phi-2

Reconstructed from its own config.json with no weights read. 1.3M 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

2.78B

2,779,520,000 parameters

In the published checkpoint

2.78B

2,779,683,840 scalars · safetensors.total, read 2025-12-08

Delta

-0.01%

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

196

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$215.87

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 2048
2	Embedding	Embedding	1 × 2048 × 2560
3	RoPE	RoPE	1 × 2048 × 2560
4	LayerNorm_1_1	LayerNorm	1 × 2048 × 2560
5	Attention_1	Grouped Query Attn	1 × 2048 × 2560
6	Add_1_attn	Add	1 × 2048 × 2560
7	LayerNorm_1_2	LayerNorm	1 × 2048 × 2560
8	FFN_1	Feed Forward	1 × 2048 × 2560
9	Add_1_ffn	Add	1 × 2048 × 2560
10	LayerNorm_2_1	LayerNorm	1 × 2048 × 2560
11	Attention_2	Grouped Query Attn	1 × 2048 × 2560
12	Add_2_attn	Add	1 × 2048 × 2560
13	LayerNorm_2_2	LayerNorm	1 × 2048 × 2560
14	FFN_2	Feed Forward	1 × 2048 × 2560
15	Add_2_ffn	Add	1 × 2048 × 2560
16	LayerNorm_3_1	LayerNorm	1 × 2048 × 2560
17	Attention_3	Grouped Query Attn	1 × 2048 × 2560
18	Add_3_attn	Add	1 × 2048 × 2560
19	LayerNorm_3_2	LayerNorm	1 × 2048 × 2560
20	FFN_3	Feed Forward	1 × 2048 × 2560
21	Add_3_ffn	Add	1 × 2048 × 2560
22	LayerNorm_4_1	LayerNorm	1 × 2048 × 2560
23	Attention_4	Grouped Query Attn	1 × 2048 × 2560
24	Add_4_attn	Add	1 × 2048 × 2560
25	LayerNorm_4_2	LayerNorm	1 × 2048 × 2560
26	FFN_4	Feed Forward	1 × 2048 × 2560
27	Add_4_ffn	Add	1 × 2048 × 2560
28	LayerNorm_5_1	LayerNorm	1 × 2048 × 2560
29	Attention_5	Grouped Query Attn	1 × 2048 × 2560
30	Add_5_attn	Add	1 × 2048 × 2560
31	LayerNorm_5_2	LayerNorm	1 × 2048 × 2560
32	FFN_5	Feed Forward	1 × 2048 × 2560
33	Add_5_ffn	Add	1 × 2048 × 2560
34	LayerNorm_6_1	LayerNorm	1 × 2048 × 2560
35	Attention_6	Grouped Query Attn	1 × 2048 × 2560
36	Add_6_attn	Add	1 × 2048 × 2560
37	LayerNorm_6_2	LayerNorm	1 × 2048 × 2560
38	FFN_6	Feed Forward	1 × 2048 × 2560
39	Add_6_ffn	Add	1 × 2048 × 2560
40	LayerNorm_7_1	LayerNorm	1 × 2048 × 2560
41	Attention_7	Grouped Query Attn	1 × 2048 × 2560
42	Add_7_attn	Add	1 × 2048 × 2560
43	LayerNorm_7_2	LayerNorm	1 × 2048 × 2560
44	FFN_7	Feed Forward	1 × 2048 × 2560
45	Add_7_ffn	Add	1 × 2048 × 2560
46	LayerNorm_8_1	LayerNorm	1 × 2048 × 2560
47	Attention_8	Grouped Query Attn	1 × 2048 × 2560
48	Add_8_attn	Add	1 × 2048 × 2560
49	LayerNorm_8_2	LayerNorm	1 × 2048 × 2560
50	FFN_8	Feed Forward	1 × 2048 × 2560
51	Add_8_ffn	Add	1 × 2048 × 2560
52	LayerNorm_9_1	LayerNorm	1 × 2048 × 2560
53	Attention_9	Grouped Query Attn	1 × 2048 × 2560
54	Add_9_attn	Add	1 × 2048 × 2560
55	LayerNorm_9_2	LayerNorm	1 × 2048 × 2560
56	FFN_9	Feed Forward	1 × 2048 × 2560
57	Add_9_ffn	Add	1 × 2048 × 2560
58	LayerNorm_10_1	LayerNorm	1 × 2048 × 2560
59	Attention_10	Grouped Query Attn	1 × 2048 × 2560
60	Add_10_attn	Add	1 × 2048 × 2560
61	LayerNorm_10_2	LayerNorm	1 × 2048 × 2560
62	FFN_10	Feed Forward	1 × 2048 × 2560
63	Add_10_ffn	Add	1 × 2048 × 2560
64	LayerNorm_11_1	LayerNorm	1 × 2048 × 2560
65	Attention_11	Grouped Query Attn	1 × 2048 × 2560
66	Add_11_attn	Add	1 × 2048 × 2560
67	LayerNorm_11_2	LayerNorm	1 × 2048 × 2560
68	FFN_11	Feed Forward	1 × 2048 × 2560
69	Add_11_ffn	Add	1 × 2048 × 2560
70	LayerNorm_12_1	LayerNorm	1 × 2048 × 2560
71	Attention_12	Grouped Query Attn	1 × 2048 × 2560
72	Add_12_attn	Add	1 × 2048 × 2560
73	LayerNorm_12_2	LayerNorm	1 × 2048 × 2560
74	FFN_12	Feed Forward	1 × 2048 × 2560
75	Add_12_ffn	Add	1 × 2048 × 2560
76	LayerNorm_13_1	LayerNorm	1 × 2048 × 2560
77	Attention_13	Grouped Query Attn	1 × 2048 × 2560
78	Add_13_attn	Add	1 × 2048 × 2560
79	LayerNorm_13_2	LayerNorm	1 × 2048 × 2560
80	FFN_13	Feed Forward	1 × 2048 × 2560
81	Add_13_ffn	Add	1 × 2048 × 2560
82	LayerNorm_14_1	LayerNorm	1 × 2048 × 2560
83	Attention_14	Grouped Query Attn	1 × 2048 × 2560
84	Add_14_attn	Add	1 × 2048 × 2560
85	LayerNorm_14_2	LayerNorm	1 × 2048 × 2560
86	FFN_14	Feed Forward	1 × 2048 × 2560
87	Add_14_ffn	Add	1 × 2048 × 2560
88	LayerNorm_15_1	LayerNorm	1 × 2048 × 2560
89	Attention_15	Grouped Query Attn	1 × 2048 × 2560
90	Add_15_attn	Add	1 × 2048 × 2560
91	LayerNorm_15_2	LayerNorm	1 × 2048 × 2560
92	FFN_15	Feed Forward	1 × 2048 × 2560
93	Add_15_ffn	Add	1 × 2048 × 2560
94	LayerNorm_16_1	LayerNorm	1 × 2048 × 2560
95	Attention_16	Grouped Query Attn	1 × 2048 × 2560
96	Add_16_attn	Add	1 × 2048 × 2560
97	LayerNorm_16_2	LayerNorm	1 × 2048 × 2560
98	FFN_16	Feed Forward	1 × 2048 × 2560
99	Add_16_ffn	Add	1 × 2048 × 2560
100	LayerNorm_17_1	LayerNorm	1 × 2048 × 2560
101	Attention_17	Grouped Query Attn	1 × 2048 × 2560
102	Add_17_attn	Add	1 × 2048 × 2560
103	LayerNorm_17_2	LayerNorm	1 × 2048 × 2560
104	FFN_17	Feed Forward	1 × 2048 × 2560
105	Add_17_ffn	Add	1 × 2048 × 2560
106	LayerNorm_18_1	LayerNorm	1 × 2048 × 2560
107	Attention_18	Grouped Query Attn	1 × 2048 × 2560
108	Add_18_attn	Add	1 × 2048 × 2560
109	LayerNorm_18_2	LayerNorm	1 × 2048 × 2560
110	FFN_18	Feed Forward	1 × 2048 × 2560
111	Add_18_ffn	Add	1 × 2048 × 2560
112	LayerNorm_19_1	LayerNorm	1 × 2048 × 2560
113	Attention_19	Grouped Query Attn	1 × 2048 × 2560
114	Add_19_attn	Add	1 × 2048 × 2560
115	LayerNorm_19_2	LayerNorm	1 × 2048 × 2560
116	FFN_19	Feed Forward	1 × 2048 × 2560
117	Add_19_ffn	Add	1 × 2048 × 2560
118	LayerNorm_20_1	LayerNorm	1 × 2048 × 2560
119	Attention_20	Grouped Query Attn	1 × 2048 × 2560
120	Add_20_attn	Add	1 × 2048 × 2560
121	LayerNorm_20_2	LayerNorm	1 × 2048 × 2560
122	FFN_20	Feed Forward	1 × 2048 × 2560
123	Add_20_ffn	Add	1 × 2048 × 2560
124	LayerNorm_21_1	LayerNorm	1 × 2048 × 2560
125	Attention_21	Grouped Query Attn	1 × 2048 × 2560
126	Add_21_attn	Add	1 × 2048 × 2560
127	LayerNorm_21_2	LayerNorm	1 × 2048 × 2560
128	FFN_21	Feed Forward	1 × 2048 × 2560
129	Add_21_ffn	Add	1 × 2048 × 2560
130	LayerNorm_22_1	LayerNorm	1 × 2048 × 2560
131	Attention_22	Grouped Query Attn	1 × 2048 × 2560
132	Add_22_attn	Add	1 × 2048 × 2560
133	LayerNorm_22_2	LayerNorm	1 × 2048 × 2560
134	FFN_22	Feed Forward	1 × 2048 × 2560
135	Add_22_ffn	Add	1 × 2048 × 2560
136	LayerNorm_23_1	LayerNorm	1 × 2048 × 2560
137	Attention_23	Grouped Query Attn	1 × 2048 × 2560
138	Add_23_attn	Add	1 × 2048 × 2560
139	LayerNorm_23_2	LayerNorm	1 × 2048 × 2560
140	FFN_23	Feed Forward	1 × 2048 × 2560
141	Add_23_ffn	Add	1 × 2048 × 2560
142	LayerNorm_24_1	LayerNorm	1 × 2048 × 2560
143	Attention_24	Grouped Query Attn	1 × 2048 × 2560
144	Add_24_attn	Add	1 × 2048 × 2560
145	LayerNorm_24_2	LayerNorm	1 × 2048 × 2560
146	FFN_24	Feed Forward	1 × 2048 × 2560
147	Add_24_ffn	Add	1 × 2048 × 2560
148	LayerNorm_25_1	LayerNorm	1 × 2048 × 2560
149	Attention_25	Grouped Query Attn	1 × 2048 × 2560
150	Add_25_attn	Add	1 × 2048 × 2560
151	LayerNorm_25_2	LayerNorm	1 × 2048 × 2560
152	FFN_25	Feed Forward	1 × 2048 × 2560
153	Add_25_ffn	Add	1 × 2048 × 2560
154	LayerNorm_26_1	LayerNorm	1 × 2048 × 2560
155	Attention_26	Grouped Query Attn	1 × 2048 × 2560
156	Add_26_attn	Add	1 × 2048 × 2560
157	LayerNorm_26_2	LayerNorm	1 × 2048 × 2560
158	FFN_26	Feed Forward	1 × 2048 × 2560
159	Add_26_ffn	Add	1 × 2048 × 2560
160	LayerNorm_27_1	LayerNorm	1 × 2048 × 2560
161	Attention_27	Grouped Query Attn	1 × 2048 × 2560
162	Add_27_attn	Add	1 × 2048 × 2560
163	LayerNorm_27_2	LayerNorm	1 × 2048 × 2560
164	FFN_27	Feed Forward	1 × 2048 × 2560
165	Add_27_ffn	Add	1 × 2048 × 2560
166	LayerNorm_28_1	LayerNorm	1 × 2048 × 2560
167	Attention_28	Grouped Query Attn	1 × 2048 × 2560
168	Add_28_attn	Add	1 × 2048 × 2560
169	LayerNorm_28_2	LayerNorm	1 × 2048 × 2560
170	FFN_28	Feed Forward	1 × 2048 × 2560
171	Add_28_ffn	Add	1 × 2048 × 2560
172	LayerNorm_29_1	LayerNorm	1 × 2048 × 2560
173	Attention_29	Grouped Query Attn	1 × 2048 × 2560
174	Add_29_attn	Add	1 × 2048 × 2560
175	LayerNorm_29_2	LayerNorm	1 × 2048 × 2560
176	FFN_29	Feed Forward	1 × 2048 × 2560
177	Add_29_ffn	Add	1 × 2048 × 2560
178	LayerNorm_30_1	LayerNorm	1 × 2048 × 2560
179	Attention_30	Grouped Query Attn	1 × 2048 × 2560
180	Add_30_attn	Add	1 × 2048 × 2560
181	LayerNorm_30_2	LayerNorm	1 × 2048 × 2560
182	FFN_30	Feed Forward	1 × 2048 × 2560
183	Add_30_ffn	Add	1 × 2048 × 2560
184	LayerNorm_31_1	LayerNorm	1 × 2048 × 2560
185	Attention_31	Grouped Query Attn	1 × 2048 × 2560
186	Add_31_attn	Add	1 × 2048 × 2560
187	LayerNorm_31_2	LayerNorm	1 × 2048 × 2560
188	FFN_31	Feed Forward	1 × 2048 × 2560
189	Add_31_ffn	Add	1 × 2048 × 2560
190	LayerNorm_32_1	LayerNorm	1 × 2048 × 2560
191	Attention_32	Grouped Query Attn	1 × 2048 × 2560
192	Add_32_attn	Add	1 × 2048 × 2560
193	LayerNorm_32_2	LayerNorm	1 × 2048 × 2560
194	FFN_32	Feed Forward	1 × 2048 × 2560
195	Add_32_ffn	Add	1 × 2048 × 2560
196	Final_LayerNorm	LayerNorm	1 × 2048 × 2560
197	LM_Head	Linear	1 × 2048 × 51200
198	Output	Output	1 × 2048 × 51200

What the verifier says

info32 attention layers at embedDim 2560 cache full per-head K/V: about 320 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 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-2 --plan --share

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