• Support normal DeepEP buffer @liz-badada [Feature] Integrate DeepEP into SGLang #4232
• Support DeepEP with async transfer @fzyzcjy Support async in DeepEP #4610
• Support low-latency DeepEP buffer
  • Single-node TP @liz-badada [Feature] Support DeepEP Low Latency #4767
    • MaskedDeepGeMM is implemented by @laixinn @sleepcoo
    • Improved by @yuleil [DeepEP] Reduce routed scaling overhead #5277
  • Multi-node TP @liz-badada [Fix] DeepEP Compatibility with Low Latency #5068
  • Support PD disaggregation @ch-wan Integrating PD disaggregation with DP attention and DeepEP #5435
• Integrate pplx-kernels @ruizhang1230 Support PPLX kernels for Deepseek models #7272
• Optimize permutation overhead
  • Implement Titon kernels @xutizhou Optimize Permute Kernel in DeepEP #4643
  • Fuse permutation with GroupedGeMM
• Extend parallelism paradigm
  • Extend DeepEP to a general TP paradigm @ch-wan @tarinkk Support (1 <= dp < tp) in the dp attention in DeepEP #4770
    • Fixed by @fzyzcjy Fix DeepSeek bug causing 2.2% MMLU drop when TP!=DP #4883
  • Support tp_size < ep_size
    • tp_size=1 @fzyzcjy Introduce moe_dense_tp_size to fix dense layer errors in DeepSeek V3 + 4x8xH100 #4836
• Overlap two batches @fzyzcjy Support overlapping two batches #4068
• Integrate continuous DeepGeMM @sleepcoo @xutizhou DeepEP normal support deepgemm-contiguous #5626
• Record expert distribution @yuhsuan-t Add endpoints to dump selected expert ids #4435
  • Improved by @fzyzcjy Expert distribution recording without overhead for EPLB #4957
• Overlap communication with shared experts’ computation @liz-badada [Feature] Overlap DeepEP Combine and Shared Experts inside same batch #5829
• Integrate EPLB @fzyzcjy EPLB #5295

Others

• The DeepSeek team is going to release a permutation kernel shortly. We may need to check their update How to use the output of DeepEP as the input of DeepGemm deepseek-ai/DeepGEMM#57 (comment)