SOTA attempt (val_bpb=1.2064) (#49)

* SOTA attempt

* Improve score on SXM

---------

Co-authored-by: spokane-way <spokane@way>

6 files changed, 1575 insertions, 0 deletions

diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md
new file mode 100644
index 0000000..735cc0c
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/README.md
@@ -0,0 +1,65 @@
+This record submission is called `Long Context Seq2048 v2`.
+
+Configuration:
+- Layout: `VOCAB_SIZE=1024 NUM_LAYERS=9 MODEL_DIM=512 NUM_HEADS=8 NUM_KV_HEADS=4 MLP_MULT=2`
+- Tied output/input embeddings: `TIE_EMBEDDINGS=1`
+- Sequence length: `TRAIN_SEQ_LEN=2048`
+- Batching: `TRAIN_BATCH_TOKENS=524288`
+- Learning rates: `TIED_EMBED_LR=0.04 MATRIX_LR=0.032 SCALAR_LR=0.032`
+
+Command:
+```bash
+NCCL_IB_DISABLE=1 \
+RUN_ID=seq2048_sxm28_full_20260319a \
+DATA_PATH=./data/datasets/fineweb10B_sp1024 \
+TOKENIZER_PATH=./data/tokenizers/fineweb_1024_bpe.model \
+VOCAB_SIZE=1024 \
+MAX_WALLCLOCK_SECONDS=600 \
+torchrun --standalone --nproc_per_node=8 \
+  records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py
+```
+
+Verification environment:
+- `8x H100 80GB HBM3`
+- all-to-all `NV18` topology
+- `torch 2.8.0+cu128`
+
+Key metrics (from `train.log` in this folder, rerun on the target SXM-class box):
+- Timed training stopped at `11564/20000` steps due to the wallclock cap.
+- Pre-quant eval at stop: `val_loss:2.0269`, `val_bpb:1.2005`
+- Post-quant roundtrip eval: `val_loss:2.0359`, `val_bpb:1.2058`
+- Exact printed metric: `final_int8_zlib_roundtrip_exact val_bpb:1.20576485`
+- Train time: `600038ms` (`step_avg:51.89ms`)
+- Peak memory: `10247 MiB allocated`, `10488 MiB reserved`
+- Serialized model int8+zlib: `15819554 bytes`
+- Code size for this standalone record script: `47716 bytes`
+- Total submission size int8+zlib: `15867270 bytes`
+
+Additional full-run reproducibility logs included in this folder:
+- `train.log`: canonical SXM rerun, `SEED=1337`, `val_bpb=1.20576485`
+- `train_seed1338.log`: SXM rerun, `SEED=1338`, `val_bpb=1.20617460`
+- `train_seed1339.log`: SXM rerun, `SEED=1339`, `val_bpb=1.20715923`
+
+Record-track significance note:
+- The public repo state for this submission has `Naive Baseline` at `1.2243657`.
+- The challenge therefore requires beating `1.2193657` to claim a new record.
+- All three included SXM full runs clear that threshold:
+  - `SEED=1337`: `1.20576485`
+  - `SEED=1338`: `1.20617460`
+  - `SEED=1339`: `1.20715923`
+- Sample mean across the three runs: `1.20636623`
+- Sample standard deviation: `0.00071667`
+- One-sided one-sample t-test against `1.2193657`: `t=31.42` with `df=2`, which gives `p=0.00051`
+
+Why this folder is standalone:
+- `train_gpt.py` compiles from inside this record folder and was used for the canonical rerun whose output is saved as `train.log`.
+- No extra Python source files are required for the training path.
+- The only inputs expected at runtime are the cached dataset and tokenizer paths described in the main repo README.
+
+Included files:
+- `train_gpt.py` (standalone winning recipe with defaults baked in)
+- `README.md` (this file)
+- `submission.json` (leaderboard metadata)
+- `train.log` (canonical full log from the standalone record script)
+- `train_seed1338.log`, `train_seed1339.log` (extra full reruns for reproducibility)
+- `logs/seq2048_sxm28_*` (raw per-run tee output and trainer text logs from the SXM verification box)
diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/submission.json b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/submission.json
new file mode 100644
index 0000000..049a8b7
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/submission.json
@@ -0,0 +1,11 @@
+{
+  "author": "Spokane Way",
+  "github_id": "spokane-way",
+  "name": "Long Context Seq2048 v2",
+  "blurb": "SP-1024 9x512 KV4 run at TRAIN_SEQ_LEN=2048 with tuned seq2048 learning rates (0.040/0.032/0.032). This standalone record script reproduces the SXM-verified 10-minute artifact under the 16,000,000-byte cap.",
+  "date": "2026-03-19T04:50:00Z",
+  "val_loss": 2.03588345,
+  "val_bpb": 1.20576485,
+  "bytes_total": 15867270,
+  "bytes_code": 47716
+}
diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train.log b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train.log
new file mode 100644
index 0000000..15a6f65
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train.log
@@ -0,0 +1,124 @@
+
+*****************************************
+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. 
+*****************************************
+logs/seq2048_sxm28_full_20260319a.txt
+val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/root/parameter-golf-sxm28/data/tokenizers/fineweb_1024_bpe.model
+train_loader:dataset:fineweb10B_sp1024 train_shards:80
+val_loader:shards pattern=/root/parameter-golf-sxm28/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632
+model_params:17059912
+world_size:8 grad_accum_steps:1
+sdp_backends:cudnn=False flash=True mem_efficient=False math=False
+attention_mode:gqa num_heads:8 num_kv_heads:4
+tie_embeddings:True embed_lr:0.04 head_lr:0.0 matrix_lr:0.032 scalar_lr:0.032
+train_batch_tokens:524288 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000
+seed:1337
+warmup_step:1/20
+warmup_step:2/20
+warmup_step:3/20
+warmup_step:4/20
+warmup_step:5/20
+warmup_step:6/20
+warmup_step:7/20
+warmup_step:8/20
+warmup_step:9/20
+warmup_step:10/20
+warmup_step:11/20
+warmup_step:12/20
+warmup_step:13/20
+warmup_step:14/20
+warmup_step:15/20
+warmup_step:16/20
+warmup_step:17/20
+warmup_step:18/20
+warmup_step:19/20
+warmup_step:20/20
+step:0/20000 val_loss:6.9357 val_bpb:4.1077 train_time:0ms step_avg:0.02ms
+step:1/20000 train_loss:6.9370 train_time:27ms step_avg:27.23ms
+step:2/20000 train_loss:14.7712 train_time:74ms step_avg:36.88ms
+step:3/20000 train_loss:8.1324 train_time:125ms step_avg:41.59ms
+step:4/20000 train_loss:6.6083 train_time:176ms step_avg:44.01ms
+step:5/20000 train_loss:6.9060 train_time:227ms step_avg:45.47ms
+step:6/20000 train_loss:7.6667 train_time:279ms step_avg:46.44ms
+step:7/20000 train_loss:6.6546 train_time:330ms step_avg:47.13ms
+step:8/20000 train_loss:6.3864 train_time:381ms step_avg:47.66ms
+step:9/20000 train_loss:6.2280 train_time:433ms step_avg:48.07ms
+step:10/20000 train_loss:6.1411 train_time:484ms step_avg:48.42ms
+step:200/20000 train_loss:2.7753 train_time:10282ms step_avg:51.41ms
+step:400/20000 train_loss:2.2990 train_time:20615ms step_avg:51.54ms
+step:600/20000 train_loss:2.5004 train_time:30958ms step_avg:51.60ms
+step:800/20000 train_loss:2.2435 train_time:41311ms step_avg:51.64ms
+step:1000/20000 train_loss:2.3383 train_time:51684ms step_avg:51.68ms
+step:1000/20000 val_loss:2.2909 val_bpb:1.3568 train_time:51717ms step_avg:51.72ms
+step:1200/20000 train_loss:2.3520 train_time:62063ms step_avg:51.72ms
+step:1400/20000 train_loss:2.3778 train_time:72454ms step_avg:51.75ms
+step:1600/20000 train_loss:2.0422 train_time:82841ms step_avg:51.78ms
+step:1800/20000 train_loss:2.1630 train_time:93248ms step_avg:51.80ms
+step:2000/20000 train_loss:2.2122 train_time:103654ms step_avg:51.83ms
+step:2000/20000 val_loss:2.1924 val_bpb:1.2984 train_time:103687ms step_avg:51.84ms
+step:2200/20000 train_loss:2.0339 train_time:114067ms step_avg:51.85ms
+step:2400/20000 train_loss:2.1666 train_time:124488ms step_avg:51.87ms
+step:2600/20000 train_loss:2.3803 train_time:134904ms step_avg:51.89ms
+step:2800/20000 train_loss:2.1944 train_time:145315ms step_avg:51.90ms
+step:3000/20000 train_loss:2.1889 train_time:155728ms step_avg:51.91ms
+step:3000/20000 val_loss:2.1524 val_bpb:1.2748 train_time:155761ms step_avg:51.92ms
+step:3200/20000 train_loss:2.1507 train_time:166139ms step_avg:51.92ms
+step:3400/20000 train_loss:2.1186 train_time:176537ms step_avg:51.92ms
+step:3600/20000 train_loss:2.0636 train_time:186950ms step_avg:51.93ms
+step:3800/20000 train_loss:2.1715 train_time:197346ms step_avg:51.93ms
+step:4000/20000 train_loss:2.1326 train_time:207738ms step_avg:51.93ms
+step:4000/20000 val_loss:2.1285 val_bpb:1.2606 train_time:207770ms step_avg:51.94ms
+step:4200/20000 train_loss:2.1300 train_time:218180ms step_avg:51.95ms
+step:4400/20000 train_loss:2.0635 train_time:228563ms step_avg:51.95ms
+step:4600/20000 train_loss:1.9340 train_time:238947ms step_avg:51.95ms
+step:4800/20000 train_loss:2.2169 train_time:249326ms step_avg:51.94ms
+step:5000/20000 train_loss:1.9728 train_time:259712ms step_avg:51.94ms
+step:5000/20000 val_loss:2.1118 val_bpb:1.2507 train_time:259745ms step_avg:51.95ms
+step:5200/20000 train_loss:2.1346 train_time:270102ms step_avg:51.94ms
+step:5400/20000 train_loss:2.1480 train_time:280489ms step_avg:51.94ms
+step:5600/20000 train_loss:2.1403 train_time:290858ms step_avg:51.94ms
+step:5800/20000 train_loss:2.0939 train_time:301230ms step_avg:51.94ms
+step:6000/20000 train_loss:2.1745 train_time:311608ms step_avg:51.93ms
+step:6000/20000 val_loss:2.1015 val_bpb:1.2446 train_time:311642ms step_avg:51.94ms
+step:6200/20000 train_loss:2.0438 train_time:321983ms step_avg:51.93ms
+step:6400/20000 train_loss:2.1272 train_time:332352ms step_avg:51.93ms
+step:6600/20000 train_loss:2.0825 train_time:342718ms step_avg:51.93ms
+step:6800/20000 train_loss:2.1436 train_time:353087ms step_avg:51.92ms
+step:7000/20000 train_loss:2.1914 train_time:363453ms step_avg:51.92ms
+step:7000/20000 val_loss:2.0907 val_bpb:1.2382 train_time:363485ms step_avg:51.93ms
+step:7200/20000 train_loss:2.1618 train_time:373813ms step_avg:51.92ms
+step:7400/20000 train_loss:2.0806 train_time:384181ms step_avg:51.92ms
+step:7600/20000 train_loss:1.9643 train_time:394550ms step_avg:51.91ms
+step:7800/20000 train_loss:2.1069 train_time:404903ms step_avg:51.91ms
+step:8000/20000 train_loss:2.0808 train_time:415270ms step_avg:51.91ms
+step:8000/20000 val_loss:2.0816 val_bpb:1.2328 train_time:415302ms step_avg:51.91ms
+step:8200/20000 train_loss:2.1517 train_time:425628ms step_avg:51.91ms
+step:8400/20000 train_loss:2.0958 train_time:436033ms step_avg:51.91ms
+step:8600/20000 train_loss:2.1052 train_time:446388ms step_avg:51.91ms
+step:8800/20000 train_loss:2.0699 train_time:456752ms step_avg:51.90ms
+step:9000/20000 train_loss:1.9858 train_time:467109ms step_avg:51.90ms
+step:9000/20000 val_loss:2.0765 val_bpb:1.2298 train_time:467142ms step_avg:51.90ms
+step:9200/20000 train_loss:2.0473 train_time:477468ms step_avg:51.90ms
+step:9400/20000 train_loss:2.0934 train_time:487824ms step_avg:51.90ms
+step:9600/20000 train_loss:2.1151 train_time:498188ms step_avg:51.89ms
+step:9800/20000 train_loss:2.0174 train_time:508551ms step_avg:51.89ms
+step:10000/20000 train_loss:2.0742 train_time:518903ms step_avg:51.89ms
+step:10000/20000 val_loss:2.0715 val_bpb:1.2268 train_time:518936ms step_avg:51.89ms
+step:10200/20000 train_loss:2.0357 train_time:529265ms step_avg:51.89ms
+step:10400/20000 train_loss:2.0548 train_time:539622ms step_avg:51.89ms
+step:10600/20000 train_loss:1.9345 train_time:549977ms step_avg:51.88ms
+step:10800/20000 train_loss:2.1369 train_time:560331ms step_avg:51.88ms
+step:11000/20000 train_loss:2.0578 train_time:570691ms step_avg:51.88ms
+step:11000/20000 val_loss:2.0447 val_bpb:1.2110 train_time:570724ms step_avg:51.88ms
+step:11200/20000 train_loss:2.0111 train_time:581136ms step_avg:51.89ms
+step:11400/20000 train_loss:1.9882 train_time:591500ms step_avg:51.89ms
+step:11564/20000 val_loss:2.0269 val_bpb:1.2005 train_time:600038ms step_avg:51.89ms
+stopping_early: wallclock_cap train_time:600038ms step:11564/20000
+peak memory allocated: 10247 MiB reserved: 10488 MiB
+Serialized model: 67224983 bytes
+Code size: 47716 bytes
+Total submission size: 67272699 bytes
+Serialized model int8+zlib: 15819554 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x)
+Total submission size int8+zlib: 15867270 bytes
+final_int8_zlib_roundtrip val_loss:2.0359 val_bpb:1.2058 eval_time:1639ms
+final_int8_zlib_roundtrip_exact val_loss:2.03588345 val_bpb:1.20576485
diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py
new file mode 100644
index 0000000..a9ceb04
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_gpt.py
@@ -0,0 +1,1127 @@
+"""
+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: `train_gpt.py` and `train_gpt_mlx.py` must never be longer than 1500 lines.
+"""
+
+from __future__ import annotations
+
+import copy
+import glob
+import io
+import math
+import os
+import random
+import subprocess
+import sys
+import time
+import uuid
+import zlib
+from pathlib import Path
+
+import numpy as np
+import sentencepiece as spm
+import torch
+import torch.distributed as dist
+import torch.nn.functional as F
+from torch import Tensor, nn
+from torch.nn.parallel import DistributedDataParallel as DDP
+
+# -----------------------------
+# HYPERPARAMETERS
+# -----------------------------
+# Default Long Context Seq2048 v2 run:
+# - 9 transformer blocks at width 512
+# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion
+# - vocab size 1024, sequence length 2048, tied embeddings
+# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap
+# - tuned seq2048 learning rates: 0.040 / 0.032 / 0.032
+
+class Hyperparameters:
+    # Data paths are shard globs produced by the existing preprocessing pipeline.
+    data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024")
+    train_files = os.path.join(data_path, "fineweb_train_*.bin")
+    val_files = os.path.join(data_path, "fineweb_val_*.bin")
+    tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_1024_bpe.model")
+    run_id = os.environ.get("RUN_ID", "long_context_seq2048_v2")
+    seed = int(os.environ.get("SEED", 1337))
+
+    # Validation cadence and batch size. Validation always uses the full fineweb_val split.
+    val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288))
+    val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000))
+    train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200))
+
+    # Training length.
+    iterations = int(os.environ.get("ITERATIONS", 20000))
+    warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200))
+    warmup_steps = int(os.environ.get("WARMUP_STEPS", 20))
+    train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288))
+    train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048))
+    max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0))
+    qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5))
+
+    # Model shape.
+    vocab_size = int(os.environ.get("VOCAB_SIZE", 1024))
+    num_layers = int(os.environ.get("NUM_LAYERS", 9))
+    num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4))
+    model_dim = int(os.environ.get("MODEL_DIM", 512))
+    num_heads = int(os.environ.get("NUM_HEADS", 8))
+    mlp_mult = int(os.environ.get("MLP_MULT", 2))
+    tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1")))
+    rope_base = float(os.environ.get("ROPE_BASE", 10000.0))
+    logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0))
+
+    # Optimizer hyperparameters.
+    embed_lr = float(os.environ.get("EMBED_LR", 0.6))
+    head_lr = float(os.environ.get("HEAD_LR", 0.008))
+    tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.04))
+    tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005))
+    matrix_lr = float(os.environ.get("MATRIX_LR", 0.032))
+    scalar_lr = float(os.environ.get("SCALAR_LR", 0.032))
+    muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95))
+    muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5))
+    muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85))
+    muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500))
+    beta1 = float(os.environ.get("BETA1", 0.9))
+    beta2 = float(os.environ.get("BETA2", 0.95))
+    adam_eps = float(os.environ.get("ADAM_EPS", 1e-8))
+    grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.0))
+
+# -----------------------------
+# MUON OPTIMIZER 
+# -----------------------------
+# 
+# As borrowed from modded-nanogpt
+# Background on Muon: https://kellerjordan.github.io/posts/muon/
+
+def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor:
+    # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration.
+    # Muon uses this to normalize matrix-shaped gradients before applying them.
+    a, b, c = (3.4445, -4.7750, 2.0315)
+    X = G.bfloat16()
+    X /= X.norm() + eps
+    transposed = G.size(0) > G.size(1)
+    if transposed:
+        X = X.T
+    for _ in range(steps):
+        A = X @ X.T
+        B = b * A + c * A @ A
+        X = a * X + B @ X
+    return X.T if transposed else X
+
+
+class Muon(torch.optim.Optimizer):
+    def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True):
+        super().__init__(
+            params,
+            dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov),
+        )
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        distributed = dist.is_available() and dist.is_initialized()
+        world_size = dist.get_world_size() if distributed else 1
+        rank = dist.get_rank() if distributed else 0
+
+        for group in self.param_groups:
+            params = group["params"]
+            if not params:
+                continue
+            lr = group["lr"]
+            momentum = group["momentum"]
+            backend_steps = group["backend_steps"]
+            nesterov = group["nesterov"]
+
+            total_params = sum(int(p.numel()) for p in params)
+            updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16)
+
+            curr = 0
+            for i, p in enumerate(params):
+                if i % world_size == rank and p.grad is not None:
+                    g = p.grad
+                    state = self.state[p]
+                    if "momentum_buffer" not in state:
+                        state["momentum_buffer"] = torch.zeros_like(g)
+                    buf = state["momentum_buffer"]
+                    buf.mul_(momentum).add_(g)
+                    if nesterov:
+                        g = g.add(buf, alpha=momentum)
+                    g = zeropower_via_newtonschulz5(g, steps=backend_steps)
+                    # Scale correction from Muon reference implementations.
+                    g *= max(1, g.size(0) / g.size(1)) ** 0.5
+                    updates_flat[curr : curr + p.numel()] = g.reshape(-1)
+                curr += p.numel()
+
+            if distributed:
+                dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM)
+
+            curr = 0
+            for p in params:
+                g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype)
+                p.add_(g, alpha=-lr)
+                curr += p.numel()
+
+        return loss
+
+
+# -----------------------------
+# TOKENIZER-AGNOSTIC EVALUATION SETUP 
+# -----------------------------
+#
+# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic.
+# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set.
+# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer.
+# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score.
+
+def build_sentencepiece_luts(
+    sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device
+) -> tuple[Tensor, Tensor, Tensor]:
+    sp_vocab_size = int(sp.vocab_size())
+    table_size = max(sp_vocab_size, vocab_size)
+    base_bytes_np = np.zeros((table_size,), dtype=np.int16)
+    has_leading_space_np = np.zeros((table_size,), dtype=np.bool_)
+    is_boundary_token_np = np.ones((table_size,), dtype=np.bool_)
+    for token_id in range(sp_vocab_size):
+        if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id):
+            continue
+        is_boundary_token_np[token_id] = False
+        if sp.is_byte(token_id):
+            base_bytes_np[token_id] = 1
+            continue
+        piece = sp.id_to_piece(token_id)
+        if piece.startswith("▁"):
+            has_leading_space_np[token_id] = True
+            piece = piece[1:]
+        base_bytes_np[token_id] = len(piece.encode("utf-8"))
+    return (
+        torch.tensor(base_bytes_np, dtype=torch.int16, device=device),
+        torch.tensor(has_leading_space_np, dtype=torch.bool, device=device),
+        torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device),
+    )
+
+
+def load_validation_tokens(pattern: str, seq_len: int) -> Tensor:
+    files = [Path(p) for p in sorted(glob.glob(pattern))]
+    if not files:
+        raise FileNotFoundError(f"No files found for pattern: {pattern}")
+    # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*.
+    tokens = torch.cat([load_data_shard(file) for file in files]).contiguous()
+    usable = ((tokens.numel() - 1) // seq_len) * seq_len
+    if usable <= 0:
+        raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}")
+    return tokens[: usable + 1]
+
+
+def eval_val(
+    args: Hyperparameters,
+    model: nn.Module,
+    rank: int,
+    world_size: int,
+    device: torch.device,
+    grad_accum_steps: int,
+    val_tokens: Tensor,
+    base_bytes_lut: Tensor,
+    has_leading_space_lut: Tensor,
+    is_boundary_token_lut: Tensor,
+) -> tuple[float, float]:
+    # Validation computes two metrics:
+    # - val_loss: token cross-entropy (natural log)
+    # - val_bpb: tokenizer-agnostic compression metric used by the challenge
+    local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps)
+    if local_batch_tokens < args.train_seq_len:
+        raise ValueError(
+            "VAL_BATCH_SIZE must provide at least one sequence per rank; "
+            f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, "
+            f"GRAD_ACCUM_STEPS={grad_accum_steps}, TRAIN_SEQ_LEN={args.train_seq_len}"
+        )
+    local_batch_seqs = local_batch_tokens // args.train_seq_len
+    total_seqs = (val_tokens.numel() - 1) // args.train_seq_len
+    seq_start = (total_seqs * rank) // world_size
+    seq_end = (total_seqs * (rank + 1)) // world_size
+    val_loss_sum = torch.zeros((), device=device, dtype=torch.float64)
+    val_token_count = torch.zeros((), device=device, dtype=torch.float64)
+    val_byte_count = torch.zeros((), device=device, dtype=torch.float64)
+
+    model.eval()
+    with torch.inference_mode():
+        for batch_seq_start in range(seq_start, seq_end, local_batch_seqs):
+            batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end)
+            raw_start = batch_seq_start * args.train_seq_len
+            raw_end = batch_seq_end * args.train_seq_len + 1
+            local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True)
+            x = local[:-1].reshape(-1, args.train_seq_len)
+            y = local[1:].reshape(-1, args.train_seq_len)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                batch_loss = model(x, y).detach()
+            batch_token_count = float(y.numel())
+            val_loss_sum += batch_loss.to(torch.float64) * batch_token_count
+            val_token_count += batch_token_count
+            prev_ids = x.reshape(-1)
+            tgt_ids = y.reshape(-1)
+            token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16)
+            token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16)
+            val_byte_count += token_bytes.to(torch.float64).sum()
+
+    if dist.is_available() and dist.is_initialized():
+        dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM)
+        dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM)
+
+    val_loss = val_loss_sum / val_token_count
+    bits_per_token = val_loss.item() / math.log(2.0)
+    tokens_per_byte = val_token_count.item() / val_byte_count.item()
+    model.train()
+    return float(val_loss.item()), float(bits_per_token * tokens_per_byte)
+
+# -----------------------------
+# POST-TRAINING QUANTIZATION
+# -----------------------------
+#
+# It's silly to export our model, which is trained in bf16 and fp32, at that same precision.
+# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing.
+# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit.
+
+CONTROL_TENSOR_NAME_PATTERNS = tuple(
+    pattern
+    for pattern in os.environ.get(
+        "CONTROL_TENSOR_NAME_PATTERNS",
+        "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights",
+    ).split(",")
+    if pattern
+)
+INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple(
+    pattern
+    for pattern in os.environ.get(
+        "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS",
+        ",".join(CONTROL_TENSOR_NAME_PATTERNS),
+    ).split(",")
+    if pattern
+)
+INT8_KEEP_FLOAT_MAX_NUMEL = 65_536
+INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16
+INT8_PER_ROW_SCALE_DTYPE = torch.float16
+INT8_CLIP_PERCENTILE = 99.99984
+INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0
+
+def tensor_nbytes(t: Tensor) -> int:
+    return int(t.numel()) * int(t.element_size())
+
+def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor:
+    if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS):
+        return t.float().contiguous()
+    if t.dtype in {torch.float32, torch.bfloat16}:
+        passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.")
+        return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous()
+    return t
+
+def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]:
+    t32 = t.float()
+    if t32.ndim == 2:
+        # Matrices get one scale per row, which usually tracks output-channel
+        # ranges much better than a single tensor-wide scale.
+        clip_abs = (
+            torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1)
+            if t32.numel()
+            else torch.empty((t32.shape[0],), dtype=torch.float32)
+        )
+        clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None])
+        scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0)
+        q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous()
+        return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous()
+
+    # Vectors / scalars use a simpler per-tensor scale.
+    clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0
+    scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32)
+    q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous()
+    return q, scale
+
+def quantize_state_dict_int8(state_dict: dict[str, Tensor]):
+    # Single supported clean-script export format:
+    # - per-row int8 for 2D float tensors
+    # - per-tensor int8 for other float tensors
+    # - exact passthrough for non-floats
+    # - passthrough for small float tensors, stored as fp16 to save bytes
+    quantized: dict[str, Tensor] = {}
+    scales: dict[str, Tensor] = {}
+    dtypes: dict[str, str] = {}
+    passthrough: dict[str, Tensor] = {}
+    passthrough_orig_dtypes: dict[str, str] = {}
+    qmeta: dict[str, dict[str, object]] = {}
+    stats = dict.fromkeys(
+        ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"),
+        0,
+    )
+
+    for name, tensor in state_dict.items():
+        t = tensor.detach().to("cpu").contiguous()
+        stats["param_count"] += int(t.numel())
+        stats["num_tensors"] += 1
+        stats["baseline_tensor_bytes"] += tensor_nbytes(t)
+
+        if not t.is_floating_point():
+            stats["num_nonfloat_tensors"] += 1
+            passthrough[name] = t
+            stats["int8_payload_bytes"] += tensor_nbytes(t)
+            continue
+
+        # Small float tensors are cheap enough to keep directly. We still downcast
+        # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size.
+        if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL:
+            kept = keep_float_tensor(name, t, passthrough_orig_dtypes)
+            passthrough[name] = kept
+            stats["int8_payload_bytes"] += tensor_nbytes(kept)
+            continue
+
+        stats["num_float_tensors"] += 1
+        q, s = quantize_float_tensor(t)
+        if s.ndim > 0:
+            qmeta[name] = {"scheme": "per_row", "axis": 0}
+        quantized[name] = q
+        scales[name] = s
+        dtypes[name] = str(t.dtype).removeprefix("torch.")
+        stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s)
+
+    obj: dict[str, object] = {
+        "__quant_format__": "int8_clean_per_row_v1",
+        "quantized": quantized,
+        "scales": scales,
+        "dtypes": dtypes,
+        "passthrough": passthrough,
+    }
+    if qmeta:
+        obj["qmeta"] = qmeta
+    if passthrough_orig_dtypes:
+        obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes
+    return obj, stats
+
+def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]:
+    out: dict[str, Tensor] = {}
+    qmeta = obj.get("qmeta", {})
+    passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {})
+    for name, q in obj["quantized"].items():
+        dtype = getattr(torch, obj["dtypes"][name])
+        s = obj["scales"][name]
+        if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0:
+            s = s.to(dtype=torch.float32)
+            # Broadcast the saved row scale back across trailing dimensions.
+            out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous()
+        else:
+            scale = float(s.item())
+            out[name] = (q.float() * scale).to(dtype=dtype).contiguous()
+    for name, t in obj["passthrough"].items():
+        # Restore small tensors, undoing the temporary fp16 storage cast if needed.
+        out_t = t.detach().to("cpu").contiguous()
+        orig_dtype = passthrough_orig_dtypes.get(name)
+        if isinstance(orig_dtype, str):
+            out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous()
+        out[name] = out_t
+    return out
+
+
+# -----------------------------
+# DATA LOADING 
+# -----------------------------
+
+def load_data_shard(file: Path) -> Tensor:
+    header_bytes = 256 * np.dtype("<i4").itemsize
+    token_bytes = np.dtype("<u2").itemsize
+    header = np.fromfile(file, dtype="<i4", count=256)
+    # SHARD HEADER INTS & SHARD_MAGIC
+    if header.size != 256 or int(header[0]) != 20240520 or int(header[1]) != 1:
+        raise ValueError(f"Unexpected shard header for {file}")
+    num_tokens = int(header[2])
+    expected_size = header_bytes + num_tokens * token_bytes
+    if file.stat().st_size != expected_size:
+        raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes")
+    tokens_np = np.fromfile(file, dtype="<u2", count=num_tokens, offset=header_bytes)
+    if tokens_np.size != num_tokens:
+        raise ValueError(f"Short read for {file}")
+    return torch.from_numpy(tokens_np.astype(np.uint16, copy=False))
+
+
+class TokenStream:
+    # Reads shards sequentially and wraps around forever. The training loop therefore
+    # has deterministic, simple streaming behavior with no sampling or workers.
+    def __init__(self, pattern: str):
+        self.files = [Path(p) for p in sorted(glob.glob(pattern))]
+        if not self.files:
+            raise FileNotFoundError(f"No files found for pattern: {pattern}")
+        self.file_idx = 0
+        self.tokens = load_data_shard(self.files[0])
+        self.pos = 0
+
+    def _advance_file(self) -> None:
+        self.file_idx = (self.file_idx + 1) % len(self.files)
+        self.tokens = load_data_shard(self.files[self.file_idx])
+        self.pos = 0
+
+    def take(self, n: int) -> Tensor:
+        chunks: list[Tensor] = []
+        remaining = n
+        while remaining > 0:
+            avail = self.tokens.numel() - self.pos
+            if avail <= 0:
+                self._advance_file()
+                continue
+            k = min(remaining, avail)
+            chunks.append(self.tokens[self.pos : self.pos + k])
+            self.pos += k
+            remaining -= k
+        return chunks[0] if len(chunks) == 1 else torch.cat(chunks)
+
+
+class DistributedTokenLoader:
+    # Each call consumes a contiguous chunk from the shared token stream, then slices out
+    # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting.
+    def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device):
+        self.rank = rank
+        self.world_size = world_size
+        self.device = device
+        self.stream = TokenStream(pattern)
+
+    def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]:
+        local_tokens = global_tokens // (self.world_size * grad_accum_steps)
+        per_rank_span = local_tokens + 1
+        chunk = self.stream.take(per_rank_span * self.world_size)
+        start = self.rank * per_rank_span
+        local = chunk[start : start + per_rank_span].to(dtype=torch.int64)
+        x = local[:-1].reshape(-1, seq_len)
+        y = local[1:].reshape(-1, seq_len)
+        return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True)
+
+# -----------------------------
+# TRANSFORMER MODULES
+# -----------------------------
+
+class RMSNorm(nn.Module):
+    def __init__(self, eps: float | None = None):
+        super().__init__()
+        self.eps = eps
+
+    def forward(self, x: Tensor) -> Tensor:
+        return F.rms_norm(x, (x.size(-1),), eps=self.eps)
+
+
+class CastedLinear(nn.Linear):
+    # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute.
+    def forward(self, x: Tensor) -> Tensor:
+        bias = self.bias.to(x.dtype) if self.bias is not None else None
+        return F.linear(x, self.weight.to(x.dtype), bias)
+
+
+def restore_low_dim_params_to_fp32(module: nn.Module) -> None:
+    # Keep small/control parameters in fp32 even when the model body runs in bf16.
+    with torch.no_grad():
+        for name, param in module.named_parameters():
+            if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32:
+                param.data = param.data.float()
+
+
+class Rotary(nn.Module):
+    # Caches cos/sin tables per sequence length on the current device.
+    def __init__(self, dim: int, base: float = 10000.0):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached: Tensor | None = None
+        self._sin_cached: Tensor | None = None
+
+    def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]:
+        if (
+            self._cos_cached is None
+            or self._sin_cached is None
+            or self._seq_len_cached != seq_len
+            or self._cos_cached.device != device
+        ):
+            t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype)
+            freqs = torch.outer(t, self.inv_freq.to(device))
+            self._cos_cached = freqs.cos()[None, None, :, :]
+            self._sin_cached = freqs.sin()[None, None, :, :]
+            self._seq_len_cached = seq_len
+        return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype)
+
+
+def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor:
+    half = x.size(-1) // 2
+    x1, x2 = x[..., :half], x[..., half:]
+    return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1)
+
+
+class CausalSelfAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_heads: int,
+        num_kv_heads: int,
+        rope_base: float,
+        qk_gain_init: float,
+    ):
+        super().__init__()
+        if dim % num_heads != 0:
+            raise ValueError("model_dim must be divisible by num_heads")
+        if num_heads % num_kv_heads != 0:
+            raise ValueError("num_heads must be divisible by num_kv_heads")
+        self.num_heads = num_heads
+        self.num_kv_heads = num_kv_heads
+        self.head_dim = dim // num_heads
+        if self.head_dim % 2 != 0:
+            raise ValueError("head_dim must be even for RoPE")
+        kv_dim = self.num_kv_heads * self.head_dim
+        self.c_q = CastedLinear(dim, dim, bias=False)
+        self.c_k = CastedLinear(dim, kv_dim, bias=False)
+        self.c_v = CastedLinear(dim, kv_dim, bias=False)
+        self.proj = CastedLinear(dim, dim, bias=False)
+        self.proj._zero_init = True
+        self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32))
+        self.rotary = Rotary(self.head_dim, base=rope_base)
+
+    def forward(self, x: Tensor) -> 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),
+        )
+        y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim)
+        return self.proj(y)
+
+
+class MLP(nn.Module):
+    # relu^2 MLP from the original modded-nanogpt setup
+    def __init__(self, dim: int, mlp_mult: int):
+        super().__init__()
+        hidden = mlp_mult * dim
+        self.fc = CastedLinear(dim, hidden, bias=False)
+        self.proj = CastedLinear(hidden, dim, bias=False)
+        self.proj._zero_init = True
+
+    def forward(self, x: Tensor) -> Tensor:
+        x = torch.relu(self.fc(x))
+        return self.proj(x.square())
+
+
+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,
+    ):
+        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())
+
+    def forward(self, x: Tensor, x0: Tensor) -> Tensor:
+        mix = self.resid_mix.to(dtype=x.dtype)
+        x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0
+        attn_out = self.attn(self.attn_norm(x))
+        x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out
+        x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x))
+        return x
+
+
+class GPT(nn.Module):
+    def __init__(
+        self,
+        vocab_size: int,
+        num_layers: int,
+        model_dim: int,
+        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,
+    ):
+        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.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,
+                )
+                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)
+        for module in self.modules():
+            if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False):
+                nn.init.zeros_(module.weight)
+
+    def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor:
+        x = self.tok_emb(input_ids)
+        x = F.rms_norm(x, (x.size(-1),))
+        x0 = x
+        skips: list[Tensor] = []
+
+        # First half stores skips; second half reuses them in reverse order.
+        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")
+
+
+# -----------------------------
+# TRAINING
+# -----------------------------
+
+def main() -> None:
+    global zeropower_via_newtonschulz5
+
+    code = Path(__file__).read_text(encoding="utf-8")
+    args = Hyperparameters()
+    zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5)
+
+    # -----------------------------
+    # DISTRIBUTED + CUDA SETUP
+    # -----------------------------
+
+    distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ
+    rank = int(os.environ.get("RANK", "0"))
+    world_size = int(os.environ.get("WORLD_SIZE", "1"))
+    local_rank = int(os.environ.get("LOCAL_RANK", "0"))
+    if world_size <= 0:
+        raise ValueError(f"WORLD_SIZE must be positive, got {world_size}")
+    if 8 % world_size != 0:
+        raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral")
+    grad_accum_steps = 8 // world_size
+    grad_scale = 1.0 / grad_accum_steps
+    if not torch.cuda.is_available():
+        raise RuntimeError("CUDA is required")
+    device = torch.device("cuda", local_rank)
+    torch.cuda.set_device(device)
+    if distributed:
+        dist.init_process_group(backend="nccl", device_id=device)
+        dist.barrier()
+    master_process = rank == 0
+
+    # Fast math knobs
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+    from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp
+
+    enable_cudnn_sdp(False)
+    enable_flash_sdp(True)
+    enable_mem_efficient_sdp(False)
+    enable_math_sdp(False)
+
+    logfile = None
+    if master_process:
+        os.makedirs("logs", exist_ok=True)
+        logfile = f"logs/{args.run_id}.txt"
+        print(logfile)
+
+    def log0(msg: str, console: bool = True) -> None:
+        if not master_process:
+            return
+        if console:
+            print(msg)
+        if logfile is not None:
+            with open(logfile, "a", encoding="utf-8") as f:
+                print(msg, file=f)
+
+    log0(code, console=False)
+    log0("=" * 100, console=False)
+    log0(f"Running Python {sys.version}", console=False)
+    log0(f"Running PyTorch {torch.__version__}", console=False)
+    log0(
+        subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout,
+        console=False,
+    )
+    log0("=" * 100, console=False)
+
+    # -----------------------------
+    # TOKENIZER + VALIDATION METRIC SETUP
+    # -----------------------------
+
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    torch.cuda.manual_seed_all(args.seed)
+
+    if not args.tokenizer_path.endswith(".model"):
+        raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}")
+    sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path)
+    if int(sp.vocab_size()) != args.vocab_size:
+        raise ValueError(
+            f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}"
+        )
+    dataset_dir = Path(args.data_path).resolve()
+    actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin")))
+    val_tokens = load_validation_tokens(args.val_files, args.train_seq_len)
+    base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts(
+        sp, args.vocab_size, device
+    )
+    log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}")
+    log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}")
+    log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}")
+
+    # -----------------------------
+    # MODEL + OPTIMIZER SETUP
+    # -----------------------------
+
+    base_model = GPT(
+        vocab_size=args.vocab_size,
+        num_layers=args.num_layers,
+        model_dim=args.model_dim,
+        num_heads=args.num_heads,
+        num_kv_heads=args.num_kv_heads,
+        mlp_mult=args.mlp_mult,
+        tie_embeddings=args.tie_embeddings,
+        tied_embed_init_std=args.tied_embed_init_std,
+        logit_softcap=args.logit_softcap,
+        rope_base=args.rope_base,
+        qk_gain_init=args.qk_gain_init,
+    ).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
+
+    # Optimizer split:
+    # - token embedding (Adam) uses EMBED_LR
+    # - untied lm_head (Adam) uses HEAD_LR
+    # - matrix params in transformer blocks use MATRIX_LR via Muon
+    # - vectors/scalars use SCALAR_LR via Adam
+    block_named_params = list(base_model.blocks.named_parameters())
+    matrix_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    scalar_params = [
+        p
+        for name, p in block_named_params
+        if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)
+    ]
+    if base_model.skip_weights.numel() > 0:
+        scalar_params.append(base_model.skip_weights)
+    token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr
+    optimizer_tok = torch.optim.Adam(
+        [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizer_muon = Muon(
+        matrix_params,
+        lr=args.matrix_lr,
+        momentum=args.muon_momentum,
+        backend_steps=args.muon_backend_steps,
+    )
+    for group in optimizer_muon.param_groups:
+        group["base_lr"] = args.matrix_lr
+    optimizer_scalar = torch.optim.Adam(
+        [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}],
+        betas=(args.beta1, args.beta2),
+        eps=args.adam_eps,
+        fused=True,
+    )
+    optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar]
+    if base_model.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("sdp_backends:cudnn=False flash=True mem_efficient=False math=False")
+    log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}")
+    log0(
+        f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} "
+        f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} "
+        f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}"
+    )
+    log0(
+        f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} "
+        f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} "
+        f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}"
+    )
+    log0(f"seed:{args.seed}")
+
+    # -----------------------------
+    # DATA LOADER & MODEL WARMUP
+    # -----------------------------
+
+    train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    def zero_grad_all() -> None:
+        for opt in optimizers:
+            opt.zero_grad(set_to_none=True)
+
+    max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None
+
+    def lr_mul(step: int, elapsed_ms: float) -> float:
+        if args.warmdown_iters <= 0:
+            return 1.0
+        if max_wallclock_ms is None:
+            warmdown_start = max(args.iterations - args.warmdown_iters, 0)
+            return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0
+        step_ms = elapsed_ms / max(step, 1)
+        warmdown_ms = args.warmdown_iters * step_ms
+        remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0)
+        return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0
+
+    # Warmup primes the compiled forward/backward/optimizer paths, then we restore the
+    # initial weights/optimizer state so measured training starts from the true init.
+    if args.warmup_steps > 0:
+        initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()}
+        initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers]
+        model.train()
+        for warmup_step in range(args.warmup_steps):
+            zero_grad_all()
+            for micro_step in range(grad_accum_steps):
+                if distributed:
+                    model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+                x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+                with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                    warmup_loss = model(x, y)
+                (warmup_loss * grad_scale).backward()
+            for opt in optimizers:
+                opt.step()
+            zero_grad_all()
+            if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps:
+                log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}")
+        base_model.load_state_dict(initial_model_state, strict=True)
+        for opt, state in zip(optimizers, initial_optimizer_states, strict=True):
+            opt.load_state_dict(state)
+        zero_grad_all()
+        if distributed:
+            model.require_backward_grad_sync = True
+        train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device)
+
+    # -----------------------------
+    # MAIN TRAINING LOOP
+    # -----------------------------
+
+    training_time_ms = 0.0
+    stop_after_step: int | None = None
+    torch.cuda.synchronize()
+    t0 = time.perf_counter()
+
+    step = 0
+    while True:
+        last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step)
+
+        should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0)
+        if should_validate:
+            torch.cuda.synchronize()
+            training_time_ms += 1000.0 * (time.perf_counter() - t0)
+            val_loss, val_bpb = eval_val(
+                args,
+                model,
+                rank,
+                world_size,
+                device,
+                grad_accum_steps,
+                val_tokens,
+                base_bytes_lut,
+                has_leading_space_lut,
+                is_boundary_token_lut,
+            )
+            log0(
+                f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} "
+                f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms"
+            )
+            torch.cuda.synchronize()
+            t0 = time.perf_counter()
+
+        if last_step:
+            if stop_after_step is not None and step < args.iterations:
+                log0(
+                    f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms "
+                    f"step:{step}/{args.iterations}"
+                )
+            break
+
+        elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        scale = lr_mul(step, elapsed_ms)
+        zero_grad_all()
+        train_loss = torch.zeros((), device=device)
+        for micro_step in range(grad_accum_steps):
+            if distributed:
+                model.require_backward_grad_sync = micro_step == grad_accum_steps - 1
+            x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps)
+            with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True):
+                loss = model(x, y)
+            train_loss += loss.detach()
+            (loss * grad_scale).backward()
+        train_loss /= grad_accum_steps
+
+        frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0
+        muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum
+        for group in optimizer_muon.param_groups:
+            group["momentum"] = muon_momentum
+
+        for opt in optimizers:
+            for group in opt.param_groups:
+                group["lr"] = group["base_lr"] * scale
+
+        if args.grad_clip_norm > 0:
+            torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm)
+        for opt in optimizers:
+            opt.step()
+        zero_grad_all()
+
+        step += 1
+        approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0)
+        should_log_train = (
+            args.train_log_every > 0
+            and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None)
+        )
+        if should_log_train:
+            log0(
+                f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} "
+                f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms"
+            )
+
+        # Needed to sync whether we've reached the wallclock cap.
+        reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms
+        if distributed and max_wallclock_ms is not None:
+            reached_cap_tensor = torch.tensor(int(reached_cap), device=device)
+            dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX)
+            reached_cap = bool(reached_cap_tensor.item())
+        if stop_after_step is None and reached_cap:
+            stop_after_step = step
+
+    log0(
+        f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB "
+        f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB"
+    )
+
+    # -----------------------------
+    # SERIALIZATION + ROUNDTRIP VALIDATION
+    # -----------------------------
+    # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce
+    # the compressed int8+zlib artifact and validate the round-tripped weights.
+
+    if master_process:
+        torch.save(base_model.state_dict(), "final_model.pt")
+        model_bytes = os.path.getsize("final_model.pt")
+        code_bytes = len(code.encode("utf-8"))
+        log0(f"Serialized model: {model_bytes} bytes")
+        log0(f"Code size: {code_bytes} bytes")
+        log0(f"Total submission size: {model_bytes + code_bytes} bytes")
+
+    quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict())
+    quant_buf = io.BytesIO()
+    torch.save(quant_obj, quant_buf)
+    quant_raw = quant_buf.getvalue()
+    quant_blob = zlib.compress(quant_raw, level=9)
+    quant_raw_bytes = len(quant_raw)
+    if master_process:
+        with open("final_model.int8.ptz", "wb") as f:
+            f.write(quant_blob)
+        quant_file_bytes = os.path.getsize("final_model.int8.ptz")
+        code_bytes = len(code.encode("utf-8"))
+        ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1)
+        log0(
+            f"Serialized model int8+zlib: {quant_file_bytes} bytes "
+            f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)"
+        )
+        log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes")
+
+    if distributed:
+        dist.barrier()
+    with open("final_model.int8.ptz", "rb") as f:
+        quant_blob_disk = f.read()
+    quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu")
+    base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True)
+    torch.cuda.synchronize()
+    t_qeval = time.perf_counter()
+    q_val_loss, q_val_bpb = eval_val(
+        args,
+        model,
+        rank,
+        world_size,
+        device,
+        grad_accum_steps,
+        val_tokens,
+        base_bytes_lut,
+        has_leading_space_lut,
+        is_boundary_token_lut,
+    )
+    torch.cuda.synchronize()
+    log0(
+        f"final_int8_zlib_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} "
+        f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms"
+    )
+    log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}")
+
+    if distributed:
+        dist.destroy_process_group()
+
+
+if __name__ == "__main__":
+    main()
diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1338.log b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1338.log
new file mode 100644
index 0000000..5851fe0
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1338.log
@@ -0,0 +1,124 @@
+
+*****************************************
+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. 
+*****************************************
+logs/seq2048_sxm28_seed1338_20260319a.txt
+val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/root/parameter-golf-sxm28/data/tokenizers/fineweb_1024_bpe.model
+train_loader:dataset:fineweb10B_sp1024 train_shards:80
+val_loader:shards pattern=/root/parameter-golf-sxm28/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632
+model_params:17059912
+world_size:8 grad_accum_steps:1
+sdp_backends:cudnn=False flash=True mem_efficient=False math=False
+attention_mode:gqa num_heads:8 num_kv_heads:4
+tie_embeddings:True embed_lr:0.04 head_lr:0.0 matrix_lr:0.032 scalar_lr:0.032
+train_batch_tokens:524288 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000
+seed:1338
+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:6.9373 val_bpb:4.1086 train_time:0ms step_avg:0.02ms
+step:1/20000 train_loss:6.9363 train_time:27ms step_avg:27.47ms
+step:2/20000 train_loss:14.7813 train_time:73ms step_avg:36.26ms
+step:3/20000 train_loss:8.1303 train_time:124ms step_avg:41.23ms
+step:4/20000 train_loss:6.6459 train_time:175ms step_avg:43.65ms
+step:5/20000 train_loss:6.9413 train_time:227ms step_avg:45.34ms
+step:6/20000 train_loss:7.8068 train_time:278ms step_avg:46.34ms
+step:7/20000 train_loss:6.8490 train_time:330ms step_avg:47.09ms
+step:8/20000 train_loss:6.4809 train_time:381ms step_avg:47.61ms
+step:9/20000 train_loss:6.2006 train_time:434ms step_avg:48.24ms
+step:10/20000 train_loss:6.0502 train_time:483ms step_avg:48.35ms
+step:200/20000 train_loss:2.7815 train_time:10297ms step_avg:51.48ms
+step:400/20000 train_loss:2.3049 train_time:20646ms step_avg:51.62ms
+step:600/20000 train_loss:2.4980 train_time:30990ms step_avg:51.65ms
+step:800/20000 train_loss:2.2491 train_time:41345ms step_avg:51.68ms
+step:1000/20000 train_loss:2.3361 train_time:51721ms step_avg:51.72ms
+step:1000/20000 val_loss:2.2898 val_bpb:1.3562 train_time:51754ms step_avg:51.75ms
+step:1200/20000 train_loss:2.3575 train_time:62108ms step_avg:51.76ms
+step:1400/20000 train_loss:2.3813 train_time:72504ms step_avg:51.79ms
+step:1600/20000 train_loss:2.0496 train_time:82894ms step_avg:51.81ms
+step:1800/20000 train_loss:2.1676 train_time:93289ms step_avg:51.83ms
+step:2000/20000 train_loss:2.2119 train_time:103684ms step_avg:51.84ms
+step:2000/20000 val_loss:2.1937 val_bpb:1.2993 train_time:103717ms step_avg:51.86ms
+step:2200/20000 train_loss:2.0320 train_time:114081ms step_avg:51.86ms
+step:2400/20000 train_loss:2.1652 train_time:124471ms step_avg:51.86ms
+step:2600/20000 train_loss:2.3821 train_time:134872ms step_avg:51.87ms
+step:2800/20000 train_loss:2.1987 train_time:145264ms step_avg:51.88ms
+step:3000/20000 train_loss:2.1912 train_time:155647ms step_avg:51.88ms
+step:3000/20000 val_loss:2.1534 val_bpb:1.2754 train_time:155680ms step_avg:51.89ms
+step:3200/20000 train_loss:2.1516 train_time:166037ms step_avg:51.89ms
+step:3400/20000 train_loss:2.1199 train_time:176417ms step_avg:51.89ms
+step:3600/20000 train_loss:2.0651 train_time:186795ms step_avg:51.89ms
+step:3800/20000 train_loss:2.1694 train_time:197173ms step_avg:51.89ms
+step:4000/20000 train_loss:2.1330 train_time:207554ms step_avg:51.89ms
+step:4000/20000 val_loss:2.1292 val_bpb:1.2610 train_time:207587ms step_avg:51.90ms
+step:4200/20000 train_loss:2.1284 train_time:217976ms step_avg:51.90ms
+step:4400/20000 train_loss:2.0686 train_time:228351ms step_avg:51.90ms
+step:4600/20000 train_loss:1.9371 train_time:238738ms step_avg:51.90ms
+step:4800/20000 train_loss:2.2171 train_time:249109ms step_avg:51.90ms
+step:5000/20000 train_loss:1.9744 train_time:259476ms step_avg:51.90ms
+step:5000/20000 val_loss:2.1127 val_bpb:1.2512 train_time:259509ms step_avg:51.90ms
+step:5200/20000 train_loss:2.1356 train_time:269848ms step_avg:51.89ms
+step:5400/20000 train_loss:2.1527 train_time:280217ms step_avg:51.89ms
+step:5600/20000 train_loss:2.1390 train_time:290578ms step_avg:51.89ms
+step:5800/20000 train_loss:2.0944 train_time:300933ms step_avg:51.89ms
+step:6000/20000 train_loss:2.1752 train_time:311294ms step_avg:51.88ms
+step:6000/20000 val_loss:2.1026 val_bpb:1.2453 train_time:311327ms step_avg:51.89ms
+step:6200/20000 train_loss:2.0458 train_time:321653ms step_avg:51.88ms
+step:6400/20000 train_loss:2.1240 train_time:332019ms step_avg:51.88ms
+step:6600/20000 train_loss:2.0830 train_time:342381ms step_avg:51.88ms
+step:6800/20000 train_loss:2.1434 train_time:352738ms step_avg:51.87ms
+step:7000/20000 train_loss:2.1907 train_time:363096ms step_avg:51.87ms
+step:7000/20000 val_loss:2.0916 val_bpb:1.2388 train_time:363129ms step_avg:51.88ms
+step:7200/20000 train_loss:2.1672 train_time:373450ms step_avg:51.87ms
+step:7400/20000 train_loss:2.0842 train_time:383806ms step_avg:51.87ms
+step:7600/20000 train_loss:1.9615 train_time:394163ms step_avg:51.86ms
+step:7800/20000 train_loss:2.1113 train_time:404518ms step_avg:51.86ms
+step:8000/20000 train_loss:2.0788 train_time:414870ms step_avg:51.86ms
+step:8000/20000 val_loss:2.0826 val_bpb:1.2334 train_time:414903ms step_avg:51.86ms
+step:8200/20000 train_loss:2.1505 train_time:425230ms step_avg:51.86ms
+step:8400/20000 train_loss:2.0933 train_time:435626ms step_avg:51.86ms
+step:8600/20000 train_loss:2.1070 train_time:445977ms step_avg:51.86ms
+step:8800/20000 train_loss:2.0708 train_time:456329ms step_avg:51.86ms
+step:9000/20000 train_loss:1.9882 train_time:466685ms step_avg:51.85ms
+step:9000/20000 val_loss:2.0772 val_bpb:1.2302 train_time:466718ms step_avg:51.86ms
+step:9200/20000 train_loss:2.0470 train_time:477091ms step_avg:51.86ms
+step:9400/20000 train_loss:2.0941 train_time:487469ms step_avg:51.86ms
+step:9600/20000 train_loss:2.1116 train_time:497817ms step_avg:51.86ms
+step:9800/20000 train_loss:2.0202 train_time:508164ms step_avg:51.85ms
+step:10000/20000 train_loss:2.0783 train_time:518510ms step_avg:51.85ms
+step:10000/20000 val_loss:2.0723 val_bpb:1.2274 train_time:518543ms step_avg:51.85ms
+step:10200/20000 train_loss:2.0342 train_time:528861ms step_avg:51.85ms
+step:10400/20000 train_loss:2.0587 train_time:539208ms step_avg:51.85ms
+step:10600/20000 train_loss:1.9323 train_time:549552ms step_avg:51.84ms
+step:10800/20000 train_loss:2.1371 train_time:559906ms step_avg:51.84ms
+step:11000/20000 train_loss:2.0567 train_time:570254ms step_avg:51.84ms
+step:11000/20000 val_loss:2.0458 val_bpb:1.2116 train_time:570286ms step_avg:51.84ms
+step:11200/20000 train_loss:2.0119 train_time:580608ms step_avg:51.84ms
+step:11400/20000 train_loss:1.9920 train_time:590954ms step_avg:51.84ms
+step:11575/20000 val_loss:2.0274 val_bpb:1.2007 train_time:600043ms step_avg:51.84ms
+stopping_early: wallclock_cap train_time:600043ms step:11575/20000
+peak memory allocated: 10247 MiB reserved: 10312 MiB
+Serialized model: 67224983 bytes
+Code size: 47716 bytes
+Total submission size: 67272699 bytes
+Serialized model int8+zlib: 15813523 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x)
+Total submission size int8+zlib: 15861239 bytes
+final_int8_zlib_roundtrip val_loss:2.0366 val_bpb:1.2062 eval_time:1638ms
+final_int8_zlib_roundtrip_exact val_loss:2.03657529 val_bpb:1.20617460
diff --git a/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1339.log b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1339.log
new file mode 100644
index 0000000..527d535
--- /dev/null
+++ b/records/track_10min_16mb/2026-03-18_LongContextSeq2048/train_seed1339.log
@@ -0,0 +1,124 @@
+
+*****************************************
+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. 
+*****************************************
+logs/seq2048_sxm28_seed1339_20260319a.txt
+val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=/root/parameter-golf-sxm28/data/tokenizers/fineweb_1024_bpe.model
+train_loader:dataset:fineweb10B_sp1024 train_shards:80
+val_loader:shards pattern=/root/parameter-golf-sxm28/data/datasets/fineweb10B_sp1024/fineweb_val_*.bin tokens:62021632
+model_params:17059912
+world_size:8 grad_accum_steps:1
+sdp_backends:cudnn=False flash=True mem_efficient=False math=False
+attention_mode:gqa num_heads:8 num_kv_heads:4
+tie_embeddings:True embed_lr:0.04 head_lr:0.0 matrix_lr:0.032 scalar_lr:0.032
+train_batch_tokens:524288 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000
+seed:1339
+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:6.9372 val_bpb:4.1086 train_time:0ms step_avg:0.02ms
+step:1/20000 train_loss:6.9375 train_time:27ms step_avg:27.41ms
+step:2/20000 train_loss:14.8195 train_time:75ms step_avg:37.30ms
+step:3/20000 train_loss:8.0864 train_time:126ms step_avg:41.96ms
+step:4/20000 train_loss:6.5534 train_time:177ms step_avg:44.18ms
+step:5/20000 train_loss:6.8329 train_time:228ms step_avg:45.62ms
+step:6/20000 train_loss:7.6678 train_time:279ms step_avg:46.55ms
+step:7/20000 train_loss:6.6834 train_time:330ms step_avg:47.21ms
+step:8/20000 train_loss:6.3318 train_time:382ms step_avg:47.72ms
+step:9/20000 train_loss:6.1684 train_time:433ms step_avg:48.12ms
+step:10/20000 train_loss:6.0971 train_time:484ms step_avg:48.42ms
+step:200/20000 train_loss:2.7828 train_time:10300ms step_avg:51.50ms
+step:400/20000 train_loss:2.2900 train_time:20640ms step_avg:51.60ms
+step:600/20000 train_loss:2.5016 train_time:30985ms step_avg:51.64ms
+step:800/20000 train_loss:2.2388 train_time:41342ms step_avg:51.68ms
+step:1000/20000 train_loss:2.3326 train_time:51710ms step_avg:51.71ms
+step:1000/20000 val_loss:2.2851 val_bpb:1.3534 train_time:51743ms step_avg:51.74ms
+step:1200/20000 train_loss:2.3498 train_time:62082ms step_avg:51.74ms
+step:1400/20000 train_loss:2.3834 train_time:72461ms step_avg:51.76ms
+step:1600/20000 train_loss:2.0438 train_time:82831ms step_avg:51.77ms
+step:1800/20000 train_loss:2.1610 train_time:93204ms step_avg:51.78ms
+step:2000/20000 train_loss:2.2097 train_time:103580ms step_avg:51.79ms
+step:2000/20000 val_loss:2.1921 val_bpb:1.2983 train_time:103613ms step_avg:51.81ms
+step:2200/20000 train_loss:2.0283 train_time:113950ms step_avg:51.80ms
+step:2400/20000 train_loss:2.1558 train_time:124322ms step_avg:51.80ms
+step:2600/20000 train_loss:2.3815 train_time:134697ms step_avg:51.81ms
+step:2800/20000 train_loss:2.1964 train_time:145067ms step_avg:51.81ms
+step:3000/20000 train_loss:2.1879 train_time:155433ms step_avg:51.81ms
+step:3000/20000 val_loss:2.1535 val_bpb:1.2754 train_time:155467ms step_avg:51.82ms
+step:3200/20000 train_loss:2.1499 train_time:165801ms step_avg:51.81ms
+step:3400/20000 train_loss:2.1196 train_time:176179ms step_avg:51.82ms
+step:3600/20000 train_loss:2.0663 train_time:186549ms step_avg:51.82ms
+step:3800/20000 train_loss:2.1720 train_time:196916ms step_avg:51.82ms
+step:4000/20000 train_loss:2.1340 train_time:207282ms step_avg:51.82ms
+step:4000/20000 val_loss:2.1294 val_bpb:1.2611 train_time:207315ms step_avg:51.83ms
+step:4200/20000 train_loss:2.1284 train_time:217691ms step_avg:51.83ms
+step:4400/20000 train_loss:2.0666 train_time:228047ms step_avg:51.83ms
+step:4600/20000 train_loss:1.9387 train_time:238412ms step_avg:51.83ms
+step:4800/20000 train_loss:2.2201 train_time:248762ms step_avg:51.83ms
+step:5000/20000 train_loss:1.9748 train_time:259127ms step_avg:51.83ms
+step:5000/20000 val_loss:2.1132 val_bpb:1.2516 train_time:259160ms step_avg:51.83ms
+step:5200/20000 train_loss:2.1342 train_time:269493ms step_avg:51.83ms
+step:5400/20000 train_loss:2.1527 train_time:279859ms step_avg:51.83ms
+step:5600/20000 train_loss:2.1413 train_time:290221ms step_avg:51.83ms
+step:5800/20000 train_loss:2.0992 train_time:300579ms step_avg:51.82ms
+step:6000/20000 train_loss:2.1790 train_time:310939ms step_avg:51.82ms
+step:6000/20000 val_loss:2.1039 val_bpb:1.2460 train_time:310972ms step_avg:51.83ms
+step:6200/20000 train_loss:2.0485 train_time:321291ms step_avg:51.82ms
+step:6400/20000 train_loss:2.1251 train_time:331655ms step_avg:51.82ms
+step:6600/20000 train_loss:2.0805 train_time:342011ms step_avg:51.82ms
+step:6800/20000 train_loss:2.1480 train_time:352372ms step_avg:51.82ms
+step:7000/20000 train_loss:2.1942 train_time:362734ms step_avg:51.82ms
+step:7000/20000 val_loss:2.0925 val_bpb:1.2393 train_time:362767ms step_avg:51.82ms
+step:7200/20000 train_loss:2.1633 train_time:373093ms step_avg:51.82ms
+step:7400/20000 train_loss:2.0834 train_time:383453ms step_avg:51.82ms
+step:7600/20000 train_loss:1.9632 train_time:393812ms step_avg:51.82ms
+step:7800/20000 train_loss:2.1106 train_time:404251ms step_avg:51.83ms
+step:8000/20000 train_loss:2.0791 train_time:414608ms step_avg:51.83ms
+step:8000/20000 val_loss:2.0836 val_bpb:1.2340 train_time:414641ms step_avg:51.83ms
+step:8200/20000 train_loss:2.1540 train_time:424958ms step_avg:51.82ms
+step:8400/20000 train_loss:2.0970 train_time:435353ms step_avg:51.83ms
+step:8600/20000 train_loss:2.1104 train_time:445701ms step_avg:51.83ms
+step:8800/20000 train_loss:2.0694 train_time:456052ms step_avg:51.82ms
+step:9000/20000 train_loss:1.9874 train_time:466402ms step_avg:51.82ms
+step:9000/20000 val_loss:2.0785 val_bpb:1.2310 train_time:466435ms step_avg:51.83ms
+step:9200/20000 train_loss:2.0479 train_time:476785ms step_avg:51.82ms
+step:9400/20000 train_loss:2.0944 train_time:487193ms step_avg:51.83ms
+step:9600/20000 train_loss:2.1114 train_time:497576ms step_avg:51.83ms
+step:9800/20000 train_loss:2.0204 train_time:507942ms step_avg:51.83ms
+step:10000/20000 train_loss:2.0788 train_time:518312ms step_avg:51.83ms
+step:10000/20000 val_loss:2.0729 val_bpb:1.2277 train_time:518345ms step_avg:51.83ms
+step:10200/20000 train_loss:2.0372 train_time:528681ms step_avg:51.83ms
+step:10400/20000 train_loss:2.0577 train_time:539056ms step_avg:51.83ms
+step:10600/20000 train_loss:1.9348 train_time:549410ms step_avg:51.83ms
+step:10800/20000 train_loss:2.1410 train_time:559767ms step_avg:51.83ms
+step:11000/20000 train_loss:2.0562 train_time:570117ms step_avg:51.83ms
+step:11000/20000 val_loss:2.0472 val_bpb:1.2125 train_time:570149ms step_avg:51.83ms
+step:11200/20000 train_loss:2.0166 train_time:580461ms step_avg:51.83ms
+step:11400/20000 train_loss:1.9939 train_time:590808ms step_avg:51.83ms
+step:11578/20000 val_loss:2.0286 val_bpb:1.2015 train_time:600051ms step_avg:51.83ms
+stopping_early: wallclock_cap train_time:600051ms step:11578/20000
+peak memory allocated: 10247 MiB reserved: 10312 MiB
+Serialized model: 67224983 bytes
+Code size: 47716 bytes
+Total submission size: 67272699 bytes
+Serialized model int8+zlib: 15814036 bytes (payload:17178912 raw_torch:17224025 payload_ratio:3.91x)
+Total submission size int8+zlib: 15861752 bytes
+final_int8_zlib_roundtrip val_loss:2.0382 val_bpb:1.2072 eval_time:1640ms
+final_int8_zlib_roundtrip_exact val_loss:2.03823779 val_bpb:1.20715923