poisonxa16/pxa v4.0.0 is a new release of a language model inference engine for Pascal·Volta GPUs. The project claims that this version delivers 3× faster decoding and 8× faster prompt processing on the same files. pxa is a fork of llama.cpp tuned for Pascal·Volta GPUs, and aims to run a model server on Tesla P100·V100 and GTX 10 series cards.
The core is the PXQ codec and kernels tuned for those GPUs. The project explains that it wrote new kernels targeting chips without DP4A and tensor cores. It supports models in its own PXQ·PXQN format and can also load ordinary GGUF files, and the documentation states it provides a planner that selects settings based on per-card measurements. It also includes splitting execution across multiple GPUs, speculative decoding, expert caching, and quantized weight handling. It offers an OpenAI-compatible API and a browser dashboard, and is licensed under MIT.
The performance numbers are claims made by the project. The performance comparison in the README is measured against the previous version on the same hardware, and the detailed figures are said to be in the separate release notes. Therefore, the 3× and 8× numbers alone do not mean the same improvement will appear with other inference engines or on other hardware. Check the comparison conditions and read them within results on the target GPUs.