ZDTaichu5.0-9B-DSpark

A DSpark speculative-decoding draft model for ZDTaichu5.0-9B target models, trained with speculators and served with vllm, primarily intended for multimodal scenarios.

Performance Evaluation of DSpark

Acceptance length and speedup evaluation of DSpark across diverse multimodal benchmarks.

• Sampling: thinking enabled, temperature 1.0
• Speedup: evaluation on a single NVIDIA H800 GPU with a concurrency of 1

Category	Dataset	Acceptance length	Speedup
Multimodal Reasoning	MathVista Mini	4.53	3.03
Multimodal Reasoning	WeMath	4.61	3.17
Multimodal Reasoning	MathVerse MINI Vision Only	4.79	3.15
Multimodal Reasoning	LogicVista	4.10	2.78
General VQA	MMStar	4.01	2.73
General VQA	RealWorldQA	3.87	2.46
Document Understanding	OCRBench	4.39	2.51
Document Understanding	AI2D	4.12	2.75
Spatial Understanding	SpatialVizBench	3.94	2.63
Spatial Understanding	MindCubeBench Tiny	4.02	2.66

Serving with vllm

    vllm serve TaichuAI/ZDTaichu5.0-9B \
        --speculative-config  '{"method": "dspark", "model": "TaichuAI/ZDTaichu5.0-9B-DSpark", "num_speculative_tokens": 7}'