Models / phimoe

Phi-tiny-MoE-instruct

Reconstructed from its own config.json with no weights read. 145K 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

3.75B

3,754,724,672 parameters

In the published checkpoint

3.76B

3,755,220,288 scalars · safetensors.total, read 2025-12-10

Delta

-0.01%

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

196

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$218.62

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 × 4096
2	Embedding	Embedding	1 × 4096 × 4096
3	RoPE	RoPE	1 × 4096 × 4096
4	RMSNorm_1_1	RMSNorm	1 × 4096 × 4096
5	Attention_1	Grouped Query Attn	1 × 4096 × 4096
6	Add_1_attn	Add	1 × 4096 × 4096
7	RMSNorm_1_2	RMSNorm	1 × 4096 × 4096
8	MoE_1	MoE Layer	1 × 4096 × 4096
9	Add_1_ffn	Add	1 × 4096 × 4096
10	RMSNorm_2_1	RMSNorm	1 × 4096 × 4096
11	Attention_2	Grouped Query Attn	1 × 4096 × 4096
12	Add_2_attn	Add	1 × 4096 × 4096
13	RMSNorm_2_2	RMSNorm	1 × 4096 × 4096
14	MoE_2	MoE Layer	1 × 4096 × 4096
15	Add_2_ffn	Add	1 × 4096 × 4096
16	RMSNorm_3_1	RMSNorm	1 × 4096 × 4096
17	Attention_3	Grouped Query Attn	1 × 4096 × 4096
18	Add_3_attn	Add	1 × 4096 × 4096
19	RMSNorm_3_2	RMSNorm	1 × 4096 × 4096
20	MoE_3	MoE Layer	1 × 4096 × 4096
21	Add_3_ffn	Add	1 × 4096 × 4096
22	RMSNorm_4_1	RMSNorm	1 × 4096 × 4096
23	Attention_4	Grouped Query Attn	1 × 4096 × 4096
24	Add_4_attn	Add	1 × 4096 × 4096
25	RMSNorm_4_2	RMSNorm	1 × 4096 × 4096
26	MoE_4	MoE Layer	1 × 4096 × 4096
27	Add_4_ffn	Add	1 × 4096 × 4096
28	RMSNorm_5_1	RMSNorm	1 × 4096 × 4096
29	Attention_5	Grouped Query Attn	1 × 4096 × 4096
30	Add_5_attn	Add	1 × 4096 × 4096
31	RMSNorm_5_2	RMSNorm	1 × 4096 × 4096
32	MoE_5	MoE Layer	1 × 4096 × 4096
33	Add_5_ffn	Add	1 × 4096 × 4096
34	RMSNorm_6_1	RMSNorm	1 × 4096 × 4096
35	Attention_6	Grouped Query Attn	1 × 4096 × 4096
36	Add_6_attn	Add	1 × 4096 × 4096
37	RMSNorm_6_2	RMSNorm	1 × 4096 × 4096
38	MoE_6	MoE Layer	1 × 4096 × 4096
39	Add_6_ffn	Add	1 × 4096 × 4096
40	RMSNorm_7_1	RMSNorm	1 × 4096 × 4096
41	Attention_7	Grouped Query Attn	1 × 4096 × 4096
42	Add_7_attn	Add	1 × 4096 × 4096
43	RMSNorm_7_2	RMSNorm	1 × 4096 × 4096
44	MoE_7	MoE Layer	1 × 4096 × 4096
45	Add_7_ffn	Add	1 × 4096 × 4096
46	RMSNorm_8_1	RMSNorm	1 × 4096 × 4096
47	Attention_8	Grouped Query Attn	1 × 4096 × 4096
48	Add_8_attn	Add	1 × 4096 × 4096
49	RMSNorm_8_2	RMSNorm	1 × 4096 × 4096
50	MoE_8	MoE Layer	1 × 4096 × 4096
51	Add_8_ffn	Add	1 × 4096 × 4096
52	RMSNorm_9_1	RMSNorm	1 × 4096 × 4096
53	Attention_9	Grouped Query Attn	1 × 4096 × 4096
54	Add_9_attn	Add	1 × 4096 × 4096
55	RMSNorm_9_2	RMSNorm	1 × 4096 × 4096
56	MoE_9	MoE Layer	1 × 4096 × 4096
57	Add_9_ffn	Add	1 × 4096 × 4096
58	RMSNorm_10_1	RMSNorm	1 × 4096 × 4096
59	Attention_10	Grouped Query Attn	1 × 4096 × 4096
60	Add_10_attn	Add	1 × 4096 × 4096
61	RMSNorm_10_2	RMSNorm	1 × 4096 × 4096
62	MoE_10	MoE Layer	1 × 4096 × 4096
63	Add_10_ffn	Add	1 × 4096 × 4096
64	RMSNorm_11_1	RMSNorm	1 × 4096 × 4096
65	Attention_11	Grouped Query Attn	1 × 4096 × 4096
66	Add_11_attn	Add	1 × 4096 × 4096
67	RMSNorm_11_2	RMSNorm	1 × 4096 × 4096
68	MoE_11	MoE Layer	1 × 4096 × 4096
69	Add_11_ffn	Add	1 × 4096 × 4096
70	RMSNorm_12_1	RMSNorm	1 × 4096 × 4096
71	Attention_12	Grouped Query Attn	1 × 4096 × 4096
72	Add_12_attn	Add	1 × 4096 × 4096
73	RMSNorm_12_2	RMSNorm	1 × 4096 × 4096
74	MoE_12	MoE Layer	1 × 4096 × 4096
75	Add_12_ffn	Add	1 × 4096 × 4096
76	RMSNorm_13_1	RMSNorm	1 × 4096 × 4096
77	Attention_13	Grouped Query Attn	1 × 4096 × 4096
78	Add_13_attn	Add	1 × 4096 × 4096
79	RMSNorm_13_2	RMSNorm	1 × 4096 × 4096
80	MoE_13	MoE Layer	1 × 4096 × 4096
81	Add_13_ffn	Add	1 × 4096 × 4096
82	RMSNorm_14_1	RMSNorm	1 × 4096 × 4096
83	Attention_14	Grouped Query Attn	1 × 4096 × 4096
84	Add_14_attn	Add	1 × 4096 × 4096
85	RMSNorm_14_2	RMSNorm	1 × 4096 × 4096
86	MoE_14	MoE Layer	1 × 4096 × 4096
87	Add_14_ffn	Add	1 × 4096 × 4096
88	RMSNorm_15_1	RMSNorm	1 × 4096 × 4096
89	Attention_15	Grouped Query Attn	1 × 4096 × 4096
90	Add_15_attn	Add	1 × 4096 × 4096
91	RMSNorm_15_2	RMSNorm	1 × 4096 × 4096
92	MoE_15	MoE Layer	1 × 4096 × 4096
93	Add_15_ffn	Add	1 × 4096 × 4096
94	RMSNorm_16_1	RMSNorm	1 × 4096 × 4096
95	Attention_16	Grouped Query Attn	1 × 4096 × 4096
96	Add_16_attn	Add	1 × 4096 × 4096
97	RMSNorm_16_2	RMSNorm	1 × 4096 × 4096
98	MoE_16	MoE Layer	1 × 4096 × 4096
99	Add_16_ffn	Add	1 × 4096 × 4096
100	RMSNorm_17_1	RMSNorm	1 × 4096 × 4096
101	Attention_17	Grouped Query Attn	1 × 4096 × 4096
102	Add_17_attn	Add	1 × 4096 × 4096
103	RMSNorm_17_2	RMSNorm	1 × 4096 × 4096
104	MoE_17	MoE Layer	1 × 4096 × 4096
105	Add_17_ffn	Add	1 × 4096 × 4096
106	RMSNorm_18_1	RMSNorm	1 × 4096 × 4096
107	Attention_18	Grouped Query Attn	1 × 4096 × 4096
108	Add_18_attn	Add	1 × 4096 × 4096
109	RMSNorm_18_2	RMSNorm	1 × 4096 × 4096
110	MoE_18	MoE Layer	1 × 4096 × 4096
111	Add_18_ffn	Add	1 × 4096 × 4096
112	RMSNorm_19_1	RMSNorm	1 × 4096 × 4096
113	Attention_19	Grouped Query Attn	1 × 4096 × 4096
114	Add_19_attn	Add	1 × 4096 × 4096
115	RMSNorm_19_2	RMSNorm	1 × 4096 × 4096
116	MoE_19	MoE Layer	1 × 4096 × 4096
117	Add_19_ffn	Add	1 × 4096 × 4096
118	RMSNorm_20_1	RMSNorm	1 × 4096 × 4096
119	Attention_20	Grouped Query Attn	1 × 4096 × 4096
120	Add_20_attn	Add	1 × 4096 × 4096
121	RMSNorm_20_2	RMSNorm	1 × 4096 × 4096
122	MoE_20	MoE Layer	1 × 4096 × 4096
123	Add_20_ffn	Add	1 × 4096 × 4096
124	RMSNorm_21_1	RMSNorm	1 × 4096 × 4096
125	Attention_21	Grouped Query Attn	1 × 4096 × 4096
126	Add_21_attn	Add	1 × 4096 × 4096
127	RMSNorm_21_2	RMSNorm	1 × 4096 × 4096
128	MoE_21	MoE Layer	1 × 4096 × 4096
129	Add_21_ffn	Add	1 × 4096 × 4096
130	RMSNorm_22_1	RMSNorm	1 × 4096 × 4096
131	Attention_22	Grouped Query Attn	1 × 4096 × 4096
132	Add_22_attn	Add	1 × 4096 × 4096
133	RMSNorm_22_2	RMSNorm	1 × 4096 × 4096
134	MoE_22	MoE Layer	1 × 4096 × 4096
135	Add_22_ffn	Add	1 × 4096 × 4096
136	RMSNorm_23_1	RMSNorm	1 × 4096 × 4096
137	Attention_23	Grouped Query Attn	1 × 4096 × 4096
138	Add_23_attn	Add	1 × 4096 × 4096
139	RMSNorm_23_2	RMSNorm	1 × 4096 × 4096
140	MoE_23	MoE Layer	1 × 4096 × 4096
141	Add_23_ffn	Add	1 × 4096 × 4096
142	RMSNorm_24_1	RMSNorm	1 × 4096 × 4096
143	Attention_24	Grouped Query Attn	1 × 4096 × 4096
144	Add_24_attn	Add	1 × 4096 × 4096
145	RMSNorm_24_2	RMSNorm	1 × 4096 × 4096
146	MoE_24	MoE Layer	1 × 4096 × 4096
147	Add_24_ffn	Add	1 × 4096 × 4096
148	RMSNorm_25_1	RMSNorm	1 × 4096 × 4096
149	Attention_25	Grouped Query Attn	1 × 4096 × 4096
150	Add_25_attn	Add	1 × 4096 × 4096
151	RMSNorm_25_2	RMSNorm	1 × 4096 × 4096
152	MoE_25	MoE Layer	1 × 4096 × 4096
153	Add_25_ffn	Add	1 × 4096 × 4096
154	RMSNorm_26_1	RMSNorm	1 × 4096 × 4096
155	Attention_26	Grouped Query Attn	1 × 4096 × 4096
156	Add_26_attn	Add	1 × 4096 × 4096
157	RMSNorm_26_2	RMSNorm	1 × 4096 × 4096
158	MoE_26	MoE Layer	1 × 4096 × 4096
159	Add_26_ffn	Add	1 × 4096 × 4096
160	RMSNorm_27_1	RMSNorm	1 × 4096 × 4096
161	Attention_27	Grouped Query Attn	1 × 4096 × 4096
162	Add_27_attn	Add	1 × 4096 × 4096
163	RMSNorm_27_2	RMSNorm	1 × 4096 × 4096
164	MoE_27	MoE Layer	1 × 4096 × 4096
165	Add_27_ffn	Add	1 × 4096 × 4096
166	RMSNorm_28_1	RMSNorm	1 × 4096 × 4096
167	Attention_28	Grouped Query Attn	1 × 4096 × 4096
168	Add_28_attn	Add	1 × 4096 × 4096
169	RMSNorm_28_2	RMSNorm	1 × 4096 × 4096
170	MoE_28	MoE Layer	1 × 4096 × 4096
171	Add_28_ffn	Add	1 × 4096 × 4096
172	RMSNorm_29_1	RMSNorm	1 × 4096 × 4096
173	Attention_29	Grouped Query Attn	1 × 4096 × 4096
174	Add_29_attn	Add	1 × 4096 × 4096
175	RMSNorm_29_2	RMSNorm	1 × 4096 × 4096
176	MoE_29	MoE Layer	1 × 4096 × 4096
177	Add_29_ffn	Add	1 × 4096 × 4096
178	RMSNorm_30_1	RMSNorm	1 × 4096 × 4096
179	Attention_30	Grouped Query Attn	1 × 4096 × 4096
180	Add_30_attn	Add	1 × 4096 × 4096
181	RMSNorm_30_2	RMSNorm	1 × 4096 × 4096
182	MoE_30	MoE Layer	1 × 4096 × 4096
183	Add_30_ffn	Add	1 × 4096 × 4096
184	RMSNorm_31_1	RMSNorm	1 × 4096 × 4096
185	Attention_31	Grouped Query Attn	1 × 4096 × 4096
186	Add_31_attn	Add	1 × 4096 × 4096
187	RMSNorm_31_2	RMSNorm	1 × 4096 × 4096
188	MoE_31	MoE Layer	1 × 4096 × 4096
189	Add_31_ffn	Add	1 × 4096 × 4096
190	RMSNorm_32_1	RMSNorm	1 × 4096 × 4096
191	Attention_32	Grouped Query Attn	1 × 4096 × 4096
192	Add_32_attn	Add	1 × 4096 × 4096
193	RMSNorm_32_2	RMSNorm	1 × 4096 × 4096
194	MoE_32	MoE Layer	1 × 4096 × 4096
195	Add_32_ffn	Add	1 × 4096 × 4096
196	Final_RMSNorm	RMSNorm	1 × 4096 × 4096
197	LM_Head	Linear	1 × 4096 × 32064
198	Output	Output	1 × 4096 × 32064

What the verifier says

infoMoE layers require an auxiliary router z-loss + load-balance loss during training to prevent expert collapse. This is not visible in the architecture diagram but must be in the training loop. Applies to all 32: MoE_1, MoE_2, MoE_3, MoE_4, MoE_5, MoE_6, MoE_7, MoE_8, MoE_9, MoE_10, MoE_11, MoE_12, MoE_13, MoE_14, MoE_15, MoE_16, MoE_17, MoE_18, MoE_19, MoE_20, MoE_21, MoE_22, MoE_23, MoE_24, MoE_25, MoE_26, MoE_27, MoE_28, MoE_29, MoE_30, MoE_31, MoE_32. Fix: Add a note on these layers. Typical aux_loss coefficient: 1e-2 (Mixtral/Switch Transformer).
moe-no-aux-loss

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-tiny-MoE-instruct --plan --share

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

Other phimoe checkpoints

Phi-3.5-MoE-instruct

41.87B derived · -0.00% against the checkpoint