Training Infrastructure
Distributed training engines, post-training systems, accelerator communication libraries, and training-efficiency infrastructure; it excludes one-command inference servers and benchmark-only datasets.
| Project | What it does | Stars | Language | License | Latest release |
|---|---|---|---|---|---|
| deepseek-ai/deepep-ascend | A low-level library that accelerates all-to-all communication for distributed MoE training and inference on Huawei Ascend NPUs. | ★ 260 | C++ | — | — |
| zihan-su/self_gradient_forcing_plus | Trains autoregressive video generators by separating context-writing and denoising parameters to reduce gradient conflict. | ★ 198 | Python | Apache-2.0 | — |