Andrzej Yanik) presented release zluda 5 , the open -up implementation of the technology cuda . The aim of the project is to provide the possibility of launching unmodified CUDA applications on systems with GPU, different from GPU NVIDIA, with performance close to the performance of applications performed without strata. The project code is written in Rust and spreads under the licenses of MIT and Apache 2.0.
zluda 5 was the second significant issue of the project formed after cleaning the code base from the code developed during the work of Andrzej in AMD. Since 2022, Andrzej worked in AMD to create a layer for compatibility GPU AMD with CUDA, but in 2024 the project was folded. In accordance with the terms of the contract and after receiving a permit for publication from an AMD representative, Andrzej opened the developments code created during operation in AMD and allowing the CUDA application on top of the stack rocm and runtime hip (heterogeneous-computing interface for Portability).
Last year, Andrzej was forced to remove the code from open access after a letter from lawyers who made it clear that the permission given during email correspondence has no legal force. After that, Andrzej began work on the new edition of Zluda, formed on the basis of the code base that existed before the work of Andrzej in AMD. In the current form, the development is focused on performing applications using CUDA to accelerate tasks related to machine learning. The project is still limited to work on the GPU AMD, but will be adapted for GPU Intel.
Key improvements in the new issue:
- The initial support of the start of the zluda frameworks llm.c , llama.cpp and
and pytorch to perform large language models using Cuda-optimizations used for GPU NVIDIA. - The initial support for the launch of applications that use to increase the productivity of the library cublas , cublaslt and nvml .
- Added layer zluda_trace for tracing applications using CUDA. Using ZLUDA_TRACE, you can diagnose problems and identify flaws in Zluda, interfering with the normal execution of the application.
- The ZOC command line (Zluda Offline Compiler) is presented, which allows compiling files with instructions nvidia ptx