UniRank: Unified Rank Allocation for Low-Rank LLM Compression

Name: UniRank: Unified Rank Allocation for Low-Rank LLM Compression
Uploaded: 2026-09-16T10:43:42.236343+00:00
Duration: 315 s
Description: The paper 'UniRank' introduces a unified method for compressing large language models. The discussion covers the Sorting-and-Truncation pipeline, which uses local energy and global functional importance to allocate rank, and Rank-Preserving Fine-Tuning (RPFT), which maintains model performance during fine-tuning without requiring expensive re-decomposition.

The paper 'UniRank' introduces a unified method for compressing large language models. The discussion covers the Sorting-and-Truncation pipeline, which uses local energy and global functional importance to allocate rank, and Rank-Preserving Fine-Tuning (RPFT), which maintains model performance during fine-tuning without requiring expensive re-decomposition.

Open the full episode