DeepEP GitHub repository (DeepSeek)

README, MIT. Facts: 'EP dispatch and combine require GPU SMs; zero-SM RDMA EP is not supported'; requires NVIDIA Hopper or newer, NVLink for intranode and RDMA for internode; bandwidth auto-detected with nvidia-smi and ibstat, and BucketBuffer needs both detectable even for a single-transport group. Low-latency kernels expose num_sms so the comm kernel does not occupy every SM. With do_cpu_sync=False (the decode path) outputs are allocated to configured capacity and the expert GEMM must consume handle.psum_num_recv_tokens_per_expert; defer_epilogue=True with async_with_compute_stream=True returns an EventOverlap so the epilogue runs on the current stream while compute proceeds. Also adds lb_prefetch_weights / lb_reduce_grads over NVLink for redundant experts.