Speedstu/CUDA-for-AMD-Windows

Run CUDA-targeted Windows applications on AMD GPUs with ZLUDA + ROCm/HIP.

What it solves

This project enables users to run CUDA-targeted applications on Windows systems equipped with AMD Radeon GPUs. It bridges the gap between NVIDIA's proprietary CUDA ecosystem and AMD's hardware by leveraging ZLUDA and the AMD ROCm/HIP stack, allowing developers and researchers to execute AI workloads on AMD hardware without needing an NVIDIA GPU.

How it works

It provides a reproducible software stack that translates CUDA API calls into AMD's HIP (Heterogeneous-compute Interface for Portability) language. The project includes automated scripts for installation, environment verification, and runtime diagnostics. It emphasizes a "fail-closed" approach, where the system is designed to report errors rather than silently producing incorrect numerical results, ensuring that users can distinguish between successful GPU execution and CPU fallbacks or crashes.

Who it’s for

This is for developers, researchers, and power users who want to run CUDA-based AI software (such as PyTorch or llama.cpp) on AMD hardware. It is particularly useful for those who need to validate whether their specific workloads can run correctly on AMD GPUs and who require a structured, evidence-based approach to compatibility testing.

Highlights

  • Provides a validated, reproducible stack for running CUDA applications on AMD Radeon GPUs.
  • Includes rigorous numerical validation tools to ensure GPU computations are accurate.
  • Supports major AI frameworks and tools like PyTorch, LibTorch, and llama.cpp.
  • Implements a strict "fail-closed" policy to prevent silent data corruption.
  • Offers experimental paths for advanced features like cuDNN-to-MIOpen bridging and CUDA Graph support.

Related

  • Project
    flagos-ai/awesome-LLM-driven-kernel-generation

    A curated list of papers, datasets, and tools exploring how large language models and agentic systems can automate the generation and optimization of high-performance computing kernels, organized by research stream and technique.

  • Project
    mattn/go-tflite

    go‑tflite is a Go binding for TensorFlow Lite’s C API, enabling Go programs to load *.tflite* models, run inference, and optionally use hardware delegates like XNNPACK or Edge‑TPU. The project provides pre‑built TensorFlow Lite libraries for major platforms, clear build instructions, and a Binder notebook demo.

  • Project
    mfoud444/ollamafreeapi

    OllamaFreeAPI is a Python client that gives free, no‑signup access to several open‑source LLMs (LLaMA, Mistral, DeepSeek, Gemma‑based models, etc.) via simple `chat` and `stream_chat` calls. It abstracts a set of community‑run API endpoints, offering model listing, info lookup, and basic load‑balancing. Ideal for quick experiments or learning, but relies on the reliability of volunteer‑hosted servers.

  • Project
    pacoxu/AI-Infra

    AI-Infra provides a visual landscape and organized learning path for AI infrastructure engineers, focusing on Kubernetes and cloud-native environments. It guides users through essential components and advanced topics, with links to detailed documentation and community resources.