Exponentially faster language modelling(arxiv.org)

270 points by born-jre 2 years ago | 132 comments

Ultra Fast Bert(arxiv.org)

227 points by gyre007 2 years ago | 57 comments

Some critical issues with the SWE-bench dataset(arxiv.org)

331 points by joshwa 1 year ago | 113 comments

arXiv paper introducing SWE-Bench+, an empirical analysis revealing data quality issues in the SWE-bench coding benchmark for LLMs, such as solution leakage and weak tests.

SplitQuantV2: Enhancing Low-Bit Quantization of LLMs Without GPUs(arxiv.org)

25 points by PaulHoule 1 year ago | 5 comments

Making LLMs Cheaper and Better via Performance-Efficiency Optimized Routing(arxiv.org)

118 points by omarsar 1 year ago | 22 comments

Lossless LLM compression for efficient GPU inference via dynamic-length float(arxiv.org)

392 points by CharlesW 1 year ago | 109 comments

arXiv paper introducing DFloat11, a lossless compression format that shrinks LLM weights by about 30% using entropy coding, with a custom GPU kernel for inference.