What problem does it solve? Migrating Mixture of Experts (MoE) code built on DeepEP from CUDA/NCCL environments to Ascend NPU platforms requires manual operator replacement, communication backend conversion, and strict constraint validation, which is error-prone and time-consuming. ## Core Features & Use Cases - DeepEP Operator Detection: Scans code for import deep_ep and dispatch/combine calls, analyzes local dependencies and communication domain setup. - Runtime Constraint Validation: Traces actual runtime parameter values (hidden_size, top_k, num_experts) instead of relying on argparse defaults, and validates them against CAM operator constraints for A2, A3, Shmem, and fused deep moe modes. - Automated Migration: Converts NCCL to HCCL, CUDA to NPU device calls, and replaces DeepEP dispatch/combine interfaces with the corresponding umdk_cam_op_lib CAM operators, including Shmem memory management. - Use Case: Given a PyTorch MoE test script using deep_ep.Buffer for expert parallelism, the skill verifies the runtime configuration fits CAM constraints, asks the user to choose between A3 normal, Shmem, or fused deep moe modes, and rewrites the file in place for Ascend NPU execution. ## Quick Start Migrate my DeepEP-based MoE dispatch and combine code in this file to CAM operators running on Ascend NPU.
