ZLUDA 5 is out today as the newest version of this open-source software for running CUDA software on non-NVIDIA GPUs.
ZLUDA 5 introduces “zoc” as an offline compiler for ZLUDA. Zoc will take a NVIDIA PTX file as input and output the LLVM intermediate representation (IR) and the RDNA Assembly for the GPU in use generated in turn by the ROCm compiler.
ZLUDA 5 has also been making progress in machine learning workloads / AI inference. ZLUDA has more of llm.c working with it as its first milestone. Plus ZLUDA is making progress on LLama.cpp with decent performance results. There also are efforts underway for ZLUDA on PyTorch.
ZLUDA developers have also been working on kernel caching, various performance improvements, and other enhancements.
ZLUDA can be downloaded from GitHub. More details on these recent ZLUDA improvements via the ZLUDA blog.
