合并deepep的auto模式分支" See merge request dcutoolkit/deeplearing/vllm!335
[feat]解决deepep auto模式 mtp >1 时卡住 See merge request dcutoolkit/deeplearing/vllm!326
[feat]解决deepep auto模式卡住问题 See merge request dcutoolkit/deeplearing/vllm!324
[fix]解决EP不启用mtp时某些size报错 See merge request dcutoolkit/deeplearing/vllm!322
[feat]解决EP开启pd_padding大输入卡住问题 See merge request dcutoolkit/deeplearing/vllm!319
[feat]低延迟模式采用int8 dispatch See merge request dcutoolkit/deeplearing/vllm!314
[fix]1.解决ep sequence parallel优化引入的mtp报错;2.解决共享专家无法和combine overlap问题 See merge request dcutoolkit/deeplearing/vllm!312
[feat]优化dp attention，减少1次allgather耗时，高吞吐提升明显 See merge request dcutoolkit/deeplearing/vllm!307
w8a8 高吞吐模式先量化再dispatch See merge request dcutoolkit/deeplearing/vllm!303
[feat]修复低延迟错误 See merge request dcutoolkit/deeplearing/vllm!301
优化高吞吐模式num_sms See merge request dcutoolkit/deeplearing/vllm!297
[feat]1.支持高吞吐模式ep_scatter+deepgemm contiguous+ep_gather方案；2.支持高吞吐模式下ETP,例如dp4 tp4 See merge request dcutoolkit/deeplearing/vllm!296