Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• Swift-SVD simultaneously guarantees theoretical optimum, practical efficiency and numerical stability in low-rank LLM compression.method 90%
• Swift-SVD achieves 3-70X speedups in end-to-end compression time compared to state-of-the-art baselines.result 90%
• Swift-SVD outperforms state-of-the-art baselines across six LLMs and eight datasets.result 85%
• Swift-SVD enables training-free, fast, and optimal layer-wise low-rank approximation.method 85%
• Swift-SVD uses effective rank to analyze local layer-wise compressibility and designs a dynamic rank allocation strategy.technical 80%
• Existing SVD-based compression methods suffer from either suboptimal reconstruction error or theoretical optimality with practical inefficiency.limitation 85%
• Swift-SVD performs incremental aggregation of covariance of output activations and a single eigenvalue decomposition after aggregation.technical 80%
• SVD-based compression provides a hardware-friendly solution to reduce memory and bandwidth demands of LLM weights and KV cache.technical 75%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card