Models / deepseek_vl_v2

DeepSeek-OCR-2

Reconstructed from its own config.json with no weights read. 963K 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

2.78B

2,777,122,560 parameters

In the published checkpoint

3.39B

3,389,119,360 scalars · safetensors.total, read 2026-02-03

Delta

-18.1%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `modeling_deepseekocr2.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

$274.99

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 8192
2	Embedding	Embedding	1 × 8192 × 1280
3	RoPE	RoPE	1 × 8192 × 1280
4	Vision input	Input	3 × 1024 × 1024
5	PatchEmbed	Patch Embed	4096 × 1280
6	Patch_Position_Embedding	Learned Pos Embed	4096 × 1280
7	Vision encoder (internals not in config)	Projection	4096 × 1280
8	Vision tokens	Reshape	1 × 4096 × 1280
9	Multimodal fusion (concat tokens)	Concatenate	1 × 12288 × 1280
10	RMSNorm_1_1	RMSNorm	1 × 12288 × 1280
11	Attention_1	Grouped Query Attn	1 × 12288 × 1280
12	Add_1_attn	Add	1 × 12288 × 1280
13	RMSNorm_1_2	RMSNorm	1 × 12288 × 1280
14	FFN_1	SwiGLU	1 × 12288 × 1280
15	Add_1_ffn	Add	1 × 12288 × 1280
16	RMSNorm_2_1	RMSNorm	1 × 12288 × 1280
17	Attention_2	Grouped Query Attn	1 × 12288 × 1280
18	Add_2_attn	Add	1 × 12288 × 1280
19	RMSNorm_2_2	RMSNorm	1 × 12288 × 1280
20	MoE_2	Shared-Expert MoE	1 × 12288 × 1280
21	Add_2_ffn	Add	1 × 12288 × 1280
22	RMSNorm_3_1	RMSNorm	1 × 12288 × 1280
23	Attention_3	Grouped Query Attn	1 × 12288 × 1280
24	Add_3_attn	Add	1 × 12288 × 1280
25	RMSNorm_3_2	RMSNorm	1 × 12288 × 1280
26	MoE_3	Shared-Expert MoE	1 × 12288 × 1280
27	Add_3_ffn	Add	1 × 12288 × 1280
28	RMSNorm_4_1	RMSNorm	1 × 12288 × 1280
29	Attention_4	Grouped Query Attn	1 × 12288 × 1280
30	Add_4_attn	Add	1 × 12288 × 1280
31	RMSNorm_4_2	RMSNorm	1 × 12288 × 1280
32	MoE_4	Shared-Expert MoE	1 × 12288 × 1280
33	Add_4_ffn	Add	1 × 12288 × 1280
34	RMSNorm_5_1	RMSNorm	1 × 12288 × 1280
35	Attention_5	Grouped Query Attn	1 × 12288 × 1280
36	Add_5_attn	Add	1 × 12288 × 1280
37	RMSNorm_5_2	RMSNorm	1 × 12288 × 1280
38	MoE_5	Shared-Expert MoE	1 × 12288 × 1280
39	Add_5_ffn	Add	1 × 12288 × 1280
40	RMSNorm_6_1	RMSNorm	1 × 12288 × 1280
41	Attention_6	Grouped Query Attn	1 × 12288 × 1280
42	Add_6_attn	Add	1 × 12288 × 1280
43	RMSNorm_6_2	RMSNorm	1 × 12288 × 1280
44	MoE_6	Shared-Expert MoE	1 × 12288 × 1280
45	Add_6_ffn	Add	1 × 12288 × 1280
46	RMSNorm_7_1	RMSNorm	1 × 12288 × 1280
47	Attention_7	Grouped Query Attn	1 × 12288 × 1280
48	Add_7_attn	Add	1 × 12288 × 1280
49	RMSNorm_7_2	RMSNorm	1 × 12288 × 1280
50	MoE_7	Shared-Expert MoE	1 × 12288 × 1280
51	Add_7_ffn	Add	1 × 12288 × 1280
52	RMSNorm_8_1	RMSNorm	1 × 12288 × 1280
53	Attention_8	Grouped Query Attn	1 × 12288 × 1280
54	Add_8_attn	Add	1 × 12288 × 1280
55	RMSNorm_8_2	RMSNorm	1 × 12288 × 1280
56	MoE_8	Shared-Expert MoE	1 × 12288 × 1280
57	Add_8_ffn	Add	1 × 12288 × 1280
58	RMSNorm_9_1	RMSNorm	1 × 12288 × 1280
59	Attention_9	Grouped Query Attn	1 × 12288 × 1280
60	Add_9_attn	Add	1 × 12288 × 1280
61	RMSNorm_9_2	RMSNorm	1 × 12288 × 1280
62	MoE_9	Shared-Expert MoE	1 × 12288 × 1280
63	Add_9_ffn	Add	1 × 12288 × 1280
64	RMSNorm_10_1	RMSNorm	1 × 12288 × 1280
65	Attention_10	Grouped Query Attn	1 × 12288 × 1280
66	Add_10_attn	Add	1 × 12288 × 1280
67	RMSNorm_10_2	RMSNorm	1 × 12288 × 1280
68	MoE_10	Shared-Expert MoE	1 × 12288 × 1280
69	Add_10_ffn	Add	1 × 12288 × 1280
70	RMSNorm_11_1	RMSNorm	1 × 12288 × 1280
71	Attention_11	Grouped Query Attn	1 × 12288 × 1280
72	Add_11_attn	Add	1 × 12288 × 1280
73	RMSNorm_11_2	RMSNorm	1 × 12288 × 1280
74	MoE_11	Shared-Expert MoE	1 × 12288 × 1280
75	Add_11_ffn	Add	1 × 12288 × 1280
76	RMSNorm_12_1	RMSNorm	1 × 12288 × 1280
77	Attention_12	Grouped Query Attn	1 × 12288 × 1280
78	Add_12_attn	Add	1 × 12288 × 1280
79	RMSNorm_12_2	RMSNorm	1 × 12288 × 1280
80	MoE_12	Shared-Expert MoE	1 × 12288 × 1280
81	Add_12_ffn	Add	1 × 12288 × 1280
82	Output	Output	1 × 12288 × 1280

What the verifier says

warn"RoPE" receives input but its output is not connected. This layer will be unreachable in the forward pass. Fix: Connect the output forward, or add an Output node if this is the final layer.
dead-end

infoLLaMA uses intermediateSize ≈ ⌊(8/3 × D) / 256⌋ × 256. Current: 6848 (5.35× embedDim). Expected: ~3328. Fix: Set intermediateSize to 3328 for embedDim=1280.
swiglu-dim-convention

infoAt 12 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 deepseek-ai/DeepSeek-OCR-2 --plan --share

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

Other deepseek_vl_v2 checkpoints

DeepSeek-OCR

2.78B derived · -16.8% against the checkpoint