Document Visual Question Answering on DocVQA 30%
135.72KV Cache Size (MB)CDPruner
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CDPrunerBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 135.72 | 2,931 | 28.01 | 2,998 | |
| IVC-PruneBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 135.73 | 3,315 | 28.64 | 3,389 | |
| ERASEBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 135.75 | 2,337 | 28.5 | 2,419 | |
| PruneSIDBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 135.76 | 3,262 | 28.96 | 3,353 | |
| DARTBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 189.74 | 2,550 | 28.22 | 2,628 | |
| iLLaVABackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 314.18 | 3,508 | 31.72 | 3,573 | |
| BaseBackbone=Qwen2.5-VL-7B, GPU=NVIDIA RTX 5880Ti, Batch size=1, FlashAttention-2=enabled, Image Resolution=4K, Vision Token Count=16K, Pruning Ratio=85.0%2026.05 | 891.27 | 3,694 | 31.03 | 3,780 |