20250908

What's Changed

• Update speculative tree ops by @MichelleWu351 in #64
• Add test low latency by @luanyundu in #65
• Solve the mlapo operator's precision anomaly in acl graph by @shengzhaotian in #68
• normal_dispatch enable quant by @zuje123 in #66
• Use Ascend name by @jia-rundong in #62
• Lint by @BourneSun0527 in #71
• Separate the buffers used by D/C and notify_dispatch to avoid conflicts by @lih827 in #67
• feat:add activerank test para by @Yael-X in #75
• torch memory saver for npu. by @lbk-sys in #74

New Contributors

• @BourneSun0527 made their first contribution in #71
• @lih827 made their first contribution in #67
• @lbk-sys made their first contribution in #74

Full Changelog: 2025090...2025090

20250901

What's Changed

• [Feat] Support Buffer::get_dispatch_layout interface by @Yael-X in #15
• support shared expert by @zuje123 in #16
• add CI workflow for Intranode test and Low-latency test by @jia-rundong in #23
• added helloworld ops as example by @xiaomingbao008 in #21
• fix deepep cmakelist bug by @RuixuanZhang06 in #24
• Pass the SOC_VERSION as a parameter in build.sh by @RuixuanZhang06 in #27
• fix SOC_VERSION not found in build.sh by @RuixuanZhang06 in #28
• Add cache_assign op and packaing functionality for sgl_kernel_npu by @RuixuanZhang06 in #31
• Add build deepep step by @jia-rundong in #32
• MTP build tree op and verify tree op by @MichelleWu351 in #35
• add dispatch normal kernel for prefill moe stage by @yuankunhan103 in #34
• add combine normal kernel for prefill moe stage by @yuankunhan103 in #36
• add notify dispatch kernel for prefill stage by @zuje123 in #37
• support normal dispatch and combine by @zuje123 in #33
• [feat] Add assign cache op by @randgun in #29
• Add UT for D_C precision by @jia-rundong in #38
• support offset: int64 and fix buffer waste by @RuixuanZhang06 in #41
• add clang-format for code style by @xiaomingbao008 in #45
• support fused cache assign op by @RuixuanZhang06 in #46
• adapter a2 with empty tp by @zuje123 in #50
• refined clang-format and disabled unused-variable check by @xiaomingbao008 in #52
• Support only build deepep by @jia-rundong in #53
• Fusion operator for MLA Preprocess by @shengzhaotian in #51
• fix dispatch return param recv_count mismatch by @zuje123 in #49
• fix combine memory problem by @zuje123 in #54
• alloc_extend for tokens slots alloc by @ranjiewen in #43
• Delete temporary performance data by @jia-rundong in #58
• [fix] rename assign cache op by @randgun in #48
• support cache loc update op by @RuixuanZhang06 in #60
• Support build option deepep/kernels by @jia-rundong in #59
• fix hostbound before notify dispatch by @zuje123 in #61
• Kernel layout by @luanyundu in #57
• combine use dedicated memory by @kip-cxj in #63

New Contributors

• @jia-rundong made their first contribution in #23
• @RuixuanZhang06 made their first contribution in #24
• @MichelleWu351 made their first contribution in #35
• @yuankunhan103 made their first contribution in #34
• @randgun made their first contribution in #29
• @shengzhaotian made their first contribution in #51
• @ranjiewen made their first contribution in #43
• @luanyundu made their first contribution in #57
• @kip-cxj made their first contribution in #63

Full Changelog: 2025080...2025090

20250802

What's Changed

• created code structure which follow the rule of sgl-kernel by @xiaomingbao008 in #1
• add deep_ep_ascend by @Yael-X in #2
• Merge br_feature/bootstrap_730 into main by @iforgetmyname in #5
• remove meaningless print by @zuje123 in #9
• add deepep test example by @zuje123 in #10
• update deepep readme by @zuje123 in #11
• solve C++ ABI incompatibility by @zuje123 in #12
• set ABI to adapt to different torch_npu versions by @zuje123 in #13

New Contributors

• @iforgetmyname made their first contribution in #5

Full Changelog: https://github.com/sgl-project/sgl-kernel-npu/commits/20250802