A working reproducible stack now allows running CUDA-targeted Windows applications on AMD GPUs through ZLUDA + ROCm/HIP. The setup supports CUDA-enabled Lib Torch workloads and has been validated on AMD Radeon RX 9060 XT (gfx1200). Key components include ZLUDA v6-preview.69, AMD HIP SDK 6.4, and Lib Torch 2.3.0 + cu118. All CUDA driver APIs including cuBLAS, cuBLASLt, cuSPARSE, and cuFFT pass cuda_check. A 2.2M-parameter PPO network successfully completed forward/inference, PPO learning, and optimizer operations on the CUDA-facing device. One clean validation iteration completed 65,536 timesteps using the runtime produced by this repository.

Installation requires AMD prerequisites: current GPU driver and HIP SDK 6.4 for Windows. Users clone the repository and run install.ps1, which detects GPU, verifies dependencies, downloads ZLUDA and Lib Torch, validates SHA-256 hashes, and runs cuda_check.exe. The runtime can be staged or launched via run-zluda.ps1. Diagnostic scripts include doctor.ps1, gpu-scan.ps1, and test-runtime.ps1.

GPU compatibility is currently limited to RX 9060 XT; other AMD GPUs are candidates but not guaranteed. The GPU scanner records model, gfx architecture, driver, and HIP info without collecting usernames or tokens. CUDA DNN is unavailable with the stable HIP SDK due to missing MIOpen, though dense/GEMM-heavy Lib Torch training may still function.