Models / unlimited-ocr

Unlimited-OCR: GPU memory to fine-tune

Fine-tuning Unlimited-OCR at batch 1 and 2048 tokens takes about 64.2 GB of GPU memory for a full fine-tune (borderline on one A100 / H100 80 GB), and 16.0 GB with LoRA r=16 on a bf16 base (borderline on one L4 24 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. 2.7M 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

2.81B

2,810,339,008 parameters

In the published checkpoint

3.34B

3,336,106,240 scalars · safetensors.total, read 2026-07-29

Delta

-10.8%

stored twice The checkpoint stores the tied embedding twice: +165.5M (`lm_head.weight` and `model.embed_tokens.weight`, 129280 x 1280). 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_unlimitedocr.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

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$113.87

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

64.2 GB

LoRA r=16, batch 1 at 2048 tokens

16.0 GB

Serve: bf16 weights

5.62 GB

Serve: KV cache, 4096 tokens

0.25 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 32
L4 24 GB	does not fit at batch 1	batch 1, borderline	batch 60
A100 40 GB	2 cards at batch 1 (47 + 32 layers)	batch 2	batch 115
A100 / H100 80 GB	batch 1, borderline	batch 6	batch 254
H200 141 GB	batch 6	batch 11	batch 464

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. "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

82 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 × 1280
3	Positional_Embedding	Learned Pos Embed	1 × 32768 × 1280
4	Vision input	Input	3 × 1024 × 1024
5	PatchEmbed	Patch Embed	4096 × 1280
6	Patch_Position_Embedding	Learned Pos Embed	4096 × 1280
7	Vision encoder (internals not in config)	Projection	4096 × 1280
8	Vision tokens	Reshape	1 × 4096 × 1280
9	Multimodal fusion (concat tokens)	Concatenate	1 × 36864 × 1280
10	LayerNorm_1_1	LayerNorm	1 × 36864 × 1280
11	Attention_1	Grouped Query Attn	1 × 36864 × 1280
12	Add_1_attn	Add	1 × 36864 × 1280
13	LayerNorm_1_2	LayerNorm	1 × 36864 × 1280
14	FFN_1	Feed Forward	1 × 36864 × 1280
15	Add_1_ffn	Add	1 × 36864 × 1280
16	LayerNorm_2_1	LayerNorm	1 × 36864 × 1280
17	Attention_2	Grouped Query Attn	1 × 36864 × 1280
18	Add_2_attn	Add	1 × 36864 × 1280
19	LayerNorm_2_2	LayerNorm	1 × 36864 × 1280
20	MoE_2	Shared-Expert MoE	1 × 36864 × 1280
21	Add_2_ffn	Add	1 × 36864 × 1280
22	LayerNorm_3_1	LayerNorm	1 × 36864 × 1280
23	Attention_3	Grouped Query Attn	1 × 36864 × 1280
24	Add_3_attn	Add	1 × 36864 × 1280
25	LayerNorm_3_2	LayerNorm	1 × 36864 × 1280
26	MoE_3	Shared-Expert MoE	1 × 36864 × 1280
27	Add_3_ffn	Add	1 × 36864 × 1280
28	LayerNorm_4_1	LayerNorm	1 × 36864 × 1280
29	Attention_4	Grouped Query Attn	1 × 36864 × 1280
30	Add_4_attn	Add	1 × 36864 × 1280
31	LayerNorm_4_2	LayerNorm	1 × 36864 × 1280
32	MoE_4	Shared-Expert MoE	1 × 36864 × 1280
33	Add_4_ffn	Add	1 × 36864 × 1280
34	LayerNorm_5_1	LayerNorm	1 × 36864 × 1280
35	Attention_5	Grouped Query Attn	1 × 36864 × 1280
36	Add_5_attn	Add	1 × 36864 × 1280
37	LayerNorm_5_2	LayerNorm	1 × 36864 × 1280
38	MoE_5	Shared-Expert MoE	1 × 36864 × 1280
39	Add_5_ffn	Add	1 × 36864 × 1280
40	LayerNorm_6_1	LayerNorm	1 × 36864 × 1280
41	Attention_6	Grouped Query Attn	1 × 36864 × 1280
42	Add_6_attn	Add	1 × 36864 × 1280
43	LayerNorm_6_2	LayerNorm	1 × 36864 × 1280
44	MoE_6	Shared-Expert MoE	1 × 36864 × 1280
45	Add_6_ffn	Add	1 × 36864 × 1280
46	LayerNorm_7_1	LayerNorm	1 × 36864 × 1280
47	Attention_7	Grouped Query Attn	1 × 36864 × 1280
48	Add_7_attn	Add	1 × 36864 × 1280
49	LayerNorm_7_2	LayerNorm	1 × 36864 × 1280
50	MoE_7	Shared-Expert MoE	1 × 36864 × 1280
51	Add_7_ffn	Add	1 × 36864 × 1280
52	LayerNorm_8_1	LayerNorm	1 × 36864 × 1280
53	Attention_8	Grouped Query Attn	1 × 36864 × 1280
54	Add_8_attn	Add	1 × 36864 × 1280
55	LayerNorm_8_2	LayerNorm	1 × 36864 × 1280
56	MoE_8	Shared-Expert MoE	1 × 36864 × 1280
57	Add_8_ffn	Add	1 × 36864 × 1280
58	LayerNorm_9_1	LayerNorm	1 × 36864 × 1280
59	Attention_9	Grouped Query Attn	1 × 36864 × 1280
60	Add_9_attn	Add	1 × 36864 × 1280
61	LayerNorm_9_2	LayerNorm	1 × 36864 × 1280
62	MoE_9	Shared-Expert MoE	1 × 36864 × 1280
63	Add_9_ffn	Add	1 × 36864 × 1280
64	LayerNorm_10_1	LayerNorm	1 × 36864 × 1280
65	Attention_10	Grouped Query Attn	1 × 36864 × 1280
66	Add_10_attn	Add	1 × 36864 × 1280
67	LayerNorm_10_2	LayerNorm	1 × 36864 × 1280
68	MoE_10	Shared-Expert MoE	1 × 36864 × 1280
69	Add_10_ffn	Add	1 × 36864 × 1280
70	LayerNorm_11_1	LayerNorm	1 × 36864 × 1280
71	Attention_11	Grouped Query Attn	1 × 36864 × 1280
72	Add_11_attn	Add	1 × 36864 × 1280
73	LayerNorm_11_2	LayerNorm	1 × 36864 × 1280
74	MoE_11	Shared-Expert MoE	1 × 36864 × 1280
75	Add_11_ffn	Add	1 × 36864 × 1280
76	LayerNorm_12_1	LayerNorm	1 × 36864 × 1280
77	Attention_12	Grouped Query Attn	1 × 36864 × 1280
78	Add_12_attn	Add	1 × 36864 × 1280
79	LayerNorm_12_2	LayerNorm	1 × 36864 × 1280
80	MoE_12	Shared-Expert MoE	1 × 36864 × 1280
81	Add_12_ffn	Add	1 × 36864 × 1280
82	Output	Output	1 × 36864 × 1280

What the verifier says

warn"Positional_Embedding" 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

infoAt 12 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 baidu/Unlimited-OCR: 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 baidu/Unlimited-OCR --plan --share

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

Other unlimited-ocr checkpoints

Unlimited-OCR-AWQ

2.94B derived · -12.3% against the checkpoint