Models / qwen2_vl

jina-reranker-m0: GPU memory to fine-tune

Fine-tuning jina-reranker-m0 at batch 1 and 2048 tokens takes about 35.4 GB of GPU memory for a full fine-tune (fits one A100 / H100 80 GB), and 9.11 GB with LoRA r=16 on a bf16 base (fits one T4 16 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. 383K 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

1.55B

1,547,500,032 parameters

In the published checkpoint

2.44B

2,444,721,665 scalars · safetensors.total, read 2026-04-09

Delta

-27.2%

stored twice The checkpoint stores the tied embedding twice: +233.4M (`lm_head.weight` and `model.embed_tokens.weight`, 151936 x 1536). The model has one such matrix and the count above has it once; the delta compares that count plus the second copy against the file. Header read 2026-10-18.

custom-code This repository ships its own modeling code (`auto_map`, e.g. `modeling.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, 2048 tokens over 50,000 samples (assumed; no dataset attached).

Layers

175

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$118.23

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

35.4 GB

LoRA r=16, batch 1 at 2048 tokens

9.11 GB

Serve: bf16 weights

3.10 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	batch 1	batch 92
L4 24 GB	2 cards at batch 1 (95 + 80 layers)	batch 3	batch 151
A100 40 GB	2 cards at batch 1 (173 + 2 layers)	batch 5	batch 269
A100 / H100 80 GB	batch 7	batch 11	batch 565
H200 141 GB	batch 16	batch 21	batch 1017

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. 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

178 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 × 1536
3	RoPE	RoPE	1 × 32768 × 1536
4	Vision input	Input	3 × 224 × 224
5	PatchEmbed	Patch Embed	196 × 1536
6	Patch_Position_Embedding	Learned Pos Embed	196 × 1536
7	Vision encoder (internals not in config)	Projection	196 × 1536
8	Vision tokens	Reshape	1 × 196 × 1536
9	Multimodal fusion (concat tokens)	Concatenate	1 × 32964 × 1536
10	RMSNorm_1_1	RMSNorm	1 × 32964 × 1536
11	Attention_1	Grouped Query Attn	1 × 32964 × 1536
12	Add_1_attn	Add	1 × 32964 × 1536
13	RMSNorm_1_2	RMSNorm	1 × 32964 × 1536
14	FFN_1	SwiGLU	1 × 32964 × 1536
15	Add_1_ffn	Add	1 × 32964 × 1536
16	RMSNorm_2_1	RMSNorm	1 × 32964 × 1536
17	Attention_2	Grouped Query Attn	1 × 32964 × 1536
18	Add_2_attn	Add	1 × 32964 × 1536
19	RMSNorm_2_2	RMSNorm	1 × 32964 × 1536
20	FFN_2	SwiGLU	1 × 32964 × 1536
21	Add_2_ffn	Add	1 × 32964 × 1536
22	RMSNorm_3_1	RMSNorm	1 × 32964 × 1536
23	Attention_3	Grouped Query Attn	1 × 32964 × 1536
24	Add_3_attn	Add	1 × 32964 × 1536
25	RMSNorm_3_2	RMSNorm	1 × 32964 × 1536
26	FFN_3	SwiGLU	1 × 32964 × 1536
27	Add_3_ffn	Add	1 × 32964 × 1536
28	RMSNorm_4_1	RMSNorm	1 × 32964 × 1536
29	Attention_4	Grouped Query Attn	1 × 32964 × 1536
30	Add_4_attn	Add	1 × 32964 × 1536
31	RMSNorm_4_2	RMSNorm	1 × 32964 × 1536
32	FFN_4	SwiGLU	1 × 32964 × 1536
33	Add_4_ffn	Add	1 × 32964 × 1536
34	RMSNorm_5_1	RMSNorm	1 × 32964 × 1536
35	Attention_5	Grouped Query Attn	1 × 32964 × 1536
36	Add_5_attn	Add	1 × 32964 × 1536
37	RMSNorm_5_2	RMSNorm	1 × 32964 × 1536
38	FFN_5	SwiGLU	1 × 32964 × 1536
39	Add_5_ffn	Add	1 × 32964 × 1536
40	RMSNorm_6_1	RMSNorm	1 × 32964 × 1536
41	Attention_6	Grouped Query Attn	1 × 32964 × 1536
42	Add_6_attn	Add	1 × 32964 × 1536
43	RMSNorm_6_2	RMSNorm	1 × 32964 × 1536
44	FFN_6	SwiGLU	1 × 32964 × 1536
45	Add_6_ffn	Add	1 × 32964 × 1536
46	RMSNorm_7_1	RMSNorm	1 × 32964 × 1536
47	Attention_7	Grouped Query Attn	1 × 32964 × 1536
48	Add_7_attn	Add	1 × 32964 × 1536
49	RMSNorm_7_2	RMSNorm	1 × 32964 × 1536
50	FFN_7	SwiGLU	1 × 32964 × 1536
51	Add_7_ffn	Add	1 × 32964 × 1536
52	RMSNorm_8_1	RMSNorm	1 × 32964 × 1536
53	Attention_8	Grouped Query Attn	1 × 32964 × 1536
54	Add_8_attn	Add	1 × 32964 × 1536
55	RMSNorm_8_2	RMSNorm	1 × 32964 × 1536
56	FFN_8	SwiGLU	1 × 32964 × 1536
57	Add_8_ffn	Add	1 × 32964 × 1536
58	RMSNorm_9_1	RMSNorm	1 × 32964 × 1536
59	Attention_9	Grouped Query Attn	1 × 32964 × 1536
60	Add_9_attn	Add	1 × 32964 × 1536
61	RMSNorm_9_2	RMSNorm	1 × 32964 × 1536
62	FFN_9	SwiGLU	1 × 32964 × 1536
63	Add_9_ffn	Add	1 × 32964 × 1536
64	RMSNorm_10_1	RMSNorm	1 × 32964 × 1536
65	Attention_10	Grouped Query Attn	1 × 32964 × 1536
66	Add_10_attn	Add	1 × 32964 × 1536
67	RMSNorm_10_2	RMSNorm	1 × 32964 × 1536
68	FFN_10	SwiGLU	1 × 32964 × 1536
69	Add_10_ffn	Add	1 × 32964 × 1536
70	RMSNorm_11_1	RMSNorm	1 × 32964 × 1536
71	Attention_11	Grouped Query Attn	1 × 32964 × 1536
72	Add_11_attn	Add	1 × 32964 × 1536
73	RMSNorm_11_2	RMSNorm	1 × 32964 × 1536
74	FFN_11	SwiGLU	1 × 32964 × 1536
75	Add_11_ffn	Add	1 × 32964 × 1536
76	RMSNorm_12_1	RMSNorm	1 × 32964 × 1536
77	Attention_12	Grouped Query Attn	1 × 32964 × 1536
78	Add_12_attn	Add	1 × 32964 × 1536
79	RMSNorm_12_2	RMSNorm	1 × 32964 × 1536
80	FFN_12	SwiGLU	1 × 32964 × 1536
81	Add_12_ffn	Add	1 × 32964 × 1536
82	RMSNorm_13_1	RMSNorm	1 × 32964 × 1536
83	Attention_13	Grouped Query Attn	1 × 32964 × 1536
84	Add_13_attn	Add	1 × 32964 × 1536
85	RMSNorm_13_2	RMSNorm	1 × 32964 × 1536
86	FFN_13	SwiGLU	1 × 32964 × 1536
87	Add_13_ffn	Add	1 × 32964 × 1536
88	RMSNorm_14_1	RMSNorm	1 × 32964 × 1536
89	Attention_14	Grouped Query Attn	1 × 32964 × 1536
90	Add_14_attn	Add	1 × 32964 × 1536
91	RMSNorm_14_2	RMSNorm	1 × 32964 × 1536
92	FFN_14	SwiGLU	1 × 32964 × 1536
93	Add_14_ffn	Add	1 × 32964 × 1536
94	RMSNorm_15_1	RMSNorm	1 × 32964 × 1536
95	Attention_15	Grouped Query Attn	1 × 32964 × 1536
96	Add_15_attn	Add	1 × 32964 × 1536
97	RMSNorm_15_2	RMSNorm	1 × 32964 × 1536
98	FFN_15	SwiGLU	1 × 32964 × 1536
99	Add_15_ffn	Add	1 × 32964 × 1536
100	RMSNorm_16_1	RMSNorm	1 × 32964 × 1536
101	Attention_16	Grouped Query Attn	1 × 32964 × 1536
102	Add_16_attn	Add	1 × 32964 × 1536
103	RMSNorm_16_2	RMSNorm	1 × 32964 × 1536
104	FFN_16	SwiGLU	1 × 32964 × 1536
105	Add_16_ffn	Add	1 × 32964 × 1536
106	RMSNorm_17_1	RMSNorm	1 × 32964 × 1536
107	Attention_17	Grouped Query Attn	1 × 32964 × 1536
108	Add_17_attn	Add	1 × 32964 × 1536
109	RMSNorm_17_2	RMSNorm	1 × 32964 × 1536
110	FFN_17	SwiGLU	1 × 32964 × 1536
111	Add_17_ffn	Add	1 × 32964 × 1536
112	RMSNorm_18_1	RMSNorm	1 × 32964 × 1536
113	Attention_18	Grouped Query Attn	1 × 32964 × 1536
114	Add_18_attn	Add	1 × 32964 × 1536
115	RMSNorm_18_2	RMSNorm	1 × 32964 × 1536
116	FFN_18	SwiGLU	1 × 32964 × 1536
117	Add_18_ffn	Add	1 × 32964 × 1536
118	RMSNorm_19_1	RMSNorm	1 × 32964 × 1536
119	Attention_19	Grouped Query Attn	1 × 32964 × 1536
120	Add_19_attn	Add	1 × 32964 × 1536
121	RMSNorm_19_2	RMSNorm	1 × 32964 × 1536
122	FFN_19	SwiGLU	1 × 32964 × 1536
123	Add_19_ffn	Add	1 × 32964 × 1536
124	RMSNorm_20_1	RMSNorm	1 × 32964 × 1536
125	Attention_20	Grouped Query Attn	1 × 32964 × 1536
126	Add_20_attn	Add	1 × 32964 × 1536
127	RMSNorm_20_2	RMSNorm	1 × 32964 × 1536
128	FFN_20	SwiGLU	1 × 32964 × 1536
129	Add_20_ffn	Add	1 × 32964 × 1536
130	RMSNorm_21_1	RMSNorm	1 × 32964 × 1536
131	Attention_21	Grouped Query Attn	1 × 32964 × 1536
132	Add_21_attn	Add	1 × 32964 × 1536
133	RMSNorm_21_2	RMSNorm	1 × 32964 × 1536
134	FFN_21	SwiGLU	1 × 32964 × 1536
135	Add_21_ffn	Add	1 × 32964 × 1536
136	RMSNorm_22_1	RMSNorm	1 × 32964 × 1536
137	Attention_22	Grouped Query Attn	1 × 32964 × 1536
138	Add_22_attn	Add	1 × 32964 × 1536
139	RMSNorm_22_2	RMSNorm	1 × 32964 × 1536
140	FFN_22	SwiGLU	1 × 32964 × 1536
141	Add_22_ffn	Add	1 × 32964 × 1536
142	RMSNorm_23_1	RMSNorm	1 × 32964 × 1536
143	Attention_23	Grouped Query Attn	1 × 32964 × 1536
144	Add_23_attn	Add	1 × 32964 × 1536
145	RMSNorm_23_2	RMSNorm	1 × 32964 × 1536
146	FFN_23	SwiGLU	1 × 32964 × 1536
147	Add_23_ffn	Add	1 × 32964 × 1536
148	RMSNorm_24_1	RMSNorm	1 × 32964 × 1536
149	Attention_24	Grouped Query Attn	1 × 32964 × 1536
150	Add_24_attn	Add	1 × 32964 × 1536
151	RMSNorm_24_2	RMSNorm	1 × 32964 × 1536
152	FFN_24	SwiGLU	1 × 32964 × 1536
153	Add_24_ffn	Add	1 × 32964 × 1536
154	RMSNorm_25_1	RMSNorm	1 × 32964 × 1536
155	Attention_25	Grouped Query Attn	1 × 32964 × 1536
156	Add_25_attn	Add	1 × 32964 × 1536
157	RMSNorm_25_2	RMSNorm	1 × 32964 × 1536
158	FFN_25	SwiGLU	1 × 32964 × 1536
159	Add_25_ffn	Add	1 × 32964 × 1536
160	RMSNorm_26_1	RMSNorm	1 × 32964 × 1536
161	Attention_26	Grouped Query Attn	1 × 32964 × 1536
162	Add_26_attn	Add	1 × 32964 × 1536
163	RMSNorm_26_2	RMSNorm	1 × 32964 × 1536
164	FFN_26	SwiGLU	1 × 32964 × 1536
165	Add_26_ffn	Add	1 × 32964 × 1536
166	RMSNorm_27_1	RMSNorm	1 × 32964 × 1536
167	Attention_27	Grouped Query Attn	1 × 32964 × 1536
168	Add_27_attn	Add	1 × 32964 × 1536
169	RMSNorm_27_2	RMSNorm	1 × 32964 × 1536
170	FFN_27	SwiGLU	1 × 32964 × 1536
171	Add_27_ffn	Add	1 × 32964 × 1536
172	RMSNorm_28_1	RMSNorm	1 × 32964 × 1536
173	Attention_28	Grouped Query Attn	1 × 32964 × 1536
174	Add_28_attn	Add	1 × 32964 × 1536
175	RMSNorm_28_2	RMSNorm	1 × 32964 × 1536
176	FFN_28	SwiGLU	1 × 32964 × 1536
177	Add_28_ffn	Add	1 × 32964 × 1536
178	Output	Output	1 × 32964 × 1536

What the verifier says

warn"RoPE" receives input but its output is not connected. This layer will be unreachable in the forward pass. Fix: Connect the output forward, or add an Output node if this is the final layer.
dead-end

infoLLaMA uses intermediateSize ≈ ⌊(8/3 × D) / 256⌋ × 256. Current: 8960 (5.83× embedDim). Expected: ~4096. Fix: Set intermediateSize to 4096 for embedDim=1536.
swiglu-dim-convention

infoAt 28 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 jinaai/jina-reranker-m0: 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 jinaai/jina-reranker-m0 --plan --share

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