Models / ultravox

ultravox-v0_5-llama-3_2-1b

Reconstructed from its own config.json with no weights read. 144K 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

2.84B

2,838,197,504 parameters

In the published checkpoint

683M

683,118,592 scalars · safetensors.total, read 2026-03-11

Delta

+315%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `ultravox_config.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

230

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$52.76

Card	Memory
T4 (16GB)	weights + activations	does not fit
A100 (40GB)	weights + activations	does not fit
H100 (80GB)	weights + activations	fits

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 512
2	Embedding	Embedding	1 × 512 × 4096
3	RoPE	RoPE	1 × 512 × 4096
4	Audio input	Input	1 × 128 × 1
5	Audio feature embed	Projection	1 × 128 × 1280
6	Audio_LN_1	LayerNorm	1 × 128 × 1280
7	Audio_Attn_1	Multi-Head Attention	1 × 128 × 1280
8	Audio_Add_1	Add	1 × 128 × 1280
9	Audio_FFN_1	Feed Forward	1 × 128 × 1280
10	Audio_LN_2	LayerNorm	1 × 128 × 1280
11	Audio_Attn_2	Multi-Head Attention	1 × 128 × 1280
12	Audio_Add_2	Add	1 × 128 × 1280
13	Audio_FFN_2	Feed Forward	1 × 128 × 1280
14	Audio_LN_3	LayerNorm	1 × 128 × 1280
15	Audio_Attn_3	Multi-Head Attention	1 × 128 × 1280
16	Audio_Add_3	Add	1 × 128 × 1280
17	Audio_FFN_3	Feed Forward	1 × 128 × 1280
18	Audio_LN_4	LayerNorm	1 × 128 × 1280
19	Audio_Attn_4	Multi-Head Attention	1 × 128 × 1280
20	Audio_Add_4	Add	1 × 128 × 1280
21	Audio_FFN_4	Feed Forward	1 × 128 × 1280
22	Audio_LN_5	LayerNorm	1 × 128 × 1280
23	Audio_Attn_5	Multi-Head Attention	1 × 128 × 1280
24	Audio_Add_5	Add	1 × 128 × 1280
25	Audio_FFN_5	Feed Forward	1 × 128 × 1280
26	Audio_LN_6	LayerNorm	1 × 128 × 1280
27	Audio_Attn_6	Multi-Head Attention	1 × 128 × 1280
28	Audio_Add_6	Add	1 × 128 × 1280
29	Audio_FFN_6	Feed Forward	1 × 128 × 1280
30	Audio_LN_7	LayerNorm	1 × 128 × 1280
31	Audio_Attn_7	Multi-Head Attention	1 × 128 × 1280
32	Audio_Add_7	Add	1 × 128 × 1280
33	Audio_FFN_7	Feed Forward	1 × 128 × 1280
34	Audio_LN_8	LayerNorm	1 × 128 × 1280
35	Audio_Attn_8	Multi-Head Attention	1 × 128 × 1280
36	Audio_Add_8	Add	1 × 128 × 1280
37	Audio_FFN_8	Feed Forward	1 × 128 × 1280
38	Audio_LN_9	LayerNorm	1 × 128 × 1280
39	Audio_Attn_9	Multi-Head Attention	1 × 128 × 1280
40	Audio_Add_9	Add	1 × 128 × 1280
41	Audio_FFN_9	Feed Forward	1 × 128 × 1280
42	Audio_LN_10	LayerNorm	1 × 128 × 1280
43	Audio_Attn_10	Multi-Head Attention	1 × 128 × 1280
44	Audio_Add_10	Add	1 × 128 × 1280
45	Audio_FFN_10	Feed Forward	1 × 128 × 1280
46	Audio_LN_11	LayerNorm	1 × 128 × 1280
47	Audio_Attn_11	Multi-Head Attention	1 × 128 × 1280
48	Audio_Add_11	Add	1 × 128 × 1280
49	Audio_FFN_11	Feed Forward	1 × 128 × 1280
50	Audio_LN_12	LayerNorm	1 × 128 × 1280
51	Audio_Attn_12	Multi-Head Attention	1 × 128 × 1280
52	Audio_Add_12	Add	1 × 128 × 1280
53	Audio_FFN_12	Feed Forward	1 × 128 × 1280
54	Audio_LN_13	LayerNorm	1 × 128 × 1280
55	Audio_Attn_13	Multi-Head Attention	1 × 128 × 1280
56	Audio_Add_13	Add	1 × 128 × 1280
57	Audio_FFN_13	Feed Forward	1 × 128 × 1280
58	Audio_LN_14	LayerNorm	1 × 128 × 1280
59	Audio_Attn_14	Multi-Head Attention	1 × 128 × 1280
60	Audio_Add_14	Add	1 × 128 × 1280
61	Audio_FFN_14	Feed Forward	1 × 128 × 1280
62	Audio_LN_15	LayerNorm	1 × 128 × 1280
63	Audio_Attn_15	Multi-Head Attention	1 × 128 × 1280
64	Audio_Add_15	Add	1 × 128 × 1280
65	Audio_FFN_15	Feed Forward	1 × 128 × 1280
66	Audio_LN_16	LayerNorm	1 × 128 × 1280
67	Audio_Attn_16	Multi-Head Attention	1 × 128 × 1280
68	Audio_Add_16	Add	1 × 128 × 1280
69	Audio_FFN_16	Feed Forward	1 × 128 × 1280
70	Audio_LN_17	LayerNorm	1 × 128 × 1280
71	Audio_Attn_17	Multi-Head Attention	1 × 128 × 1280
72	Audio_Add_17	Add	1 × 128 × 1280
73	Audio_FFN_17	Feed Forward	1 × 128 × 1280
74	Audio_LN_18	LayerNorm	1 × 128 × 1280
75	Audio_Attn_18	Multi-Head Attention	1 × 128 × 1280
76	Audio_Add_18	Add	1 × 128 × 1280
77	Audio_FFN_18	Feed Forward	1 × 128 × 1280
78	Audio_LN_19	LayerNorm	1 × 128 × 1280
79	Audio_Attn_19	Multi-Head Attention	1 × 128 × 1280
80	Audio_Add_19	Add	1 × 128 × 1280
81	Audio_FFN_19	Feed Forward	1 × 128 × 1280
82	Audio_LN_20	LayerNorm	1 × 128 × 1280
83	Audio_Attn_20	Multi-Head Attention	1 × 128 × 1280
84	Audio_Add_20	Add	1 × 128 × 1280
85	Audio_FFN_20	Feed Forward	1 × 128 × 1280
86	Audio_LN_21	LayerNorm	1 × 128 × 1280
87	Audio_Attn_21	Multi-Head Attention	1 × 128 × 1280
88	Audio_Add_21	Add	1 × 128 × 1280
89	Audio_FFN_21	Feed Forward	1 × 128 × 1280
90	Audio_LN_22	LayerNorm	1 × 128 × 1280
91	Audio_Attn_22	Multi-Head Attention	1 × 128 × 1280
92	Audio_Add_22	Add	1 × 128 × 1280
93	Audio_FFN_22	Feed Forward	1 × 128 × 1280
94	Audio_LN_23	LayerNorm	1 × 128 × 1280
95	Audio_Attn_23	Multi-Head Attention	1 × 128 × 1280
96	Audio_Add_23	Add	1 × 128 × 1280
97	Audio_FFN_23	Feed Forward	1 × 128 × 1280
98	Audio_LN_24	LayerNorm	1 × 128 × 1280
99	Audio_Attn_24	Multi-Head Attention	1 × 128 × 1280
100	Audio_Add_24	Add	1 × 128 × 1280
101	Audio_FFN_24	Feed Forward	1 × 128 × 1280
102	Audio_LN_25	LayerNorm	1 × 128 × 1280
103	Audio_Attn_25	Multi-Head Attention	1 × 128 × 1280
104	Audio_Add_25	Add	1 × 128 × 1280
105	Audio_FFN_25	Feed Forward	1 × 128 × 1280
106	Audio_LN_26	LayerNorm	1 × 128 × 1280
107	Audio_Attn_26	Multi-Head Attention	1 × 128 × 1280
108	Audio_Add_26	Add	1 × 128 × 1280
109	Audio_FFN_26	Feed Forward	1 × 128 × 1280
110	Audio_LN_27	LayerNorm	1 × 128 × 1280
111	Audio_Attn_27	Multi-Head Attention	1 × 128 × 1280
112	Audio_Add_27	Add	1 × 128 × 1280
113	Audio_FFN_27	Feed Forward	1 × 128 × 1280
114	Audio_LN_28	LayerNorm	1 × 128 × 1280
115	Audio_Attn_28	Multi-Head Attention	1 × 128 × 1280
116	Audio_Add_28	Add	1 × 128 × 1280
117	Audio_FFN_28	Feed Forward	1 × 128 × 1280
118	Audio_LN_29	LayerNorm	1 × 128 × 1280
119	Audio_Attn_29	Multi-Head Attention	1 × 128 × 1280
120	Audio_Add_29	Add	1 × 128 × 1280
121	Audio_FFN_29	Feed Forward	1 × 128 × 1280
122	Audio_LN_30	LayerNorm	1 × 128 × 1280
123	Audio_Attn_30	Multi-Head Attention	1 × 128 × 1280
124	Audio_Add_30	Add	1 × 128 × 1280
125	Audio_FFN_30	Feed Forward	1 × 128 × 1280
126	Audio_LN_31	LayerNorm	1 × 128 × 1280
127	Audio_Attn_31	Multi-Head Attention	1 × 128 × 1280
128	Audio_Add_31	Add	1 × 128 × 1280
129	Audio_FFN_31	Feed Forward	1 × 128 × 1280
130	Audio_LN_32	LayerNorm	1 × 128 × 1280
131	Audio_Attn_32	Multi-Head Attention	1 × 128 × 1280
132	Audio_Add_32	Add	1 × 128 × 1280
133	Audio_FFN_32	Feed Forward	1 × 128 × 1280
134	Audio projector	Projection	1 × 128 × 4096
135	Audio tokens	Reshape	1 × 256 × 4096
136	Multimodal fusion (concat tokens)	Concatenate	1 × 768 × 4096
137	RMSNorm_1_1	RMSNorm	1 × 768 × 4096
138	Attention_1	Grouped Query Attn	1 × 768 × 4096
139	Add_1_attn	Add	1 × 768 × 4096
140	RMSNorm_1_2	RMSNorm	1 × 768 × 4096
141	FFN_1	SwiGLU	1 × 768 × 4096
142	Add_1_ffn	Add	1 × 768 × 4096
143	RMSNorm_2_1	RMSNorm	1 × 768 × 4096
144	Attention_2	Grouped Query Attn	1 × 768 × 4096
145	Add_2_attn	Add	1 × 768 × 4096
146	RMSNorm_2_2	RMSNorm	1 × 768 × 4096
147	FFN_2	SwiGLU	1 × 768 × 4096
148	Add_2_ffn	Add	1 × 768 × 4096
149	RMSNorm_3_1	RMSNorm	1 × 768 × 4096
150	Attention_3	Grouped Query Attn	1 × 768 × 4096
151	Add_3_attn	Add	1 × 768 × 4096
152	RMSNorm_3_2	RMSNorm	1 × 768 × 4096
153	FFN_3	SwiGLU	1 × 768 × 4096
154	Add_3_ffn	Add	1 × 768 × 4096
155	RMSNorm_4_1	RMSNorm	1 × 768 × 4096
156	Attention_4	Grouped Query Attn	1 × 768 × 4096
157	Add_4_attn	Add	1 × 768 × 4096
158	RMSNorm_4_2	RMSNorm	1 × 768 × 4096
159	FFN_4	SwiGLU	1 × 768 × 4096
160	Add_4_ffn	Add	1 × 768 × 4096
161	RMSNorm_5_1	RMSNorm	1 × 768 × 4096
162	Attention_5	Grouped Query Attn	1 × 768 × 4096
163	Add_5_attn	Add	1 × 768 × 4096
164	RMSNorm_5_2	RMSNorm	1 × 768 × 4096
165	FFN_5	SwiGLU	1 × 768 × 4096
166	Add_5_ffn	Add	1 × 768 × 4096
167	RMSNorm_6_1	RMSNorm	1 × 768 × 4096
168	Attention_6	Grouped Query Attn	1 × 768 × 4096
169	Add_6_attn	Add	1 × 768 × 4096
170	RMSNorm_6_2	RMSNorm	1 × 768 × 4096
171	FFN_6	SwiGLU	1 × 768 × 4096
172	Add_6_ffn	Add	1 × 768 × 4096
173	RMSNorm_7_1	RMSNorm	1 × 768 × 4096
174	Attention_7	Grouped Query Attn	1 × 768 × 4096
175	Add_7_attn	Add	1 × 768 × 4096
176	RMSNorm_7_2	RMSNorm	1 × 768 × 4096
177	FFN_7	SwiGLU	1 × 768 × 4096
178	Add_7_ffn	Add	1 × 768 × 4096
179	RMSNorm_8_1	RMSNorm	1 × 768 × 4096
180	Attention_8	Grouped Query Attn	1 × 768 × 4096
181	Add_8_attn	Add	1 × 768 × 4096
182	RMSNorm_8_2	RMSNorm	1 × 768 × 4096
183	FFN_8	SwiGLU	1 × 768 × 4096
184	Add_8_ffn	Add	1 × 768 × 4096
185	RMSNorm_9_1	RMSNorm	1 × 768 × 4096
186	Attention_9	Grouped Query Attn	1 × 768 × 4096
187	Add_9_attn	Add	1 × 768 × 4096
188	RMSNorm_9_2	RMSNorm	1 × 768 × 4096
189	FFN_9	SwiGLU	1 × 768 × 4096
190	Add_9_ffn	Add	1 × 768 × 4096
191	RMSNorm_10_1	RMSNorm	1 × 768 × 4096
192	Attention_10	Grouped Query Attn	1 × 768 × 4096
193	Add_10_attn	Add	1 × 768 × 4096
194	RMSNorm_10_2	RMSNorm	1 × 768 × 4096
195	FFN_10	SwiGLU	1 × 768 × 4096
196	Add_10_ffn	Add	1 × 768 × 4096
197	RMSNorm_11_1	RMSNorm	1 × 768 × 4096
198	Attention_11	Grouped Query Attn	1 × 768 × 4096
199	Add_11_attn	Add	1 × 768 × 4096
200	RMSNorm_11_2	RMSNorm	1 × 768 × 4096
201	FFN_11	SwiGLU	1 × 768 × 4096
202	Add_11_ffn	Add	1 × 768 × 4096
203	RMSNorm_12_1	RMSNorm	1 × 768 × 4096
204	Attention_12	Grouped Query Attn	1 × 768 × 4096
205	Add_12_attn	Add	1 × 768 × 4096
206	RMSNorm_12_2	RMSNorm	1 × 768 × 4096
207	FFN_12	SwiGLU	1 × 768 × 4096
208	Add_12_ffn	Add	1 × 768 × 4096
209	RMSNorm_13_1	RMSNorm	1 × 768 × 4096
210	Attention_13	Grouped Query Attn	1 × 768 × 4096
211	Add_13_attn	Add	1 × 768 × 4096
212	RMSNorm_13_2	RMSNorm	1 × 768 × 4096
213	FFN_13	SwiGLU	1 × 768 × 4096
214	Add_13_ffn	Add	1 × 768 × 4096
215	RMSNorm_14_1	RMSNorm	1 × 768 × 4096
216	Attention_14	Grouped Query Attn	1 × 768 × 4096
217	Add_14_attn	Add	1 × 768 × 4096
218	RMSNorm_14_2	RMSNorm	1 × 768 × 4096
219	FFN_14	SwiGLU	1 × 768 × 4096
220	Add_14_ffn	Add	1 × 768 × 4096
221	RMSNorm_15_1	RMSNorm	1 × 768 × 4096
222	Attention_15	Grouped Query Attn	1 × 768 × 4096
223	Add_15_attn	Add	1 × 768 × 4096
224	RMSNorm_15_2	RMSNorm	1 × 768 × 4096
225	FFN_15	SwiGLU	1 × 768 × 4096
226	Add_15_ffn	Add	1 × 768 × 4096
227	RMSNorm_16_1	RMSNorm	1 × 768 × 4096
228	Attention_16	Grouped Query Attn	1 × 768 × 4096
229	Add_16_attn	Add	1 × 768 × 4096
230	RMSNorm_16_2	RMSNorm	1 × 768 × 4096
231	FFN_16	SwiGLU	1 × 768 × 4096
232	Add_16_ffn	Add	1 × 768 × 4096
233	Output	Output	1 × 768 × 4096

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: 3072 (0.75× embedDim). Expected: ~11008. Fix: Set intermediateSize to 11008 for embedDim=4096.
swiglu-dim-convention

info48 attention layers at embedDim 2219 cache full per-head K/V: about 416 KB per token at fp16, which dominates memory at long context. Grouped-query attention (e.g. 8:1) would cut this ~8×; multi-head latent attention (MLA) shrinks it ~10× or more. This is the move production LLMs make; it does not change the parameter count. Fix: Switch attention to groupedQueryAttention (set numKVHeads below numHeads, e.g. numHeads/4) or mla (a low-rank cached latent).
full-mha-serving-cost

infoAt 48 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

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 fixie-ai/ultravox-v0_5-llama-3_2-1b --plan --share

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