Models / tipsv2

tipsv2-so400m14: GPU memory to fine-tune

Fine-tuning tipsv2-so400m14 at batch 1 and 64 tokens takes about 8.67 GB of GPU memory for a full fine-tune (fits one T4 16 GB), and 1.13 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. 270K 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

448M

448,342,128 parameters

In the published checkpoint

862M

861,726,816 scalars · safetensors.total, read 2026-08-17

Delta

-48.0%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_tips.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, 64 tokens over 50,000 samples (assumed; no dataset attached).

Layers

110

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$1.03

Memory: what fits on which card

Largest batch per card. Training is sized at 64 tokens per sample, serving at 64 tokens per sequence.

Full fine-tune, batch 1 at 64 tokens

8.67 GB

LoRA r=16, batch 1 at 64 tokens

1.13 GB

Serve: bf16 weights

0.90 GB

Serve: KV cache, 64 tokens

8.0 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 64 tokens
T4 16 GB	batch 81	batch 197	batch 1634
L4 24 GB	batch 188	batch 304	batch 2508
A100 40 GB	batch 401	batch 517	batch 4255
A100 / H100 80 GB	batch 935	batch 1051	batch 8623
H200 141 GB	batch 1749	batch 1865	batch 15285

Training: 64 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 64 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

112 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 64 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 64
2	Embedding	Embedding	1 × 64 × 1152
3	Positional_Embedding	Learned Pos Embed	1 × 64 × 1152
4	Attention_1	Multi-Head Attention	1 × 64 × 1152
5	Add_1	Add	1 × 64 × 1152
6	LayerNorm_1_1	LayerNorm	1 × 64 × 1152
7	FFN_1	Feed Forward	1 × 64 × 1152
8	Attention_2	Multi-Head Attention	1 × 64 × 1152
9	Add_2	Add	1 × 64 × 1152
10	LayerNorm_2_1	LayerNorm	1 × 64 × 1152
11	FFN_2	Feed Forward	1 × 64 × 1152
12	Attention_3	Multi-Head Attention	1 × 64 × 1152
13	Add_3	Add	1 × 64 × 1152
14	LayerNorm_3_1	LayerNorm	1 × 64 × 1152
15	FFN_3	Feed Forward	1 × 64 × 1152
16	Attention_4	Multi-Head Attention	1 × 64 × 1152
17	Add_4	Add	1 × 64 × 1152
18	LayerNorm_4_1	LayerNorm	1 × 64 × 1152
19	FFN_4	Feed Forward	1 × 64 × 1152
20	Attention_5	Multi-Head Attention	1 × 64 × 1152
21	Add_5	Add	1 × 64 × 1152
22	LayerNorm_5_1	LayerNorm	1 × 64 × 1152
23	FFN_5	Feed Forward	1 × 64 × 1152
24	Attention_6	Multi-Head Attention	1 × 64 × 1152
25	Add_6	Add	1 × 64 × 1152
26	LayerNorm_6_1	LayerNorm	1 × 64 × 1152
27	FFN_6	Feed Forward	1 × 64 × 1152
28	Attention_7	Multi-Head Attention	1 × 64 × 1152
29	Add_7	Add	1 × 64 × 1152
30	LayerNorm_7_1	LayerNorm	1 × 64 × 1152
31	FFN_7	Feed Forward	1 × 64 × 1152
32	Attention_8	Multi-Head Attention	1 × 64 × 1152
33	Add_8	Add	1 × 64 × 1152
34	LayerNorm_8_1	LayerNorm	1 × 64 × 1152
35	FFN_8	Feed Forward	1 × 64 × 1152
36	Attention_9	Multi-Head Attention	1 × 64 × 1152
37	Add_9	Add	1 × 64 × 1152
38	LayerNorm_9_1	LayerNorm	1 × 64 × 1152
39	FFN_9	Feed Forward	1 × 64 × 1152
40	Attention_10	Multi-Head Attention	1 × 64 × 1152
41	Add_10	Add	1 × 64 × 1152
42	LayerNorm_10_1	LayerNorm	1 × 64 × 1152
43	FFN_10	Feed Forward	1 × 64 × 1152
44	Attention_11	Multi-Head Attention	1 × 64 × 1152
45	Add_11	Add	1 × 64 × 1152
46	LayerNorm_11_1	LayerNorm	1 × 64 × 1152
47	FFN_11	Feed Forward	1 × 64 × 1152
48	Attention_12	Multi-Head Attention	1 × 64 × 1152
49	Add_12	Add	1 × 64 × 1152
50	LayerNorm_12_1	LayerNorm	1 × 64 × 1152
51	FFN_12	Feed Forward	1 × 64 × 1152
52	Attention_13	Multi-Head Attention	1 × 64 × 1152
53	Add_13	Add	1 × 64 × 1152
54	LayerNorm_13_1	LayerNorm	1 × 64 × 1152
55	FFN_13	Feed Forward	1 × 64 × 1152
56	Attention_14	Multi-Head Attention	1 × 64 × 1152
57	Add_14	Add	1 × 64 × 1152
58	LayerNorm_14_1	LayerNorm	1 × 64 × 1152
59	FFN_14	Feed Forward	1 × 64 × 1152
60	Attention_15	Multi-Head Attention	1 × 64 × 1152
61	Add_15	Add	1 × 64 × 1152
62	LayerNorm_15_1	LayerNorm	1 × 64 × 1152
63	FFN_15	Feed Forward	1 × 64 × 1152
64	Attention_16	Multi-Head Attention	1 × 64 × 1152
65	Add_16	Add	1 × 64 × 1152
66	LayerNorm_16_1	LayerNorm	1 × 64 × 1152
67	FFN_16	Feed Forward	1 × 64 × 1152
68	Attention_17	Multi-Head Attention	1 × 64 × 1152
69	Add_17	Add	1 × 64 × 1152
70	LayerNorm_17_1	LayerNorm	1 × 64 × 1152
71	FFN_17	Feed Forward	1 × 64 × 1152
72	Attention_18	Multi-Head Attention	1 × 64 × 1152
73	Add_18	Add	1 × 64 × 1152
74	LayerNorm_18_1	LayerNorm	1 × 64 × 1152
75	FFN_18	Feed Forward	1 × 64 × 1152
76	Attention_19	Multi-Head Attention	1 × 64 × 1152
77	Add_19	Add	1 × 64 × 1152
78	LayerNorm_19_1	LayerNorm	1 × 64 × 1152
79	FFN_19	Feed Forward	1 × 64 × 1152
80	Attention_20	Multi-Head Attention	1 × 64 × 1152
81	Add_20	Add	1 × 64 × 1152
82	LayerNorm_20_1	LayerNorm	1 × 64 × 1152
83	FFN_20	Feed Forward	1 × 64 × 1152
84	Attention_21	Multi-Head Attention	1 × 64 × 1152
85	Add_21	Add	1 × 64 × 1152
86	LayerNorm_21_1	LayerNorm	1 × 64 × 1152
87	FFN_21	Feed Forward	1 × 64 × 1152
88	Attention_22	Multi-Head Attention	1 × 64 × 1152
89	Add_22	Add	1 × 64 × 1152
90	LayerNorm_22_1	LayerNorm	1 × 64 × 1152
91	FFN_22	Feed Forward	1 × 64 × 1152
92	Attention_23	Multi-Head Attention	1 × 64 × 1152
93	Add_23	Add	1 × 64 × 1152
94	LayerNorm_23_1	LayerNorm	1 × 64 × 1152
95	FFN_23	Feed Forward	1 × 64 × 1152
96	Attention_24	Multi-Head Attention	1 × 64 × 1152
97	Add_24	Add	1 × 64 × 1152
98	LayerNorm_24_1	LayerNorm	1 × 64 × 1152
99	FFN_24	Feed Forward	1 × 64 × 1152
100	Attention_25	Multi-Head Attention	1 × 64 × 1152
101	Add_25	Add	1 × 64 × 1152
102	LayerNorm_25_1	LayerNorm	1 × 64 × 1152
103	FFN_25	Feed Forward	1 × 64 × 1152
104	Attention_26	Multi-Head Attention	1 × 64 × 1152
105	Add_26	Add	1 × 64 × 1152
106	LayerNorm_26_1	LayerNorm	1 × 64 × 1152
107	FFN_26	Feed Forward	1 × 64 × 1152
108	Attention_27	Multi-Head Attention	1 × 64 × 1152
109	Add_27	Add	1 × 64 × 1152
110	LayerNorm_27_1	LayerNorm	1 × 64 × 1152
111	FFN_27	Feed Forward	1 × 64 × 1152
112	Output	Output	1 × 64 × 1152

What the verifier says

infoAt 27 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 google/tipsv2-so400m14: 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 google/tipsv2-so400m14 --plan --share

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

Other tipsv2 checkpoints

tipsv2-b14

110M derived · -44.0% against the checkpoint