Models / modernbert

ModernBERT-base: GPU memory to fine-tune

Fine-tuning ModernBERT-base at batch 1 and 2048 tokens takes about 3.72 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 1.24 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. 4.8M 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

149M

149,081,088 parameters

In the published checkpoint

150M

149,655,232 scalars · safetensors.total, read 2025-01-15

Delta

-0.38%

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

$11.15

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

3.72 GB

LoRA r=16, batch 1 at 2048 tokens

1.24 GB

Serve: bf16 weights

0.30 GB

Serve: KV cache, 4096 tokens

0.28 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	batch 12	batch 15	batch 49
L4 24 GB	batch 20	batch 23	batch 74
A100 40 GB	batch 37	batch 40	batch 124
A100 / H100 80 GB	batch 77	batch 80	batch 250
H200 141 GB	batch 139	batch 142	batch 441

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 8192), 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

92 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 8192 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 8192
2	Embedding	Embedding	1 × 8192 × 768
3	RoPE	RoPE	1 × 8192 × 768
4	Attention_1	Multi-Head Attention	1 × 8192 × 768
5	Add_1	Add	1 × 8192 × 768
6	LayerNorm_1_1	LayerNorm	1 × 8192 × 768
7	FFN_1	GeGLU	1 × 8192 × 768
8	Attention_2	Multi-Head Attention	1 × 8192 × 768
9	Add_2	Add	1 × 8192 × 768
10	LayerNorm_2_1	LayerNorm	1 × 8192 × 768
11	FFN_2	GeGLU	1 × 8192 × 768
12	Attention_3	Multi-Head Attention	1 × 8192 × 768
13	Add_3	Add	1 × 8192 × 768
14	LayerNorm_3_1	LayerNorm	1 × 8192 × 768
15	FFN_3	GeGLU	1 × 8192 × 768
16	Attention_4	Multi-Head Attention	1 × 8192 × 768
17	Add_4	Add	1 × 8192 × 768
18	LayerNorm_4_1	LayerNorm	1 × 8192 × 768
19	FFN_4	GeGLU	1 × 8192 × 768
20	Attention_5	Multi-Head Attention	1 × 8192 × 768
21	Add_5	Add	1 × 8192 × 768
22	LayerNorm_5_1	LayerNorm	1 × 8192 × 768
23	FFN_5	GeGLU	1 × 8192 × 768
24	Attention_6	Multi-Head Attention	1 × 8192 × 768
25	Add_6	Add	1 × 8192 × 768
26	LayerNorm_6_1	LayerNorm	1 × 8192 × 768
27	FFN_6	GeGLU	1 × 8192 × 768
28	Attention_7	Multi-Head Attention	1 × 8192 × 768
29	Add_7	Add	1 × 8192 × 768
30	LayerNorm_7_1	LayerNorm	1 × 8192 × 768
31	FFN_7	GeGLU	1 × 8192 × 768
32	Attention_8	Multi-Head Attention	1 × 8192 × 768
33	Add_8	Add	1 × 8192 × 768
34	LayerNorm_8_1	LayerNorm	1 × 8192 × 768
35	FFN_8	GeGLU	1 × 8192 × 768
36	Attention_9	Multi-Head Attention	1 × 8192 × 768
37	Add_9	Add	1 × 8192 × 768
38	LayerNorm_9_1	LayerNorm	1 × 8192 × 768
39	FFN_9	GeGLU	1 × 8192 × 768
40	Attention_10	Multi-Head Attention	1 × 8192 × 768
41	Add_10	Add	1 × 8192 × 768
42	LayerNorm_10_1	LayerNorm	1 × 8192 × 768
43	FFN_10	GeGLU	1 × 8192 × 768
44	Attention_11	Multi-Head Attention	1 × 8192 × 768
45	Add_11	Add	1 × 8192 × 768
46	LayerNorm_11_1	LayerNorm	1 × 8192 × 768
47	FFN_11	GeGLU	1 × 8192 × 768
48	Attention_12	Multi-Head Attention	1 × 8192 × 768
49	Add_12	Add	1 × 8192 × 768
50	LayerNorm_12_1	LayerNorm	1 × 8192 × 768
51	FFN_12	GeGLU	1 × 8192 × 768
52	Attention_13	Multi-Head Attention	1 × 8192 × 768
53	Add_13	Add	1 × 8192 × 768
54	LayerNorm_13_1	LayerNorm	1 × 8192 × 768
55	FFN_13	GeGLU	1 × 8192 × 768
56	Attention_14	Multi-Head Attention	1 × 8192 × 768
57	Add_14	Add	1 × 8192 × 768
58	LayerNorm_14_1	LayerNorm	1 × 8192 × 768
59	FFN_14	GeGLU	1 × 8192 × 768
60	Attention_15	Multi-Head Attention	1 × 8192 × 768
61	Add_15	Add	1 × 8192 × 768
62	LayerNorm_15_1	LayerNorm	1 × 8192 × 768
63	FFN_15	GeGLU	1 × 8192 × 768
64	Attention_16	Multi-Head Attention	1 × 8192 × 768
65	Add_16	Add	1 × 8192 × 768
66	LayerNorm_16_1	LayerNorm	1 × 8192 × 768
67	FFN_16	GeGLU	1 × 8192 × 768
68	Attention_17	Multi-Head Attention	1 × 8192 × 768
69	Add_17	Add	1 × 8192 × 768
70	LayerNorm_17_1	LayerNorm	1 × 8192 × 768
71	FFN_17	GeGLU	1 × 8192 × 768
72	Attention_18	Multi-Head Attention	1 × 8192 × 768
73	Add_18	Add	1 × 8192 × 768
74	LayerNorm_18_1	LayerNorm	1 × 8192 × 768
75	FFN_18	GeGLU	1 × 8192 × 768
76	Attention_19	Multi-Head Attention	1 × 8192 × 768
77	Add_19	Add	1 × 8192 × 768
78	LayerNorm_19_1	LayerNorm	1 × 8192 × 768
79	FFN_19	GeGLU	1 × 8192 × 768
80	Attention_20	Multi-Head Attention	1 × 8192 × 768
81	Add_20	Add	1 × 8192 × 768
82	LayerNorm_20_1	LayerNorm	1 × 8192 × 768
83	FFN_20	GeGLU	1 × 8192 × 768
84	Attention_21	Multi-Head Attention	1 × 8192 × 768
85	Add_21	Add	1 × 8192 × 768
86	LayerNorm_21_1	LayerNorm	1 × 8192 × 768
87	FFN_21	GeGLU	1 × 8192 × 768
88	Attention_22	Multi-Head Attention	1 × 8192 × 768
89	Add_22	Add	1 × 8192 × 768
90	LayerNorm_22_1	LayerNorm	1 × 8192 × 768
91	FFN_22	GeGLU	1 × 8192 × 768
92	Output	Output	1 × 8192 × 768

What the verifier says

infoAt 22 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 answerdotai/ModernBERT-base: 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 answerdotai/ModernBERT-base --plan --share

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

Other modernbert checkpoints

granite-embedding-small-english-r2

47.7M derived · +0.04% against the checkpoint

gte-reranker-modernbert-base

149M derived · -0.35% against the checkpoint