Models / gemma2

gemma-2-9b-it: GPU memory to fine-tune

Fine-tuning gemma-2-9b-it at batch 1 and 2048 tokens takes about 192 GB of GPU memory for a full fine-tune (2 x H200 141 GB, split by layer), and 33.9 GB with LoRA r=16 on a bf16 base (borderline on one A100 40 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 711K 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

9.24B

9,241,401,344 parameters

In the published checkpoint

9.24B

9,241,705,984 scalars · safetensors.total, read 2024-08-27

Delta

-0.00%

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

254

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$665.15

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

192 GB

LoRA r=16, batch 1 at 2048 tokens

33.9 GB

Serve: bf16 weights

18.5 GB

Serve: KV cache, 4096 tokens

1.41 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
L4 24 GB	does not fit at batch 1	does not fit at batch 1	batch 1
A100 40 GB	does not fit at batch 1	batch 1, borderline	batch 11
A100 / H100 80 GB	does not fit at batch 1	batch 3	batch 36
H200 141 GB	2 cards at batch 1 (156 + 98 layers)	batch 7	batch 73

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. "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 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

256 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 × 3584
3	RoPE	RoPE	1 × 8192 × 3584
4	RMSNorm_1_1	RMSNorm	1 × 8192 × 3584
5	Attention_1	Local Attention	1 × 8192 × 3584
6	Add_1_attn	Add	1 × 8192 × 3584
7	RMSNorm_1_2	RMSNorm	1 × 8192 × 3584
8	FFN_1	SwiGLU	1 × 8192 × 3584
9	Add_1_ffn	Add	1 × 8192 × 3584
10	RMSNorm_2_1	RMSNorm	1 × 8192 × 3584
11	Attention_2	Grouped Query Attn	1 × 8192 × 3584
12	Add_2_attn	Add	1 × 8192 × 3584
13	RMSNorm_2_2	RMSNorm	1 × 8192 × 3584
14	FFN_2	SwiGLU	1 × 8192 × 3584
15	Add_2_ffn	Add	1 × 8192 × 3584
16	RMSNorm_3_1	RMSNorm	1 × 8192 × 3584
17	Attention_3	Local Attention	1 × 8192 × 3584
18	Add_3_attn	Add	1 × 8192 × 3584
19	RMSNorm_3_2	RMSNorm	1 × 8192 × 3584
20	FFN_3	SwiGLU	1 × 8192 × 3584
21	Add_3_ffn	Add	1 × 8192 × 3584
22	RMSNorm_4_1	RMSNorm	1 × 8192 × 3584
23	Attention_4	Grouped Query Attn	1 × 8192 × 3584
24	Add_4_attn	Add	1 × 8192 × 3584
25	RMSNorm_4_2	RMSNorm	1 × 8192 × 3584
26	FFN_4	SwiGLU	1 × 8192 × 3584
27	Add_4_ffn	Add	1 × 8192 × 3584
28	RMSNorm_5_1	RMSNorm	1 × 8192 × 3584
29	Attention_5	Local Attention	1 × 8192 × 3584
30	Add_5_attn	Add	1 × 8192 × 3584
31	RMSNorm_5_2	RMSNorm	1 × 8192 × 3584
32	FFN_5	SwiGLU	1 × 8192 × 3584
33	Add_5_ffn	Add	1 × 8192 × 3584
34	RMSNorm_6_1	RMSNorm	1 × 8192 × 3584
35	Attention_6	Grouped Query Attn	1 × 8192 × 3584
36	Add_6_attn	Add	1 × 8192 × 3584
37	RMSNorm_6_2	RMSNorm	1 × 8192 × 3584
38	FFN_6	SwiGLU	1 × 8192 × 3584
39	Add_6_ffn	Add	1 × 8192 × 3584
40	RMSNorm_7_1	RMSNorm	1 × 8192 × 3584
41	Attention_7	Local Attention	1 × 8192 × 3584
42	Add_7_attn	Add	1 × 8192 × 3584
43	RMSNorm_7_2	RMSNorm	1 × 8192 × 3584
44	FFN_7	SwiGLU	1 × 8192 × 3584
45	Add_7_ffn	Add	1 × 8192 × 3584
46	RMSNorm_8_1	RMSNorm	1 × 8192 × 3584
47	Attention_8	Grouped Query Attn	1 × 8192 × 3584
48	Add_8_attn	Add	1 × 8192 × 3584
49	RMSNorm_8_2	RMSNorm	1 × 8192 × 3584
50	FFN_8	SwiGLU	1 × 8192 × 3584
51	Add_8_ffn	Add	1 × 8192 × 3584
52	RMSNorm_9_1	RMSNorm	1 × 8192 × 3584
53	Attention_9	Local Attention	1 × 8192 × 3584
54	Add_9_attn	Add	1 × 8192 × 3584
55	RMSNorm_9_2	RMSNorm	1 × 8192 × 3584
56	FFN_9	SwiGLU	1 × 8192 × 3584
57	Add_9_ffn	Add	1 × 8192 × 3584
58	RMSNorm_10_1	RMSNorm	1 × 8192 × 3584
59	Attention_10	Grouped Query Attn	1 × 8192 × 3584
60	Add_10_attn	Add	1 × 8192 × 3584
61	RMSNorm_10_2	RMSNorm	1 × 8192 × 3584
62	FFN_10	SwiGLU	1 × 8192 × 3584
63	Add_10_ffn	Add	1 × 8192 × 3584
64	RMSNorm_11_1	RMSNorm	1 × 8192 × 3584
65	Attention_11	Local Attention	1 × 8192 × 3584
66	Add_11_attn	Add	1 × 8192 × 3584
67	RMSNorm_11_2	RMSNorm	1 × 8192 × 3584
68	FFN_11	SwiGLU	1 × 8192 × 3584
69	Add_11_ffn	Add	1 × 8192 × 3584
70	RMSNorm_12_1	RMSNorm	1 × 8192 × 3584
71	Attention_12	Grouped Query Attn	1 × 8192 × 3584
72	Add_12_attn	Add	1 × 8192 × 3584
73	RMSNorm_12_2	RMSNorm	1 × 8192 × 3584
74	FFN_12	SwiGLU	1 × 8192 × 3584
75	Add_12_ffn	Add	1 × 8192 × 3584
76	RMSNorm_13_1	RMSNorm	1 × 8192 × 3584
77	Attention_13	Local Attention	1 × 8192 × 3584
78	Add_13_attn	Add	1 × 8192 × 3584
79	RMSNorm_13_2	RMSNorm	1 × 8192 × 3584
80	FFN_13	SwiGLU	1 × 8192 × 3584
81	Add_13_ffn	Add	1 × 8192 × 3584
82	RMSNorm_14_1	RMSNorm	1 × 8192 × 3584
83	Attention_14	Grouped Query Attn	1 × 8192 × 3584
84	Add_14_attn	Add	1 × 8192 × 3584
85	RMSNorm_14_2	RMSNorm	1 × 8192 × 3584
86	FFN_14	SwiGLU	1 × 8192 × 3584
87	Add_14_ffn	Add	1 × 8192 × 3584
88	RMSNorm_15_1	RMSNorm	1 × 8192 × 3584
89	Attention_15	Local Attention	1 × 8192 × 3584
90	Add_15_attn	Add	1 × 8192 × 3584
91	RMSNorm_15_2	RMSNorm	1 × 8192 × 3584
92	FFN_15	SwiGLU	1 × 8192 × 3584
93	Add_15_ffn	Add	1 × 8192 × 3584
94	RMSNorm_16_1	RMSNorm	1 × 8192 × 3584
95	Attention_16	Grouped Query Attn	1 × 8192 × 3584
96	Add_16_attn	Add	1 × 8192 × 3584
97	RMSNorm_16_2	RMSNorm	1 × 8192 × 3584
98	FFN_16	SwiGLU	1 × 8192 × 3584
99	Add_16_ffn	Add	1 × 8192 × 3584
100	RMSNorm_17_1	RMSNorm	1 × 8192 × 3584
101	Attention_17	Local Attention	1 × 8192 × 3584
102	Add_17_attn	Add	1 × 8192 × 3584
103	RMSNorm_17_2	RMSNorm	1 × 8192 × 3584
104	FFN_17	SwiGLU	1 × 8192 × 3584
105	Add_17_ffn	Add	1 × 8192 × 3584
106	RMSNorm_18_1	RMSNorm	1 × 8192 × 3584
107	Attention_18	Grouped Query Attn	1 × 8192 × 3584
108	Add_18_attn	Add	1 × 8192 × 3584
109	RMSNorm_18_2	RMSNorm	1 × 8192 × 3584
110	FFN_18	SwiGLU	1 × 8192 × 3584
111	Add_18_ffn	Add	1 × 8192 × 3584
112	RMSNorm_19_1	RMSNorm	1 × 8192 × 3584
113	Attention_19	Local Attention	1 × 8192 × 3584
114	Add_19_attn	Add	1 × 8192 × 3584
115	RMSNorm_19_2	RMSNorm	1 × 8192 × 3584
116	FFN_19	SwiGLU	1 × 8192 × 3584
117	Add_19_ffn	Add	1 × 8192 × 3584
118	RMSNorm_20_1	RMSNorm	1 × 8192 × 3584
119	Attention_20	Grouped Query Attn	1 × 8192 × 3584
120	Add_20_attn	Add	1 × 8192 × 3584
121	RMSNorm_20_2	RMSNorm	1 × 8192 × 3584
122	FFN_20	SwiGLU	1 × 8192 × 3584
123	Add_20_ffn	Add	1 × 8192 × 3584
124	RMSNorm_21_1	RMSNorm	1 × 8192 × 3584
125	Attention_21	Local Attention	1 × 8192 × 3584
126	Add_21_attn	Add	1 × 8192 × 3584
127	RMSNorm_21_2	RMSNorm	1 × 8192 × 3584
128	FFN_21	SwiGLU	1 × 8192 × 3584
129	Add_21_ffn	Add	1 × 8192 × 3584
130	RMSNorm_22_1	RMSNorm	1 × 8192 × 3584
131	Attention_22	Grouped Query Attn	1 × 8192 × 3584
132	Add_22_attn	Add	1 × 8192 × 3584
133	RMSNorm_22_2	RMSNorm	1 × 8192 × 3584
134	FFN_22	SwiGLU	1 × 8192 × 3584
135	Add_22_ffn	Add	1 × 8192 × 3584
136	RMSNorm_23_1	RMSNorm	1 × 8192 × 3584
137	Attention_23	Local Attention	1 × 8192 × 3584
138	Add_23_attn	Add	1 × 8192 × 3584
139	RMSNorm_23_2	RMSNorm	1 × 8192 × 3584
140	FFN_23	SwiGLU	1 × 8192 × 3584
141	Add_23_ffn	Add	1 × 8192 × 3584
142	RMSNorm_24_1	RMSNorm	1 × 8192 × 3584
143	Attention_24	Grouped Query Attn	1 × 8192 × 3584
144	Add_24_attn	Add	1 × 8192 × 3584
145	RMSNorm_24_2	RMSNorm	1 × 8192 × 3584
146	FFN_24	SwiGLU	1 × 8192 × 3584
147	Add_24_ffn	Add	1 × 8192 × 3584
148	RMSNorm_25_1	RMSNorm	1 × 8192 × 3584
149	Attention_25	Local Attention	1 × 8192 × 3584
150	Add_25_attn	Add	1 × 8192 × 3584
151	RMSNorm_25_2	RMSNorm	1 × 8192 × 3584
152	FFN_25	SwiGLU	1 × 8192 × 3584
153	Add_25_ffn	Add	1 × 8192 × 3584
154	RMSNorm_26_1	RMSNorm	1 × 8192 × 3584
155	Attention_26	Grouped Query Attn	1 × 8192 × 3584
156	Add_26_attn	Add	1 × 8192 × 3584
157	RMSNorm_26_2	RMSNorm	1 × 8192 × 3584
158	FFN_26	SwiGLU	1 × 8192 × 3584
159	Add_26_ffn	Add	1 × 8192 × 3584
160	RMSNorm_27_1	RMSNorm	1 × 8192 × 3584
161	Attention_27	Local Attention	1 × 8192 × 3584
162	Add_27_attn	Add	1 × 8192 × 3584
163	RMSNorm_27_2	RMSNorm	1 × 8192 × 3584
164	FFN_27	SwiGLU	1 × 8192 × 3584
165	Add_27_ffn	Add	1 × 8192 × 3584
166	RMSNorm_28_1	RMSNorm	1 × 8192 × 3584
167	Attention_28	Grouped Query Attn	1 × 8192 × 3584
168	Add_28_attn	Add	1 × 8192 × 3584
169	RMSNorm_28_2	RMSNorm	1 × 8192 × 3584
170	FFN_28	SwiGLU	1 × 8192 × 3584
171	Add_28_ffn	Add	1 × 8192 × 3584
172	RMSNorm_29_1	RMSNorm	1 × 8192 × 3584
173	Attention_29	Local Attention	1 × 8192 × 3584
174	Add_29_attn	Add	1 × 8192 × 3584
175	RMSNorm_29_2	RMSNorm	1 × 8192 × 3584
176	FFN_29	SwiGLU	1 × 8192 × 3584
177	Add_29_ffn	Add	1 × 8192 × 3584
178	RMSNorm_30_1	RMSNorm	1 × 8192 × 3584
179	Attention_30	Grouped Query Attn	1 × 8192 × 3584
180	Add_30_attn	Add	1 × 8192 × 3584
181	RMSNorm_30_2	RMSNorm	1 × 8192 × 3584
182	FFN_30	SwiGLU	1 × 8192 × 3584
183	Add_30_ffn	Add	1 × 8192 × 3584
184	RMSNorm_31_1	RMSNorm	1 × 8192 × 3584
185	Attention_31	Local Attention	1 × 8192 × 3584
186	Add_31_attn	Add	1 × 8192 × 3584
187	RMSNorm_31_2	RMSNorm	1 × 8192 × 3584
188	FFN_31	SwiGLU	1 × 8192 × 3584
189	Add_31_ffn	Add	1 × 8192 × 3584
190	RMSNorm_32_1	RMSNorm	1 × 8192 × 3584
191	Attention_32	Grouped Query Attn	1 × 8192 × 3584
192	Add_32_attn	Add	1 × 8192 × 3584
193	RMSNorm_32_2	RMSNorm	1 × 8192 × 3584
194	FFN_32	SwiGLU	1 × 8192 × 3584
195	Add_32_ffn	Add	1 × 8192 × 3584
196	RMSNorm_33_1	RMSNorm	1 × 8192 × 3584
197	Attention_33	Local Attention	1 × 8192 × 3584
198	Add_33_attn	Add	1 × 8192 × 3584
199	RMSNorm_33_2	RMSNorm	1 × 8192 × 3584
200	FFN_33	SwiGLU	1 × 8192 × 3584
201	Add_33_ffn	Add	1 × 8192 × 3584
202	RMSNorm_34_1	RMSNorm	1 × 8192 × 3584
203	Attention_34	Grouped Query Attn	1 × 8192 × 3584
204	Add_34_attn	Add	1 × 8192 × 3584
205	RMSNorm_34_2	RMSNorm	1 × 8192 × 3584
206	FFN_34	SwiGLU	1 × 8192 × 3584
207	Add_34_ffn	Add	1 × 8192 × 3584
208	RMSNorm_35_1	RMSNorm	1 × 8192 × 3584
209	Attention_35	Local Attention	1 × 8192 × 3584
210	Add_35_attn	Add	1 × 8192 × 3584
211	RMSNorm_35_2	RMSNorm	1 × 8192 × 3584
212	FFN_35	SwiGLU	1 × 8192 × 3584
213	Add_35_ffn	Add	1 × 8192 × 3584
214	RMSNorm_36_1	RMSNorm	1 × 8192 × 3584
215	Attention_36	Grouped Query Attn	1 × 8192 × 3584
216	Add_36_attn	Add	1 × 8192 × 3584
217	RMSNorm_36_2	RMSNorm	1 × 8192 × 3584
218	FFN_36	SwiGLU	1 × 8192 × 3584
219	Add_36_ffn	Add	1 × 8192 × 3584
220	RMSNorm_37_1	RMSNorm	1 × 8192 × 3584
221	Attention_37	Local Attention	1 × 8192 × 3584
222	Add_37_attn	Add	1 × 8192 × 3584
223	RMSNorm_37_2	RMSNorm	1 × 8192 × 3584
224	FFN_37	SwiGLU	1 × 8192 × 3584
225	Add_37_ffn	Add	1 × 8192 × 3584
226	RMSNorm_38_1	RMSNorm	1 × 8192 × 3584
227	Attention_38	Grouped Query Attn	1 × 8192 × 3584
228	Add_38_attn	Add	1 × 8192 × 3584
229	RMSNorm_38_2	RMSNorm	1 × 8192 × 3584
230	FFN_38	SwiGLU	1 × 8192 × 3584
231	Add_38_ffn	Add	1 × 8192 × 3584
232	RMSNorm_39_1	RMSNorm	1 × 8192 × 3584
233	Attention_39	Local Attention	1 × 8192 × 3584
234	Add_39_attn	Add	1 × 8192 × 3584
235	RMSNorm_39_2	RMSNorm	1 × 8192 × 3584
236	FFN_39	SwiGLU	1 × 8192 × 3584
237	Add_39_ffn	Add	1 × 8192 × 3584
238	RMSNorm_40_1	RMSNorm	1 × 8192 × 3584
239	Attention_40	Grouped Query Attn	1 × 8192 × 3584
240	Add_40_attn	Add	1 × 8192 × 3584
241	RMSNorm_40_2	RMSNorm	1 × 8192 × 3584
242	FFN_40	SwiGLU	1 × 8192 × 3584
243	Add_40_ffn	Add	1 × 8192 × 3584
244	RMSNorm_41_1	RMSNorm	1 × 8192 × 3584
245	Attention_41	Local Attention	1 × 8192 × 3584
246	Add_41_attn	Add	1 × 8192 × 3584
247	RMSNorm_41_2	RMSNorm	1 × 8192 × 3584
248	FFN_41	SwiGLU	1 × 8192 × 3584
249	Add_41_ffn	Add	1 × 8192 × 3584
250	RMSNorm_42_1	RMSNorm	1 × 8192 × 3584
251	Attention_42	Grouped Query Attn	1 × 8192 × 3584
252	Add_42_attn	Add	1 × 8192 × 3584
253	RMSNorm_42_2	RMSNorm	1 × 8192 × 3584
254	FFN_42	SwiGLU	1 × 8192 × 3584
255	Add_42_ffn	Add	1 × 8192 × 3584
256	Output	Output	1 × 8192 × 3584

What the verifier says

infoAt 21 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/gemma-2-9b-it: 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/gemma-2-9b-it --plan --share

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