MoE-200M (Qwen3) — step 160 000 / 100B-run intermediate checkpoint (5.24 B tokens)

Mid-training snapshot of the in-flight moe-200m-qwen3-100b pretrain run, captured at step 159 999 ≈ 5.24 B tokens consumed (5.2 % of the 100 B-token budget, +74 % more training than the step 70 000 / 2.4 B snapshot). Trained autonomously by the ml-intern Claude Code skill on 2× Tesla V100-SXM2 32 GB.

Source code, training log, full eval bundle: AlexWortega/moe-200m-qwen3-100b- (GitHub repo, in progress).

The model is not yet converged — the 100 B run continues. The final checkpoint will land at AlexWortega/moe-200m-qwen3-100b-* once the run completes.

Architecture

DeepSeekMoE-style MoE with a Qwen3 tokenizer:


total params	616.5 M
active params	203.6 M / token
vocab	151 936 (Qwen3)
d_model	640
n_layers	16 (layer 0 dense, layers 1–15 MoE)
attention	GQA — 10 Q heads / 2 KV heads, head_dim 64, partial RoPE (32 dims)
experts	16 routed + 1 shared, top-2 sigmoid router
d_ff	1024 (per expert)
tied embed/lm_head	yes
µP base_d	512
precision (training)	fp16 AMP, Muon + AdamW, WSD schedule

The MoE dispatch uses a token-permuted, capacity-padded grouped-bmm kernel (moe_backend="grouped") — stacked expert weights of shape [E, d_ff, d] (gate, up) and [E, d, d_ff] (down). State-dict keys are flat blocks.{i}.ffn.{gate,up,down} rather than per-expert ModuleList entries; the legacy per-expert layout is also accepted (auto-stacked) by MoEModel.load_state_dict.

The router is a sigmoid + top-k pick with aux_coef=1e-3, z_coef=1e-3, and an additive bias controller (symmetric, with starved-expert boost, clamped to ±10). See model.py for the full SigmoidRouter + _moe_dispatch_grouped implementation.

Training state @ step 160 000


tokens_seen	5 242 880 000 (≈ 5.2 % of 100 B target)
step	159 999
train lm_loss (200-step window)	≈ 3.47
eval loss (held-out)	3.458
router CV	≈ 0.64
router entropy	≈ 3.61 bits
throughput	≈ 26.5 k tokens/s on 2× V100
hardware	2× Tesla V100-SXM2 32 GB (GPUs 2, 3)

Zero-shot lm-evaluation-harness results

6 standard tasks, num_fewshot=0, batch_size=8, dtype=float16, Qwen3 tokenizer, single seed, no bootstrap. Δ is vs the step 90 000 / 3.0 B snapshot of the same run.

task	metric	random	gpt2-124M	pythia-160m	our-100M@21B	LFM2-350M	qwen2.5-0.5B	our-200M@3B	our-200M@5.2B (this)	Δ vs 3B
boolq	acc	50.0	48.7	55.2	58.1	64.2	62.5	44.1	56.9	+12.9
hellaswag	acc	25.0	28.9	28.4	31.7	38.4	40.6	28.7	29.7	+1.0
hellaswag	acc_norm	25.0	31.2	30.3	35.9	49.0	52.2	31.1	32.2	+1.1
piqa	acc_norm	50.0	62.5	61.4	64.2	69.5	69.9	57.6	57.1	−0.5
winogrande	acc	50.0	52.4	51.0	50.0	55.7	56.5	49.4	50.2	+0.8
arc_easy	acc	25.0	43.6	43.8	54.9	69.4	64.5	44.7	42.8	−1.9
arc_easy	acc_norm	25.0	39.6	39.9	48.9	66.2	58.6	42.2	42.1	~0
lambada_openai	acc	0.0	32.2	32.7	23.5	40.2	52.5	17.8	16.6	−1.2
lambada_openai	ppl ↓	—	40.1	38.1	overflow	27.4	10.6	1064.3	984.6	−7.5 %

Average Δ over the 6 headline tasks vs the 3 B snapshot: +1.9 pt (3 of 6 positive). Boolq alone added +12.9 pt; piqa / arc_easy / lambada slipped by 0.5–1.9 pt (within single-seed noise band on these set sizes).

The headline qualitative event in this 2.24 B-extra-tokens window is boolq crossing chance (44 → 57) — eval_loss is still dropping monotonically (3.55 → 3.46) and lambada perplexity is now finite (was overflowing fp16 on our 100M @ 21B run), but most multi-choice tasks haven't moved much yet.

See EVAL_5B.md in the run-dir / GitHub repo for the full honesty pass.

How to load

A self-contained, runnable test ships with the repo:

pip install transformers safetensors huggingface_hub torch
python load_test.py

load_test.py does the full reload: snapshot-downloads this repo, imports MoEModel from the bundled model.py, builds the config from config.json, loads model.safetensors strictly, runs one forward pass, and samples 60 tokens from "Once upon a time,". A successful run prints LOAD_TEST: PASS.

Python:

from huggingface_hub import snapshot_download
import importlib.util, json, torch
from safetensors.torch import load_file
from pathlib import Path

local = Path(snapshot_download("AlexWortega/moe-200m-qwen3-step160000-5B-20260522-2256"))
spec = importlib.util.spec_from_file_location("_mdl", local / "model.py")
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)

cfg = json.loads((local / "config.json").read_text())
cfg = {k: v for k, v in cfg.items() if not k.startswith("_")}
config = mod.MoEModelConfig(**cfg)
config.router_noise_std = 0.0
config.use_liger_ce = False
config.use_chunked_ce = False

model = mod.MoEModel(config).to("cuda").to(torch.float32)
sd = load_file(local / "model.safetensors", device="cuda")
missing, unexpected = model.load_state_dict(sd, strict=False)
assert not unexpected
model.eval()

Reproducibility

model.py is the verbatim training-time file. config.json is generated from the checkpoint's cfg dict (asdict on the MoEModelConfig dataclass) with _model_class, _ckpt_step, and _tokens_seen metadata appended. The Qwen3 tokenizer is the unmodified Qwen/Qwen3-0.6B-Base tokenizer.

Caveats

Mid-training snapshot — outputs are coherent at the sentence level but not yet competitive on multi-choice benchmarks vs reference 124–500 M models.
Sigmoid router + grouped-bmm dispatch is the only supported backend at this scale; the legacy per-expert bmm backend would fall over on V100 throughput.
router_noise_std / use_liger_ce / use_chunked_ce are forced false in config.json for inference; the training-time AMP / chunked-CE / Liger fused-CE paths are documented in the bundled model.py but are not exercised by load_test.py.