Models / deepseek_v3

Instella-MoE-16B-A3B-Think-AWQ-INT4: GPU memory to fine-tune

Fine-tuning Instella-MoE-16B-A3B-Think-AWQ-INT4 at batch 1 and 2048 tokens takes about 310 GB of GPU memory for a full fine-tune (at least 3 x H200 sharded), and 39.5 GB with LoRA r=16 on a bf16 base (fits one A100 / H100 80 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 117K 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

15.75B

15,749,670,272 parameters

In the published checkpoint

15.86B

15,862,789,248 scalars · safetensors.total, read 2026-07-30

Delta

-0.71%

quantized This checkpoint is stored quantized (compressed-tensors). The tensor count in the file counts stored elements under a packing scheme, not logical parameters, so the two numbers below are not measuring the same thing in either direction.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

166

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$195.80

Memory: what fits on which card

Largest batch per card. Training is sized at 2048 tokens per sample, serving at 4096 tokens per sequence.

Full fine-tune, batch 1 at 2048 tokens

310 GB

LoRA r=16, batch 1 at 2048 tokens

39.5 GB

Serve: bf16 weights

31.5 GB

Serve: KV cache, 4096 tokens

0.12 GB

Card	Full fine-tune	LoRA r=16	Serve bf16, 4096 tokens
T4 16 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
L4 24 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
A100 40 GB	does not fit at batch 1	does not fit at batch 1	batch 27
A100 / H100 80 GB	does not fit at batch 1	batch 4	batch 316
H200 141 GB	does not fit at batch 1	batch 11	batch 757
Sharded, by memory alone	at least 3 x H200 or 5 x 80 GB

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 32768), AdamW with bf16 autocast, SDPA attention (no score matrix stored), 1.15x headroom on every card. Full fine-tune holds weights, gradients and optimizer state (1.2 x 16 bytes per parameter) plus activations. LoRA is r=16 on every linear layer but the head, base frozen in bf16 (2 bytes per parameter), adapters at full AdamW state, same activations. A two-card split is a contiguous layer cut (pipeline parallel), computed only where one card cannot hold batch 1. "Sharded" is a floor: the batch-1 peak with headroom divided across cards as if it split perfectly evenly (FSDP / ZeRO-3 for training, tensor parallel for serving), with no communication buffers counted. Serving: bf16 weights plus the KV cache of 4096 tokens per sequence; prefill working set, CUDA context and framework reserve not counted.

These are estimates from the graph, not measurements. Against measured runs, held out by architecture, the estimate flagged 113 of the 131 jobs that did not fit their card and said "fits" on the other 18. Every measured run.

Structure

168 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 32768 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 32768
2	Embedding	Embedding	1 × 32768 × 2048
3	RoPE	RoPE	1 × 32768 × 2048
4	RMSNorm_1_1	RMSNorm	1 × 32768 × 2048
5	Attention_1	Multi-Head Latent Attention	1 × 32768 × 2048
6	Add_1_attn	Add	1 × 32768 × 2048
7	RMSNorm_1_2	RMSNorm	1 × 32768 × 2048
8	FFN_1	SwiGLU	1 × 32768 × 2048
9	Add_1_ffn	Add	1 × 32768 × 2048
10	RMSNorm_2_1	RMSNorm	1 × 32768 × 2048
11	Attention_2	Multi-Head Latent Attention	1 × 32768 × 2048
12	Add_2_attn	Add	1 × 32768 × 2048
13	RMSNorm_2_2	RMSNorm	1 × 32768 × 2048
14	MoE_2	Shared-Expert MoE	1 × 32768 × 2048
15	Add_2_ffn	Add	1 × 32768 × 2048
16	RMSNorm_3_1	RMSNorm	1 × 32768 × 2048
17	Attention_3	Multi-Head Latent Attention	1 × 32768 × 2048
18	Add_3_attn	Add	1 × 32768 × 2048
19	RMSNorm_3_2	RMSNorm	1 × 32768 × 2048
20	MoE_3	Shared-Expert MoE	1 × 32768 × 2048
21	Add_3_ffn	Add	1 × 32768 × 2048
22	RMSNorm_4_1	RMSNorm	1 × 32768 × 2048
23	Attention_4	Multi-Head Latent Attention	1 × 32768 × 2048
24	Add_4_attn	Add	1 × 32768 × 2048
25	RMSNorm_4_2	RMSNorm	1 × 32768 × 2048
26	MoE_4	Shared-Expert MoE	1 × 32768 × 2048
27	Add_4_ffn	Add	1 × 32768 × 2048
28	RMSNorm_5_1	RMSNorm	1 × 32768 × 2048
29	Attention_5	Multi-Head Latent Attention	1 × 32768 × 2048
30	Add_5_attn	Add	1 × 32768 × 2048
31	RMSNorm_5_2	RMSNorm	1 × 32768 × 2048
32	MoE_5	Shared-Expert MoE	1 × 32768 × 2048
33	Add_5_ffn	Add	1 × 32768 × 2048
34	RMSNorm_6_1	RMSNorm	1 × 32768 × 2048
35	Attention_6	Multi-Head Latent Attention	1 × 32768 × 2048
36	Add_6_attn	Add	1 × 32768 × 2048
37	RMSNorm_6_2	RMSNorm	1 × 32768 × 2048
38	MoE_6	Shared-Expert MoE	1 × 32768 × 2048
39	Add_6_ffn	Add	1 × 32768 × 2048
40	RMSNorm_7_1	RMSNorm	1 × 32768 × 2048
41	Attention_7	Multi-Head Latent Attention	1 × 32768 × 2048
42	Add_7_attn	Add	1 × 32768 × 2048
43	RMSNorm_7_2	RMSNorm	1 × 32768 × 2048
44	MoE_7	Shared-Expert MoE	1 × 32768 × 2048
45	Add_7_ffn	Add	1 × 32768 × 2048
46	RMSNorm_8_1	RMSNorm	1 × 32768 × 2048
47	Attention_8	Multi-Head Latent Attention	1 × 32768 × 2048
48	Add_8_attn	Add	1 × 32768 × 2048
49	RMSNorm_8_2	RMSNorm	1 × 32768 × 2048
50	MoE_8	Shared-Expert MoE	1 × 32768 × 2048
51	Add_8_ffn	Add	1 × 32768 × 2048
52	RMSNorm_9_1	RMSNorm	1 × 32768 × 2048
53	Attention_9	Multi-Head Latent Attention	1 × 32768 × 2048
54	Add_9_attn	Add	1 × 32768 × 2048
55	RMSNorm_9_2	RMSNorm	1 × 32768 × 2048
56	MoE_9	Shared-Expert MoE	1 × 32768 × 2048
57	Add_9_ffn	Add	1 × 32768 × 2048
58	RMSNorm_10_1	RMSNorm	1 × 32768 × 2048
59	Attention_10	Multi-Head Latent Attention	1 × 32768 × 2048
60	Add_10_attn	Add	1 × 32768 × 2048
61	RMSNorm_10_2	RMSNorm	1 × 32768 × 2048
62	MoE_10	Shared-Expert MoE	1 × 32768 × 2048
63	Add_10_ffn	Add	1 × 32768 × 2048
64	RMSNorm_11_1	RMSNorm	1 × 32768 × 2048
65	Attention_11	Multi-Head Latent Attention	1 × 32768 × 2048
66	Add_11_attn	Add	1 × 32768 × 2048
67	RMSNorm_11_2	RMSNorm	1 × 32768 × 2048
68	MoE_11	Shared-Expert MoE	1 × 32768 × 2048
69	Add_11_ffn	Add	1 × 32768 × 2048
70	RMSNorm_12_1	RMSNorm	1 × 32768 × 2048
71	Attention_12	Multi-Head Latent Attention	1 × 32768 × 2048
72	Add_12_attn	Add	1 × 32768 × 2048
73	RMSNorm_12_2	RMSNorm	1 × 32768 × 2048
74	MoE_12	Shared-Expert MoE	1 × 32768 × 2048
75	Add_12_ffn	Add	1 × 32768 × 2048
76	RMSNorm_13_1	RMSNorm	1 × 32768 × 2048
77	Attention_13	Multi-Head Latent Attention	1 × 32768 × 2048
78	Add_13_attn	Add	1 × 32768 × 2048
79	RMSNorm_13_2	RMSNorm	1 × 32768 × 2048
80	MoE_13	Shared-Expert MoE	1 × 32768 × 2048
81	Add_13_ffn	Add	1 × 32768 × 2048
82	RMSNorm_14_1	RMSNorm	1 × 32768 × 2048
83	Attention_14	Multi-Head Latent Attention	1 × 32768 × 2048
84	Add_14_attn	Add	1 × 32768 × 2048
85	RMSNorm_14_2	RMSNorm	1 × 32768 × 2048
86	MoE_14	Shared-Expert MoE	1 × 32768 × 2048
87	Add_14_ffn	Add	1 × 32768 × 2048
88	RMSNorm_15_1	RMSNorm	1 × 32768 × 2048
89	Attention_15	Multi-Head Latent Attention	1 × 32768 × 2048
90	Add_15_attn	Add	1 × 32768 × 2048
91	RMSNorm_15_2	RMSNorm	1 × 32768 × 2048
92	MoE_15	Shared-Expert MoE	1 × 32768 × 2048
93	Add_15_ffn	Add	1 × 32768 × 2048
94	RMSNorm_16_1	RMSNorm	1 × 32768 × 2048
95	Attention_16	Multi-Head Latent Attention	1 × 32768 × 2048
96	Add_16_attn	Add	1 × 32768 × 2048
97	RMSNorm_16_2	RMSNorm	1 × 32768 × 2048
98	MoE_16	Shared-Expert MoE	1 × 32768 × 2048
99	Add_16_ffn	Add	1 × 32768 × 2048
100	RMSNorm_17_1	RMSNorm	1 × 32768 × 2048
101	Attention_17	Multi-Head Latent Attention	1 × 32768 × 2048
102	Add_17_attn	Add	1 × 32768 × 2048
103	RMSNorm_17_2	RMSNorm	1 × 32768 × 2048
104	MoE_17	Shared-Expert MoE	1 × 32768 × 2048
105	Add_17_ffn	Add	1 × 32768 × 2048
106	RMSNorm_18_1	RMSNorm	1 × 32768 × 2048
107	Attention_18	Multi-Head Latent Attention	1 × 32768 × 2048
108	Add_18_attn	Add	1 × 32768 × 2048
109	RMSNorm_18_2	RMSNorm	1 × 32768 × 2048
110	MoE_18	Shared-Expert MoE	1 × 32768 × 2048
111	Add_18_ffn	Add	1 × 32768 × 2048
112	RMSNorm_19_1	RMSNorm	1 × 32768 × 2048
113	Attention_19	Multi-Head Latent Attention	1 × 32768 × 2048
114	Add_19_attn	Add	1 × 32768 × 2048
115	RMSNorm_19_2	RMSNorm	1 × 32768 × 2048
116	MoE_19	Shared-Expert MoE	1 × 32768 × 2048
117	Add_19_ffn	Add	1 × 32768 × 2048
118	RMSNorm_20_1	RMSNorm	1 × 32768 × 2048
119	Attention_20	Multi-Head Latent Attention	1 × 32768 × 2048
120	Add_20_attn	Add	1 × 32768 × 2048
121	RMSNorm_20_2	RMSNorm	1 × 32768 × 2048
122	MoE_20	Shared-Expert MoE	1 × 32768 × 2048
123	Add_20_ffn	Add	1 × 32768 × 2048
124	RMSNorm_21_1	RMSNorm	1 × 32768 × 2048
125	Attention_21	Multi-Head Latent Attention	1 × 32768 × 2048
126	Add_21_attn	Add	1 × 32768 × 2048
127	RMSNorm_21_2	RMSNorm	1 × 32768 × 2048
128	MoE_21	Shared-Expert MoE	1 × 32768 × 2048
129	Add_21_ffn	Add	1 × 32768 × 2048
130	RMSNorm_22_1	RMSNorm	1 × 32768 × 2048
131	Attention_22	Multi-Head Latent Attention	1 × 32768 × 2048
132	Add_22_attn	Add	1 × 32768 × 2048
133	RMSNorm_22_2	RMSNorm	1 × 32768 × 2048
134	MoE_22	Shared-Expert MoE	1 × 32768 × 2048
135	Add_22_ffn	Add	1 × 32768 × 2048
136	RMSNorm_23_1	RMSNorm	1 × 32768 × 2048
137	Attention_23	Multi-Head Latent Attention	1 × 32768 × 2048
138	Add_23_attn	Add	1 × 32768 × 2048
139	RMSNorm_23_2	RMSNorm	1 × 32768 × 2048
140	MoE_23	Shared-Expert MoE	1 × 32768 × 2048
141	Add_23_ffn	Add	1 × 32768 × 2048
142	RMSNorm_24_1	RMSNorm	1 × 32768 × 2048
143	Attention_24	Multi-Head Latent Attention	1 × 32768 × 2048
144	Add_24_attn	Add	1 × 32768 × 2048
145	RMSNorm_24_2	RMSNorm	1 × 32768 × 2048
146	MoE_24	Shared-Expert MoE	1 × 32768 × 2048
147	Add_24_ffn	Add	1 × 32768 × 2048
148	RMSNorm_25_1	RMSNorm	1 × 32768 × 2048
149	Attention_25	Multi-Head Latent Attention	1 × 32768 × 2048
150	Add_25_attn	Add	1 × 32768 × 2048
151	RMSNorm_25_2	RMSNorm	1 × 32768 × 2048
152	MoE_25	Shared-Expert MoE	1 × 32768 × 2048
153	Add_25_ffn	Add	1 × 32768 × 2048
154	RMSNorm_26_1	RMSNorm	1 × 32768 × 2048
155	Attention_26	Multi-Head Latent Attention	1 × 32768 × 2048
156	Add_26_attn	Add	1 × 32768 × 2048
157	RMSNorm_26_2	RMSNorm	1 × 32768 × 2048
158	MoE_26	Shared-Expert MoE	1 × 32768 × 2048
159	Add_26_ffn	Add	1 × 32768 × 2048
160	RMSNorm_27_1	RMSNorm	1 × 32768 × 2048
161	Attention_27	Multi-Head Latent Attention	1 × 32768 × 2048
162	Add_27_attn	Add	1 × 32768 × 2048
163	RMSNorm_27_2	RMSNorm	1 × 32768 × 2048
164	MoE_27	Shared-Expert MoE	1 × 32768 × 2048
165	Add_27_ffn	Add	1 × 32768 × 2048
166	Final_RMSNorm	RMSNorm	1 × 32768 × 2048
167	LM_Head	Linear	1 × 32768 × 128896
168	Output	Output	1 × 32768 × 128896

What the verifier says

infoLLaMA uses intermediateSize ≈ ⌊(8/3 × D) / 256⌋ × 256. Current: 10944 (5.34× embedDim). Expected: ~5376. Fix: Set intermediateSize to 5376 for embedDim=2048.
swiglu-dim-convention

infoAt 27 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

What changed upstream

The change tracker follows cyankiwi/Instella-MoE-16B-A3B-Think-AWQ-INT4: every revision it has read, what changed in the config, tokenizer, chat template or weights between them, and an OSV advisory per change.

Revision history and changes Badge JSON OSV advisories All tracked changes

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 cyankiwi/Instella-MoE-16B-A3B-Think-AWQ-INT4 --plan --share

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

Other deepseek_v3 checkpoints

DeepSeek-R1

684B derived · +0.00% against the checkpoint

DeepSeek-R1-0528

684B derived · -0.01% against the checkpoint

DeepSeek-R1-MXFP4

684B derived · +84.6% against the checkpoint

DeepSeek-V3

684B derived · -0.01% against the checkpoint