The authors developed activation denoising for post-training quantization, which compresses model weights to a few bits without retraining.
Noise-robust layers narrow the accuracy gap in parallel LLM quantization
Activation denoising makes independently quantized layers less sensitive to upstream errors, retaining parallel processing while recovering much of sequential quantization’s benefit.
Big Tech
Yan Scholten · Rachel Lawrence · James Hensman · Stephan Günnemann · Alicia Curth · Riccardo Grazzi
Munich Data Science Institute · Technical University of Munich · Microsoft Research Cambridge
Research Digest··2 min read
Scholten et al.
Why this paper
From Microsoft Research Cambridge and 2 others
In one line
Parallel LLM quantization with activation denoising recovers much of sequential quantization's accuracy benefit without serial overhead.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§