Models / qwen3

voyage-4-nano: GPU memory to fine-tune

Fine-tuning voyage-4-nano at batch 1 and 2048 tokens takes about 7.77 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 1.93 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. 292K 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

344M

344,350,720 parameters

In the published checkpoint

346M

346,451,968 scalars · safetensors.total, read 2026-03-02

Delta

-0.61%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `modeling_qwen3_bidirectional.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

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$18.37

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

7.77 GB

LoRA r=16, batch 1 at 2048 tokens

1.93 GB

Serve: bf16 weights

0.69 GB

Serve: KV cache, 4096 tokens

0.20 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	batch 6	batch 11	batch 65
L4 24 GB	batch 12	batch 17	batch 100
A100 40 GB	batch 24	batch 29	batch 169
A100 / H100 80 GB	batch 54	batch 59	batch 342
H200 141 GB	batch 100	batch 105	batch 605

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 40960), 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

76 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 40960 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 40960
2	Embedding	Embedding	1 × 40960 × 1024
3	RoPE	RoPE	1 × 40960 × 1024
4	RMSNorm_1_1	RMSNorm	1 × 40960 × 1024
5	Attention_1	Grouped Query Attn	1 × 40960 × 1024
6	Add_1_attn	Add	1 × 40960 × 1024
7	RMSNorm_1_2	RMSNorm	1 × 40960 × 1024
8	FFN_1	SwiGLU	1 × 40960 × 1024
9	Add_1_ffn	Add	1 × 40960 × 1024
10	RMSNorm_2_1	RMSNorm	1 × 40960 × 1024
11	Attention_2	Grouped Query Attn	1 × 40960 × 1024
12	Add_2_attn	Add	1 × 40960 × 1024
13	RMSNorm_2_2	RMSNorm	1 × 40960 × 1024
14	FFN_2	SwiGLU	1 × 40960 × 1024
15	Add_2_ffn	Add	1 × 40960 × 1024
16	RMSNorm_3_1	RMSNorm	1 × 40960 × 1024
17	Attention_3	Grouped Query Attn	1 × 40960 × 1024
18	Add_3_attn	Add	1 × 40960 × 1024
19	RMSNorm_3_2	RMSNorm	1 × 40960 × 1024
20	FFN_3	SwiGLU	1 × 40960 × 1024
21	Add_3_ffn	Add	1 × 40960 × 1024
22	RMSNorm_4_1	RMSNorm	1 × 40960 × 1024
23	Attention_4	Grouped Query Attn	1 × 40960 × 1024
24	Add_4_attn	Add	1 × 40960 × 1024
25	RMSNorm_4_2	RMSNorm	1 × 40960 × 1024
26	FFN_4	SwiGLU	1 × 40960 × 1024
27	Add_4_ffn	Add	1 × 40960 × 1024
28	RMSNorm_5_1	RMSNorm	1 × 40960 × 1024
29	Attention_5	Grouped Query Attn	1 × 40960 × 1024
30	Add_5_attn	Add	1 × 40960 × 1024
31	RMSNorm_5_2	RMSNorm	1 × 40960 × 1024
32	FFN_5	SwiGLU	1 × 40960 × 1024
33	Add_5_ffn	Add	1 × 40960 × 1024
34	RMSNorm_6_1	RMSNorm	1 × 40960 × 1024
35	Attention_6	Grouped Query Attn	1 × 40960 × 1024
36	Add_6_attn	Add	1 × 40960 × 1024
37	RMSNorm_6_2	RMSNorm	1 × 40960 × 1024
38	FFN_6	SwiGLU	1 × 40960 × 1024
39	Add_6_ffn	Add	1 × 40960 × 1024
40	RMSNorm_7_1	RMSNorm	1 × 40960 × 1024
41	Attention_7	Grouped Query Attn	1 × 40960 × 1024
42	Add_7_attn	Add	1 × 40960 × 1024
43	RMSNorm_7_2	RMSNorm	1 × 40960 × 1024
44	FFN_7	SwiGLU	1 × 40960 × 1024
45	Add_7_ffn	Add	1 × 40960 × 1024
46	RMSNorm_8_1	RMSNorm	1 × 40960 × 1024
47	Attention_8	Grouped Query Attn	1 × 40960 × 1024
48	Add_8_attn	Add	1 × 40960 × 1024
49	RMSNorm_8_2	RMSNorm	1 × 40960 × 1024
50	FFN_8	SwiGLU	1 × 40960 × 1024
51	Add_8_ffn	Add	1 × 40960 × 1024
52	RMSNorm_9_1	RMSNorm	1 × 40960 × 1024
53	Attention_9	Grouped Query Attn	1 × 40960 × 1024
54	Add_9_attn	Add	1 × 40960 × 1024
55	RMSNorm_9_2	RMSNorm	1 × 40960 × 1024
56	FFN_9	SwiGLU	1 × 40960 × 1024
57	Add_9_ffn	Add	1 × 40960 × 1024
58	RMSNorm_10_1	RMSNorm	1 × 40960 × 1024
59	Attention_10	Grouped Query Attn	1 × 40960 × 1024
60	Add_10_attn	Add	1 × 40960 × 1024
61	RMSNorm_10_2	RMSNorm	1 × 40960 × 1024
62	FFN_10	SwiGLU	1 × 40960 × 1024
63	Add_10_ffn	Add	1 × 40960 × 1024
64	RMSNorm_11_1	RMSNorm	1 × 40960 × 1024
65	Attention_11	Grouped Query Attn	1 × 40960 × 1024
66	Add_11_attn	Add	1 × 40960 × 1024
67	RMSNorm_11_2	RMSNorm	1 × 40960 × 1024
68	FFN_11	SwiGLU	1 × 40960 × 1024
69	Add_11_ffn	Add	1 × 40960 × 1024
70	RMSNorm_12_1	RMSNorm	1 × 40960 × 1024
71	Attention_12	Grouped Query Attn	1 × 40960 × 1024
72	Add_12_attn	Add	1 × 40960 × 1024
73	RMSNorm_12_2	RMSNorm	1 × 40960 × 1024
74	FFN_12	SwiGLU	1 × 40960 × 1024
75	Add_12_ffn	Add	1 × 40960 × 1024
76	Output	Output	1 × 40960 × 1024

What the verifier says

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 voyageai/voyage-4-nano: 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 voyageai/voyage-4-nano --plan --share

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

Other qwen3 checkpoints

jina-reranker-v3

596M derived · -0.13% against the checkpoint

Kimi-K3-DSpark

4.26B derived · +89.3% against the checkpoint

Qwen3-0.6B

596M derived · -0.00% against the checkpoint

Qwen3-1.7B

1.72B derived · -0.00% against the checkpoint