esm2_t33_650M_UR50D

Reconstructed from its own config.json with no weights read. 1.4M 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

651M

650,626,560 parameters

In the published checkpoint

652M

652,358,616 scalars · safetensors.total, read 2026-09-06

Delta

-0.27%

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

134

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$26.62

Card	Memory
T4 (16GB)	weights + activations	fits
A100 (40GB)	weights + activations	fits
H100 (80GB)	weights + activations	fits

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 1026
2	Embedding	Embedding	1 × 1026 × 1280
3	Positional_Embedding	Learned Pos Embed	1 × 1026 × 1280
4	Attention_1	Multi-Head Attention	1 × 1026 × 1280
5	Add_1	Add	1 × 1026 × 1280
6	LayerNorm_1_1	LayerNorm	1 × 1026 × 1280
7	FFN_1	Feed Forward	1 × 1026 × 1280
8	Attention_2	Multi-Head Attention	1 × 1026 × 1280
9	Add_2	Add	1 × 1026 × 1280
10	LayerNorm_2_1	LayerNorm	1 × 1026 × 1280
11	FFN_2	Feed Forward	1 × 1026 × 1280
12	Attention_3	Multi-Head Attention	1 × 1026 × 1280
13	Add_3	Add	1 × 1026 × 1280
14	LayerNorm_3_1	LayerNorm	1 × 1026 × 1280
15	FFN_3	Feed Forward	1 × 1026 × 1280
16	Attention_4	Multi-Head Attention	1 × 1026 × 1280
17	Add_4	Add	1 × 1026 × 1280
18	LayerNorm_4_1	LayerNorm	1 × 1026 × 1280
19	FFN_4	Feed Forward	1 × 1026 × 1280
20	Attention_5	Multi-Head Attention	1 × 1026 × 1280
21	Add_5	Add	1 × 1026 × 1280
22	LayerNorm_5_1	LayerNorm	1 × 1026 × 1280
23	FFN_5	Feed Forward	1 × 1026 × 1280
24	Attention_6	Multi-Head Attention	1 × 1026 × 1280
25	Add_6	Add	1 × 1026 × 1280
26	LayerNorm_6_1	LayerNorm	1 × 1026 × 1280
27	FFN_6	Feed Forward	1 × 1026 × 1280
28	Attention_7	Multi-Head Attention	1 × 1026 × 1280
29	Add_7	Add	1 × 1026 × 1280
30	LayerNorm_7_1	LayerNorm	1 × 1026 × 1280
31	FFN_7	Feed Forward	1 × 1026 × 1280
32	Attention_8	Multi-Head Attention	1 × 1026 × 1280
33	Add_8	Add	1 × 1026 × 1280
34	LayerNorm_8_1	LayerNorm	1 × 1026 × 1280
35	FFN_8	Feed Forward	1 × 1026 × 1280
36	Attention_9	Multi-Head Attention	1 × 1026 × 1280
37	Add_9	Add	1 × 1026 × 1280
38	LayerNorm_9_1	LayerNorm	1 × 1026 × 1280
39	FFN_9	Feed Forward	1 × 1026 × 1280
40	Attention_10	Multi-Head Attention	1 × 1026 × 1280
41	Add_10	Add	1 × 1026 × 1280
42	LayerNorm_10_1	LayerNorm	1 × 1026 × 1280
43	FFN_10	Feed Forward	1 × 1026 × 1280
44	Attention_11	Multi-Head Attention	1 × 1026 × 1280
45	Add_11	Add	1 × 1026 × 1280
46	LayerNorm_11_1	LayerNorm	1 × 1026 × 1280
47	FFN_11	Feed Forward	1 × 1026 × 1280
48	Attention_12	Multi-Head Attention	1 × 1026 × 1280
49	Add_12	Add	1 × 1026 × 1280
50	LayerNorm_12_1	LayerNorm	1 × 1026 × 1280
51	FFN_12	Feed Forward	1 × 1026 × 1280
52	Attention_13	Multi-Head Attention	1 × 1026 × 1280
53	Add_13	Add	1 × 1026 × 1280
54	LayerNorm_13_1	LayerNorm	1 × 1026 × 1280
55	FFN_13	Feed Forward	1 × 1026 × 1280
56	Attention_14	Multi-Head Attention	1 × 1026 × 1280
57	Add_14	Add	1 × 1026 × 1280
58	LayerNorm_14_1	LayerNorm	1 × 1026 × 1280
59	FFN_14	Feed Forward	1 × 1026 × 1280
60	Attention_15	Multi-Head Attention	1 × 1026 × 1280
61	Add_15	Add	1 × 1026 × 1280
62	LayerNorm_15_1	LayerNorm	1 × 1026 × 1280
63	FFN_15	Feed Forward	1 × 1026 × 1280
64	Attention_16	Multi-Head Attention	1 × 1026 × 1280
65	Add_16	Add	1 × 1026 × 1280
66	LayerNorm_16_1	LayerNorm	1 × 1026 × 1280
67	FFN_16	Feed Forward	1 × 1026 × 1280
68	Attention_17	Multi-Head Attention	1 × 1026 × 1280
69	Add_17	Add	1 × 1026 × 1280
70	LayerNorm_17_1	LayerNorm	1 × 1026 × 1280
71	FFN_17	Feed Forward	1 × 1026 × 1280
72	Attention_18	Multi-Head Attention	1 × 1026 × 1280
73	Add_18	Add	1 × 1026 × 1280
74	LayerNorm_18_1	LayerNorm	1 × 1026 × 1280
75	FFN_18	Feed Forward	1 × 1026 × 1280
76	Attention_19	Multi-Head Attention	1 × 1026 × 1280
77	Add_19	Add	1 × 1026 × 1280
78	LayerNorm_19_1	LayerNorm	1 × 1026 × 1280
79	FFN_19	Feed Forward	1 × 1026 × 1280
80	Attention_20	Multi-Head Attention	1 × 1026 × 1280
81	Add_20	Add	1 × 1026 × 1280
82	LayerNorm_20_1	LayerNorm	1 × 1026 × 1280
83	FFN_20	Feed Forward	1 × 1026 × 1280
84	Attention_21	Multi-Head Attention	1 × 1026 × 1280
85	Add_21	Add	1 × 1026 × 1280
86	LayerNorm_21_1	LayerNorm	1 × 1026 × 1280
87	FFN_21	Feed Forward	1 × 1026 × 1280
88	Attention_22	Multi-Head Attention	1 × 1026 × 1280
89	Add_22	Add	1 × 1026 × 1280
90	LayerNorm_22_1	LayerNorm	1 × 1026 × 1280
91	FFN_22	Feed Forward	1 × 1026 × 1280
92	Attention_23	Multi-Head Attention	1 × 1026 × 1280
93	Add_23	Add	1 × 1026 × 1280
94	LayerNorm_23_1	LayerNorm	1 × 1026 × 1280
95	FFN_23	Feed Forward	1 × 1026 × 1280
96	Attention_24	Multi-Head Attention	1 × 1026 × 1280
97	Add_24	Add	1 × 1026 × 1280
98	LayerNorm_24_1	LayerNorm	1 × 1026 × 1280
99	FFN_24	Feed Forward	1 × 1026 × 1280
100	Attention_25	Multi-Head Attention	1 × 1026 × 1280
101	Add_25	Add	1 × 1026 × 1280
102	LayerNorm_25_1	LayerNorm	1 × 1026 × 1280
103	FFN_25	Feed Forward	1 × 1026 × 1280
104	Attention_26	Multi-Head Attention	1 × 1026 × 1280
105	Add_26	Add	1 × 1026 × 1280
106	LayerNorm_26_1	LayerNorm	1 × 1026 × 1280
107	FFN_26	Feed Forward	1 × 1026 × 1280
108	Attention_27	Multi-Head Attention	1 × 1026 × 1280
109	Add_27	Add	1 × 1026 × 1280
110	LayerNorm_27_1	LayerNorm	1 × 1026 × 1280
111	FFN_27	Feed Forward	1 × 1026 × 1280
112	Attention_28	Multi-Head Attention	1 × 1026 × 1280
113	Add_28	Add	1 × 1026 × 1280
114	LayerNorm_28_1	LayerNorm	1 × 1026 × 1280
115	FFN_28	Feed Forward	1 × 1026 × 1280
116	Attention_29	Multi-Head Attention	1 × 1026 × 1280
117	Add_29	Add	1 × 1026 × 1280
118	LayerNorm_29_1	LayerNorm	1 × 1026 × 1280
119	FFN_29	Feed Forward	1 × 1026 × 1280
120	Attention_30	Multi-Head Attention	1 × 1026 × 1280
121	Add_30	Add	1 × 1026 × 1280
122	LayerNorm_30_1	LayerNorm	1 × 1026 × 1280
123	FFN_30	Feed Forward	1 × 1026 × 1280
124	Attention_31	Multi-Head Attention	1 × 1026 × 1280
125	Add_31	Add	1 × 1026 × 1280
126	LayerNorm_31_1	LayerNorm	1 × 1026 × 1280
127	FFN_31	Feed Forward	1 × 1026 × 1280
128	Attention_32	Multi-Head Attention	1 × 1026 × 1280
129	Add_32	Add	1 × 1026 × 1280
130	LayerNorm_32_1	LayerNorm	1 × 1026 × 1280
131	FFN_32	Feed Forward	1 × 1026 × 1280
132	Attention_33	Multi-Head Attention	1 × 1026 × 1280
133	Add_33	Add	1 × 1026 × 1280
134	LayerNorm_33_1	LayerNorm	1 × 1026 × 1280
135	FFN_33	Feed Forward	1 × 1026 × 1280
136	Output	Output	1 × 1026 × 1280

What the verifier says

infoAt 33 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 facebook/esm2_t33_650M_UR50D --plan --share

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