RocmPilot Studio — CUDA→AMD Migration Cockpit

AMD's MI300X wins on price/performance, but what keeps enterprises on CUDA isn't silicon — it's migration uncertainty. A typical AI codebase is CUDA-first top to bottom: hardcoded cuda device logic, nvidia/cuda base images, CUDA-pinned wheels, and the part everyone fears — custom kernels with warp-level assumptions. Tools like hipify translate the easy majority, but nobody can tell you up front how ready a whole codebase is, or where the hard work hides. RocmPilot Studio is the command center for that migration. Paste a repo URL and it runs a six-stage pipeline: SCAN → PLAN → PATCH → VALIDATE → REPORT. • Scan: a deterministic engine (no LLM — findings can't be hallucinated) detects blockers with file, line and severity, plus a kernel-risk classifier that names the hard 20% in AMD vocabulary: 32-lane warp shuffles vs 64-lane wavefronts, warpSize hazards, WMMA→rocWMMA, CUTLASS→Composable Kernel, cuBLAS→hipBLAS, NCCL→RCCL — at repository scale. • Plan: a multi-model agent orchestra (all AMD-hosted via Fireworks AI) — DeepSeek drafts the prioritized plan, then GLM, a different model, independently critiques it before you see it. Live-streamed agent trace in the UI. • Patch: safe auto-fixes as patch.diff, plus a ready-to-run ROCm Dockerfile, smoke test and benchmark — every change explained from the real diff. Download the fully patched repo as a zip. • Validate: smoke test + benchmark executed on real AMD hardware (passed on Radeon gfx1100, ROCm 7.2, PyTorch 2.9.1, ~0.84 ms inference; replays are always labeled). On failure, a research agent (Kimi) returns a cited, RAG-grounded fix from a ROCm/HIP knowledge base in Qdrant. • Report: an honest, blocker-weighted readiness score (sample repo: 37 → 72 → 86) and an exportable report. Fallback-safe at every layer — no API key, no vector DB, no GPU still yields a working, honest run. 160+ deterministic tests. Built by a two-person team with a protected-branch PR workflow.