Models / qwen3

zerank-2-reranker: GPU memory to fine-tune

Fine-tuning zerank-2-reranker at batch 1 and 2048 tokens takes about 86.0 GB of GPU memory for a full fine-tune (borderline on one H200 141 GB), and 17.5 GB with LoRA r=16 on a bf16 base (borderline on one L4 24 GB). Estimated from the model's structure, not measured. What fits on which card.

Reconstructed from its own config.json with no weights read. 638K 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

4.02B

4,022,456,320 parameters

In the published checkpoint

4.02B

4,022,468,096 scalars · safetensors.total, read 2026-07-24

Delta

-0.00%

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

218

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$301.74

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

86.0 GB

LoRA r=16, batch 1 at 2048 tokens

17.5 GB

Serve: bf16 weights

8.04 GB

Serve: KV cache, 4096 tokens

0.60 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	batch 9
L4 24 GB	does not fit at batch 1	batch 1, borderline	batch 21
A100 40 GB	does not fit at batch 1	batch 2	batch 44
A100 / H100 80 GB	2 cards at batch 1 (174 + 44 layers)	batch 6	batch 101
H200 141 GB	batch 5, borderline	batch 12	batch 189

Training: 2048 tokens per sample (2048, or the model's context when shorter; the structure table is at the config's maximum, 40960), 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. 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

220 nodes. Output shapes are propagated from the input shape, batch dimension excluded, at 40960 tokens: the config's maximum context, not the 2048 the memory above is sized at.

	Layer	Type	Output shape
1	Input	Input	1 × 40960
2	Embedding	Embedding	1 × 40960 × 2560
3	RoPE	RoPE	1 × 40960 × 2560
4	RMSNorm_1_1	RMSNorm	1 × 40960 × 2560
5	Attention_1	Grouped Query Attn	1 × 40960 × 2560
6	Add_1_attn	Add	1 × 40960 × 2560
7	RMSNorm_1_2	RMSNorm	1 × 40960 × 2560
8	FFN_1	SwiGLU	1 × 40960 × 2560
9	Add_1_ffn	Add	1 × 40960 × 2560
10	RMSNorm_2_1	RMSNorm	1 × 40960 × 2560
11	Attention_2	Grouped Query Attn	1 × 40960 × 2560
12	Add_2_attn	Add	1 × 40960 × 2560
13	RMSNorm_2_2	RMSNorm	1 × 40960 × 2560
14	FFN_2	SwiGLU	1 × 40960 × 2560
15	Add_2_ffn	Add	1 × 40960 × 2560
16	RMSNorm_3_1	RMSNorm	1 × 40960 × 2560
17	Attention_3	Grouped Query Attn	1 × 40960 × 2560
18	Add_3_attn	Add	1 × 40960 × 2560
19	RMSNorm_3_2	RMSNorm	1 × 40960 × 2560
20	FFN_3	SwiGLU	1 × 40960 × 2560
21	Add_3_ffn	Add	1 × 40960 × 2560
22	RMSNorm_4_1	RMSNorm	1 × 40960 × 2560
23	Attention_4	Grouped Query Attn	1 × 40960 × 2560
24	Add_4_attn	Add	1 × 40960 × 2560
25	RMSNorm_4_2	RMSNorm	1 × 40960 × 2560
26	FFN_4	SwiGLU	1 × 40960 × 2560
27	Add_4_ffn	Add	1 × 40960 × 2560
28	RMSNorm_5_1	RMSNorm	1 × 40960 × 2560
29	Attention_5	Grouped Query Attn	1 × 40960 × 2560
30	Add_5_attn	Add	1 × 40960 × 2560
31	RMSNorm_5_2	RMSNorm	1 × 40960 × 2560
32	FFN_5	SwiGLU	1 × 40960 × 2560
33	Add_5_ffn	Add	1 × 40960 × 2560
34	RMSNorm_6_1	RMSNorm	1 × 40960 × 2560
35	Attention_6	Grouped Query Attn	1 × 40960 × 2560
36	Add_6_attn	Add	1 × 40960 × 2560
37	RMSNorm_6_2	RMSNorm	1 × 40960 × 2560
38	FFN_6	SwiGLU	1 × 40960 × 2560
39	Add_6_ffn	Add	1 × 40960 × 2560
40	RMSNorm_7_1	RMSNorm	1 × 40960 × 2560
41	Attention_7	Grouped Query Attn	1 × 40960 × 2560
42	Add_7_attn	Add	1 × 40960 × 2560
43	RMSNorm_7_2	RMSNorm	1 × 40960 × 2560
44	FFN_7	SwiGLU	1 × 40960 × 2560
45	Add_7_ffn	Add	1 × 40960 × 2560
46	RMSNorm_8_1	RMSNorm	1 × 40960 × 2560
47	Attention_8	Grouped Query Attn	1 × 40960 × 2560
48	Add_8_attn	Add	1 × 40960 × 2560
49	RMSNorm_8_2	RMSNorm	1 × 40960 × 2560
50	FFN_8	SwiGLU	1 × 40960 × 2560
51	Add_8_ffn	Add	1 × 40960 × 2560
52	RMSNorm_9_1	RMSNorm	1 × 40960 × 2560
53	Attention_9	Grouped Query Attn	1 × 40960 × 2560
54	Add_9_attn	Add	1 × 40960 × 2560
55	RMSNorm_9_2	RMSNorm	1 × 40960 × 2560
56	FFN_9	SwiGLU	1 × 40960 × 2560
57	Add_9_ffn	Add	1 × 40960 × 2560
58	RMSNorm_10_1	RMSNorm	1 × 40960 × 2560
59	Attention_10	Grouped Query Attn	1 × 40960 × 2560
60	Add_10_attn	Add	1 × 40960 × 2560
61	RMSNorm_10_2	RMSNorm	1 × 40960 × 2560
62	FFN_10	SwiGLU	1 × 40960 × 2560
63	Add_10_ffn	Add	1 × 40960 × 2560
64	RMSNorm_11_1	RMSNorm	1 × 40960 × 2560
65	Attention_11	Grouped Query Attn	1 × 40960 × 2560
66	Add_11_attn	Add	1 × 40960 × 2560
67	RMSNorm_11_2	RMSNorm	1 × 40960 × 2560
68	FFN_11	SwiGLU	1 × 40960 × 2560
69	Add_11_ffn	Add	1 × 40960 × 2560
70	RMSNorm_12_1	RMSNorm	1 × 40960 × 2560
71	Attention_12	Grouped Query Attn	1 × 40960 × 2560
72	Add_12_attn	Add	1 × 40960 × 2560
73	RMSNorm_12_2	RMSNorm	1 × 40960 × 2560
74	FFN_12	SwiGLU	1 × 40960 × 2560
75	Add_12_ffn	Add	1 × 40960 × 2560
76	RMSNorm_13_1	RMSNorm	1 × 40960 × 2560
77	Attention_13	Grouped Query Attn	1 × 40960 × 2560
78	Add_13_attn	Add	1 × 40960 × 2560
79	RMSNorm_13_2	RMSNorm	1 × 40960 × 2560
80	FFN_13	SwiGLU	1 × 40960 × 2560
81	Add_13_ffn	Add	1 × 40960 × 2560
82	RMSNorm_14_1	RMSNorm	1 × 40960 × 2560
83	Attention_14	Grouped Query Attn	1 × 40960 × 2560
84	Add_14_attn	Add	1 × 40960 × 2560
85	RMSNorm_14_2	RMSNorm	1 × 40960 × 2560
86	FFN_14	SwiGLU	1 × 40960 × 2560
87	Add_14_ffn	Add	1 × 40960 × 2560
88	RMSNorm_15_1	RMSNorm	1 × 40960 × 2560
89	Attention_15	Grouped Query Attn	1 × 40960 × 2560
90	Add_15_attn	Add	1 × 40960 × 2560
91	RMSNorm_15_2	RMSNorm	1 × 40960 × 2560
92	FFN_15	SwiGLU	1 × 40960 × 2560
93	Add_15_ffn	Add	1 × 40960 × 2560
94	RMSNorm_16_1	RMSNorm	1 × 40960 × 2560
95	Attention_16	Grouped Query Attn	1 × 40960 × 2560
96	Add_16_attn	Add	1 × 40960 × 2560
97	RMSNorm_16_2	RMSNorm	1 × 40960 × 2560
98	FFN_16	SwiGLU	1 × 40960 × 2560
99	Add_16_ffn	Add	1 × 40960 × 2560
100	RMSNorm_17_1	RMSNorm	1 × 40960 × 2560
101	Attention_17	Grouped Query Attn	1 × 40960 × 2560
102	Add_17_attn	Add	1 × 40960 × 2560
103	RMSNorm_17_2	RMSNorm	1 × 40960 × 2560
104	FFN_17	SwiGLU	1 × 40960 × 2560
105	Add_17_ffn	Add	1 × 40960 × 2560
106	RMSNorm_18_1	RMSNorm	1 × 40960 × 2560
107	Attention_18	Grouped Query Attn	1 × 40960 × 2560
108	Add_18_attn	Add	1 × 40960 × 2560
109	RMSNorm_18_2	RMSNorm	1 × 40960 × 2560
110	FFN_18	SwiGLU	1 × 40960 × 2560
111	Add_18_ffn	Add	1 × 40960 × 2560
112	RMSNorm_19_1	RMSNorm	1 × 40960 × 2560
113	Attention_19	Grouped Query Attn	1 × 40960 × 2560
114	Add_19_attn	Add	1 × 40960 × 2560
115	RMSNorm_19_2	RMSNorm	1 × 40960 × 2560
116	FFN_19	SwiGLU	1 × 40960 × 2560
117	Add_19_ffn	Add	1 × 40960 × 2560
118	RMSNorm_20_1	RMSNorm	1 × 40960 × 2560
119	Attention_20	Grouped Query Attn	1 × 40960 × 2560
120	Add_20_attn	Add	1 × 40960 × 2560
121	RMSNorm_20_2	RMSNorm	1 × 40960 × 2560
122	FFN_20	SwiGLU	1 × 40960 × 2560
123	Add_20_ffn	Add	1 × 40960 × 2560
124	RMSNorm_21_1	RMSNorm	1 × 40960 × 2560
125	Attention_21	Grouped Query Attn	1 × 40960 × 2560
126	Add_21_attn	Add	1 × 40960 × 2560
127	RMSNorm_21_2	RMSNorm	1 × 40960 × 2560
128	FFN_21	SwiGLU	1 × 40960 × 2560
129	Add_21_ffn	Add	1 × 40960 × 2560
130	RMSNorm_22_1	RMSNorm	1 × 40960 × 2560
131	Attention_22	Grouped Query Attn	1 × 40960 × 2560
132	Add_22_attn	Add	1 × 40960 × 2560
133	RMSNorm_22_2	RMSNorm	1 × 40960 × 2560
134	FFN_22	SwiGLU	1 × 40960 × 2560
135	Add_22_ffn	Add	1 × 40960 × 2560
136	RMSNorm_23_1	RMSNorm	1 × 40960 × 2560
137	Attention_23	Grouped Query Attn	1 × 40960 × 2560
138	Add_23_attn	Add	1 × 40960 × 2560
139	RMSNorm_23_2	RMSNorm	1 × 40960 × 2560
140	FFN_23	SwiGLU	1 × 40960 × 2560
141	Add_23_ffn	Add	1 × 40960 × 2560
142	RMSNorm_24_1	RMSNorm	1 × 40960 × 2560
143	Attention_24	Grouped Query Attn	1 × 40960 × 2560
144	Add_24_attn	Add	1 × 40960 × 2560
145	RMSNorm_24_2	RMSNorm	1 × 40960 × 2560
146	FFN_24	SwiGLU	1 × 40960 × 2560
147	Add_24_ffn	Add	1 × 40960 × 2560
148	RMSNorm_25_1	RMSNorm	1 × 40960 × 2560
149	Attention_25	Grouped Query Attn	1 × 40960 × 2560
150	Add_25_attn	Add	1 × 40960 × 2560
151	RMSNorm_25_2	RMSNorm	1 × 40960 × 2560
152	FFN_25	SwiGLU	1 × 40960 × 2560
153	Add_25_ffn	Add	1 × 40960 × 2560
154	RMSNorm_26_1	RMSNorm	1 × 40960 × 2560
155	Attention_26	Grouped Query Attn	1 × 40960 × 2560
156	Add_26_attn	Add	1 × 40960 × 2560
157	RMSNorm_26_2	RMSNorm	1 × 40960 × 2560
158	FFN_26	SwiGLU	1 × 40960 × 2560
159	Add_26_ffn	Add	1 × 40960 × 2560
160	RMSNorm_27_1	RMSNorm	1 × 40960 × 2560
161	Attention_27	Grouped Query Attn	1 × 40960 × 2560
162	Add_27_attn	Add	1 × 40960 × 2560
163	RMSNorm_27_2	RMSNorm	1 × 40960 × 2560
164	FFN_27	SwiGLU	1 × 40960 × 2560
165	Add_27_ffn	Add	1 × 40960 × 2560
166	RMSNorm_28_1	RMSNorm	1 × 40960 × 2560
167	Attention_28	Grouped Query Attn	1 × 40960 × 2560
168	Add_28_attn	Add	1 × 40960 × 2560
169	RMSNorm_28_2	RMSNorm	1 × 40960 × 2560
170	FFN_28	SwiGLU	1 × 40960 × 2560
171	Add_28_ffn	Add	1 × 40960 × 2560
172	RMSNorm_29_1	RMSNorm	1 × 40960 × 2560
173	Attention_29	Grouped Query Attn	1 × 40960 × 2560
174	Add_29_attn	Add	1 × 40960 × 2560
175	RMSNorm_29_2	RMSNorm	1 × 40960 × 2560
176	FFN_29	SwiGLU	1 × 40960 × 2560
177	Add_29_ffn	Add	1 × 40960 × 2560
178	RMSNorm_30_1	RMSNorm	1 × 40960 × 2560
179	Attention_30	Grouped Query Attn	1 × 40960 × 2560
180	Add_30_attn	Add	1 × 40960 × 2560
181	RMSNorm_30_2	RMSNorm	1 × 40960 × 2560
182	FFN_30	SwiGLU	1 × 40960 × 2560
183	Add_30_ffn	Add	1 × 40960 × 2560
184	RMSNorm_31_1	RMSNorm	1 × 40960 × 2560
185	Attention_31	Grouped Query Attn	1 × 40960 × 2560
186	Add_31_attn	Add	1 × 40960 × 2560
187	RMSNorm_31_2	RMSNorm	1 × 40960 × 2560
188	FFN_31	SwiGLU	1 × 40960 × 2560
189	Add_31_ffn	Add	1 × 40960 × 2560
190	RMSNorm_32_1	RMSNorm	1 × 40960 × 2560
191	Attention_32	Grouped Query Attn	1 × 40960 × 2560
192	Add_32_attn	Add	1 × 40960 × 2560
193	RMSNorm_32_2	RMSNorm	1 × 40960 × 2560
194	FFN_32	SwiGLU	1 × 40960 × 2560
195	Add_32_ffn	Add	1 × 40960 × 2560
196	RMSNorm_33_1	RMSNorm	1 × 40960 × 2560
197	Attention_33	Grouped Query Attn	1 × 40960 × 2560
198	Add_33_attn	Add	1 × 40960 × 2560
199	RMSNorm_33_2	RMSNorm	1 × 40960 × 2560
200	FFN_33	SwiGLU	1 × 40960 × 2560
201	Add_33_ffn	Add	1 × 40960 × 2560
202	RMSNorm_34_1	RMSNorm	1 × 40960 × 2560
203	Attention_34	Grouped Query Attn	1 × 40960 × 2560
204	Add_34_attn	Add	1 × 40960 × 2560
205	RMSNorm_34_2	RMSNorm	1 × 40960 × 2560
206	FFN_34	SwiGLU	1 × 40960 × 2560
207	Add_34_ffn	Add	1 × 40960 × 2560
208	RMSNorm_35_1	RMSNorm	1 × 40960 × 2560
209	Attention_35	Grouped Query Attn	1 × 40960 × 2560
210	Add_35_attn	Add	1 × 40960 × 2560
211	RMSNorm_35_2	RMSNorm	1 × 40960 × 2560
212	FFN_35	SwiGLU	1 × 40960 × 2560
213	Add_35_ffn	Add	1 × 40960 × 2560
214	RMSNorm_36_1	RMSNorm	1 × 40960 × 2560
215	Attention_36	Grouped Query Attn	1 × 40960 × 2560
216	Add_36_attn	Add	1 × 40960 × 2560
217	RMSNorm_36_2	RMSNorm	1 × 40960 × 2560
218	FFN_36	SwiGLU	1 × 40960 × 2560
219	Add_36_ffn	Add	1 × 40960 × 2560
220	Output	Output	1 × 40960 × 2560

What the verifier says

infoAt 36 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 zeroentropy/zerank-2-reranker: 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 zeroentropy/zerank-2-reranker --plan --share

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

Other qwen3 checkpoints

jina-reranker-v3

596M derived · -0.13% against the checkpoint

Kimi-K3-DSpark

4.26B derived · +89.3% against the checkpoint

Qwen3-0.6B

596M derived · -0.00% against the checkpoint

Qwen3-1.7B

1.72B derived · -0.00% against the checkpoint