MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head
Open paper
Benchmark
Model
Rank
Results
text-to-video-generation-on-vbench
Wan-MHLA-H (Wan2.1-1.3B)
#20
VBench Total Score: 83.82
image-classification-on-imagenet
MHLA-VLT-S (27M)
#329
Top 1 Accuracy: 84.6%
Rank counts only results with a code link.