ROCmify - Automated CUDA to AMD ROCm Migration

Most AI and HPC software is written in CUDA, which only runs on NVIDIA GPUs. Porting it to AMD is usually a manual process that takes weeks, introduces subtle bugs, and keeps teams locked into NVIDIA even when AMD hardware is a better fit. The biggest barrier is not the hardware. It's the codebase you already have. ROCmify automates that migration with a cost-efficient, reliability-first pipeline. It starts by running AMD's HIPIFY tool and compiling with `hipcc`. If compilation fails, ROCmify classifies the compiler error and applies a deterministic repair plugin. If no rule matches, it checks a SQLite cache of solved patches. Both steps are free. Only then does it escalate to a language model on Fireworks. If the model fails three times, ROCmify restores the original file with `git checkout` and flags it for human review. Your code is never left half-modified. ROCmify also fixes problems compilers cannot detect. For example, code that hardcodes `threadIdx.x % 32` compiles on AMD but produces incorrect results because CDNA GPUs use 64-thread wavefronts instead of 32-thread warps. Using a tree-sitter AST, ROCmify rewrites these cases to the `warpSize` builtin, but only when the intent is unambiguous. If `32` could represent a tile size, it refuses to guess. Just as importantly, ROCmify knows when not to automate. Inline PTX, NVIDIA's low-level assembly, has no mechanical HIP equivalent. Instead of sending it to an LLM that might generate an incorrect translation, ROCmify quarantines it for manual review. An honest refusal is safer than a confident wrong answer. On real AMD hardware (ROCm 7.2, `gfx1100`), ROCmify performed one deterministic repair at $0.00, quarantined one PTX block, and required one LLM call costing $0.000084. A second run cost $0.00 because the fix came from cache. ROCmify includes 257 automated tests, is fully containerized, and runs on laptops without a GPU because `hipcc` supports cross-compilation.