Models / sarvam_moe

sarvam-30b

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

32.15B

32,152,907,776 parameters

In the published checkpoint

32.15B

32,152,650,368 scalars · safetensors.total, read 2026-03-23

Delta

+0.00%

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

118

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$66950.39

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

Structure

120 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 × 4096
3	RoPE	RoPE	1 × 131072 × 4096
4	RMSNorm_1_1	RMSNorm	1 × 131072 × 4096
5	Attention_1	Grouped Query Attn	1 × 131072 × 4096
6	Add_1_attn	Add	1 × 131072 × 4096
7	RMSNorm_1_2	RMSNorm	1 × 131072 × 4096
8	FFN_1	SwiGLU	1 × 131072 × 4096
9	Add_1_ffn	Add	1 × 131072 × 4096
10	RMSNorm_2_1	RMSNorm	1 × 131072 × 4096
11	Attention_2	Grouped Query Attn	1 × 131072 × 4096
12	Add_2_attn	Add	1 × 131072 × 4096
13	RMSNorm_2_2	RMSNorm	1 × 131072 × 4096
14	MoE_2	Shared-Expert MoE	1 × 131072 × 4096
15	Add_2_ffn	Add	1 × 131072 × 4096
16	RMSNorm_3_1	RMSNorm	1 × 131072 × 4096
17	Attention_3	Grouped Query Attn	1 × 131072 × 4096
18	Add_3_attn	Add	1 × 131072 × 4096
19	RMSNorm_3_2	RMSNorm	1 × 131072 × 4096
20	MoE_3	Shared-Expert MoE	1 × 131072 × 4096
21	Add_3_ffn	Add	1 × 131072 × 4096
22	RMSNorm_4_1	RMSNorm	1 × 131072 × 4096
23	Attention_4	Grouped Query Attn	1 × 131072 × 4096
24	Add_4_attn	Add	1 × 131072 × 4096
25	RMSNorm_4_2	RMSNorm	1 × 131072 × 4096
26	MoE_4	Shared-Expert MoE	1 × 131072 × 4096
27	Add_4_ffn	Add	1 × 131072 × 4096
28	RMSNorm_5_1	RMSNorm	1 × 131072 × 4096
29	Attention_5	Grouped Query Attn	1 × 131072 × 4096
30	Add_5_attn	Add	1 × 131072 × 4096
31	RMSNorm_5_2	RMSNorm	1 × 131072 × 4096
32	MoE_5	Shared-Expert MoE	1 × 131072 × 4096
33	Add_5_ffn	Add	1 × 131072 × 4096
34	RMSNorm_6_1	RMSNorm	1 × 131072 × 4096
35	Attention_6	Grouped Query Attn	1 × 131072 × 4096
36	Add_6_attn	Add	1 × 131072 × 4096
37	RMSNorm_6_2	RMSNorm	1 × 131072 × 4096
38	MoE_6	Shared-Expert MoE	1 × 131072 × 4096
39	Add_6_ffn	Add	1 × 131072 × 4096
40	RMSNorm_7_1	RMSNorm	1 × 131072 × 4096
41	Attention_7	Grouped Query Attn	1 × 131072 × 4096
42	Add_7_attn	Add	1 × 131072 × 4096
43	RMSNorm_7_2	RMSNorm	1 × 131072 × 4096
44	MoE_7	Shared-Expert MoE	1 × 131072 × 4096
45	Add_7_ffn	Add	1 × 131072 × 4096
46	RMSNorm_8_1	RMSNorm	1 × 131072 × 4096
47	Attention_8	Grouped Query Attn	1 × 131072 × 4096
48	Add_8_attn	Add	1 × 131072 × 4096
49	RMSNorm_8_2	RMSNorm	1 × 131072 × 4096
50	MoE_8	Shared-Expert MoE	1 × 131072 × 4096
51	Add_8_ffn	Add	1 × 131072 × 4096
52	RMSNorm_9_1	RMSNorm	1 × 131072 × 4096
53	Attention_9	Grouped Query Attn	1 × 131072 × 4096
54	Add_9_attn	Add	1 × 131072 × 4096
55	RMSNorm_9_2	RMSNorm	1 × 131072 × 4096
56	MoE_9	Shared-Expert MoE	1 × 131072 × 4096
57	Add_9_ffn	Add	1 × 131072 × 4096
58	RMSNorm_10_1	RMSNorm	1 × 131072 × 4096
59	Attention_10	Grouped Query Attn	1 × 131072 × 4096
60	Add_10_attn	Add	1 × 131072 × 4096
61	RMSNorm_10_2	RMSNorm	1 × 131072 × 4096
62	MoE_10	Shared-Expert MoE	1 × 131072 × 4096
63	Add_10_ffn	Add	1 × 131072 × 4096
64	RMSNorm_11_1	RMSNorm	1 × 131072 × 4096
65	Attention_11	Grouped Query Attn	1 × 131072 × 4096
66	Add_11_attn	Add	1 × 131072 × 4096
67	RMSNorm_11_2	RMSNorm	1 × 131072 × 4096
68	MoE_11	Shared-Expert MoE	1 × 131072 × 4096
69	Add_11_ffn	Add	1 × 131072 × 4096
70	RMSNorm_12_1	RMSNorm	1 × 131072 × 4096
71	Attention_12	Grouped Query Attn	1 × 131072 × 4096
72	Add_12_attn	Add	1 × 131072 × 4096
73	RMSNorm_12_2	RMSNorm	1 × 131072 × 4096
74	MoE_12	Shared-Expert MoE	1 × 131072 × 4096
75	Add_12_ffn	Add	1 × 131072 × 4096
76	RMSNorm_13_1	RMSNorm	1 × 131072 × 4096
77	Attention_13	Grouped Query Attn	1 × 131072 × 4096
78	Add_13_attn	Add	1 × 131072 × 4096
79	RMSNorm_13_2	RMSNorm	1 × 131072 × 4096
80	MoE_13	Shared-Expert MoE	1 × 131072 × 4096
81	Add_13_ffn	Add	1 × 131072 × 4096
82	RMSNorm_14_1	RMSNorm	1 × 131072 × 4096
83	Attention_14	Grouped Query Attn	1 × 131072 × 4096
84	Add_14_attn	Add	1 × 131072 × 4096
85	RMSNorm_14_2	RMSNorm	1 × 131072 × 4096
86	MoE_14	Shared-Expert MoE	1 × 131072 × 4096
87	Add_14_ffn	Add	1 × 131072 × 4096
88	RMSNorm_15_1	RMSNorm	1 × 131072 × 4096
89	Attention_15	Grouped Query Attn	1 × 131072 × 4096
90	Add_15_attn	Add	1 × 131072 × 4096
91	RMSNorm_15_2	RMSNorm	1 × 131072 × 4096
92	MoE_15	Shared-Expert MoE	1 × 131072 × 4096
93	Add_15_ffn	Add	1 × 131072 × 4096
94	RMSNorm_16_1	RMSNorm	1 × 131072 × 4096
95	Attention_16	Grouped Query Attn	1 × 131072 × 4096
96	Add_16_attn	Add	1 × 131072 × 4096
97	RMSNorm_16_2	RMSNorm	1 × 131072 × 4096
98	MoE_16	Shared-Expert MoE	1 × 131072 × 4096
99	Add_16_ffn	Add	1 × 131072 × 4096
100	RMSNorm_17_1	RMSNorm	1 × 131072 × 4096
101	Attention_17	Grouped Query Attn	1 × 131072 × 4096
102	Add_17_attn	Add	1 × 131072 × 4096
103	RMSNorm_17_2	RMSNorm	1 × 131072 × 4096
104	MoE_17	Shared-Expert MoE	1 × 131072 × 4096
105	Add_17_ffn	Add	1 × 131072 × 4096
106	RMSNorm_18_1	RMSNorm	1 × 131072 × 4096
107	Attention_18	Grouped Query Attn	1 × 131072 × 4096
108	Add_18_attn	Add	1 × 131072 × 4096
109	RMSNorm_18_2	RMSNorm	1 × 131072 × 4096
110	MoE_18	Shared-Expert MoE	1 × 131072 × 4096
111	Add_18_ffn	Add	1 × 131072 × 4096
112	RMSNorm_19_1	RMSNorm	1 × 131072 × 4096
113	Attention_19	Grouped Query Attn	1 × 131072 × 4096
114	Add_19_attn	Add	1 × 131072 × 4096
115	RMSNorm_19_2	RMSNorm	1 × 131072 × 4096
116	MoE_19	Shared-Expert MoE	1 × 131072 × 4096
117	Add_19_ffn	Add	1 × 131072 × 4096
118	Final_RMSNorm	RMSNorm	1 × 131072 × 4096
119	LM_Head	Linear	1 × 131072 × 262144
120	Output	Output	1 × 131072 × 262144

What the verifier says

warnLinear "LM_Head" is 4096 × 262144 = 1074M parameters (~4.0 GB float32). A single dense layer this large usually means a feature map was flattened without pooling first; embedding / vocab-projection heads are the expected exception. Fix: Add a Global Average Pool or more downsampling before the Linear, or factorize it (low-rank / bottleneck projection).
huge-linear-params

infoAt 19 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 sarvamai/sarvam-30b --plan --share

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