] | None = None
+    swa_count = 0
+    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)
+        if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled:
+            CastedLinear._qat_enabled = True
+            log0(f"late_qat:enabled step:{step} scale:{scale:.4f}")
+        zero_grad_all()
+        train_loss = torch.zeros((), device=device)
+        for micro_step in range(grad_accum_steps):
+            if distributed:
+                model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+            x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                loss = model(x, y)
+            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
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0:
+            if swa_state is None:
+                swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}
+                swa_count = 1
+                log0(f"swa:start step:{step}")
+            else:
+                for name, t in base_model.state_dict().items():
+                    swa_state[name] += t.detach().cpu()
+                swa_count += 1
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        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"
+            )
+        reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms
+        if distributed and max_wallclock_ms is not None:
+            reached_cap_tensor = torch.tensor(int(reached_cap), device=device)
+            dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX)
+            reached_cap = bool(reached_cap_tensor.item())
+        if stop_after_step is None and reached_cap:
+            stop_after_step = step
+    log0(
+        f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+        f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB"
+    )
+    if args.swa_enabled and swa_state is not None and swa_count > 1:
+        log0(f"swa:applying averaged {swa_count} checkpoints")
+        avg_state = {name: (t / swa_count).to(dtype=base_model.state_dict()[name].dtype)
+                     for name, t in swa_state.items()}
+        base_model.load_state_dict(avg_state, strict=True)
+        torch.cuda.synchronize()
+        t_diag = time.perf_counter()
+        diag_val_loss, diag_val_bpb = eval_val(
+            args, compiled_model, rank, world_size, device, grad_accum_steps,
+            val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut,
+        )
+        torch.cuda.synchronize()
+        log0(
+            f"DIAGNOSTIC post_swa val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} "
+            f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms"
+        )
+    # -----------------------------
+    # -----------------------------
+    export_sd = base_model.state_dict()
+    if master_process:
+        torch.save(export_sd, "final_model.pt")
+        model_bytes = os.path.getsize("final_model.pt")
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model: {model_bytes} bytes")
+        log0(f"Code size: {code_bytes} bytes")
+    sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()}
+    if args.prune_pct > 0:
+        for k, v in sd_cpu.items():
+            if v.ndim == 2 and v.numel() > 65536:
+                thresh = torch.quantile(v.abs().float(), args.prune_pct)
+                v[v.abs() < thresh] = 0.0
+        if master_process:
+            log0(f"pruning:{args.prune_pct*100:.1f}% magnitude pruning applied")
+    quant_result, quant_meta = mixed_quantize_int6(sd_cpu, {"mlp", "attn"})
+    quant_buf = io.BytesIO()
+    torch.save({"w": quant_result, "m": quant_meta}, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    quant_blob = zstandard.ZstdCompressor(level=22).compress(quant_raw) if _COMPRESSOR == "zstd" else zlib.compress(quant_raw, 9)
+    if master_process:
+        with open("final_model.int6.ptz", "wb") as f:
+            f.write(quant_blob)
+        quant_file_bytes = len(quant_blob)
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model int6+{_COMPRESSOR}: {quant_file_bytes} bytes")
+        log0(f"Total submission size int6+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes")
+    if distributed:
+        dist.barrier()
+    with open("final_model.int6.ptz", "rb") as f:
+        quant_blob_disk = f.read()
+    quant_state = torch.load(
+        io.BytesIO(zstandard.ZstdDecompressor().decompress(quant_blob_disk) if _COMPRESSOR == "zstd" else zlib.decompress(quant_blob_disk)),
+        map_location="cpu",
+    )
+    deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu)
+    eval_model = GPT(
+        vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim,
+        num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult,
+        tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std,
+        logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init,
+        bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim,
+        xsa_last_n=args.xsa_last_n,  # must match training model
+        rope_dims=args.rope_dims, ln_scale=args.ln_scale, dtg=args.dtg_enabled,
+        ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers,
+    ).to(device).bfloat16()
+    for m in eval_model.modules():
+        if isinstance(m, CastedLinear):
+            m.float()
+    restore_low_dim_params_to_fp32(eval_model)
+    eval_model.load_state_dict(deq_state, strict=True)
+    ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1")))
+    ttt_batch_seqs = int(os.environ.get("TTT_BATCH_SEQS", 64))
+    # Phase 1: norm-only (quant repair)
+    ttt_p1_epochs = int(os.environ.get("TTT_P1_EPOCHS", 100))
+    ttt_p1_lr = float(os.environ.get("TTT_P1_LR", 0.01))
+    # Phase 2: selective-freeze (block adaptation)
+    ttt_p2_epochs = int(os.environ.get("TTT_P2_EPOCHS", 25))
+    ttt_p2_lr = float(os.environ.get("TTT_P2_LR", 0.005))
+    ttt_p2_unfreeze_last = int(os.environ.get("TTT_P2_UNFREEZE_LAST", 3))
+    ttt_p2_momentum = float(os.environ.get("TTT_P2_MOMENTUM", 0.9))
+
+    if ttt_enabled:
+        torch.cuda.synchronize()
+        t_ttt = time.perf_counter()
+        n_blocks = len(eval_model.blocks)
+        seq_len = effective_eval_seq_len
+        n_val = val_tokens.numel() - 1
+        n_seqs = n_val // seq_len
+        chunk = ttt_batch_seqs * world_size
+        usable = (n_seqs // chunk) * chunk
+        if usable == 0:
+            usable = n_seqs
+
+        def run_ttt_phase(phase_name, epochs, lr, param_selector, opt_fn):
+            """Run one phase of TTT with given params and optimizer."""
+            eval_model.train()
+            ttt_params = []
+            frozen_count, unfrozen_count = 0, 0
+            for name, p in eval_model.named_parameters():
+                if param_selector(name):
+                    p.requires_grad_(True)
+                    ttt_params.append(p)
+                    unfrozen_count += p.numel()
+                else:
+                    p.requires_grad_(False)
+                    frozen_count += p.numel()
+            log0(f"ttt:{phase_name}:start epochs:{epochs} lr:{lr} unfrozen:{unfrozen_count} frozen:{frozen_count}")
+            ttt_compiled = torch.compile(eval_model, dynamic=False, fullgraph=True)
+            ttt_opt = opt_fn(ttt_params, lr)
+            for ttt_ep in range(epochs):
+                cos_lr = lr * 0.5 * (1.0 + math.cos(math.pi * ttt_ep / epochs))
+                for pg in ttt_opt.param_groups:
+                    pg['lr'] = cos_lr
+                perm = torch.randperm(usable, device=device)
+                if distributed:
+                    dist.broadcast(perm, src=0)
+                perm = perm.cpu()
+                ep_loss, ep_count = 0.0, 0
+                for bs in range(0, usable, chunk):
+                    local_bs = bs + rank * ttt_batch_seqs
+                    local_be = local_bs + ttt_batch_seqs
+                    idx = perm[local_bs:local_be]
+                    bx = torch.stack([val_tokens[i * seq_len : i * seq_len + seq_len] for i in idx]).to(device=device, dtype=torch.int64)
+                    by = torch.stack([val_tokens[i * seq_len + 1 : i * seq_len + seq_len + 1] for i in idx]).to(device=device, dtype=torch.int64)
+                    ttt_opt.zero_grad(set_to_none=True)
+                    with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                        loss = ttt_compiled(bx, by)
+                    loss.backward()
+                    if distributed:
+                        for p in ttt_params:
+                            if p.grad is not None:
+                                dist.all_reduce(p.grad, op=dist.ReduceOp.SUM)
+                                p.grad /= world_size
+                    ttt_opt.step()
+                    ep_loss += loss.item() * ttt_batch_seqs
+                    ep_count += ttt_batch_seqs
+                log0(f"ttt:{phase_name}:epoch:{ttt_ep + 1}/{epochs} lr:{cos_lr:.6f} loss:{ep_loss / max(ep_count, 1):.4f}")
+            eval_model.eval()
+            # clear compiled graph for next phase
+            torch._dynamo.reset()
+            log0(f"ttt:{phase_name}:done")
+
+        # Phase 1: Norm-only (quantization repair)
+        def norm_selector(name):
+            return "norm" in name or "ln_scale" in name or "attn_scale" in name or "mlp_scale" in name or "resid_mix" in name or "final_norm" in name
+        run_ttt_phase("phase1_norm", ttt_p1_epochs, ttt_p1_lr, norm_selector,
+                      lambda params, lr: torch.optim.Adam(params, lr=lr))
+
+        # Phase 2: Selective-freeze (block adaptation)
+        def block_selector(name):
+            is_last_blocks = any(f"blocks.{i}." in name for i in range(n_blocks - ttt_p2_unfreeze_last, n_blocks))
+            is_norm = "norm" in name or "ln_scale" in name
+            is_scale = "attn_scale" in name or "mlp_scale" in name or "resid_mix" in name
+            is_head = "final_norm" in name or "tok_emb" in name or "lm_head" in name
+            return is_last_blocks or is_norm or is_scale or is_head
+        run_ttt_phase("phase2_block", ttt_p2_epochs, ttt_p2_lr, block_selector,
+                      lambda params, lr: torch.optim.SGD(params, lr=lr, momentum=ttt_p2_momentum))
+
+        torch.cuda.synchronize()
+        log0(f"ttt:all_done time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms")
+    sw_seq_len = effective_eval_seq_len
+    if args.eval_stride > 0 and args.eval_stride < sw_seq_len:
+        torch.cuda.synchronize()
+        t_slide = time.perf_counter()
+        sw_val_loss, sw_val_bpb = eval_val_sliding(
+            args, eval_model, rank, world_size, device,
+            val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut,
+            stride=args.eval_stride,
+            eval_seq_len=sw_seq_len,
+        )
+        torch.cuda.synchronize()
+        log0(
+            f"final_int6_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} "
+            f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms"
+        )
+        log0(f"final_int6_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}")
+    if distributed:
+        dist.destroy_process_group()
+if __name__ == "__main__":
+    main()