Speculative Decoding on MVBench
3.87Tau (τ)Sparrow
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SparrowTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.87 | 2.06 | |
| SpecVLMTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.8 | 1.09 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.6 | 1.92 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=8, Source=Original Weights2026.02 | 2.83 | 1.52 | |
| SparrowTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.55 | 1.36 | |
| ViSpecTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.44 | 1.31 | |
| MSDTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.86 | 0.95 | |
| MSDTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.63 | 0.7 |