Speculative Decoding on LongVideoBench ~15k visual tokens
3.82Tau (τ)SpecVLM
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SpecVLMTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.82 | 115 | |
| SparrowTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.78 | 202 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 3.55 | 191 | |
| ViSpecTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=8, Source=Original Weights2026.02 | 2.75 | 147 | |
| SparrowTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.54 | 137 | |
| ViSpecTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 2.46 | 134 | |
| MSDTemperature=0, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.75 | 88 | |
| MSDTemperature=1, Target Model=Qwen2.5-VL-7B, Hardware=NVIDIA L20 GPUs, Batch Size=162026.02 | 1.52 | 74 |