Models / kimi_linear

Kimi-Linear-48B-A3B-Instruct: GPU memory to fine-tune

Fine-tuning Kimi-Linear-48B-A3B-Instruct at batch 1 and 512 tokens takes about 941 GB of GPU memory for a full fine-tune (at least 8 x H200 sharded), and 100.0 GB with LoRA r=16 on a bf16 base (borderline on one H200 141 GB). Estimated from the model's structure, reading the config under transformers semantics, not measured. What fits on which card.

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

48.91B

48,914,851,072 parameters

In the published checkpoint

49.12B

49,122,681,728 scalars · safetensors.total, read 2025-12-16

Delta

-0.42%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `configuration_kimi.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, 512 tokens over 50,000 samples (assumed; no dataset attached).

Layers

166

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$57.23

Memory: what fits on which card

Largest batch per card. Training is sized at 512 tokens per sample, serving at 512 tokens per sequence.

Full fine-tune, batch 1 at 512 tokens

941 GB

LoRA r=16, batch 1 at 512 tokens

100.0 GB

Serve: bf16 weights

97.8 GB

Serve: KV cache, 512 tokens

16 MB

Card	Full fine-tune	LoRA r=16	Serve bf16, 512 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	does not fit at batch 1
A100 / H100 80 GB	does not fit at batch 1	does not fit at batch 1	does not fit at batch 1
H200 141 GB	does not fit at batch 1	batch 11, borderline	batch 1555
Sharded, by memory alone	at least 8 x H200 or 14 x 80 GB

Training: 512 tokens per sample (2048, or the model's context when shorter), 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. "Borderline" means the card holds batch 1 on the estimate but not on the conservative peak (the estimate x 1.43, at or above the measured peak on 95% of held-out runs): measure one step before reserving it. "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 512 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 512 tokens: the config's maximum context.

	Layer	Type	Output shape
1	Input	Input	1 × 512
2	Embedding	Embedding	1 × 512 × 2304
3	RoPE	RoPE	1 × 512 × 2304
4	RMSNorm_1_1	RMSNorm	1 × 512 × 2304
5	Attention_1	Multi-Head Latent Attention	1 × 512 × 2304
6	Add_1_attn	Add	1 × 512 × 2304
7	RMSNorm_1_2	RMSNorm	1 × 512 × 2304
8	FFN_1	SwiGLU	1 × 512 × 2304
9	Add_1_ffn	Add	1 × 512 × 2304
10	RMSNorm_2_1	RMSNorm	1 × 512 × 2304
11	Attention_2	Multi-Head Latent Attention	1 × 512 × 2304
12	Add_2_attn	Add	1 × 512 × 2304
13	RMSNorm_2_2	RMSNorm	1 × 512 × 2304
14	MoE_2	Shared-Expert MoE	1 × 512 × 2304
15	Add_2_ffn	Add	1 × 512 × 2304
16	RMSNorm_3_1	RMSNorm	1 × 512 × 2304
17	Attention_3	Multi-Head Latent Attention	1 × 512 × 2304
18	Add_3_attn	Add	1 × 512 × 2304
19	RMSNorm_3_2	RMSNorm	1 × 512 × 2304
20	MoE_3	Shared-Expert MoE	1 × 512 × 2304
21	Add_3_ffn	Add	1 × 512 × 2304
22	RMSNorm_4_1	RMSNorm	1 × 512 × 2304
23	Attention_4	Multi-Head Latent Attention	1 × 512 × 2304
24	Add_4_attn	Add	1 × 512 × 2304
25	RMSNorm_4_2	RMSNorm	1 × 512 × 2304
26	MoE_4	Shared-Expert MoE	1 × 512 × 2304
27	Add_4_ffn	Add	1 × 512 × 2304
28	RMSNorm_5_1	RMSNorm	1 × 512 × 2304
29	Attention_5	Multi-Head Latent Attention	1 × 512 × 2304
30	Add_5_attn	Add	1 × 512 × 2304
31	RMSNorm_5_2	RMSNorm	1 × 512 × 2304
32	MoE_5	Shared-Expert MoE	1 × 512 × 2304
33	Add_5_ffn	Add	1 × 512 × 2304
34	RMSNorm_6_1	RMSNorm	1 × 512 × 2304
35	Attention_6	Multi-Head Latent Attention	1 × 512 × 2304
36	Add_6_attn	Add	1 × 512 × 2304
37	RMSNorm_6_2	RMSNorm	1 × 512 × 2304
38	MoE_6	Shared-Expert MoE	1 × 512 × 2304
39	Add_6_ffn	Add	1 × 512 × 2304
40	RMSNorm_7_1	RMSNorm	1 × 512 × 2304
41	Attention_7	Multi-Head Latent Attention	1 × 512 × 2304
42	Add_7_attn	Add	1 × 512 × 2304
43	RMSNorm_7_2	RMSNorm	1 × 512 × 2304
44	MoE_7	Shared-Expert MoE	1 × 512 × 2304
45	Add_7_ffn	Add	1 × 512 × 2304
46	RMSNorm_8_1	RMSNorm	1 × 512 × 2304
47	Attention_8	Multi-Head Latent Attention	1 × 512 × 2304
48	Add_8_attn	Add	1 × 512 × 2304
49	RMSNorm_8_2	RMSNorm	1 × 512 × 2304
50	MoE_8	Shared-Expert MoE	1 × 512 × 2304
51	Add_8_ffn	Add	1 × 512 × 2304
52	RMSNorm_9_1	RMSNorm	1 × 512 × 2304
53	Attention_9	Multi-Head Latent Attention	1 × 512 × 2304
54	Add_9_attn	Add	1 × 512 × 2304
55	RMSNorm_9_2	RMSNorm	1 × 512 × 2304
56	MoE_9	Shared-Expert MoE	1 × 512 × 2304
57	Add_9_ffn	Add	1 × 512 × 2304
58	RMSNorm_10_1	RMSNorm	1 × 512 × 2304
59	Attention_10	Multi-Head Latent Attention	1 × 512 × 2304
60	Add_10_attn	Add	1 × 512 × 2304
61	RMSNorm_10_2	RMSNorm	1 × 512 × 2304
62	MoE_10	Shared-Expert MoE	1 × 512 × 2304
63	Add_10_ffn	Add	1 × 512 × 2304
64	RMSNorm_11_1	RMSNorm	1 × 512 × 2304
65	Attention_11	Multi-Head Latent Attention	1 × 512 × 2304
66	Add_11_attn	Add	1 × 512 × 2304
67	RMSNorm_11_2	RMSNorm	1 × 512 × 2304
68	MoE_11	Shared-Expert MoE	1 × 512 × 2304
69	Add_11_ffn	Add	1 × 512 × 2304
70	RMSNorm_12_1	RMSNorm	1 × 512 × 2304
71	Attention_12	Multi-Head Latent Attention	1 × 512 × 2304
72	Add_12_attn	Add	1 × 512 × 2304
73	RMSNorm_12_2	RMSNorm	1 × 512 × 2304
74	MoE_12	Shared-Expert MoE	1 × 512 × 2304
75	Add_12_ffn	Add	1 × 512 × 2304
76	RMSNorm_13_1	RMSNorm	1 × 512 × 2304
77	Attention_13	Multi-Head Latent Attention	1 × 512 × 2304
78	Add_13_attn	Add	1 × 512 × 2304
79	RMSNorm_13_2	RMSNorm	1 × 512 × 2304
80	MoE_13	Shared-Expert MoE	1 × 512 × 2304
81	Add_13_ffn	Add	1 × 512 × 2304
82	RMSNorm_14_1	RMSNorm	1 × 512 × 2304
83	Attention_14	Multi-Head Latent Attention	1 × 512 × 2304
84	Add_14_attn	Add	1 × 512 × 2304
85	RMSNorm_14_2	RMSNorm	1 × 512 × 2304
86	MoE_14	Shared-Expert MoE	1 × 512 × 2304
87	Add_14_ffn	Add	1 × 512 × 2304
88	RMSNorm_15_1	RMSNorm	1 × 512 × 2304
89	Attention_15	Multi-Head Latent Attention	1 × 512 × 2304
90	Add_15_attn	Add	1 × 512 × 2304
91	RMSNorm_15_2	RMSNorm	1 × 512 × 2304
92	MoE_15	Shared-Expert MoE	1 × 512 × 2304
93	Add_15_ffn	Add	1 × 512 × 2304
94	RMSNorm_16_1	RMSNorm	1 × 512 × 2304
95	Attention_16	Multi-Head Latent Attention	1 × 512 × 2304
96	Add_16_attn	Add	1 × 512 × 2304
97	RMSNorm_16_2	RMSNorm	1 × 512 × 2304
98	MoE_16	Shared-Expert MoE	1 × 512 × 2304
99	Add_16_ffn	Add	1 × 512 × 2304
100	RMSNorm_17_1	RMSNorm	1 × 512 × 2304
101	Attention_17	Multi-Head Latent Attention	1 × 512 × 2304
102	Add_17_attn	Add	1 × 512 × 2304
103	RMSNorm_17_2	RMSNorm	1 × 512 × 2304
104	MoE_17	Shared-Expert MoE	1 × 512 × 2304
105	Add_17_ffn	Add	1 × 512 × 2304
106	RMSNorm_18_1	RMSNorm	1 × 512 × 2304
107	Attention_18	Multi-Head Latent Attention	1 × 512 × 2304
108	Add_18_attn	Add	1 × 512 × 2304
109	RMSNorm_18_2	RMSNorm	1 × 512 × 2304
110	MoE_18	Shared-Expert MoE	1 × 512 × 2304
111	Add_18_ffn	Add	1 × 512 × 2304
112	RMSNorm_19_1	RMSNorm	1 × 512 × 2304
113	Attention_19	Multi-Head Latent Attention	1 × 512 × 2304
114	Add_19_attn	Add	1 × 512 × 2304
115	RMSNorm_19_2	RMSNorm	1 × 512 × 2304
116	MoE_19	Shared-Expert MoE	1 × 512 × 2304
117	Add_19_ffn	Add	1 × 512 × 2304
118	RMSNorm_20_1	RMSNorm	1 × 512 × 2304
119	Attention_20	Multi-Head Latent Attention	1 × 512 × 2304
120	Add_20_attn	Add	1 × 512 × 2304
121	RMSNorm_20_2	RMSNorm	1 × 512 × 2304
122	MoE_20	Shared-Expert MoE	1 × 512 × 2304
123	Add_20_ffn	Add	1 × 512 × 2304
124	RMSNorm_21_1	RMSNorm	1 × 512 × 2304
125	Attention_21	Multi-Head Latent Attention	1 × 512 × 2304
126	Add_21_attn	Add	1 × 512 × 2304
127	RMSNorm_21_2	RMSNorm	1 × 512 × 2304
128	MoE_21	Shared-Expert MoE	1 × 512 × 2304
129	Add_21_ffn	Add	1 × 512 × 2304
130	RMSNorm_22_1	RMSNorm	1 × 512 × 2304
131	Attention_22	Multi-Head Latent Attention	1 × 512 × 2304
132	Add_22_attn	Add	1 × 512 × 2304
133	RMSNorm_22_2	RMSNorm	1 × 512 × 2304
134	MoE_22	Shared-Expert MoE	1 × 512 × 2304
135	Add_22_ffn	Add	1 × 512 × 2304
136	RMSNorm_23_1	RMSNorm	1 × 512 × 2304
137	Attention_23	Multi-Head Latent Attention	1 × 512 × 2304
138	Add_23_attn	Add	1 × 512 × 2304
139	RMSNorm_23_2	RMSNorm	1 × 512 × 2304
140	MoE_23	Shared-Expert MoE	1 × 512 × 2304
141	Add_23_ffn	Add	1 × 512 × 2304
142	RMSNorm_24_1	RMSNorm	1 × 512 × 2304
143	Attention_24	Multi-Head Latent Attention	1 × 512 × 2304
144	Add_24_attn	Add	1 × 512 × 2304
145	RMSNorm_24_2	RMSNorm	1 × 512 × 2304
146	MoE_24	Shared-Expert MoE	1 × 512 × 2304
147	Add_24_ffn	Add	1 × 512 × 2304
148	RMSNorm_25_1	RMSNorm	1 × 512 × 2304
149	Attention_25	Multi-Head Latent Attention	1 × 512 × 2304
150	Add_25_attn	Add	1 × 512 × 2304
151	RMSNorm_25_2	RMSNorm	1 × 512 × 2304
152	MoE_25	Shared-Expert MoE	1 × 512 × 2304
153	Add_25_ffn	Add	1 × 512 × 2304
154	RMSNorm_26_1	RMSNorm	1 × 512 × 2304
155	Attention_26	Multi-Head Latent Attention	1 × 512 × 2304
156	Add_26_attn	Add	1 × 512 × 2304
157	RMSNorm_26_2	RMSNorm	1 × 512 × 2304
158	MoE_26	Shared-Expert MoE	1 × 512 × 2304
159	Add_26_ffn	Add	1 × 512 × 2304
160	RMSNorm_27_1	RMSNorm	1 × 512 × 2304
161	Attention_27	Multi-Head Latent Attention	1 × 512 × 2304
162	Add_27_attn	Add	1 × 512 × 2304
163	RMSNorm_27_2	RMSNorm	1 × 512 × 2304
164	MoE_27	Shared-Expert MoE	1 × 512 × 2304
165	Add_27_ffn	Add	1 × 512 × 2304
166	Final_RMSNorm	RMSNorm	1 × 512 × 2304
167	LM_Head	Linear	1 × 512 × 163840
168	Output	Output	1 × 512 × 163840

What the verifier says

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 moonshotai/Kimi-Linear-48B-A3B-Instruct: 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 moonshotai/Kimi-Linear-48B-A3B-Instruct --plan --share

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