Models / bert

paraphrase-multilingual-minilm-l12-v3-mn: GPU memory to fine-tune

Fine-tuning paraphrase-multilingual-minilm-l12-v3-mn at batch 1 and 512 tokens takes about 2.34 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 0.34 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. 149K 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

117M

117,495,168 parameters

In the published checkpoint

118M

117,653,760 scalars · safetensors.total, read 2025-01-07

Delta

-0.13%

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 512 tokens over 50,000 samples (assumed; no dataset attached).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$0.49

Memory: what fits on which card

Largest batch per card. Training is sized at 512 tokens per sample, serving at 512 tokens per sequence.

Full fine-tune, batch 1 at 512 tokens

2.34 GB

LoRA r=16, batch 1 at 512 tokens

0.34 GB

Serve: bf16 weights

0.23 GB

Serve: KV cache, 512 tokens

9.4 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 512 tokens
T4 16 GB	batch 143	batch 167	batch 1449
L4 24 GB	batch 228	batch 252	batch 2186
A100 40 GB	batch 399	batch 423	batch 3660
A100 / H100 80 GB	batch 826	batch 850	batch 7346
H200 141 GB	batch 1477	batch 1501	batch 12967

Training: 512 tokens per sample (2048, or the model's context when shorter), 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 512 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

52 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 512 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 512
2	Embedding	Embedding	1 × 512 × 384
3	Positional_Embedding	Learned Pos Embed	1 × 512 × 384
4	Attention_1	Multi-Head Attention	1 × 512 × 384
5	Add_1	Add	1 × 512 × 384
6	LayerNorm_1_1	LayerNorm	1 × 512 × 384
7	FFN_1	Feed Forward	1 × 512 × 384
8	Attention_2	Multi-Head Attention	1 × 512 × 384
9	Add_2	Add	1 × 512 × 384
10	LayerNorm_2_1	LayerNorm	1 × 512 × 384
11	FFN_2	Feed Forward	1 × 512 × 384
12	Attention_3	Multi-Head Attention	1 × 512 × 384
13	Add_3	Add	1 × 512 × 384
14	LayerNorm_3_1	LayerNorm	1 × 512 × 384
15	FFN_3	Feed Forward	1 × 512 × 384
16	Attention_4	Multi-Head Attention	1 × 512 × 384
17	Add_4	Add	1 × 512 × 384
18	LayerNorm_4_1	LayerNorm	1 × 512 × 384
19	FFN_4	Feed Forward	1 × 512 × 384
20	Attention_5	Multi-Head Attention	1 × 512 × 384
21	Add_5	Add	1 × 512 × 384
22	LayerNorm_5_1	LayerNorm	1 × 512 × 384
23	FFN_5	Feed Forward	1 × 512 × 384
24	Attention_6	Multi-Head Attention	1 × 512 × 384
25	Add_6	Add	1 × 512 × 384
26	LayerNorm_6_1	LayerNorm	1 × 512 × 384
27	FFN_6	Feed Forward	1 × 512 × 384
28	Attention_7	Multi-Head Attention	1 × 512 × 384
29	Add_7	Add	1 × 512 × 384
30	LayerNorm_7_1	LayerNorm	1 × 512 × 384
31	FFN_7	Feed Forward	1 × 512 × 384
32	Attention_8	Multi-Head Attention	1 × 512 × 384
33	Add_8	Add	1 × 512 × 384
34	LayerNorm_8_1	LayerNorm	1 × 512 × 384
35	FFN_8	Feed Forward	1 × 512 × 384
36	Attention_9	Multi-Head Attention	1 × 512 × 384
37	Add_9	Add	1 × 512 × 384
38	LayerNorm_9_1	LayerNorm	1 × 512 × 384
39	FFN_9	Feed Forward	1 × 512 × 384
40	Attention_10	Multi-Head Attention	1 × 512 × 384
41	Add_10	Add	1 × 512 × 384
42	LayerNorm_10_1	LayerNorm	1 × 512 × 384
43	FFN_10	Feed Forward	1 × 512 × 384
44	Attention_11	Multi-Head Attention	1 × 512 × 384
45	Add_11	Add	1 × 512 × 384
46	LayerNorm_11_1	LayerNorm	1 × 512 × 384
47	FFN_11	Feed Forward	1 × 512 × 384
48	Attention_12	Multi-Head Attention	1 × 512 × 384
49	Add_12	Add	1 × 512 × 384
50	LayerNorm_12_1	LayerNorm	1 × 512 × 384
51	FFN_12	Feed Forward	1 × 512 × 384
52	Output	Output	1 × 512 × 384

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 gmunkhtur/paraphrase-multilingual-minilm-l12-v3-mn: 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 gmunkhtur/paraphrase-multilingual-minilm-l12-v3-mn --plan --share

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

Other bert checkpoints

all-MiniLM-L12-v2

33.2M derived · -0.48% against the checkpoint

all-MiniLM-L6-v2

22.6M derived · -0.68% against the checkpoint

bert-base-cased

108M derived · -1.13% against the checkpoint

bert-base-multilingual-uncased

167M derived · -0.78% against the checkpoint