Models / xlm-roberta

multilingual-e5-large: GPU memory to fine-tune

Fine-tuning multilingual-e5-large at batch 1 and 514 tokens takes about 11.2 GB of GPU memory for a full fine-tune (borderline on one T4 16 GB), and 1.69 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. 7.0M 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

559M

558,788,608 parameters

In the published checkpoint

560M

559,890,946 scalars · safetensors.total, read 2026-04-02

Delta

-0.20%

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 514 tokens over 50,000 samples (assumed; no dataset attached).

Layers

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$6.10

Memory: what fits on which card

Largest batch per card. Training is sized at 514 tokens per sample, serving at 514 tokens per sequence.

Full fine-tune, batch 1 at 514 tokens

11.2 GB

LoRA r=16, batch 1 at 514 tokens

1.69 GB

Serve: bf16 weights

1.12 GB

Serve: KV cache, 514 tokens

51 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 514 tokens
T4 16 GB	batch 7, borderline	batch 29	batch 253
L4 24 GB	batch 23	batch 45	batch 390
A100 40 GB	batch 55	batch 77	batch 666
A100 / H100 80 GB	batch 136	batch 158	batch 1354
H200 141 GB	batch 259	batch 281	batch 2404

Training: 514 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. "Borderline" means the card holds batch 1 on the estimate but not on the conservative peak (the estimate x 1.43, at or above the measured peak on 95% of held-out runs): measure one step before reserving it. Serving: bf16 weights plus the KV cache of 514 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

100 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 514 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 514
2	Embedding	Embedding	1 × 514 × 1024
3	Positional_Embedding	Learned Pos Embed	1 × 514 × 1024
4	Attention_1	Multi-Head Attention	1 × 514 × 1024
5	Add_1	Add	1 × 514 × 1024
6	LayerNorm_1_1	LayerNorm	1 × 514 × 1024
7	FFN_1	Feed Forward	1 × 514 × 1024
8	Attention_2	Multi-Head Attention	1 × 514 × 1024
9	Add_2	Add	1 × 514 × 1024
10	LayerNorm_2_1	LayerNorm	1 × 514 × 1024
11	FFN_2	Feed Forward	1 × 514 × 1024
12	Attention_3	Multi-Head Attention	1 × 514 × 1024
13	Add_3	Add	1 × 514 × 1024
14	LayerNorm_3_1	LayerNorm	1 × 514 × 1024
15	FFN_3	Feed Forward	1 × 514 × 1024
16	Attention_4	Multi-Head Attention	1 × 514 × 1024
17	Add_4	Add	1 × 514 × 1024
18	LayerNorm_4_1	LayerNorm	1 × 514 × 1024
19	FFN_4	Feed Forward	1 × 514 × 1024
20	Attention_5	Multi-Head Attention	1 × 514 × 1024
21	Add_5	Add	1 × 514 × 1024
22	LayerNorm_5_1	LayerNorm	1 × 514 × 1024
23	FFN_5	Feed Forward	1 × 514 × 1024
24	Attention_6	Multi-Head Attention	1 × 514 × 1024
25	Add_6	Add	1 × 514 × 1024
26	LayerNorm_6_1	LayerNorm	1 × 514 × 1024
27	FFN_6	Feed Forward	1 × 514 × 1024
28	Attention_7	Multi-Head Attention	1 × 514 × 1024
29	Add_7	Add	1 × 514 × 1024
30	LayerNorm_7_1	LayerNorm	1 × 514 × 1024
31	FFN_7	Feed Forward	1 × 514 × 1024
32	Attention_8	Multi-Head Attention	1 × 514 × 1024
33	Add_8	Add	1 × 514 × 1024
34	LayerNorm_8_1	LayerNorm	1 × 514 × 1024
35	FFN_8	Feed Forward	1 × 514 × 1024
36	Attention_9	Multi-Head Attention	1 × 514 × 1024
37	Add_9	Add	1 × 514 × 1024
38	LayerNorm_9_1	LayerNorm	1 × 514 × 1024
39	FFN_9	Feed Forward	1 × 514 × 1024
40	Attention_10	Multi-Head Attention	1 × 514 × 1024
41	Add_10	Add	1 × 514 × 1024
42	LayerNorm_10_1	LayerNorm	1 × 514 × 1024
43	FFN_10	Feed Forward	1 × 514 × 1024
44	Attention_11	Multi-Head Attention	1 × 514 × 1024
45	Add_11	Add	1 × 514 × 1024
46	LayerNorm_11_1	LayerNorm	1 × 514 × 1024
47	FFN_11	Feed Forward	1 × 514 × 1024
48	Attention_12	Multi-Head Attention	1 × 514 × 1024
49	Add_12	Add	1 × 514 × 1024
50	LayerNorm_12_1	LayerNorm	1 × 514 × 1024
51	FFN_12	Feed Forward	1 × 514 × 1024
52	Attention_13	Multi-Head Attention	1 × 514 × 1024
53	Add_13	Add	1 × 514 × 1024
54	LayerNorm_13_1	LayerNorm	1 × 514 × 1024
55	FFN_13	Feed Forward	1 × 514 × 1024
56	Attention_14	Multi-Head Attention	1 × 514 × 1024
57	Add_14	Add	1 × 514 × 1024
58	LayerNorm_14_1	LayerNorm	1 × 514 × 1024
59	FFN_14	Feed Forward	1 × 514 × 1024
60	Attention_15	Multi-Head Attention	1 × 514 × 1024
61	Add_15	Add	1 × 514 × 1024
62	LayerNorm_15_1	LayerNorm	1 × 514 × 1024
63	FFN_15	Feed Forward	1 × 514 × 1024
64	Attention_16	Multi-Head Attention	1 × 514 × 1024
65	Add_16	Add	1 × 514 × 1024
66	LayerNorm_16_1	LayerNorm	1 × 514 × 1024
67	FFN_16	Feed Forward	1 × 514 × 1024
68	Attention_17	Multi-Head Attention	1 × 514 × 1024
69	Add_17	Add	1 × 514 × 1024
70	LayerNorm_17_1	LayerNorm	1 × 514 × 1024
71	FFN_17	Feed Forward	1 × 514 × 1024
72	Attention_18	Multi-Head Attention	1 × 514 × 1024
73	Add_18	Add	1 × 514 × 1024
74	LayerNorm_18_1	LayerNorm	1 × 514 × 1024
75	FFN_18	Feed Forward	1 × 514 × 1024
76	Attention_19	Multi-Head Attention	1 × 514 × 1024
77	Add_19	Add	1 × 514 × 1024
78	LayerNorm_19_1	LayerNorm	1 × 514 × 1024
79	FFN_19	Feed Forward	1 × 514 × 1024
80	Attention_20	Multi-Head Attention	1 × 514 × 1024
81	Add_20	Add	1 × 514 × 1024
82	LayerNorm_20_1	LayerNorm	1 × 514 × 1024
83	FFN_20	Feed Forward	1 × 514 × 1024
84	Attention_21	Multi-Head Attention	1 × 514 × 1024
85	Add_21	Add	1 × 514 × 1024
86	LayerNorm_21_1	LayerNorm	1 × 514 × 1024
87	FFN_21	Feed Forward	1 × 514 × 1024
88	Attention_22	Multi-Head Attention	1 × 514 × 1024
89	Add_22	Add	1 × 514 × 1024
90	LayerNorm_22_1	LayerNorm	1 × 514 × 1024
91	FFN_22	Feed Forward	1 × 514 × 1024
92	Attention_23	Multi-Head Attention	1 × 514 × 1024
93	Add_23	Add	1 × 514 × 1024
94	LayerNorm_23_1	LayerNorm	1 × 514 × 1024
95	FFN_23	Feed Forward	1 × 514 × 1024
96	Attention_24	Multi-Head Attention	1 × 514 × 1024
97	Add_24	Add	1 × 514 × 1024
98	LayerNorm_24_1	LayerNorm	1 × 514 × 1024
99	FFN_24	Feed Forward	1 × 514 × 1024
100	Output	Output	1 × 514 × 1024

What the verifier says

infoAt 24 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 intfloat/multilingual-e5-large: 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 intfloat/multilingual-e5-large --plan --share

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

Other xlm-roberta checkpoints

bge-reranker-base

277M derived · -0.22% against the checkpoint

bge-reranker-v2-m3

567M derived · -0.19% against the checkpoint

multilingual-e5-base

277M derived · -0.22% against the checkpoint

paraphrase-multilingual-mpnet-base-v2

277M derived · -0.22% against the checkpoint