deepseek-ai/DeepEP ? reverse-engineered prompt

Build me a Python package for fast GPU communication in large model training, mainly for expert parallel MoE dispatch and combine. It should work with PyTorch distributed, run on Hopper class GPUs, and be built around just in time compiled CUDA kernels so it installs without a full CUDA build step.

I want a clean deep_ep API that lets me create a reusable communication buffer, send tokens to experts across ranks, bring results back, and support low precision modes like FP8 and BF16. If it makes sense, include the newer unified buffer style for both high throughput and low latency paths, plus basic support for the extra experimental pieces like pipeline parallel, context parallel, and remote memory access.

Please include a simple install and test flow, a few examples, and make sure it can run in a typical multi GPU distributed setup. If you need current NCCL or PyTorch docs, look them up online.