pmadinei/Interlace-Qwen3-VL-8B-10pc
Image-Text-to-Text • 745k • Updated • 31
INTERLACE: Interleaved Layer Pruning in VLMs (CVPR 2025). Pruned Qwen3-VL models retaining up to 94% performance.
Note Qwen3-VL-8B with 10% layers pruned. 94.0% relative performance.
Note Qwen3-VL-8B with 15% layers pruned. 92.1% relative performance.
Note Qwen3-VL-8B with 20% layers pruned. 86.9% relative performance.
Note Qwen3-VL-8B with 25% layers pruned. 86.1% relative performance. 1.18x TTFT speedup.
Note Qwen3-VL-4B with 10% layers pruned. 93.9% relative performance.
Note Qwen3-VL-4B with 15% layers pruned. 91.9% relative performance.
Note Qwen3-VL-4B with 20% layers pruned. 88.0% relative performance.
Note Qwen3-VL-4B with 25% layers pruned. 81.7% relative performance.