Models / jina_embeddings_v5

jina-embeddings-v5-text-small: GPU memory to fine-tune

Fine-tuning jina-embeddings-v5-text-small at batch 1 and 2048 tokens takes about 13.9 GB of GPU memory for a full fine-tune (fits one L4 24 GB), and 3.86 GB with LoRA r=16 on a bf16 base (fits one T4 16 GB). Estimated from the model's structure, reading the config under transformers semantics, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 276K 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

596M

596,041,728 parameters

In the published checkpoint

596M

596,049,920 scalars · safetensors.total, read 2026-04-15

Delta

-0.00%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_jina_embeddings_v5.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, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

114

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$42.81

Memory: what fits on which card

Largest batch per card. Training is sized at 2048 tokens per sample, serving at 4096 tokens per sequence.

Full fine-tune, batch 1 at 2048 tokens

13.9 GB

LoRA r=16, batch 1 at 2048 tokens

3.86 GB

Serve: bf16 weights

1.19 GB

Serve: KV cache, 4096 tokens

0.47 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	2 cards at batch 1 (113 + 1 layers)	batch 5	batch 27
L4 24 GB	batch 3	batch 7	batch 41
A100 40 GB	batch 9	batch 13	batch 71
A100 / H100 80 GB	batch 23	batch 27	batch 145
H200 141 GB	batch 44	batch 48	batch 258

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 32768), 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. Serving: bf16 weights plus the KV cache of 4096 tokens per sequence; prefill working set, CUDA context and framework reserve not counted.

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

116 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 32768 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 32768
2	Embedding	Embedding	1 × 32768 × 1024
3	RoPE	RoPE	1 × 32768 × 1024
4	Attention_1	Grouped Query Attn	1 × 32768 × 1024
5	Add_1	Add	1 × 32768 × 1024
6	LayerNorm_1_1	LayerNorm	1 × 32768 × 1024
7	FFN_1	SwiGLU	1 × 32768 × 1024
8	Attention_2	Grouped Query Attn	1 × 32768 × 1024
9	Add_2	Add	1 × 32768 × 1024
10	LayerNorm_2_1	LayerNorm	1 × 32768 × 1024
11	FFN_2	SwiGLU	1 × 32768 × 1024
12	Attention_3	Grouped Query Attn	1 × 32768 × 1024
13	Add_3	Add	1 × 32768 × 1024
14	LayerNorm_3_1	LayerNorm	1 × 32768 × 1024
15	FFN_3	SwiGLU	1 × 32768 × 1024
16	Attention_4	Grouped Query Attn	1 × 32768 × 1024
17	Add_4	Add	1 × 32768 × 1024
18	LayerNorm_4_1	LayerNorm	1 × 32768 × 1024
19	FFN_4	SwiGLU	1 × 32768 × 1024
20	Attention_5	Grouped Query Attn	1 × 32768 × 1024
21	Add_5	Add	1 × 32768 × 1024
22	LayerNorm_5_1	LayerNorm	1 × 32768 × 1024
23	FFN_5	SwiGLU	1 × 32768 × 1024
24	Attention_6	Grouped Query Attn	1 × 32768 × 1024
25	Add_6	Add	1 × 32768 × 1024
26	LayerNorm_6_1	LayerNorm	1 × 32768 × 1024
27	FFN_6	SwiGLU	1 × 32768 × 1024
28	Attention_7	Grouped Query Attn	1 × 32768 × 1024
29	Add_7	Add	1 × 32768 × 1024
30	LayerNorm_7_1	LayerNorm	1 × 32768 × 1024
31	FFN_7	SwiGLU	1 × 32768 × 1024
32	Attention_8	Grouped Query Attn	1 × 32768 × 1024
33	Add_8	Add	1 × 32768 × 1024
34	LayerNorm_8_1	LayerNorm	1 × 32768 × 1024
35	FFN_8	SwiGLU	1 × 32768 × 1024
36	Attention_9	Grouped Query Attn	1 × 32768 × 1024
37	Add_9	Add	1 × 32768 × 1024
38	LayerNorm_9_1	LayerNorm	1 × 32768 × 1024
39	FFN_9	SwiGLU	1 × 32768 × 1024
40	Attention_10	Grouped Query Attn	1 × 32768 × 1024
41	Add_10	Add	1 × 32768 × 1024
42	LayerNorm_10_1	LayerNorm	1 × 32768 × 1024
43	FFN_10	SwiGLU	1 × 32768 × 1024
44	Attention_11	Grouped Query Attn	1 × 32768 × 1024
45	Add_11	Add	1 × 32768 × 1024
46	LayerNorm_11_1	LayerNorm	1 × 32768 × 1024
47	FFN_11	SwiGLU	1 × 32768 × 1024
48	Attention_12	Grouped Query Attn	1 × 32768 × 1024
49	Add_12	Add	1 × 32768 × 1024
50	LayerNorm_12_1	LayerNorm	1 × 32768 × 1024
51	FFN_12	SwiGLU	1 × 32768 × 1024
52	Attention_13	Grouped Query Attn	1 × 32768 × 1024
53	Add_13	Add	1 × 32768 × 1024
54	LayerNorm_13_1	LayerNorm	1 × 32768 × 1024
55	FFN_13	SwiGLU	1 × 32768 × 1024
56	Attention_14	Grouped Query Attn	1 × 32768 × 1024
57	Add_14	Add	1 × 32768 × 1024
58	LayerNorm_14_1	LayerNorm	1 × 32768 × 1024
59	FFN_14	SwiGLU	1 × 32768 × 1024
60	Attention_15	Grouped Query Attn	1 × 32768 × 1024
61	Add_15	Add	1 × 32768 × 1024
62	LayerNorm_15_1	LayerNorm	1 × 32768 × 1024
63	FFN_15	SwiGLU	1 × 32768 × 1024
64	Attention_16	Grouped Query Attn	1 × 32768 × 1024
65	Add_16	Add	1 × 32768 × 1024
66	LayerNorm_16_1	LayerNorm	1 × 32768 × 1024
67	FFN_16	SwiGLU	1 × 32768 × 1024
68	Attention_17	Grouped Query Attn	1 × 32768 × 1024
69	Add_17	Add	1 × 32768 × 1024
70	LayerNorm_17_1	LayerNorm	1 × 32768 × 1024
71	FFN_17	SwiGLU	1 × 32768 × 1024
72	Attention_18	Grouped Query Attn	1 × 32768 × 1024
73	Add_18	Add	1 × 32768 × 1024
74	LayerNorm_18_1	LayerNorm	1 × 32768 × 1024
75	FFN_18	SwiGLU	1 × 32768 × 1024
76	Attention_19	Grouped Query Attn	1 × 32768 × 1024
77	Add_19	Add	1 × 32768 × 1024
78	LayerNorm_19_1	LayerNorm	1 × 32768 × 1024
79	FFN_19	SwiGLU	1 × 32768 × 1024
80	Attention_20	Grouped Query Attn	1 × 32768 × 1024
81	Add_20	Add	1 × 32768 × 1024
82	LayerNorm_20_1	LayerNorm	1 × 32768 × 1024
83	FFN_20	SwiGLU	1 × 32768 × 1024
84	Attention_21	Grouped Query Attn	1 × 32768 × 1024
85	Add_21	Add	1 × 32768 × 1024
86	LayerNorm_21_1	LayerNorm	1 × 32768 × 1024
87	FFN_21	SwiGLU	1 × 32768 × 1024
88	Attention_22	Grouped Query Attn	1 × 32768 × 1024
89	Add_22	Add	1 × 32768 × 1024
90	LayerNorm_22_1	LayerNorm	1 × 32768 × 1024
91	FFN_22	SwiGLU	1 × 32768 × 1024
92	Attention_23	Grouped Query Attn	1 × 32768 × 1024
93	Add_23	Add	1 × 32768 × 1024
94	LayerNorm_23_1	LayerNorm	1 × 32768 × 1024
95	FFN_23	SwiGLU	1 × 32768 × 1024
96	Attention_24	Grouped Query Attn	1 × 32768 × 1024
97	Add_24	Add	1 × 32768 × 1024
98	LayerNorm_24_1	LayerNorm	1 × 32768 × 1024
99	FFN_24	SwiGLU	1 × 32768 × 1024
100	Attention_25	Grouped Query Attn	1 × 32768 × 1024
101	Add_25	Add	1 × 32768 × 1024
102	LayerNorm_25_1	LayerNorm	1 × 32768 × 1024
103	FFN_25	SwiGLU	1 × 32768 × 1024
104	Attention_26	Grouped Query Attn	1 × 32768 × 1024
105	Add_26	Add	1 × 32768 × 1024
106	LayerNorm_26_1	LayerNorm	1 × 32768 × 1024
107	FFN_26	SwiGLU	1 × 32768 × 1024
108	Attention_27	Grouped Query Attn	1 × 32768 × 1024
109	Add_27	Add	1 × 32768 × 1024
110	LayerNorm_27_1	LayerNorm	1 × 32768 × 1024
111	FFN_27	SwiGLU	1 × 32768 × 1024
112	Attention_28	Grouped Query Attn	1 × 32768 × 1024
113	Add_28	Add	1 × 32768 × 1024
114	LayerNorm_28_1	LayerNorm	1 × 32768 × 1024
115	FFN_28	SwiGLU	1 × 32768 × 1024
116	Output	Output	1 × 32768 × 1024

What the verifier says

infoAt 28 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 jinaai/jina-embeddings-v5-text-small: 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 jinaai/jina-embeddings-v5-text-small --plan --share

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

Other jina_embeddings_v5 checkpoints

jina-embeddings-v5-text-nano

212M derived · +0.02% against the checkpoint