Models / granite4_vision

granite-vision-4.1-4b

Reconstructed from its own config.json with no weights read. 115K 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

3.32B

3,319,787,520 parameters

In the published checkpoint

4.00B

3,997,203,904 scalars · safetensors.total, read 2026-07-13

Delta

-16.9%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration.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

162

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$79217.40

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 131072
2	Embedding	Embedding	1 × 131072 × 2560
3	Positional_Embedding	Learned Pos Embed	1 × 131072 × 2560
4	Attention_1	Multi-Head Attention	1 × 131072 × 2560
5	Add_1	Add	1 × 131072 × 2560
6	LayerNorm_1_1	LayerNorm	1 × 131072 × 2560
7	FFN_1	Feed Forward	1 × 131072 × 2560
8	Attention_2	Multi-Head Attention	1 × 131072 × 2560
9	Add_2	Add	1 × 131072 × 2560
10	LayerNorm_2_1	LayerNorm	1 × 131072 × 2560
11	FFN_2	Feed Forward	1 × 131072 × 2560
12	Attention_3	Multi-Head Attention	1 × 131072 × 2560
13	Add_3	Add	1 × 131072 × 2560
14	LayerNorm_3_1	LayerNorm	1 × 131072 × 2560
15	FFN_3	Feed Forward	1 × 131072 × 2560
16	Attention_4	Multi-Head Attention	1 × 131072 × 2560
17	Add_4	Add	1 × 131072 × 2560
18	LayerNorm_4_1	LayerNorm	1 × 131072 × 2560
19	FFN_4	Feed Forward	1 × 131072 × 2560
20	Attention_5	Multi-Head Attention	1 × 131072 × 2560
21	Add_5	Add	1 × 131072 × 2560
22	LayerNorm_5_1	LayerNorm	1 × 131072 × 2560
23	FFN_5	Feed Forward	1 × 131072 × 2560
24	Attention_6	Multi-Head Attention	1 × 131072 × 2560
25	Add_6	Add	1 × 131072 × 2560
26	LayerNorm_6_1	LayerNorm	1 × 131072 × 2560
27	FFN_6	Feed Forward	1 × 131072 × 2560
28	Attention_7	Multi-Head Attention	1 × 131072 × 2560
29	Add_7	Add	1 × 131072 × 2560
30	LayerNorm_7_1	LayerNorm	1 × 131072 × 2560
31	FFN_7	Feed Forward	1 × 131072 × 2560
32	Attention_8	Multi-Head Attention	1 × 131072 × 2560
33	Add_8	Add	1 × 131072 × 2560
34	LayerNorm_8_1	LayerNorm	1 × 131072 × 2560
35	FFN_8	Feed Forward	1 × 131072 × 2560
36	Attention_9	Multi-Head Attention	1 × 131072 × 2560
37	Add_9	Add	1 × 131072 × 2560
38	LayerNorm_9_1	LayerNorm	1 × 131072 × 2560
39	FFN_9	Feed Forward	1 × 131072 × 2560
40	Attention_10	Multi-Head Attention	1 × 131072 × 2560
41	Add_10	Add	1 × 131072 × 2560
42	LayerNorm_10_1	LayerNorm	1 × 131072 × 2560
43	FFN_10	Feed Forward	1 × 131072 × 2560
44	Attention_11	Multi-Head Attention	1 × 131072 × 2560
45	Add_11	Add	1 × 131072 × 2560
46	LayerNorm_11_1	LayerNorm	1 × 131072 × 2560
47	FFN_11	Feed Forward	1 × 131072 × 2560
48	Attention_12	Multi-Head Attention	1 × 131072 × 2560
49	Add_12	Add	1 × 131072 × 2560
50	LayerNorm_12_1	LayerNorm	1 × 131072 × 2560
51	FFN_12	Feed Forward	1 × 131072 × 2560
52	Attention_13	Multi-Head Attention	1 × 131072 × 2560
53	Add_13	Add	1 × 131072 × 2560
54	LayerNorm_13_1	LayerNorm	1 × 131072 × 2560
55	FFN_13	Feed Forward	1 × 131072 × 2560
56	Attention_14	Multi-Head Attention	1 × 131072 × 2560
57	Add_14	Add	1 × 131072 × 2560
58	LayerNorm_14_1	LayerNorm	1 × 131072 × 2560
59	FFN_14	Feed Forward	1 × 131072 × 2560
60	Attention_15	Multi-Head Attention	1 × 131072 × 2560
61	Add_15	Add	1 × 131072 × 2560
62	LayerNorm_15_1	LayerNorm	1 × 131072 × 2560
63	FFN_15	Feed Forward	1 × 131072 × 2560
64	Attention_16	Multi-Head Attention	1 × 131072 × 2560
65	Add_16	Add	1 × 131072 × 2560
66	LayerNorm_16_1	LayerNorm	1 × 131072 × 2560
67	FFN_16	Feed Forward	1 × 131072 × 2560
68	Attention_17	Multi-Head Attention	1 × 131072 × 2560
69	Add_17	Add	1 × 131072 × 2560
70	LayerNorm_17_1	LayerNorm	1 × 131072 × 2560
71	FFN_17	Feed Forward	1 × 131072 × 2560
72	Attention_18	Multi-Head Attention	1 × 131072 × 2560
73	Add_18	Add	1 × 131072 × 2560
74	LayerNorm_18_1	LayerNorm	1 × 131072 × 2560
75	FFN_18	Feed Forward	1 × 131072 × 2560
76	Attention_19	Multi-Head Attention	1 × 131072 × 2560
77	Add_19	Add	1 × 131072 × 2560
78	LayerNorm_19_1	LayerNorm	1 × 131072 × 2560
79	FFN_19	Feed Forward	1 × 131072 × 2560
80	Attention_20	Multi-Head Attention	1 × 131072 × 2560
81	Add_20	Add	1 × 131072 × 2560
82	LayerNorm_20_1	LayerNorm	1 × 131072 × 2560
83	FFN_20	Feed Forward	1 × 131072 × 2560
84	Attention_21	Multi-Head Attention	1 × 131072 × 2560
85	Add_21	Add	1 × 131072 × 2560
86	LayerNorm_21_1	LayerNorm	1 × 131072 × 2560
87	FFN_21	Feed Forward	1 × 131072 × 2560
88	Attention_22	Multi-Head Attention	1 × 131072 × 2560
89	Add_22	Add	1 × 131072 × 2560
90	LayerNorm_22_1	LayerNorm	1 × 131072 × 2560
91	FFN_22	Feed Forward	1 × 131072 × 2560
92	Attention_23	Multi-Head Attention	1 × 131072 × 2560
93	Add_23	Add	1 × 131072 × 2560
94	LayerNorm_23_1	LayerNorm	1 × 131072 × 2560
95	FFN_23	Feed Forward	1 × 131072 × 2560
96	Attention_24	Multi-Head Attention	1 × 131072 × 2560
97	Add_24	Add	1 × 131072 × 2560
98	LayerNorm_24_1	LayerNorm	1 × 131072 × 2560
99	FFN_24	Feed Forward	1 × 131072 × 2560
100	Attention_25	Multi-Head Attention	1 × 131072 × 2560
101	Add_25	Add	1 × 131072 × 2560
102	LayerNorm_25_1	LayerNorm	1 × 131072 × 2560
103	FFN_25	Feed Forward	1 × 131072 × 2560
104	Attention_26	Multi-Head Attention	1 × 131072 × 2560
105	Add_26	Add	1 × 131072 × 2560
106	LayerNorm_26_1	LayerNorm	1 × 131072 × 2560
107	FFN_26	Feed Forward	1 × 131072 × 2560
108	Attention_27	Multi-Head Attention	1 × 131072 × 2560
109	Add_27	Add	1 × 131072 × 2560
110	LayerNorm_27_1	LayerNorm	1 × 131072 × 2560
111	FFN_27	Feed Forward	1 × 131072 × 2560
112	Attention_28	Multi-Head Attention	1 × 131072 × 2560
113	Add_28	Add	1 × 131072 × 2560
114	LayerNorm_28_1	LayerNorm	1 × 131072 × 2560
115	FFN_28	Feed Forward	1 × 131072 × 2560
116	Attention_29	Multi-Head Attention	1 × 131072 × 2560
117	Add_29	Add	1 × 131072 × 2560
118	LayerNorm_29_1	LayerNorm	1 × 131072 × 2560
119	FFN_29	Feed Forward	1 × 131072 × 2560
120	Attention_30	Multi-Head Attention	1 × 131072 × 2560
121	Add_30	Add	1 × 131072 × 2560
122	LayerNorm_30_1	LayerNorm	1 × 131072 × 2560
123	FFN_30	Feed Forward	1 × 131072 × 2560
124	Attention_31	Multi-Head Attention	1 × 131072 × 2560
125	Add_31	Add	1 × 131072 × 2560
126	LayerNorm_31_1	LayerNorm	1 × 131072 × 2560
127	FFN_31	Feed Forward	1 × 131072 × 2560
128	Attention_32	Multi-Head Attention	1 × 131072 × 2560
129	Add_32	Add	1 × 131072 × 2560
130	LayerNorm_32_1	LayerNorm	1 × 131072 × 2560
131	FFN_32	Feed Forward	1 × 131072 × 2560
132	Attention_33	Multi-Head Attention	1 × 131072 × 2560
133	Add_33	Add	1 × 131072 × 2560
134	LayerNorm_33_1	LayerNorm	1 × 131072 × 2560
135	FFN_33	Feed Forward	1 × 131072 × 2560
136	Attention_34	Multi-Head Attention	1 × 131072 × 2560
137	Add_34	Add	1 × 131072 × 2560
138	LayerNorm_34_1	LayerNorm	1 × 131072 × 2560
139	FFN_34	Feed Forward	1 × 131072 × 2560
140	Attention_35	Multi-Head Attention	1 × 131072 × 2560
141	Add_35	Add	1 × 131072 × 2560
142	LayerNorm_35_1	LayerNorm	1 × 131072 × 2560
143	FFN_35	Feed Forward	1 × 131072 × 2560
144	Attention_36	Multi-Head Attention	1 × 131072 × 2560
145	Add_36	Add	1 × 131072 × 2560
146	LayerNorm_36_1	LayerNorm	1 × 131072 × 2560
147	FFN_36	Feed Forward	1 × 131072 × 2560
148	Attention_37	Multi-Head Attention	1 × 131072 × 2560
149	Add_37	Add	1 × 131072 × 2560
150	LayerNorm_37_1	LayerNorm	1 × 131072 × 2560
151	FFN_37	Feed Forward	1 × 131072 × 2560
152	Attention_38	Multi-Head Attention	1 × 131072 × 2560
153	Add_38	Add	1 × 131072 × 2560
154	LayerNorm_38_1	LayerNorm	1 × 131072 × 2560
155	FFN_38	Feed Forward	1 × 131072 × 2560
156	Attention_39	Multi-Head Attention	1 × 131072 × 2560
157	Add_39	Add	1 × 131072 × 2560
158	LayerNorm_39_1	LayerNorm	1 × 131072 × 2560
159	FFN_39	Feed Forward	1 × 131072 × 2560
160	Attention_40	Multi-Head Attention	1 × 131072 × 2560
161	Add_40	Add	1 × 131072 × 2560
162	LayerNorm_40_1	LayerNorm	1 × 131072 × 2560
163	FFN_40	Feed Forward	1 × 131072 × 2560
164	Output	Output	1 × 131072 × 2560

What the verifier says

info40 attention layers at embedDim 2560 cache full per-head K/V: about 400 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 40 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 ibm-granite/granite-vision-4.1-4b --plan --share

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