Phi-4-mini-instruct

Reconstructed from its own config.json with no weights read. 461K 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

3.84B

3,836,018,688 parameters

In the published checkpoint

3.84B

3,836,021,760 scalars · safetensors.total, read 2025-12-10

Delta

-0.00%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_phi3.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

194

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$79013.86

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 131072
2	Embedding	Embedding	1 × 131072 × 3072
3	RoPE	RoPE	1 × 131072 × 3072
4	RMSNorm_1_1	RMSNorm	1 × 131072 × 3072
5	Attention_1	Grouped Query Attn	1 × 131072 × 3072
6	Add_1_attn	Add	1 × 131072 × 3072
7	RMSNorm_1_2	RMSNorm	1 × 131072 × 3072
8	FFN_1	SwiGLU	1 × 131072 × 3072
9	Add_1_ffn	Add	1 × 131072 × 3072
10	RMSNorm_2_1	RMSNorm	1 × 131072 × 3072
11	Attention_2	Grouped Query Attn	1 × 131072 × 3072
12	Add_2_attn	Add	1 × 131072 × 3072
13	RMSNorm_2_2	RMSNorm	1 × 131072 × 3072
14	FFN_2	SwiGLU	1 × 131072 × 3072
15	Add_2_ffn	Add	1 × 131072 × 3072
16	RMSNorm_3_1	RMSNorm	1 × 131072 × 3072
17	Attention_3	Grouped Query Attn	1 × 131072 × 3072
18	Add_3_attn	Add	1 × 131072 × 3072
19	RMSNorm_3_2	RMSNorm	1 × 131072 × 3072
20	FFN_3	SwiGLU	1 × 131072 × 3072
21	Add_3_ffn	Add	1 × 131072 × 3072
22	RMSNorm_4_1	RMSNorm	1 × 131072 × 3072
23	Attention_4	Grouped Query Attn	1 × 131072 × 3072
24	Add_4_attn	Add	1 × 131072 × 3072
25	RMSNorm_4_2	RMSNorm	1 × 131072 × 3072
26	FFN_4	SwiGLU	1 × 131072 × 3072
27	Add_4_ffn	Add	1 × 131072 × 3072
28	RMSNorm_5_1	RMSNorm	1 × 131072 × 3072
29	Attention_5	Grouped Query Attn	1 × 131072 × 3072
30	Add_5_attn	Add	1 × 131072 × 3072
31	RMSNorm_5_2	RMSNorm	1 × 131072 × 3072
32	FFN_5	SwiGLU	1 × 131072 × 3072
33	Add_5_ffn	Add	1 × 131072 × 3072
34	RMSNorm_6_1	RMSNorm	1 × 131072 × 3072
35	Attention_6	Grouped Query Attn	1 × 131072 × 3072
36	Add_6_attn	Add	1 × 131072 × 3072
37	RMSNorm_6_2	RMSNorm	1 × 131072 × 3072
38	FFN_6	SwiGLU	1 × 131072 × 3072
39	Add_6_ffn	Add	1 × 131072 × 3072
40	RMSNorm_7_1	RMSNorm	1 × 131072 × 3072
41	Attention_7	Grouped Query Attn	1 × 131072 × 3072
42	Add_7_attn	Add	1 × 131072 × 3072
43	RMSNorm_7_2	RMSNorm	1 × 131072 × 3072
44	FFN_7	SwiGLU	1 × 131072 × 3072
45	Add_7_ffn	Add	1 × 131072 × 3072
46	RMSNorm_8_1	RMSNorm	1 × 131072 × 3072
47	Attention_8	Grouped Query Attn	1 × 131072 × 3072
48	Add_8_attn	Add	1 × 131072 × 3072
49	RMSNorm_8_2	RMSNorm	1 × 131072 × 3072
50	FFN_8	SwiGLU	1 × 131072 × 3072
51	Add_8_ffn	Add	1 × 131072 × 3072
52	RMSNorm_9_1	RMSNorm	1 × 131072 × 3072
53	Attention_9	Grouped Query Attn	1 × 131072 × 3072
54	Add_9_attn	Add	1 × 131072 × 3072
55	RMSNorm_9_2	RMSNorm	1 × 131072 × 3072
56	FFN_9	SwiGLU	1 × 131072 × 3072
57	Add_9_ffn	Add	1 × 131072 × 3072
58	RMSNorm_10_1	RMSNorm	1 × 131072 × 3072
59	Attention_10	Grouped Query Attn	1 × 131072 × 3072
60	Add_10_attn	Add	1 × 131072 × 3072
61	RMSNorm_10_2	RMSNorm	1 × 131072 × 3072
62	FFN_10	SwiGLU	1 × 131072 × 3072
63	Add_10_ffn	Add	1 × 131072 × 3072
64	RMSNorm_11_1	RMSNorm	1 × 131072 × 3072
65	Attention_11	Grouped Query Attn	1 × 131072 × 3072
66	Add_11_attn	Add	1 × 131072 × 3072
67	RMSNorm_11_2	RMSNorm	1 × 131072 × 3072
68	FFN_11	SwiGLU	1 × 131072 × 3072
69	Add_11_ffn	Add	1 × 131072 × 3072
70	RMSNorm_12_1	RMSNorm	1 × 131072 × 3072
71	Attention_12	Grouped Query Attn	1 × 131072 × 3072
72	Add_12_attn	Add	1 × 131072 × 3072
73	RMSNorm_12_2	RMSNorm	1 × 131072 × 3072
74	FFN_12	SwiGLU	1 × 131072 × 3072
75	Add_12_ffn	Add	1 × 131072 × 3072
76	RMSNorm_13_1	RMSNorm	1 × 131072 × 3072
77	Attention_13	Grouped Query Attn	1 × 131072 × 3072
78	Add_13_attn	Add	1 × 131072 × 3072
79	RMSNorm_13_2	RMSNorm	1 × 131072 × 3072
80	FFN_13	SwiGLU	1 × 131072 × 3072
81	Add_13_ffn	Add	1 × 131072 × 3072
82	RMSNorm_14_1	RMSNorm	1 × 131072 × 3072
83	Attention_14	Grouped Query Attn	1 × 131072 × 3072
84	Add_14_attn	Add	1 × 131072 × 3072
85	RMSNorm_14_2	RMSNorm	1 × 131072 × 3072
86	FFN_14	SwiGLU	1 × 131072 × 3072
87	Add_14_ffn	Add	1 × 131072 × 3072
88	RMSNorm_15_1	RMSNorm	1 × 131072 × 3072
89	Attention_15	Grouped Query Attn	1 × 131072 × 3072
90	Add_15_attn	Add	1 × 131072 × 3072
91	RMSNorm_15_2	RMSNorm	1 × 131072 × 3072
92	FFN_15	SwiGLU	1 × 131072 × 3072
93	Add_15_ffn	Add	1 × 131072 × 3072
94	RMSNorm_16_1	RMSNorm	1 × 131072 × 3072
95	Attention_16	Grouped Query Attn	1 × 131072 × 3072
96	Add_16_attn	Add	1 × 131072 × 3072
97	RMSNorm_16_2	RMSNorm	1 × 131072 × 3072
98	FFN_16	SwiGLU	1 × 131072 × 3072
99	Add_16_ffn	Add	1 × 131072 × 3072
100	RMSNorm_17_1	RMSNorm	1 × 131072 × 3072
101	Attention_17	Grouped Query Attn	1 × 131072 × 3072
102	Add_17_attn	Add	1 × 131072 × 3072
103	RMSNorm_17_2	RMSNorm	1 × 131072 × 3072
104	FFN_17	SwiGLU	1 × 131072 × 3072
105	Add_17_ffn	Add	1 × 131072 × 3072
106	RMSNorm_18_1	RMSNorm	1 × 131072 × 3072
107	Attention_18	Grouped Query Attn	1 × 131072 × 3072
108	Add_18_attn	Add	1 × 131072 × 3072
109	RMSNorm_18_2	RMSNorm	1 × 131072 × 3072
110	FFN_18	SwiGLU	1 × 131072 × 3072
111	Add_18_ffn	Add	1 × 131072 × 3072
112	RMSNorm_19_1	RMSNorm	1 × 131072 × 3072
113	Attention_19	Grouped Query Attn	1 × 131072 × 3072
114	Add_19_attn	Add	1 × 131072 × 3072
115	RMSNorm_19_2	RMSNorm	1 × 131072 × 3072
116	FFN_19	SwiGLU	1 × 131072 × 3072
117	Add_19_ffn	Add	1 × 131072 × 3072
118	RMSNorm_20_1	RMSNorm	1 × 131072 × 3072
119	Attention_20	Grouped Query Attn	1 × 131072 × 3072
120	Add_20_attn	Add	1 × 131072 × 3072
121	RMSNorm_20_2	RMSNorm	1 × 131072 × 3072
122	FFN_20	SwiGLU	1 × 131072 × 3072
123	Add_20_ffn	Add	1 × 131072 × 3072
124	RMSNorm_21_1	RMSNorm	1 × 131072 × 3072
125	Attention_21	Grouped Query Attn	1 × 131072 × 3072
126	Add_21_attn	Add	1 × 131072 × 3072
127	RMSNorm_21_2	RMSNorm	1 × 131072 × 3072
128	FFN_21	SwiGLU	1 × 131072 × 3072
129	Add_21_ffn	Add	1 × 131072 × 3072
130	RMSNorm_22_1	RMSNorm	1 × 131072 × 3072
131	Attention_22	Grouped Query Attn	1 × 131072 × 3072
132	Add_22_attn	Add	1 × 131072 × 3072
133	RMSNorm_22_2	RMSNorm	1 × 131072 × 3072
134	FFN_22	SwiGLU	1 × 131072 × 3072
135	Add_22_ffn	Add	1 × 131072 × 3072
136	RMSNorm_23_1	RMSNorm	1 × 131072 × 3072
137	Attention_23	Grouped Query Attn	1 × 131072 × 3072
138	Add_23_attn	Add	1 × 131072 × 3072
139	RMSNorm_23_2	RMSNorm	1 × 131072 × 3072
140	FFN_23	SwiGLU	1 × 131072 × 3072
141	Add_23_ffn	Add	1 × 131072 × 3072
142	RMSNorm_24_1	RMSNorm	1 × 131072 × 3072
143	Attention_24	Grouped Query Attn	1 × 131072 × 3072
144	Add_24_attn	Add	1 × 131072 × 3072
145	RMSNorm_24_2	RMSNorm	1 × 131072 × 3072
146	FFN_24	SwiGLU	1 × 131072 × 3072
147	Add_24_ffn	Add	1 × 131072 × 3072
148	RMSNorm_25_1	RMSNorm	1 × 131072 × 3072
149	Attention_25	Grouped Query Attn	1 × 131072 × 3072
150	Add_25_attn	Add	1 × 131072 × 3072
151	RMSNorm_25_2	RMSNorm	1 × 131072 × 3072
152	FFN_25	SwiGLU	1 × 131072 × 3072
153	Add_25_ffn	Add	1 × 131072 × 3072
154	RMSNorm_26_1	RMSNorm	1 × 131072 × 3072
155	Attention_26	Grouped Query Attn	1 × 131072 × 3072
156	Add_26_attn	Add	1 × 131072 × 3072
157	RMSNorm_26_2	RMSNorm	1 × 131072 × 3072
158	FFN_26	SwiGLU	1 × 131072 × 3072
159	Add_26_ffn	Add	1 × 131072 × 3072
160	RMSNorm_27_1	RMSNorm	1 × 131072 × 3072
161	Attention_27	Grouped Query Attn	1 × 131072 × 3072
162	Add_27_attn	Add	1 × 131072 × 3072
163	RMSNorm_27_2	RMSNorm	1 × 131072 × 3072
164	FFN_27	SwiGLU	1 × 131072 × 3072
165	Add_27_ffn	Add	1 × 131072 × 3072
166	RMSNorm_28_1	RMSNorm	1 × 131072 × 3072
167	Attention_28	Grouped Query Attn	1 × 131072 × 3072
168	Add_28_attn	Add	1 × 131072 × 3072
169	RMSNorm_28_2	RMSNorm	1 × 131072 × 3072
170	FFN_28	SwiGLU	1 × 131072 × 3072
171	Add_28_ffn	Add	1 × 131072 × 3072
172	RMSNorm_29_1	RMSNorm	1 × 131072 × 3072
173	Attention_29	Grouped Query Attn	1 × 131072 × 3072
174	Add_29_attn	Add	1 × 131072 × 3072
175	RMSNorm_29_2	RMSNorm	1 × 131072 × 3072
176	FFN_29	SwiGLU	1 × 131072 × 3072
177	Add_29_ffn	Add	1 × 131072 × 3072
178	RMSNorm_30_1	RMSNorm	1 × 131072 × 3072
179	Attention_30	Grouped Query Attn	1 × 131072 × 3072
180	Add_30_attn	Add	1 × 131072 × 3072
181	RMSNorm_30_2	RMSNorm	1 × 131072 × 3072
182	FFN_30	SwiGLU	1 × 131072 × 3072
183	Add_30_ffn	Add	1 × 131072 × 3072
184	RMSNorm_31_1	RMSNorm	1 × 131072 × 3072
185	Attention_31	Grouped Query Attn	1 × 131072 × 3072
186	Add_31_attn	Add	1 × 131072 × 3072
187	RMSNorm_31_2	RMSNorm	1 × 131072 × 3072
188	FFN_31	SwiGLU	1 × 131072 × 3072
189	Add_31_ffn	Add	1 × 131072 × 3072
190	RMSNorm_32_1	RMSNorm	1 × 131072 × 3072
191	Attention_32	Grouped Query Attn	1 × 131072 × 3072
192	Add_32_attn	Add	1 × 131072 × 3072
193	RMSNorm_32_2	RMSNorm	1 × 131072 × 3072
194	FFN_32	SwiGLU	1 × 131072 × 3072
195	Add_32_ffn	Add	1 × 131072 × 3072
196	Output	Output	1 × 131072 × 3072

What the verifier says

infoAt 32 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 microsoft/Phi-4-mini-instruct --plan --share

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

Other phi3 checkpoints

Phi-3.5-mini-instruct

3.82B derived · +0.00% against the checkpoint

Phi-3-mini-128k-instruct

3.82B derived · +0.00% against the checkpoint

Phi-3-mini-4k-instruct

3.82B derived · +0.00% against the checkpoint

phi-4

14.66B derived · +0.00% against the checkpoint