Ghost-Coder: Autonomous CUDA-to-HIP Agent

The migration of high-performance computing workloads from NVIDIA's CUDA to AMD's ROCm ecosystem is a critical bottleneck for developers adopting new hardware. Ghost-Coder eliminates this friction by serving as an autonomous, self-healing translation agent. Powered by the Qwen2.5-Coder-32B model running on an AMD Instinct™ MI300X (192GB) instance, Ghost-Coder doesn't just do static text replacement—it deeply analyzes device-side kernels and host-side memory API calls to generate native, highly optimized HIP code. Architecture & Workflow: The system is built with a decoupled frontend/backend architecture to maximize GPU resource allocation. The UI is a reactive Gradio web application hosted via Hugging Face Spaces, designed with an agentic interface that visualizes the translation and verification steps in real-time. This frontend communicates via a secure tunneling protocol to a custom FastAPI bridge running directly on the AMD Developer Cloud droplet. Inside the droplet, the AI engine operates under strict hardware guardrails. To achieve stable, high-throughput inference on the MI300X architecture, we engineered a specialized execution environment. By overriding the ISA to gfx942, strictly serializing kernel execution, and disabling experimental Triton memory paths, we stabilized the Qwen model to prevent memory aperture violations during continuous generative decoding. Self-Healing & Translation: When a user submits a CUDA kernel (e.g., a tiled matrix multiplication), Ghost-Coder analyzes the computational logic. It dynamically maps NVIDIA-specific host calls like cudaMalloc and cudaMemcpy to their hipMalloc and hipMemcpy equivalents, while restructuring execution macros (like transforming <<<blocks, threads>>> into hipLaunchKernelGGL). The result is drop-in ready C++ code tailored for the AMD stack. Ghost-Coder accelerates the transition to open-source GPU computing by turning days of manual kernel porting into a seamless, automated seconds-long process.