Models / llada2_moe

LLaDA2.1-mini

Reconstructed from its own config.json with no weights read. 110K downloads on Hugging Face.

Our count against the checkpoint

The left number comes from the graph. The right one is the number of scalars in the published weight files. Nothing on this page was tuned to make them agree.

Derived from structure

1.14B

1,144,385,536 parameters

In the published checkpoint

16.26B

16,255,643,392 scalars · safetensors.total, read 2026-04-13

Delta

-93.0%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_llada2_moe.py`), so `config.json` names a class in the repo rather than an architecture `transformers` defines. The graph below is what those config keys mean under `transformers` semantics, which is not necessarily what the repo's own file builds. A gap here is a statement about what we read, not about the checkpoint.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32 over 50,000 samples (assumed; no dataset attached). GPU fit is fp32 weights plus gradients plus two Adam moments (16 bytes per parameter) with 1.3x headroom; activations are not included and grow with batch size.

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$2572.10

Card	Memory
T4 (16GB)	weights + activations	does not fit
A100 (40GB)	weights + activations	fits
H100 (80GB)	weights + activations	fits

Structure

84 nodes. Output shapes are propagated from the input shape, batch dimension excluded.

	Layer	Type	Output shape
1	Input	Input	1 × 32768
2	Embedding	Embedding	1 × 32768 × 2048
3	Positional_Embedding	Learned Pos Embed	1 × 32768 × 2048
4	Attention_1	Multi-Head Attention	1 × 32768 × 2048
5	Add_1	Add	1 × 32768 × 2048
6	LayerNorm_1_1	LayerNorm	1 × 32768 × 2048
7	FFN_1	Feed Forward	1 × 32768 × 2048
8	Attention_2	Multi-Head Attention	1 × 32768 × 2048
9	Add_2	Add	1 × 32768 × 2048
10	LayerNorm_2_1	LayerNorm	1 × 32768 × 2048
11	FFN_2	Feed Forward	1 × 32768 × 2048
12	Attention_3	Multi-Head Attention	1 × 32768 × 2048
13	Add_3	Add	1 × 32768 × 2048
14	LayerNorm_3_1	LayerNorm	1 × 32768 × 2048
15	FFN_3	Feed Forward	1 × 32768 × 2048
16	Attention_4	Multi-Head Attention	1 × 32768 × 2048
17	Add_4	Add	1 × 32768 × 2048
18	LayerNorm_4_1	LayerNorm	1 × 32768 × 2048
19	FFN_4	Feed Forward	1 × 32768 × 2048
20	Attention_5	Multi-Head Attention	1 × 32768 × 2048
21	Add_5	Add	1 × 32768 × 2048
22	LayerNorm_5_1	LayerNorm	1 × 32768 × 2048
23	FFN_5	Feed Forward	1 × 32768 × 2048
24	Attention_6	Multi-Head Attention	1 × 32768 × 2048
25	Add_6	Add	1 × 32768 × 2048
26	LayerNorm_6_1	LayerNorm	1 × 32768 × 2048
27	FFN_6	Feed Forward	1 × 32768 × 2048
28	Attention_7	Multi-Head Attention	1 × 32768 × 2048
29	Add_7	Add	1 × 32768 × 2048
30	LayerNorm_7_1	LayerNorm	1 × 32768 × 2048
31	FFN_7	Feed Forward	1 × 32768 × 2048
32	Attention_8	Multi-Head Attention	1 × 32768 × 2048
33	Add_8	Add	1 × 32768 × 2048
34	LayerNorm_8_1	LayerNorm	1 × 32768 × 2048
35	FFN_8	Feed Forward	1 × 32768 × 2048
36	Attention_9	Multi-Head Attention	1 × 32768 × 2048
37	Add_9	Add	1 × 32768 × 2048
38	LayerNorm_9_1	LayerNorm	1 × 32768 × 2048
39	FFN_9	Feed Forward	1 × 32768 × 2048
40	Attention_10	Multi-Head Attention	1 × 32768 × 2048
41	Add_10	Add	1 × 32768 × 2048
42	LayerNorm_10_1	LayerNorm	1 × 32768 × 2048
43	FFN_10	Feed Forward	1 × 32768 × 2048
44	Attention_11	Multi-Head Attention	1 × 32768 × 2048
45	Add_11	Add	1 × 32768 × 2048
46	LayerNorm_11_1	LayerNorm	1 × 32768 × 2048
47	FFN_11	Feed Forward	1 × 32768 × 2048
48	Attention_12	Multi-Head Attention	1 × 32768 × 2048
49	Add_12	Add	1 × 32768 × 2048
50	LayerNorm_12_1	LayerNorm	1 × 32768 × 2048
51	FFN_12	Feed Forward	1 × 32768 × 2048
52	Attention_13	Multi-Head Attention	1 × 32768 × 2048
53	Add_13	Add	1 × 32768 × 2048
54	LayerNorm_13_1	LayerNorm	1 × 32768 × 2048
55	FFN_13	Feed Forward	1 × 32768 × 2048
56	Attention_14	Multi-Head Attention	1 × 32768 × 2048
57	Add_14	Add	1 × 32768 × 2048
58	LayerNorm_14_1	LayerNorm	1 × 32768 × 2048
59	FFN_14	Feed Forward	1 × 32768 × 2048
60	Attention_15	Multi-Head Attention	1 × 32768 × 2048
61	Add_15	Add	1 × 32768 × 2048
62	LayerNorm_15_1	LayerNorm	1 × 32768 × 2048
63	FFN_15	Feed Forward	1 × 32768 × 2048
64	Attention_16	Multi-Head Attention	1 × 32768 × 2048
65	Add_16	Add	1 × 32768 × 2048
66	LayerNorm_16_1	LayerNorm	1 × 32768 × 2048
67	FFN_16	Feed Forward	1 × 32768 × 2048
68	Attention_17	Multi-Head Attention	1 × 32768 × 2048
69	Add_17	Add	1 × 32768 × 2048
70	LayerNorm_17_1	LayerNorm	1 × 32768 × 2048
71	FFN_17	Feed Forward	1 × 32768 × 2048
72	Attention_18	Multi-Head Attention	1 × 32768 × 2048
73	Add_18	Add	1 × 32768 × 2048
74	LayerNorm_18_1	LayerNorm	1 × 32768 × 2048
75	FFN_18	Feed Forward	1 × 32768 × 2048
76	Attention_19	Multi-Head Attention	1 × 32768 × 2048
77	Add_19	Add	1 × 32768 × 2048
78	LayerNorm_19_1	LayerNorm	1 × 32768 × 2048
79	FFN_19	Feed Forward	1 × 32768 × 2048
80	Attention_20	Multi-Head Attention	1 × 32768 × 2048
81	Add_20	Add	1 × 32768 × 2048
82	LayerNorm_20_1	LayerNorm	1 × 32768 × 2048
83	FFN_20	Feed Forward	1 × 32768 × 2048
84	Output	Output	1 × 32768 × 2048

What the verifier says

info20 attention layers at embedDim 2048 cache full per-head K/V: about 160 KB per token at fp16, which dominates memory at long context. Grouped-query attention (e.g. 8:1) would cut this ~8×; multi-head latent attention (MLA) shrinks it ~10× or more. This is the move production LLMs make; it does not change the parameter count. Fix: Switch attention to groupedQueryAttention (set numKVHeads below numHeads, e.g. numHeads/4) or mla (a low-rank cached latent).
full-mha-serving-cost

infoAt 20 stacked attention layers, residual-branch outputs add up; unscaled init lets activation variance grow with depth. GPT-2/LLaMA-family models scale the residual projections by depth (N(0, 0.02 / √(2L))). Fix: Scale residual output projections by depth: nn.init.normal_(w, std=0.02 / math.sqrt(2 * n_layers))
deep-attention-default-init

Do this to your own model

Same numbers, on a model in your repo, in one command. No account.

pip install neurarch-trace
neurarch-trace inclusionAI/LLaDA2.1-mini --plan --share

Paste a model instead Markdown twin Index JSON POST /api/v1/plan

Other llada2_moe checkpoints

LLaDA2.0-mini

1.14B derived · -93.0% against the checkpoint