wtf is project-blackbeard?

Project Blackbeard is a fork of the well known llama.cpp project, which lets you run large language models locally using GGUF format model files. The fork exists for one narrow purpose: making that inference engine run faster specifically on NVIDIA's newest Blackwell generation graphics cards, sometimes called SM100 hardware. The README is clear that this is not meant to replace the original llama.cpp project, and it credits Georgi Gerganov and the ggml org community for the underlying work this fork builds on. Because the project's whole focus is Blackwell hardware, older or different GPU architectures are explicitly out of scope, and contributions targeting non Blackwell hardware are not accepted. In practice this means the fork only supports CUDA, with extra optimized code paths for Blackwell chips, plus a CPU fallback mode for debugging and checking correctness. It does not attempt to support the wide range of backends the original llama.cpp offers, such as Metal, Vulkan, or several other specialized compute platforms. Day to day, using this fork looks the same as using regular llama.cpp. It includes the same core command line tools: llama-cli for local chat and text completion, llama-server for running a lightweight server that mimics the OpenAI API, llama-bench for measuring how fast the model generates tokens and processes prompts, and llama-perplexity for checking model quality. You can load models directly from Hugging Face using the -hf flag, and quantized GGUF models are supported the same way they are in the upstream project. To use it you need an NVIDIA Blackwell GPU, a compatible CUDA toolkit, and a build setup using cmake and ninja or a similar tool. Anyone contributing changes is expected to target Blackwell hardware specifically, provide a reproducible benchmark showing the improvement, and be able to explain how the change affects scheduling or memory latency. AI assisted contributions are allowed, but only if they come with real profiling done on actual Blackwell hardware. The project is released under the MIT license, matching the license of upstream llama.cpp.