Scholar Feed

LLM quantization

PB-LLM

PB-LLM: Partially Binarized Large Language Models

Superseded baseline#11 of 80 most-superseded · first seen Sep 29, 2023

Superseded — cited as a baseline and beaten by newer methods

4 papers critique it · 3 beat it on benchmarks

What papers say

Verbatim critique sentences, each from a paper that cites PB-LLM as a baseline.

However, it reveals the challenge of retaining a significant portion of weights, typically over 30%, at INT8 precision to maintain acceptable performance.
— Bi-VLM: Pushing Ultra-Low Precision Post-Training Quantization Boundaries in Vision-Language Models
“both methods introduce additional unstructured fine-grained masks to distinguish salient weights which requires additional 1-bit per weight to store the mask and leads the memory of the quantized model to exceeding 2-bit per weight, where PB-LLM with 2.7-bit and BiLLM with 2.1-bit respectively”
— PTQ1.61: Push the Real Limit of Extremely Low-Bit Post-Training Quantization Methods for Large Language Models
“Even the recent binary PTQ method for LLMs, PB-LLM~shang2023pb, only maintains a perplexity metric of around 800 with an average weight of 1.7 bits.”
— BiLLM: Pushing the Limit of Post-Training Quantization for LLMs
“PB-LLM~shang2023pb applies fixed saliency thresholds across layers”
— Minimizing the Hidden Cost of Scales: Graph-Guided Ultra-Low-Bit Quantization for Large Language Models

Beaten on benchmarks

Head-to-head results where a newer method reports beating PB-LLM. Values are copied from the source paper's tables — verify against the cited paper.

What to use instead

Recent methods in the same sub-problem, not yet superseded in the knowledge base — arXiv benchmark leaders, not vetted production recommendations.