Models / vit

vit-base-patch16-224: GPU memory to fine-tune

Fine-tuning vit-base-patch16-224 at batch 1 takes about 1.72 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 0.29 GB with LoRA r=16 on a bf16 base (fits one T4 16 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 5.1M 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

85.8M

85,797,120 parameters

In the published checkpoint

86.6M

86,567,656 scalars · safetensors.total, read 2026-09-06

Delta

-0.89%

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).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$0.69

Memory: what fits on which card

Largest batch per card. Training is sized at the input shape below per sample.

Full fine-tune, batch 1

1.72 GB

LoRA r=16, batch 1

0.29 GB

Card	Full fine-tune	LoRA r=16
T4 16 GB	batch 177	batch 198
L4 24 GB	batch 278	batch 299
A100 40 GB	batch 480	batch 500
A100 / H100 80 GB	batch 983	batch 1004
H200 141 GB	batch 1752	batch 1772

Training: the input shape in the structure table, AdamW with bf16 autocast, SDPA attention (no score matrix stored), 1.15x headroom on every card. Full fine-tune holds weights, gradients and optimizer state (1.2 x 16 bytes per parameter) plus activations. LoRA is r=16 on every linear layer but the head, base frozen in bf16 (2 bytes per parameter), adapters at full AdamW state, same activations. A two-card split is a contiguous layer cut (pipeline parallel), computed only where one card cannot hold batch 1.

These are estimates from the graph, not measurements. Against measured runs, held out by architecture, the estimate flagged 113 of the 131 jobs that did not fit their card and said "fits" on the other 18. Every measured run.

Structure

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

	Layer	Type	Output shape
1	Input	Input	3 × 224 × 224
2	PatchEmbedding	Patch Embed	197 × 768
3	Patch_Position_Embedding	Learned Pos Embed	197 × 768
4	Attention_1	Multi-Head Attention	197 × 768
5	Add_1	Add	197 × 768
6	LayerNorm_1_1	LayerNorm	197 × 768
7	FFN_1	Feed Forward	197 × 768
8	Add_1_2	Add	197 × 768
9	LayerNorm_1_2	LayerNorm	197 × 768
10	Attention_2	Multi-Head Attention	197 × 768
11	Add_2	Add	197 × 768
12	LayerNorm_2_1	LayerNorm	197 × 768
13	FFN_2	Feed Forward	197 × 768
14	Add_2_2	Add	197 × 768
15	LayerNorm_2_2	LayerNorm	197 × 768
16	Attention_3	Multi-Head Attention	197 × 768
17	Add_3	Add	197 × 768
18	LayerNorm_3_1	LayerNorm	197 × 768
19	FFN_3	Feed Forward	197 × 768
20	Add_3_2	Add	197 × 768
21	LayerNorm_3_2	LayerNorm	197 × 768
22	Attention_4	Multi-Head Attention	197 × 768
23	Add_4	Add	197 × 768
24	LayerNorm_4_1	LayerNorm	197 × 768
25	FFN_4	Feed Forward	197 × 768
26	Add_4_2	Add	197 × 768
27	LayerNorm_4_2	LayerNorm	197 × 768
28	Attention_5	Multi-Head Attention	197 × 768
29	Add_5	Add	197 × 768
30	LayerNorm_5_1	LayerNorm	197 × 768
31	FFN_5	Feed Forward	197 × 768
32	Add_5_2	Add	197 × 768
33	LayerNorm_5_2	LayerNorm	197 × 768
34	Attention_6	Multi-Head Attention	197 × 768
35	Add_6	Add	197 × 768
36	LayerNorm_6_1	LayerNorm	197 × 768
37	FFN_6	Feed Forward	197 × 768
38	Add_6_2	Add	197 × 768
39	LayerNorm_6_2	LayerNorm	197 × 768
40	Attention_7	Multi-Head Attention	197 × 768
41	Add_7	Add	197 × 768
42	LayerNorm_7_1	LayerNorm	197 × 768
43	FFN_7	Feed Forward	197 × 768
44	Add_7_2	Add	197 × 768
45	LayerNorm_7_2	LayerNorm	197 × 768
46	Attention_8	Multi-Head Attention	197 × 768
47	Add_8	Add	197 × 768
48	LayerNorm_8_1	LayerNorm	197 × 768
49	FFN_8	Feed Forward	197 × 768
50	Add_8_2	Add	197 × 768
51	LayerNorm_8_2	LayerNorm	197 × 768
52	Attention_9	Multi-Head Attention	197 × 768
53	Add_9	Add	197 × 768
54	LayerNorm_9_1	LayerNorm	197 × 768
55	FFN_9	Feed Forward	197 × 768
56	Add_9_2	Add	197 × 768
57	LayerNorm_9_2	LayerNorm	197 × 768
58	Attention_10	Multi-Head Attention	197 × 768
59	Add_10	Add	197 × 768
60	LayerNorm_10_1	LayerNorm	197 × 768
61	FFN_10	Feed Forward	197 × 768
62	Add_10_2	Add	197 × 768
63	LayerNorm_10_2	LayerNorm	197 × 768
64	Attention_11	Multi-Head Attention	197 × 768
65	Add_11	Add	197 × 768
66	LayerNorm_11_1	LayerNorm	197 × 768
67	FFN_11	Feed Forward	197 × 768
68	Add_11_2	Add	197 × 768
69	LayerNorm_11_2	LayerNorm	197 × 768
70	Attention_12	Multi-Head Attention	197 × 768
71	Add_12	Add	197 × 768
72	LayerNorm_12_1	LayerNorm	197 × 768
73	FFN_12	Feed Forward	197 × 768
74	Add_12_2	Add	197 × 768
75	LayerNorm_12_2	LayerNorm	197 × 768
76	Output	Output	197 × 768

What the verifier says

warn"LayerNorm_12_2" (layerNorm) is the last layer before Output. Normalizing the raw logits constrains the output range and breaks standard loss functions. Fix: Move normalization before the final Linear/Conv layer.
bn-at-output

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

What changed upstream

The change tracker follows google/vit-base-patch16-224: every revision it has read, what changed in the config, tokenizer, chat template or weights between them, and an OSV advisory per change.

Revision history and changes Badge JSON OSV advisories All tracked changes

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 google/vit-base-patch16-224 --plan --share

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

Other vit checkpoints

nsfw_image_detection

124M derived · +45.0% against the checkpoint

vit-base-oxford-iiit-pets

85.8M derived · -0.03% against the checkpoint