deepep-to-cam-converter

What problem does it solve?

DeepEP-based Mixture-of-Experts (MoE) code is difficult to port to Huawei Ascend CAM operators safely and correctly, especially when communication domains and operator constraints must match real runtime values.

Core Features & Use Cases

• DeepEP MoE dispatch/combine detection: Scans for DeepEP imports and dispatch/combine patterns to confirm the code path is eligible for migration.
• Runtime-parameter constraint verification: Enforces that operator constraints are validated using actual runtime values (not argparse defaults).
• NCCL→HCCL and CUDA→NPU migration: Converts communication backend and device/tensor calls to the target Ascend execution environment.
• A2/A3 operator targeting with interactive mode selection: Guides environment selection (A2 vs A3) and, in ambiguous A3 cases, forces the user to choose between compatible CAM modes.
• Safety-first unsupported-feature handling: Produces a feature support report and asks the user whether to stop, partially migrate, or fully delete unsupported code paths.

Quick Start

Use this skill to convert a DeepEP MoE project into Ascend CAM operators by migrating dispatch & combine logic and communication/device handling for NCCL/HCCL and CUDA/NPU.