Models / locateanything

LocateAnything-3B: GPU memory to fine-tune

Fine-tuning LocateAnything-3B at batch 1 and 2048 tokens takes about 76.8 GB of GPU memory for a full fine-tune (fits one H200 141 GB), and 17.3 GB with LoRA r=16 on a bf16 base (borderline on one L4 24 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. 105K 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

3.50B

3,502,110,576 parameters

In the published checkpoint

3.83B

3,830,665,968 scalars · safetensors.total, read 2026-06-12

Delta

-0.41%

stored twice The checkpoint stores the tied embedding twice: +312.7M (`language_model.lm_head.weight` and `language_model.model.embed_tokens.weight`, 152681 x 2048). The model has one such matrix and the count above has it once; the delta compares that count plus the second copy against the file. Header read 2026-10-18.

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_locateanything.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, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

331

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$258.02

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

76.8 GB

LoRA r=16, batch 1 at 2048 tokens

17.3 GB

Serve: bf16 weights

7.00 GB

Serve: KV cache, 4096 tokens

0.66 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	batch 10
L4 24 GB	does not fit at batch 1	batch 1, borderline	batch 20
A100 40 GB	does not fit at batch 1	batch 2	batch 42
A100 / H100 80 GB	2 cards at batch 1 (305 + 26 layers)	batch 6	batch 94
H200 141 GB	batch 5	batch 12	batch 174

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 32768), 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

334 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 32768 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 32768
2	Embedding	Embedding	1 × 32768 × 2048
3	RoPE	RoPE	1 × 32768 × 2048
4	Vision input	Input	3 × 224 × 224
5	PatchEmbed	Patch Embed	256 × 1152
6	Patch_Position_Embedding	Learned Pos Embed	256 × 1152
7	Vision_LN_1	LayerNorm	256 × 1152
8	Vision_Attn_1	Multi-Head Attention	256 × 1152
9	Vision_Add_1	Add	256 × 1152
10	Vision_FFN_1	Feed Forward	256 × 1152
11	Vision_LN_2	LayerNorm	256 × 1152
12	Vision_Attn_2	Multi-Head Attention	256 × 1152
13	Vision_Add_2	Add	256 × 1152
14	Vision_FFN_2	Feed Forward	256 × 1152
15	Vision_LN_3	LayerNorm	256 × 1152
16	Vision_Attn_3	Multi-Head Attention	256 × 1152
17	Vision_Add_3	Add	256 × 1152
18	Vision_FFN_3	Feed Forward	256 × 1152
19	Vision_LN_4	LayerNorm	256 × 1152
20	Vision_Attn_4	Multi-Head Attention	256 × 1152
21	Vision_Add_4	Add	256 × 1152
22	Vision_FFN_4	Feed Forward	256 × 1152
23	Vision_LN_5	LayerNorm	256 × 1152
24	Vision_Attn_5	Multi-Head Attention	256 × 1152
25	Vision_Add_5	Add	256 × 1152
26	Vision_FFN_5	Feed Forward	256 × 1152
27	Vision_LN_6	LayerNorm	256 × 1152
28	Vision_Attn_6	Multi-Head Attention	256 × 1152
29	Vision_Add_6	Add	256 × 1152
30	Vision_FFN_6	Feed Forward	256 × 1152
31	Vision_LN_7	LayerNorm	256 × 1152
32	Vision_Attn_7	Multi-Head Attention	256 × 1152
33	Vision_Add_7	Add	256 × 1152
34	Vision_FFN_7	Feed Forward	256 × 1152
35	Vision_LN_8	LayerNorm	256 × 1152
36	Vision_Attn_8	Multi-Head Attention	256 × 1152
37	Vision_Add_8	Add	256 × 1152
38	Vision_FFN_8	Feed Forward	256 × 1152
39	Vision_LN_9	LayerNorm	256 × 1152
40	Vision_Attn_9	Multi-Head Attention	256 × 1152
41	Vision_Add_9	Add	256 × 1152
42	Vision_FFN_9	Feed Forward	256 × 1152
43	Vision_LN_10	LayerNorm	256 × 1152
44	Vision_Attn_10	Multi-Head Attention	256 × 1152
45	Vision_Add_10	Add	256 × 1152
46	Vision_FFN_10	Feed Forward	256 × 1152
47	Vision_LN_11	LayerNorm	256 × 1152
48	Vision_Attn_11	Multi-Head Attention	256 × 1152
49	Vision_Add_11	Add	256 × 1152
50	Vision_FFN_11	Feed Forward	256 × 1152
51	Vision_LN_12	LayerNorm	256 × 1152
52	Vision_Attn_12	Multi-Head Attention	256 × 1152
53	Vision_Add_12	Add	256 × 1152
54	Vision_FFN_12	Feed Forward	256 × 1152
55	Vision_LN_13	LayerNorm	256 × 1152
56	Vision_Attn_13	Multi-Head Attention	256 × 1152
57	Vision_Add_13	Add	256 × 1152
58	Vision_FFN_13	Feed Forward	256 × 1152
59	Vision_LN_14	LayerNorm	256 × 1152
60	Vision_Attn_14	Multi-Head Attention	256 × 1152
61	Vision_Add_14	Add	256 × 1152
62	Vision_FFN_14	Feed Forward	256 × 1152
63	Vision_LN_15	LayerNorm	256 × 1152
64	Vision_Attn_15	Multi-Head Attention	256 × 1152
65	Vision_Add_15	Add	256 × 1152
66	Vision_FFN_15	Feed Forward	256 × 1152
67	Vision_LN_16	LayerNorm	256 × 1152
68	Vision_Attn_16	Multi-Head Attention	256 × 1152
69	Vision_Add_16	Add	256 × 1152
70	Vision_FFN_16	Feed Forward	256 × 1152
71	Vision_LN_17	LayerNorm	256 × 1152
72	Vision_Attn_17	Multi-Head Attention	256 × 1152
73	Vision_Add_17	Add	256 × 1152
74	Vision_FFN_17	Feed Forward	256 × 1152
75	Vision_LN_18	LayerNorm	256 × 1152
76	Vision_Attn_18	Multi-Head Attention	256 × 1152
77	Vision_Add_18	Add	256 × 1152
78	Vision_FFN_18	Feed Forward	256 × 1152
79	Vision_LN_19	LayerNorm	256 × 1152
80	Vision_Attn_19	Multi-Head Attention	256 × 1152
81	Vision_Add_19	Add	256 × 1152
82	Vision_FFN_19	Feed Forward	256 × 1152
83	Vision_LN_20	LayerNorm	256 × 1152
84	Vision_Attn_20	Multi-Head Attention	256 × 1152
85	Vision_Add_20	Add	256 × 1152
86	Vision_FFN_20	Feed Forward	256 × 1152
87	Vision_LN_21	LayerNorm	256 × 1152
88	Vision_Attn_21	Multi-Head Attention	256 × 1152
89	Vision_Add_21	Add	256 × 1152
90	Vision_FFN_21	Feed Forward	256 × 1152
91	Vision_LN_22	LayerNorm	256 × 1152
92	Vision_Attn_22	Multi-Head Attention	256 × 1152
93	Vision_Add_22	Add	256 × 1152
94	Vision_FFN_22	Feed Forward	256 × 1152
95	Vision_LN_23	LayerNorm	256 × 1152
96	Vision_Attn_23	Multi-Head Attention	256 × 1152
97	Vision_Add_23	Add	256 × 1152
98	Vision_FFN_23	Feed Forward	256 × 1152
99	Vision_LN_24	LayerNorm	256 × 1152
100	Vision_Attn_24	Multi-Head Attention	256 × 1152
101	Vision_Add_24	Add	256 × 1152
102	Vision_FFN_24	Feed Forward	256 × 1152
103	Vision_LN_25	LayerNorm	256 × 1152
104	Vision_Attn_25	Multi-Head Attention	256 × 1152
105	Vision_Add_25	Add	256 × 1152
106	Vision_FFN_25	Feed Forward	256 × 1152
107	Vision_LN_26	LayerNorm	256 × 1152
108	Vision_Attn_26	Multi-Head Attention	256 × 1152
109	Vision_Add_26	Add	256 × 1152
110	Vision_FFN_26	Feed Forward	256 × 1152
111	Vision_LN_27	LayerNorm	256 × 1152
112	Vision_Attn_27	Multi-Head Attention	256 × 1152
113	Vision_Add_27	Add	256 × 1152
114	Vision_FFN_27	Feed Forward	256 × 1152
115	Vision projector	Projection	256 × 2048
116	Vision tokens	Reshape	1 × 256 × 2048
117	Multimodal fusion (concat tokens)	Concatenate	1 × 33024 × 2048
118	RMSNorm_1_1	RMSNorm	1 × 33024 × 2048
119	Attention_1	Grouped Query Attn	1 × 33024 × 2048
120	Add_1_attn	Add	1 × 33024 × 2048
121	RMSNorm_1_2	RMSNorm	1 × 33024 × 2048
122	FFN_1	SwiGLU	1 × 33024 × 2048
123	Add_1_ffn	Add	1 × 33024 × 2048
124	RMSNorm_2_1	RMSNorm	1 × 33024 × 2048
125	Attention_2	Grouped Query Attn	1 × 33024 × 2048
126	Add_2_attn	Add	1 × 33024 × 2048
127	RMSNorm_2_2	RMSNorm	1 × 33024 × 2048
128	FFN_2	SwiGLU	1 × 33024 × 2048
129	Add_2_ffn	Add	1 × 33024 × 2048
130	RMSNorm_3_1	RMSNorm	1 × 33024 × 2048
131	Attention_3	Grouped Query Attn	1 × 33024 × 2048
132	Add_3_attn	Add	1 × 33024 × 2048
133	RMSNorm_3_2	RMSNorm	1 × 33024 × 2048
134	FFN_3	SwiGLU	1 × 33024 × 2048
135	Add_3_ffn	Add	1 × 33024 × 2048
136	RMSNorm_4_1	RMSNorm	1 × 33024 × 2048
137	Attention_4	Grouped Query Attn	1 × 33024 × 2048
138	Add_4_attn	Add	1 × 33024 × 2048
139	RMSNorm_4_2	RMSNorm	1 × 33024 × 2048
140	FFN_4	SwiGLU	1 × 33024 × 2048
141	Add_4_ffn	Add	1 × 33024 × 2048
142	RMSNorm_5_1	RMSNorm	1 × 33024 × 2048
143	Attention_5	Grouped Query Attn	1 × 33024 × 2048
144	Add_5_attn	Add	1 × 33024 × 2048
145	RMSNorm_5_2	RMSNorm	1 × 33024 × 2048
146	FFN_5	SwiGLU	1 × 33024 × 2048
147	Add_5_ffn	Add	1 × 33024 × 2048
148	RMSNorm_6_1	RMSNorm	1 × 33024 × 2048
149	Attention_6	Grouped Query Attn	1 × 33024 × 2048
150	Add_6_attn	Add	1 × 33024 × 2048
151	RMSNorm_6_2	RMSNorm	1 × 33024 × 2048
152	FFN_6	SwiGLU	1 × 33024 × 2048
153	Add_6_ffn	Add	1 × 33024 × 2048
154	RMSNorm_7_1	RMSNorm	1 × 33024 × 2048
155	Attention_7	Grouped Query Attn	1 × 33024 × 2048
156	Add_7_attn	Add	1 × 33024 × 2048
157	RMSNorm_7_2	RMSNorm	1 × 33024 × 2048
158	FFN_7	SwiGLU	1 × 33024 × 2048
159	Add_7_ffn	Add	1 × 33024 × 2048
160	RMSNorm_8_1	RMSNorm	1 × 33024 × 2048
161	Attention_8	Grouped Query Attn	1 × 33024 × 2048
162	Add_8_attn	Add	1 × 33024 × 2048
163	RMSNorm_8_2	RMSNorm	1 × 33024 × 2048
164	FFN_8	SwiGLU	1 × 33024 × 2048
165	Add_8_ffn	Add	1 × 33024 × 2048
166	RMSNorm_9_1	RMSNorm	1 × 33024 × 2048
167	Attention_9	Grouped Query Attn	1 × 33024 × 2048
168	Add_9_attn	Add	1 × 33024 × 2048
169	RMSNorm_9_2	RMSNorm	1 × 33024 × 2048
170	FFN_9	SwiGLU	1 × 33024 × 2048
171	Add_9_ffn	Add	1 × 33024 × 2048
172	RMSNorm_10_1	RMSNorm	1 × 33024 × 2048
173	Attention_10	Grouped Query Attn	1 × 33024 × 2048
174	Add_10_attn	Add	1 × 33024 × 2048
175	RMSNorm_10_2	RMSNorm	1 × 33024 × 2048
176	FFN_10	SwiGLU	1 × 33024 × 2048
177	Add_10_ffn	Add	1 × 33024 × 2048
178	RMSNorm_11_1	RMSNorm	1 × 33024 × 2048
179	Attention_11	Grouped Query Attn	1 × 33024 × 2048
180	Add_11_attn	Add	1 × 33024 × 2048
181	RMSNorm_11_2	RMSNorm	1 × 33024 × 2048
182	FFN_11	SwiGLU	1 × 33024 × 2048
183	Add_11_ffn	Add	1 × 33024 × 2048
184	RMSNorm_12_1	RMSNorm	1 × 33024 × 2048
185	Attention_12	Grouped Query Attn	1 × 33024 × 2048
186	Add_12_attn	Add	1 × 33024 × 2048
187	RMSNorm_12_2	RMSNorm	1 × 33024 × 2048
188	FFN_12	SwiGLU	1 × 33024 × 2048
189	Add_12_ffn	Add	1 × 33024 × 2048
190	RMSNorm_13_1	RMSNorm	1 × 33024 × 2048
191	Attention_13	Grouped Query Attn	1 × 33024 × 2048
192	Add_13_attn	Add	1 × 33024 × 2048
193	RMSNorm_13_2	RMSNorm	1 × 33024 × 2048
194	FFN_13	SwiGLU	1 × 33024 × 2048
195	Add_13_ffn	Add	1 × 33024 × 2048
196	RMSNorm_14_1	RMSNorm	1 × 33024 × 2048
197	Attention_14	Grouped Query Attn	1 × 33024 × 2048
198	Add_14_attn	Add	1 × 33024 × 2048
199	RMSNorm_14_2	RMSNorm	1 × 33024 × 2048
200	FFN_14	SwiGLU	1 × 33024 × 2048
201	Add_14_ffn	Add	1 × 33024 × 2048
202	RMSNorm_15_1	RMSNorm	1 × 33024 × 2048
203	Attention_15	Grouped Query Attn	1 × 33024 × 2048
204	Add_15_attn	Add	1 × 33024 × 2048
205	RMSNorm_15_2	RMSNorm	1 × 33024 × 2048
206	FFN_15	SwiGLU	1 × 33024 × 2048
207	Add_15_ffn	Add	1 × 33024 × 2048
208	RMSNorm_16_1	RMSNorm	1 × 33024 × 2048
209	Attention_16	Grouped Query Attn	1 × 33024 × 2048
210	Add_16_attn	Add	1 × 33024 × 2048
211	RMSNorm_16_2	RMSNorm	1 × 33024 × 2048
212	FFN_16	SwiGLU	1 × 33024 × 2048
213	Add_16_ffn	Add	1 × 33024 × 2048
214	RMSNorm_17_1	RMSNorm	1 × 33024 × 2048
215	Attention_17	Grouped Query Attn	1 × 33024 × 2048
216	Add_17_attn	Add	1 × 33024 × 2048
217	RMSNorm_17_2	RMSNorm	1 × 33024 × 2048
218	FFN_17	SwiGLU	1 × 33024 × 2048
219	Add_17_ffn	Add	1 × 33024 × 2048
220	RMSNorm_18_1	RMSNorm	1 × 33024 × 2048
221	Attention_18	Grouped Query Attn	1 × 33024 × 2048
222	Add_18_attn	Add	1 × 33024 × 2048
223	RMSNorm_18_2	RMSNorm	1 × 33024 × 2048
224	FFN_18	SwiGLU	1 × 33024 × 2048
225	Add_18_ffn	Add	1 × 33024 × 2048
226	RMSNorm_19_1	RMSNorm	1 × 33024 × 2048
227	Attention_19	Grouped Query Attn	1 × 33024 × 2048
228	Add_19_attn	Add	1 × 33024 × 2048
229	RMSNorm_19_2	RMSNorm	1 × 33024 × 2048
230	FFN_19	SwiGLU	1 × 33024 × 2048
231	Add_19_ffn	Add	1 × 33024 × 2048
232	RMSNorm_20_1	RMSNorm	1 × 33024 × 2048
233	Attention_20	Grouped Query Attn	1 × 33024 × 2048
234	Add_20_attn	Add	1 × 33024 × 2048
235	RMSNorm_20_2	RMSNorm	1 × 33024 × 2048
236	FFN_20	SwiGLU	1 × 33024 × 2048
237	Add_20_ffn	Add	1 × 33024 × 2048
238	RMSNorm_21_1	RMSNorm	1 × 33024 × 2048
239	Attention_21	Grouped Query Attn	1 × 33024 × 2048
240	Add_21_attn	Add	1 × 33024 × 2048
241	RMSNorm_21_2	RMSNorm	1 × 33024 × 2048
242	FFN_21	SwiGLU	1 × 33024 × 2048
243	Add_21_ffn	Add	1 × 33024 × 2048
244	RMSNorm_22_1	RMSNorm	1 × 33024 × 2048
245	Attention_22	Grouped Query Attn	1 × 33024 × 2048
246	Add_22_attn	Add	1 × 33024 × 2048
247	RMSNorm_22_2	RMSNorm	1 × 33024 × 2048
248	FFN_22	SwiGLU	1 × 33024 × 2048
249	Add_22_ffn	Add	1 × 33024 × 2048
250	RMSNorm_23_1	RMSNorm	1 × 33024 × 2048
251	Attention_23	Grouped Query Attn	1 × 33024 × 2048
252	Add_23_attn	Add	1 × 33024 × 2048
253	RMSNorm_23_2	RMSNorm	1 × 33024 × 2048
254	FFN_23	SwiGLU	1 × 33024 × 2048
255	Add_23_ffn	Add	1 × 33024 × 2048
256	RMSNorm_24_1	RMSNorm	1 × 33024 × 2048
257	Attention_24	Grouped Query Attn	1 × 33024 × 2048
258	Add_24_attn	Add	1 × 33024 × 2048
259	RMSNorm_24_2	RMSNorm	1 × 33024 × 2048
260	FFN_24	SwiGLU	1 × 33024 × 2048
261	Add_24_ffn	Add	1 × 33024 × 2048
262	RMSNorm_25_1	RMSNorm	1 × 33024 × 2048
263	Attention_25	Grouped Query Attn	1 × 33024 × 2048
264	Add_25_attn	Add	1 × 33024 × 2048
265	RMSNorm_25_2	RMSNorm	1 × 33024 × 2048
266	FFN_25	SwiGLU	1 × 33024 × 2048
267	Add_25_ffn	Add	1 × 33024 × 2048
268	RMSNorm_26_1	RMSNorm	1 × 33024 × 2048
269	Attention_26	Grouped Query Attn	1 × 33024 × 2048
270	Add_26_attn	Add	1 × 33024 × 2048
271	RMSNorm_26_2	RMSNorm	1 × 33024 × 2048
272	FFN_26	SwiGLU	1 × 33024 × 2048
273	Add_26_ffn	Add	1 × 33024 × 2048
274	RMSNorm_27_1	RMSNorm	1 × 33024 × 2048
275	Attention_27	Grouped Query Attn	1 × 33024 × 2048
276	Add_27_attn	Add	1 × 33024 × 2048
277	RMSNorm_27_2	RMSNorm	1 × 33024 × 2048
278	FFN_27	SwiGLU	1 × 33024 × 2048
279	Add_27_ffn	Add	1 × 33024 × 2048
280	RMSNorm_28_1	RMSNorm	1 × 33024 × 2048
281	Attention_28	Grouped Query Attn	1 × 33024 × 2048
282	Add_28_attn	Add	1 × 33024 × 2048
283	RMSNorm_28_2	RMSNorm	1 × 33024 × 2048
284	FFN_28	SwiGLU	1 × 33024 × 2048
285	Add_28_ffn	Add	1 × 33024 × 2048
286	RMSNorm_29_1	RMSNorm	1 × 33024 × 2048
287	Attention_29	Grouped Query Attn	1 × 33024 × 2048
288	Add_29_attn	Add	1 × 33024 × 2048
289	RMSNorm_29_2	RMSNorm	1 × 33024 × 2048
290	FFN_29	SwiGLU	1 × 33024 × 2048
291	Add_29_ffn	Add	1 × 33024 × 2048
292	RMSNorm_30_1	RMSNorm	1 × 33024 × 2048
293	Attention_30	Grouped Query Attn	1 × 33024 × 2048
294	Add_30_attn	Add	1 × 33024 × 2048
295	RMSNorm_30_2	RMSNorm	1 × 33024 × 2048
296	FFN_30	SwiGLU	1 × 33024 × 2048
297	Add_30_ffn	Add	1 × 33024 × 2048
298	RMSNorm_31_1	RMSNorm	1 × 33024 × 2048
299	Attention_31	Grouped Query Attn	1 × 33024 × 2048
300	Add_31_attn	Add	1 × 33024 × 2048
301	RMSNorm_31_2	RMSNorm	1 × 33024 × 2048
302	FFN_31	SwiGLU	1 × 33024 × 2048
303	Add_31_ffn	Add	1 × 33024 × 2048
304	RMSNorm_32_1	RMSNorm	1 × 33024 × 2048
305	Attention_32	Grouped Query Attn	1 × 33024 × 2048
306	Add_32_attn	Add	1 × 33024 × 2048
307	RMSNorm_32_2	RMSNorm	1 × 33024 × 2048
308	FFN_32	SwiGLU	1 × 33024 × 2048
309	Add_32_ffn	Add	1 × 33024 × 2048
310	RMSNorm_33_1	RMSNorm	1 × 33024 × 2048
311	Attention_33	Grouped Query Attn	1 × 33024 × 2048
312	Add_33_attn	Add	1 × 33024 × 2048
313	RMSNorm_33_2	RMSNorm	1 × 33024 × 2048
314	FFN_33	SwiGLU	1 × 33024 × 2048
315	Add_33_ffn	Add	1 × 33024 × 2048
316	RMSNorm_34_1	RMSNorm	1 × 33024 × 2048
317	Attention_34	Grouped Query Attn	1 × 33024 × 2048
318	Add_34_attn	Add	1 × 33024 × 2048
319	RMSNorm_34_2	RMSNorm	1 × 33024 × 2048
320	FFN_34	SwiGLU	1 × 33024 × 2048
321	Add_34_ffn	Add	1 × 33024 × 2048
322	RMSNorm_35_1	RMSNorm	1 × 33024 × 2048
323	Attention_35	Grouped Query Attn	1 × 33024 × 2048
324	Add_35_attn	Add	1 × 33024 × 2048
325	RMSNorm_35_2	RMSNorm	1 × 33024 × 2048
326	FFN_35	SwiGLU	1 × 33024 × 2048
327	Add_35_ffn	Add	1 × 33024 × 2048
328	RMSNorm_36_1	RMSNorm	1 × 33024 × 2048
329	Attention_36	Grouped Query Attn	1 × 33024 × 2048
330	Add_36_attn	Add	1 × 33024 × 2048
331	RMSNorm_36_2	RMSNorm	1 × 33024 × 2048
332	FFN_36	SwiGLU	1 × 33024 × 2048
333	Add_36_ffn	Add	1 × 33024 × 2048
334	Output	Output	1 × 33024 × 2048

What the verifier says

warn"RoPE" receives input but its output is not connected. This layer will be unreachable in the forward pass. Fix: Connect the output forward, or add an Output node if this is the final layer.
dead-end

infoLLaMA uses intermediateSize ≈ ⌊(8/3 × D) / 256⌋ × 256. Current: 11008 (5.38× embedDim). Expected: ~5376. Fix: Set intermediateSize to 5376 for embedDim=2048.
swiglu-dim-convention

infoAt 63 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 nvidia/LocateAnything-3B: 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 nvidia/LocateAnything-3B --plan --share

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