| --- |
| library_name: kernels |
| license: mit |
| --- |
| |
| This is the repository card of kernels-community/deep-ep that has been pushed on the Hub. It was built to be used with the [`kernels` library](https://github.com/huggingface/kernels). This card was automatically generated. |
|
|
| ## How to use |
|
|
| Usage example not available. |
|
|
| ## Available functions |
|
|
| Function list not available. |
|
|
| ## Requirements |
|
|
| This is [DeepEP](https://github.com/deepseek-ai/DeepEP) 2.5.0 (`def8651`), with |
| [DeepJIT](https://github.com/deepseek-ai/DeepJIT) `cba802d`. Only DeepEP's host |
| extension is prebuilt; its communication kernels are compiled at runtime. On |
| top of the Hub kernel's own requirements, using it needs: |
|
|
| - NVIDIA Hopper or newer GPUs, connected with NVLink (RDMA for multi-node). |
| - NCCL >= 2.32.3, loaded by PyTorch. The kernel does not ship NCCL: it uses the |
| NCCL library that PyTorch loads, and compiles its kernels against that |
| library's headers. PyTorch 2.14 and 2.15 ship NCCL 2.30.7; upgrade it with |
| `pip install "nvidia-nccl-cu13>=2.32.3"` (or use a PyTorch build that ships |
| 2.32.3 or newer). |
| - A CUDA toolkit >= 13.1 (`nvcc`) and a C++20 host compiler at runtime, found |
| through `CUDA_HOME`, `CUDA_PATH`, `nvcc` on `PATH` or `/usr/local/cuda`. |
|
|
| These are checked when the first buffer is created, with an error explaining |
| what is missing. Compiled kernels are cached in `$HOME/.dj` (`EP_JIT_CACHE_DIR`). |
|
|
| ## Benchmarks |
|
|
| No benchmark available yet. |
|
|
| ## Source code |
|
|
| Source code of this kernel originally comes from https://github.com/deepseek-ai/DeepEP and it was repurposed for compatibility with `kernels`. |
|
|
