面壁 MiniCPM-OCR 1.0 发布 DSpark 投机解码草稿模型，端到端效率提升 55%

This repository hosts the DSpark draft (speculative decoding) head for MiniCPM-OCR 1.0. It is a draft model, not a standalone OCR model — it must be paired with the matching main checkpoint.

DSpark speeds up decoding with a small draft model that proposes tokens the main model then verifies in one pass. With this head, MiniCPM-OCR 1.0 improves end-to-end processing efficiency by 55%; throughput reaches about 1.49× PaddleOCR-VL-1.6 and 3.62× MinerU2.5 Pro.

Specification

Property	Value
Target model	MiniCPM-OCR 1.0 (1.3B)
Draft parameters	296,837,889
Draft layers	5
Draft tokens per forward pass	15
Target hidden-state layers	`[3, 7, 11, 15, 19]`
Checkpoint precision	BF16
Input vocabulary	248,100
Draft output vocabulary	114,688

The output LM head and the Markov output rows are cropped to 114,688 tokens. The checkpoint keeps its d2t offset mapping and the full input vocabulary; local-to-target.json holds the corresponding absolute target token IDs.

Note on d2t: in this checkpoint the tensor is delta-encoded (target_id - draft_id). It must be converted to absolute IDs (d2t[i] + i) before use — the GGUF build already does this, and local-to-target.json is the ground truth to verify against.

Usage with vLLM

vllm serve openbmb/MiniCPM-OCR-1.0 \
  --served-model-name MiniCPM-OCR \
  --trust-remote-code \
  --max-model-len 40960 \
  --max-num-batched-tokens 40960 \
  --max-num-seqs 1 \
  --trust-request-mm-kwargs \
  --mm-processor-kwargs '{"downsample_mode":"4x","max_slice_nums":9}' \
  --speculative-config '{"method":"dspark",
                         "model":"openbmb/MiniCPM-OCR-1.0-DSpark",
                         "num_speculative_tokens":15,
                         "enable_adaptive_verification":false}' \
  --default-chat-template-kwargs '{"enable_thinking":false}'

Selection and tuning: DSpark acceleration.

Files

File	Notes
`model.safetensors`	64 tensors — 5 draft layers, `fc` projector, `lm_head`, Markov heads, confidence head, `d2t`
`config.json`	`architectures: ["Qwen3DSparkModel"]`, `dflash_config` with target layer ids and Markov rank
`local-to-target.json`	114,688 absolute target token IDs, the reference for `d2t`
`export-audit.json`	export-time audit: weight/mapping hashes and vocabulary coverage

Related checkpoints

Format	HF Mirror	ModelScope
BF16 (original)	🤗	ModelScope
GGUF (llama.cpp)	🤗	ModelScope
AWQ W4A16	🤗	ModelScope
GPTQ W4A16	🤗	ModelScope
DSpark draft (BF16) (this repo)	🤗	ModelScope
DSpark draft (GGUF)	🤗	ModelScope

DSpark requires a matching main and draft model from the same release; do not pair this draft with a different main checkpoint.

FAQ

Installation, model downloads, GPU memory, input formats, and unexpected outputs: FAQ.

License

Model weights, code, and documentation are released under the Apache License 2.0.