N Neurarch Architectures Models Checks Data Docs Open the app

Models / internvl_chat

InternVL2-26B

Reconstructed from its own config.json with no weights read. 177K 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
25.42B
25,417,368,585 parameters
In the published checkpoint
25.51B
25,514,186,112 scalars · safetensors.total, read 2025-03-25
Delta
-0.38%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_internvl_chat.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 over 50,000 samples (assumed; no dataset attached). GPU fit is fp32 weights plus gradients plus two Adam moments (16 bytes per parameter) with 1.3x headroom; activations are not included and grow with batch size.

Layers
477
Will it forward-pass
Yes
Priced on
A10G (24GB)
Est. one run
$37223.31
CardMemory
T4 (16GB)weights + activationsdoes not fit
A100 (40GB)weights + activationsdoes not fit
H100 (80GB)weights + activationsdoes not fit

Structure

480 nodes. Output shapes are propagated from the input shape, batch dimension excluded.

LayerTypeOutput shape
1InputInput1 × 32768
2EmbeddingEmbedding1 × 32768 × 6144
3RoPERoPE1 × 32768 × 6144
4Vision inputInput3 × 448 × 448
5PatchEmbedPatch Embed1024 × 3200
6Patch_Position_EmbeddingLearned Pos Embed1024 × 3200
7Vision_LN_1LayerNorm1024 × 3200
8Vision_Attn_1Multi-Head Attention1024 × 3200
9Vision_Add_1Add1024 × 3200
10Vision_FFN_1Feed Forward1024 × 3200
11Vision_LN_2LayerNorm1024 × 3200
12Vision_Attn_2Multi-Head Attention1024 × 3200
13Vision_Add_2Add1024 × 3200
14Vision_FFN_2Feed Forward1024 × 3200
15Vision_LN_3LayerNorm1024 × 3200
16Vision_Attn_3Multi-Head Attention1024 × 3200
17Vision_Add_3Add1024 × 3200
18Vision_FFN_3Feed Forward1024 × 3200
19Vision_LN_4LayerNorm1024 × 3200
20Vision_Attn_4Multi-Head Attention1024 × 3200
21Vision_Add_4Add1024 × 3200
22Vision_FFN_4Feed Forward1024 × 3200
23Vision_LN_5LayerNorm1024 × 3200
24Vision_Attn_5Multi-Head Attention1024 × 3200
25Vision_Add_5Add1024 × 3200
26Vision_FFN_5Feed Forward1024 × 3200
27Vision_LN_6LayerNorm1024 × 3200
28Vision_Attn_6Multi-Head Attention1024 × 3200
29Vision_Add_6Add1024 × 3200
30Vision_FFN_6Feed Forward1024 × 3200
31Vision_LN_7LayerNorm1024 × 3200
32Vision_Attn_7Multi-Head Attention1024 × 3200
33Vision_Add_7Add1024 × 3200
34Vision_FFN_7Feed Forward1024 × 3200
35Vision_LN_8LayerNorm1024 × 3200
36Vision_Attn_8Multi-Head Attention1024 × 3200
37Vision_Add_8Add1024 × 3200
38Vision_FFN_8Feed Forward1024 × 3200
39Vision_LN_9LayerNorm1024 × 3200
40Vision_Attn_9Multi-Head Attention1024 × 3200
41Vision_Add_9Add1024 × 3200
42Vision_FFN_9Feed Forward1024 × 3200
43Vision_LN_10LayerNorm1024 × 3200
44Vision_Attn_10Multi-Head Attention1024 × 3200
45Vision_Add_10Add1024 × 3200
46Vision_FFN_10Feed Forward1024 × 3200
47Vision_LN_11LayerNorm1024 × 3200
48Vision_Attn_11Multi-Head Attention1024 × 3200
49Vision_Add_11Add1024 × 3200
50Vision_FFN_11Feed Forward1024 × 3200
51Vision_LN_12LayerNorm1024 × 3200
52Vision_Attn_12Multi-Head Attention1024 × 3200
53Vision_Add_12Add1024 × 3200
54Vision_FFN_12Feed Forward1024 × 3200
55Vision_LN_13LayerNorm1024 × 3200
56Vision_Attn_13Multi-Head Attention1024 × 3200
57Vision_Add_13Add1024 × 3200
58Vision_FFN_13Feed Forward1024 × 3200
59Vision_LN_14LayerNorm1024 × 3200
60Vision_Attn_14Multi-Head Attention1024 × 3200
61Vision_Add_14Add1024 × 3200
62Vision_FFN_14Feed Forward1024 × 3200
63Vision_LN_15LayerNorm1024 × 3200
64Vision_Attn_15Multi-Head Attention1024 × 3200
65Vision_Add_15Add1024 × 3200
66Vision_FFN_15Feed Forward1024 × 3200
67Vision_LN_16LayerNorm1024 × 3200
68Vision_Attn_16Multi-Head Attention1024 × 3200
69Vision_Add_16Add1024 × 3200
70Vision_FFN_16Feed Forward1024 × 3200
71Vision_LN_17LayerNorm1024 × 3200
72Vision_Attn_17Multi-Head Attention1024 × 3200
73Vision_Add_17Add1024 × 3200
74Vision_FFN_17Feed Forward1024 × 3200
75Vision_LN_18LayerNorm1024 × 3200
76Vision_Attn_18Multi-Head Attention1024 × 3200
77Vision_Add_18Add1024 × 3200
78Vision_FFN_18Feed Forward1024 × 3200
79Vision_LN_19LayerNorm1024 × 3200
80Vision_Attn_19Multi-Head Attention1024 × 3200
81Vision_Add_19Add1024 × 3200
82Vision_FFN_19Feed Forward1024 × 3200
83Vision_LN_20LayerNorm1024 × 3200
84Vision_Attn_20Multi-Head Attention1024 × 3200
85Vision_Add_20Add1024 × 3200
86Vision_FFN_20Feed Forward1024 × 3200
87Vision_LN_21LayerNorm1024 × 3200
88Vision_Attn_21Multi-Head Attention1024 × 3200
89Vision_Add_21Add1024 × 3200
90Vision_FFN_21Feed Forward1024 × 3200
91Vision_LN_22LayerNorm1024 × 3200
92Vision_Attn_22Multi-Head Attention1024 × 3200
93Vision_Add_22Add1024 × 3200
94Vision_FFN_22Feed Forward1024 × 3200
95Vision_LN_23LayerNorm1024 × 3200
96Vision_Attn_23Multi-Head Attention1024 × 3200
97Vision_Add_23Add1024 × 3200
98Vision_FFN_23Feed Forward1024 × 3200
99Vision_LN_24LayerNorm1024 × 3200
100Vision_Attn_24Multi-Head Attention1024 × 3200
101Vision_Add_24Add1024 × 3200
102Vision_FFN_24Feed Forward1024 × 3200
103Vision_LN_25LayerNorm1024 × 3200
104Vision_Attn_25Multi-Head Attention1024 × 3200
105Vision_Add_25Add1024 × 3200
106Vision_FFN_25Feed Forward1024 × 3200
107Vision_LN_26LayerNorm1024 × 3200
108Vision_Attn_26Multi-Head Attention1024 × 3200
109Vision_Add_26Add1024 × 3200
110Vision_FFN_26Feed Forward1024 × 3200
111Vision_LN_27LayerNorm1024 × 3200
112Vision_Attn_27Multi-Head Attention1024 × 3200
113Vision_Add_27Add1024 × 3200
114Vision_FFN_27Feed Forward1024 × 3200
115Vision_LN_28LayerNorm1024 × 3200
116Vision_Attn_28Multi-Head Attention1024 × 3200
117Vision_Add_28Add1024 × 3200
118Vision_FFN_28Feed Forward1024 × 3200
119Vision_LN_29LayerNorm1024 × 3200
120Vision_Attn_29Multi-Head Attention1024 × 3200
121Vision_Add_29Add1024 × 3200
122Vision_FFN_29Feed Forward1024 × 3200
123Vision_LN_30LayerNorm1024 × 3200
124Vision_Attn_30Multi-Head Attention1024 × 3200
125Vision_Add_30Add1024 × 3200
126Vision_FFN_30Feed Forward1024 × 3200
127Vision_LN_31LayerNorm1024 × 3200
128Vision_Attn_31Multi-Head Attention1024 × 3200
129Vision_Add_31Add1024 × 3200
130Vision_FFN_31Feed Forward1024 × 3200
131Vision_LN_32LayerNorm1024 × 3200
132Vision_Attn_32Multi-Head Attention1024 × 3200
133Vision_Add_32Add1024 × 3200
134Vision_FFN_32Feed Forward1024 × 3200
135Vision_LN_33LayerNorm1024 × 3200
136Vision_Attn_33Multi-Head Attention1024 × 3200
137Vision_Add_33Add1024 × 3200
138Vision_FFN_33Feed Forward1024 × 3200
139Vision_LN_34LayerNorm1024 × 3200
140Vision_Attn_34Multi-Head Attention1024 × 3200
141Vision_Add_34Add1024 × 3200
142Vision_FFN_34Feed Forward1024 × 3200
143Vision_LN_35LayerNorm1024 × 3200
144Vision_Attn_35Multi-Head Attention1024 × 3200
145Vision_Add_35Add1024 × 3200
146Vision_FFN_35Feed Forward1024 × 3200
147Vision_LN_36LayerNorm1024 × 3200
148Vision_Attn_36Multi-Head Attention1024 × 3200
149Vision_Add_36Add1024 × 3200
150Vision_FFN_36Feed Forward1024 × 3200
151Vision_LN_37LayerNorm1024 × 3200
152Vision_Attn_37Multi-Head Attention1024 × 3200
153Vision_Add_37Add1024 × 3200
154Vision_FFN_37Feed Forward1024 × 3200
155Vision_LN_38LayerNorm1024 × 3200
156Vision_Attn_38Multi-Head Attention1024 × 3200
157Vision_Add_38Add1024 × 3200
158Vision_FFN_38Feed Forward1024 × 3200
159Vision_LN_39LayerNorm1024 × 3200
160Vision_Attn_39Multi-Head Attention1024 × 3200
161Vision_Add_39Add1024 × 3200
162Vision_FFN_39Feed Forward1024 × 3200
163Vision_LN_40LayerNorm1024 × 3200
164Vision_Attn_40Multi-Head Attention1024 × 3200
165Vision_Add_40Add1024 × 3200
166Vision_FFN_40Feed Forward1024 × 3200
167Vision_LN_41LayerNorm1024 × 3200
168Vision_Attn_41Multi-Head Attention1024 × 3200
169Vision_Add_41Add1024 × 3200
170Vision_FFN_41Feed Forward1024 × 3200
171Vision_LN_42LayerNorm1024 × 3200
172Vision_Attn_42Multi-Head Attention1024 × 3200
173Vision_Add_42Add1024 × 3200
174Vision_FFN_42Feed Forward1024 × 3200
175Vision_LN_43LayerNorm1024 × 3200
176Vision_Attn_43Multi-Head Attention1024 × 3200
177Vision_Add_43Add1024 × 3200
178Vision_FFN_43Feed Forward1024 × 3200
179Vision_LN_44LayerNorm1024 × 3200
180Vision_Attn_44Multi-Head Attention1024 × 3200
181Vision_Add_44Add1024 × 3200
182Vision_FFN_44Feed Forward1024 × 3200
183Vision_LN_45LayerNorm1024 × 3200
184Vision_Attn_45Multi-Head Attention1024 × 3200
185Vision_Add_45Add1024 × 3200
186Vision_FFN_45Feed Forward1024 × 3200
187Vision projectorProjection1024 × 6144
188Vision tokensReshape1 × 1024 × 6144
189Multimodal fusion (concat tokens)Concatenate1 × 33792 × 6144
190RMSNorm_1_1RMSNorm1 × 33792 × 6144
191Attention_1Grouped Query Attn1 × 33792 × 6144
192Add_1_attnAdd1 × 33792 × 6144
193RMSNorm_1_2RMSNorm1 × 33792 × 6144
194FFN_1SwiGLU1 × 33792 × 6144
195Add_1_ffnAdd1 × 33792 × 6144
196RMSNorm_2_1RMSNorm1 × 33792 × 6144
197Attention_2Grouped Query Attn1 × 33792 × 6144
198Add_2_attnAdd1 × 33792 × 6144
199RMSNorm_2_2RMSNorm1 × 33792 × 6144
200FFN_2SwiGLU1 × 33792 × 6144
201Add_2_ffnAdd1 × 33792 × 6144
202RMSNorm_3_1RMSNorm1 × 33792 × 6144
203Attention_3Grouped Query Attn1 × 33792 × 6144
204Add_3_attnAdd1 × 33792 × 6144
205RMSNorm_3_2RMSNorm1 × 33792 × 6144
206FFN_3SwiGLU1 × 33792 × 6144
207Add_3_ffnAdd1 × 33792 × 6144
208RMSNorm_4_1RMSNorm1 × 33792 × 6144
209Attention_4Grouped Query Attn1 × 33792 × 6144
210Add_4_attnAdd1 × 33792 × 6144
211RMSNorm_4_2RMSNorm1 × 33792 × 6144
212FFN_4SwiGLU1 × 33792 × 6144
213Add_4_ffnAdd1 × 33792 × 6144
214RMSNorm_5_1RMSNorm1 × 33792 × 6144
215Attention_5Grouped Query Attn1 × 33792 × 6144
216Add_5_attnAdd1 × 33792 × 6144
217RMSNorm_5_2RMSNorm1 × 33792 × 6144
218FFN_5SwiGLU1 × 33792 × 6144
219Add_5_ffnAdd1 × 33792 × 6144
220RMSNorm_6_1RMSNorm1 × 33792 × 6144
221Attention_6Grouped Query Attn1 × 33792 × 6144
222Add_6_attnAdd1 × 33792 × 6144
223RMSNorm_6_2RMSNorm1 × 33792 × 6144
224FFN_6SwiGLU1 × 33792 × 6144
225Add_6_ffnAdd1 × 33792 × 6144
226RMSNorm_7_1RMSNorm1 × 33792 × 6144
227Attention_7Grouped Query Attn1 × 33792 × 6144
228Add_7_attnAdd1 × 33792 × 6144
229RMSNorm_7_2RMSNorm1 × 33792 × 6144
230FFN_7SwiGLU1 × 33792 × 6144
231Add_7_ffnAdd1 × 33792 × 6144
232RMSNorm_8_1RMSNorm1 × 33792 × 6144
233Attention_8Grouped Query Attn1 × 33792 × 6144
234Add_8_attnAdd1 × 33792 × 6144
235RMSNorm_8_2RMSNorm1 × 33792 × 6144
236FFN_8SwiGLU1 × 33792 × 6144
237Add_8_ffnAdd1 × 33792 × 6144
238RMSNorm_9_1RMSNorm1 × 33792 × 6144
239Attention_9Grouped Query Attn1 × 33792 × 6144
240Add_9_attnAdd1 × 33792 × 6144
241RMSNorm_9_2RMSNorm1 × 33792 × 6144
242FFN_9SwiGLU1 × 33792 × 6144
243Add_9_ffnAdd1 × 33792 × 6144
244RMSNorm_10_1RMSNorm1 × 33792 × 6144
245Attention_10Grouped Query Attn1 × 33792 × 6144
246Add_10_attnAdd1 × 33792 × 6144
247RMSNorm_10_2RMSNorm1 × 33792 × 6144
248FFN_10SwiGLU1 × 33792 × 6144
249Add_10_ffnAdd1 × 33792 × 6144
250RMSNorm_11_1RMSNorm1 × 33792 × 6144
251Attention_11Grouped Query Attn1 × 33792 × 6144
252Add_11_attnAdd1 × 33792 × 6144
253RMSNorm_11_2RMSNorm1 × 33792 × 6144
254FFN_11SwiGLU1 × 33792 × 6144
255Add_11_ffnAdd1 × 33792 × 6144
256RMSNorm_12_1RMSNorm1 × 33792 × 6144
257Attention_12Grouped Query Attn1 × 33792 × 6144
258Add_12_attnAdd1 × 33792 × 6144
259RMSNorm_12_2RMSNorm1 × 33792 × 6144
260FFN_12SwiGLU1 × 33792 × 6144
261Add_12_ffnAdd1 × 33792 × 6144
262RMSNorm_13_1RMSNorm1 × 33792 × 6144
263Attention_13Grouped Query Attn1 × 33792 × 6144
264Add_13_attnAdd1 × 33792 × 6144
265RMSNorm_13_2RMSNorm1 × 33792 × 6144
266FFN_13SwiGLU1 × 33792 × 6144
267Add_13_ffnAdd1 × 33792 × 6144
268RMSNorm_14_1RMSNorm1 × 33792 × 6144
269Attention_14Grouped Query Attn1 × 33792 × 6144
270Add_14_attnAdd1 × 33792 × 6144
271RMSNorm_14_2RMSNorm1 × 33792 × 6144
272FFN_14SwiGLU1 × 33792 × 6144
273Add_14_ffnAdd1 × 33792 × 6144
274RMSNorm_15_1RMSNorm1 × 33792 × 6144
275Attention_15Grouped Query Attn1 × 33792 × 6144
276Add_15_attnAdd1 × 33792 × 6144
277RMSNorm_15_2RMSNorm1 × 33792 × 6144
278FFN_15SwiGLU1 × 33792 × 6144
279Add_15_ffnAdd1 × 33792 × 6144
280RMSNorm_16_1RMSNorm1 × 33792 × 6144
281Attention_16Grouped Query Attn1 × 33792 × 6144
282Add_16_attnAdd1 × 33792 × 6144
283RMSNorm_16_2RMSNorm1 × 33792 × 6144
284FFN_16SwiGLU1 × 33792 × 6144
285Add_16_ffnAdd1 × 33792 × 6144
286RMSNorm_17_1RMSNorm1 × 33792 × 6144
287Attention_17Grouped Query Attn1 × 33792 × 6144
288Add_17_attnAdd1 × 33792 × 6144
289RMSNorm_17_2RMSNorm1 × 33792 × 6144
290FFN_17SwiGLU1 × 33792 × 6144
291Add_17_ffnAdd1 × 33792 × 6144
292RMSNorm_18_1RMSNorm1 × 33792 × 6144
293Attention_18Grouped Query Attn1 × 33792 × 6144
294Add_18_attnAdd1 × 33792 × 6144
295RMSNorm_18_2RMSNorm1 × 33792 × 6144
296FFN_18SwiGLU1 × 33792 × 6144
297Add_18_ffnAdd1 × 33792 × 6144
298RMSNorm_19_1RMSNorm1 × 33792 × 6144
299Attention_19Grouped Query Attn1 × 33792 × 6144
300Add_19_attnAdd1 × 33792 × 6144
301RMSNorm_19_2RMSNorm1 × 33792 × 6144
302FFN_19SwiGLU1 × 33792 × 6144
303Add_19_ffnAdd1 × 33792 × 6144
304RMSNorm_20_1RMSNorm1 × 33792 × 6144
305Attention_20Grouped Query Attn1 × 33792 × 6144
306Add_20_attnAdd1 × 33792 × 6144
307RMSNorm_20_2RMSNorm1 × 33792 × 6144
308FFN_20SwiGLU1 × 33792 × 6144
309Add_20_ffnAdd1 × 33792 × 6144
310RMSNorm_21_1RMSNorm1 × 33792 × 6144
311Attention_21Grouped Query Attn1 × 33792 × 6144
312Add_21_attnAdd1 × 33792 × 6144
313RMSNorm_21_2RMSNorm1 × 33792 × 6144
314FFN_21SwiGLU1 × 33792 × 6144
315Add_21_ffnAdd1 × 33792 × 6144
316RMSNorm_22_1RMSNorm1 × 33792 × 6144
317Attention_22Grouped Query Attn1 × 33792 × 6144
318Add_22_attnAdd1 × 33792 × 6144
319RMSNorm_22_2RMSNorm1 × 33792 × 6144
320FFN_22SwiGLU1 × 33792 × 6144
321Add_22_ffnAdd1 × 33792 × 6144
322RMSNorm_23_1RMSNorm1 × 33792 × 6144
323Attention_23Grouped Query Attn1 × 33792 × 6144
324Add_23_attnAdd1 × 33792 × 6144
325RMSNorm_23_2RMSNorm1 × 33792 × 6144
326FFN_23SwiGLU1 × 33792 × 6144
327Add_23_ffnAdd1 × 33792 × 6144
328RMSNorm_24_1RMSNorm1 × 33792 × 6144
329Attention_24Grouped Query Attn1 × 33792 × 6144
330Add_24_attnAdd1 × 33792 × 6144
331RMSNorm_24_2RMSNorm1 × 33792 × 6144
332FFN_24SwiGLU1 × 33792 × 6144
333Add_24_ffnAdd1 × 33792 × 6144
334RMSNorm_25_1RMSNorm1 × 33792 × 6144
335Attention_25Grouped Query Attn1 × 33792 × 6144
336Add_25_attnAdd1 × 33792 × 6144
337RMSNorm_25_2RMSNorm1 × 33792 × 6144
338FFN_25SwiGLU1 × 33792 × 6144
339Add_25_ffnAdd1 × 33792 × 6144
340RMSNorm_26_1RMSNorm1 × 33792 × 6144
341Attention_26Grouped Query Attn1 × 33792 × 6144
342Add_26_attnAdd1 × 33792 × 6144
343RMSNorm_26_2RMSNorm1 × 33792 × 6144
344FFN_26SwiGLU1 × 33792 × 6144
345Add_26_ffnAdd1 × 33792 × 6144
346RMSNorm_27_1RMSNorm1 × 33792 × 6144
347Attention_27Grouped Query Attn1 × 33792 × 6144
348Add_27_attnAdd1 × 33792 × 6144
349RMSNorm_27_2RMSNorm1 × 33792 × 6144
350FFN_27SwiGLU1 × 33792 × 6144
351Add_27_ffnAdd1 × 33792 × 6144
352RMSNorm_28_1RMSNorm1 × 33792 × 6144
353Attention_28Grouped Query Attn1 × 33792 × 6144
354Add_28_attnAdd1 × 33792 × 6144
355RMSNorm_28_2RMSNorm1 × 33792 × 6144
356FFN_28SwiGLU1 × 33792 × 6144
357Add_28_ffnAdd1 × 33792 × 6144
358RMSNorm_29_1RMSNorm1 × 33792 × 6144
359Attention_29Grouped Query Attn1 × 33792 × 6144
360Add_29_attnAdd1 × 33792 × 6144
361RMSNorm_29_2RMSNorm1 × 33792 × 6144
362FFN_29SwiGLU1 × 33792 × 6144
363Add_29_ffnAdd1 × 33792 × 6144
364RMSNorm_30_1RMSNorm1 × 33792 × 6144
365Attention_30Grouped Query Attn1 × 33792 × 6144
366Add_30_attnAdd1 × 33792 × 6144
367RMSNorm_30_2RMSNorm1 × 33792 × 6144
368FFN_30SwiGLU1 × 33792 × 6144
369Add_30_ffnAdd1 × 33792 × 6144
370RMSNorm_31_1RMSNorm1 × 33792 × 6144
371Attention_31Grouped Query Attn1 × 33792 × 6144
372Add_31_attnAdd1 × 33792 × 6144
373RMSNorm_31_2RMSNorm1 × 33792 × 6144
374FFN_31SwiGLU1 × 33792 × 6144
375Add_31_ffnAdd1 × 33792 × 6144
376RMSNorm_32_1RMSNorm1 × 33792 × 6144
377Attention_32Grouped Query Attn1 × 33792 × 6144
378Add_32_attnAdd1 × 33792 × 6144
379RMSNorm_32_2RMSNorm1 × 33792 × 6144
380FFN_32SwiGLU1 × 33792 × 6144
381Add_32_ffnAdd1 × 33792 × 6144
382RMSNorm_33_1RMSNorm1 × 33792 × 6144
383Attention_33Grouped Query Attn1 × 33792 × 6144
384Add_33_attnAdd1 × 33792 × 6144
385RMSNorm_33_2RMSNorm1 × 33792 × 6144
386FFN_33SwiGLU1 × 33792 × 6144
387Add_33_ffnAdd1 × 33792 × 6144
388RMSNorm_34_1RMSNorm1 × 33792 × 6144
389Attention_34Grouped Query Attn1 × 33792 × 6144
390Add_34_attnAdd1 × 33792 × 6144
391RMSNorm_34_2RMSNorm1 × 33792 × 6144
392FFN_34SwiGLU1 × 33792 × 6144
393Add_34_ffnAdd1 × 33792 × 6144
394RMSNorm_35_1RMSNorm1 × 33792 × 6144
395Attention_35Grouped Query Attn1 × 33792 × 6144
396Add_35_attnAdd1 × 33792 × 6144
397RMSNorm_35_2RMSNorm1 × 33792 × 6144
398FFN_35SwiGLU1 × 33792 × 6144
399Add_35_ffnAdd1 × 33792 × 6144
400RMSNorm_36_1RMSNorm1 × 33792 × 6144
401Attention_36Grouped Query Attn1 × 33792 × 6144
402Add_36_attnAdd1 × 33792 × 6144
403RMSNorm_36_2RMSNorm1 × 33792 × 6144
404FFN_36SwiGLU1 × 33792 × 6144
405Add_36_ffnAdd1 × 33792 × 6144
406RMSNorm_37_1RMSNorm1 × 33792 × 6144
407Attention_37Grouped Query Attn1 × 33792 × 6144
408Add_37_attnAdd1 × 33792 × 6144
409RMSNorm_37_2RMSNorm1 × 33792 × 6144
410FFN_37SwiGLU1 × 33792 × 6144
411Add_37_ffnAdd1 × 33792 × 6144
412RMSNorm_38_1RMSNorm1 × 33792 × 6144
413Attention_38Grouped Query Attn1 × 33792 × 6144
414Add_38_attnAdd1 × 33792 × 6144
415RMSNorm_38_2RMSNorm1 × 33792 × 6144
416FFN_38SwiGLU1 × 33792 × 6144
417Add_38_ffnAdd1 × 33792 × 6144
418RMSNorm_39_1RMSNorm1 × 33792 × 6144
419Attention_39Grouped Query Attn1 × 33792 × 6144
420Add_39_attnAdd1 × 33792 × 6144
421RMSNorm_39_2RMSNorm1 × 33792 × 6144
422FFN_39SwiGLU1 × 33792 × 6144
423Add_39_ffnAdd1 × 33792 × 6144
424RMSNorm_40_1RMSNorm1 × 33792 × 6144
425Attention_40Grouped Query Attn1 × 33792 × 6144
426Add_40_attnAdd1 × 33792 × 6144
427RMSNorm_40_2RMSNorm1 × 33792 × 6144
428FFN_40SwiGLU1 × 33792 × 6144
429Add_40_ffnAdd1 × 33792 × 6144
430RMSNorm_41_1RMSNorm1 × 33792 × 6144
431Attention_41Grouped Query Attn1 × 33792 × 6144
432Add_41_attnAdd1 × 33792 × 6144
433RMSNorm_41_2RMSNorm1 × 33792 × 6144
434FFN_41SwiGLU1 × 33792 × 6144
435Add_41_ffnAdd1 × 33792 × 6144
436RMSNorm_42_1RMSNorm1 × 33792 × 6144
437Attention_42Grouped Query Attn1 × 33792 × 6144
438Add_42_attnAdd1 × 33792 × 6144
439RMSNorm_42_2RMSNorm1 × 33792 × 6144
440FFN_42SwiGLU1 × 33792 × 6144
441Add_42_ffnAdd1 × 33792 × 6144
442RMSNorm_43_1RMSNorm1 × 33792 × 6144
443Attention_43Grouped Query Attn1 × 33792 × 6144
444Add_43_attnAdd1 × 33792 × 6144
445RMSNorm_43_2RMSNorm1 × 33792 × 6144
446FFN_43SwiGLU1 × 33792 × 6144
447Add_43_ffnAdd1 × 33792 × 6144
448RMSNorm_44_1RMSNorm1 × 33792 × 6144
449Attention_44Grouped Query Attn1 × 33792 × 6144
450Add_44_attnAdd1 × 33792 × 6144
451RMSNorm_44_2RMSNorm1 × 33792 × 6144
452FFN_44SwiGLU1 × 33792 × 6144
453Add_44_ffnAdd1 × 33792 × 6144
454RMSNorm_45_1RMSNorm1 × 33792 × 6144
455Attention_45Grouped Query Attn1 × 33792 × 6144
456Add_45_attnAdd1 × 33792 × 6144
457RMSNorm_45_2RMSNorm1 × 33792 × 6144
458FFN_45SwiGLU1 × 33792 × 6144
459Add_45_ffnAdd1 × 33792 × 6144
460RMSNorm_46_1RMSNorm1 × 33792 × 6144
461Attention_46Grouped Query Attn1 × 33792 × 6144
462Add_46_attnAdd1 × 33792 × 6144
463RMSNorm_46_2RMSNorm1 × 33792 × 6144
464FFN_46SwiGLU1 × 33792 × 6144
465Add_46_ffnAdd1 × 33792 × 6144
466RMSNorm_47_1RMSNorm1 × 33792 × 6144
467Attention_47Grouped Query Attn1 × 33792 × 6144
468Add_47_attnAdd1 × 33792 × 6144
469RMSNorm_47_2RMSNorm1 × 33792 × 6144
470FFN_47SwiGLU1 × 33792 × 6144
471Add_47_ffnAdd1 × 33792 × 6144
472RMSNorm_48_1RMSNorm1 × 33792 × 6144
473Attention_48Grouped Query Attn1 × 33792 × 6144
474Add_48_attnAdd1 × 33792 × 6144
475RMSNorm_48_2RMSNorm1 × 33792 × 6144
476FFN_48SwiGLU1 × 33792 × 6144
477Add_48_ffnAdd1 × 33792 × 6144
478Final_RMSNormRMSNorm1 × 33792 × 6144
479LM_HeadLinear1 × 33792 × 92553
480OutputOutput1 × 33792 × 92553

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
infoAt 93 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
warnAcross 93 attention layers this design caches 755 KB per token, so a single 8,192-token sequence needs ~6.3 GB of KV cache before weights or activations. That exceeds the 4 GB budget this rule assumes for serving headroom. Fix: Cut KV width: raise the GQA ratio (fewer numKVHeads), switch to MLA, reduce depth or embedDim, or accept a shorter serving context.
kv-cache-context-budget

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 OpenGVLab/InternVL2-26B --plan --share

Other internvl_chat checkpoints

InternVL2-1B
935M derived · -0.38% against the checkpoint
InternVL2-2B
2.20B derived · -0.48% against the checkpoint
InternVL2_5-4B
3.70B derived · -0.28% against the checkpoint
InternVL3-1B
935M derived · -0.38% against the checkpoint