Speculative Decoding on VideoDetailCaption (~17k visual tokens)
3.91Tau (τ)SpecVLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecVLMTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.91 | 1.2 | |
| SparrowTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.89 | 2.04 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.52 | 1.87 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=8, Source=Original Weights2026.02 | 2.79 | 1.53 | |
| SparrowTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.37 | 1.27 | |
| ViSpecTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.24 | 1.21 | |
| MSDTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.11 | 0.52 | |
| MSDTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.1 | 0.48 |