2026.03.01

What's Changed

• Build the deepep package with the chip model included. by @oagniqgnat in #274
• fix:buffer control by @Yael-X in #361
• Revert " Build the deepep package with the chip model included." by @kaniel-outis in #363
• reset ci -- run test mixed running for experts on a2. by @zhuyutong332 in #365
• adapt ant moving to A2 single machine by @luanyundu in #362
• Fix the bug that total expert num greater than 256 or local expert num is less than 8 by @luanyundu in #364
• CI execution requirements for separating a2 and a3 by @zhuyutong332 in #367
• support qwen3.5 by @chenxu214 in #377
• Update layernorm_gated.py by @chenxu214 in #378
• GLM5 optimize by @cen121212 in #382
• [fix] Handle transposed w13_weight by @gjsheu in #357
• Fix the bug that the layout kernel crashed when the num of experts is no less than 384 by @luanyundu in #383
• revise causal_conv1d: bugfix and enhance accuracy for model kimilinear by @McZyWu in #370
• feat:[fused_sigmoid_gating_delta_rule_update_npu_kernel] support kda feature--to be aligned with sgl-kernel, for model kimi-linear by @McZyWu in #371
• Bump version to 2026.03.01 by @iforgetmyname in #388

New Contributors

• @chenxu214 made their first contribution in #377
• @gjsheu made their first contribution in #357
• @McZyWu made their first contribution in #370

Full Changelog: 2026.02.01...2026.03.01

2026.02.01.post2

What's Changed

• Build the deepep package with the chip model included. by @oagniqgnat in #274
• fix:buffer control by @Yael-X in #361
• Revert " Build the deepep package with the chip model included." by @kaniel-outis in #363
• reset ci -- run test mixed running for experts on a2. by @zhuyutong332 in #365
• adapt ant moving to A2 single machine by @luanyundu in #362
• Fix the bug that total expert num greater than 256 or local expert num is less than 8 by @luanyundu in #364
• CI execution requirements for separating a2 and a3 by @zhuyutong332 in #367
• support qwen3.5 by @chenxu214 in #377
• Update layernorm_gated.py by @chenxu214 in #378

New Contributors

• @chenxu214 made their first contribution in #377

Full Changelog: 2026.02.01...2026.02.01.post2

2026.02.01.post1

What's Changed

• Build the deepep package with the chip model included. by @oagniqgnat in #274
• fix:buffer control by @Yael-X in #361
• Revert " Build the deepep package with the chip model included." by @kaniel-outis in #363
• reset ci -- run test mixed running for experts on a2. by @zhuyutong332 in #365
• adapt ant moving to A2 single machine by @luanyundu in #362
• Fix the bug that total expert num greater than 256 or local expert num is less than 8 by @luanyundu in #364
• CI execution requirements for separating a2 and a3 by @zhuyutong332 in #367
• 【WIP】GLM by @cen121212 in #373

New Contributors

• @cen121212 made their first contribution in #373

Full Changelog: 2026.02.01...2026.02.01.post1

2026.02.01

What's Changed

• Fix notify dispatch cann8.3 by @oagniqgnat in #245
• fix normal and low_latency layerd rdma_data_size when mixed running by @zuje123 in #246
• fixing release ci by @BourneSun0527 in #248
• add a script for generalize test by @goosj in #131
• sgl-kernel-npu add release version by @iforgetmyname in #253
• modify md file by @BourneSun0527 in #255
• add long sequence feature for normal deep_ep by @oagniqgnat in #254
• [fix] fixup bug in conv1d_update_fn by @zhuyijie88 in #259
• qwen3-next op optimize by @shengzhaotian in #257
• [Bugfix] fix TorchNpuHelper rename bugs by @ltcs11 in #265
• Fixing Chinese character encoding issues by @oagniqgnat in #275
• Add the long-sequence ant migration feature for the prefill combine operator. by @oagniqgnat in #267
• add deepep a2 doc by @zuje123 in #277
• prepare build and release for a2 by @iforgetmyname in #273
• fix a2 deepep doc by @zuje123 in #279
• Fix the issue of HCCL buffer tiling verification failure during one round of testing. by @oagniqgnat in #280
• l2 norm const parameter change by @shengzhaotian in #276
• bump version to 2025.12.25 by @iforgetmyname in #281
• modify split_qkv_rmsnorm_rope by @Liwansi in #282
• Fix the performance degradation issue of the single-wheel operation in Ant Moving. by @oagniqgnat in #287
• Optimize prepare_lens by removing device transfer by @shengzhaotian in #289
• split_qkv_rmsnorm_rope bugfix by @Liwansi in #290
• fix notify magic auto-increment bug by @zuje123 in #291
• Resolving the UB out-of-bounds issue caused by A2 dual-machine mixed operation by @oagniqgnat in #288
• fix a2 single combine aclnn params by @zuje123 in #292
• LoRA: Optimization LoRA kernels and refactoring by @vlserov in #284
• Support build with cann 8.5 by @BourneSun0527 in #283
• Added an environment variable to control whether to enable the Combine Ant Migration feature. by @oagniqgnat in #304
• Supplement A2 doc, software and hardware compatibility info by @zuje123 in #294
• fix layout numTokensPerExpertTensor partial Initialization bug by @zuje123 in #303
• optimize gdn gating and fused_qkvzba_split_reshape_cat by @RuixuanZhang06 in #306
• support add_gemma_rms_norm by @RuixuanZhang06 in #310
• feat:add performance compare by @Yael-X in #311
• [chore] version bump to 2026.01.12 by @iforgetmyname in #312
• Add swiglu_oai_triton for GPTOSS by @Todobe in #270
• Optimize sinks attention for prefix cache by @Todobe in #260
• fix little batchsize and int8 quant on ci by @zhuyutong332 in #302
• fix bmm transpose in cann 8.5 by @randgun in #316
• Modify contribution guide by @BourneSun0527 in #315
• Integrate ccache for faster compilation by @randgun in #318
• add dfx for operator FusedDeepMoe by @wangyibo1005 in #317
• [Chore] CANN version bump to 8.5.0 by @iforgetmyname in #326
• Deepep adapt custom cann installation path by @BourneSun0527 in #327
• Optimize the performance of the Combine Ant Moving function and the use of HCCL buffer by @oagniqgnat in #314
• remove the limit that A2 internode only support topk 8 by @luanyundu in #323
• add deepep normal api doc by @zuje123 in #336
• 【Doc】add fused deep moe doc by @kaniel-outis in #335
• Document get_dispatch_layout API by @luanyundu in #338
• Added the verification of num_max_dispatch_tokens_per_rank to the decode operator adaptation layer. by @oagniqgnat in #330
• Added the low_latency operator API documentation. by @oagniqgnat in #337
• The environment variable DEEPEP_HCCL_BUFFSIZE is added by @zzx-study in #329
• chunk_gated_delta_rule_npu output final state by @RuixuanZhang06 in #341
• support the situation that topk maybe -1 on machine A3 by @luanyundu in #313
• Add AscendC triangular inverse by @zouzias in #332
• (test) add solve_tril from upstream by @zouzias in #339
• [Doc] Improved README.md content and English grammar and integrated the DeepWiki badge for Ask AI by @Mitchell-xiyunfeng in #345
• add function for deep-ep tests by @zhuyutong332 in #301
• Support x86_64 and aarch64 binary release by @iforgetmyname in #325
• Add scripts for building CMake files by @1329009851 in #344
• Revert "Add scripts for building CMake files" by @1329009851 in #353
• Modify the description of DeepEP in the README file. by @oagniqgnat in #348
• [Bugfix] Fix build script working with cann 8.5.0 by @iforgetmyname in #354
• fix the hanging bug by @luanyundu in #355
• Modify notifydispatch to support DEEPEP_NORMAL_LONG_SEQ_ROUND up to 128. by @WSEmma in #352
• release follows naming convention by @iforgetmyname in #356
• Cover the workflows cases on a3 by @zhuyutong332 in #321
• [Chore] Bump sgl-kernel-npu version to 2026.02.01 by @iforgetmyname in #359
• [Bugfix] Fix mismatched package building directory by @iforgetmyname in #360

New Contributors

• @zhuyijie88 made their first contribution in #259
• @zhuyutong332 made their first contribution in #302
• @zzx-study made their first contribution in #329
• @zouzias made their first contribution in #332
• @Mitchell-xiyunfeng made their first contribution in #345
• @1329009851 made their first contribution in #344
• @WSEmma made their first contribution in #352

Full Changelog: 2025120...2026.02.01

2026.01.28

What's Changed

• Added the low_latency operator API documentation. by @oagniqgnat in #337
• The environment variable DEEPEP_HCCL_BUFFSIZE is added by @zzx-study in #329
• chunk_gated_delta_rule_npu output final state by @RuixuanZhang06 in #341
• support the situation that topk maybe -1 on machine A3 by @luanyundu in #313
• Add AscendC triangular inverse by @zouzias in #332
• (test) add solve_tril from upstream by @zouzias in #339
• [Doc] Improved README.md content and English grammar and integrated the DeepWiki badge for Ask AI by @Mitchell-xiyunfeng in #345
• add function for deep-ep tests by @zhuyutong332 in #301
• Support x86_64 and aarch64 binary release by @iforgetmyname in #325

New Contributors

• @zzx-study made their first contribution in #329
• @zouzias made their first contribution in #332
• @Mitchell-xiyunfeng made their first contribution in #345

Full Changelog: 2026.01.21...2026.01.28

2026.01.21

Added the verification of num_max_dispatch_tokens_per_rank to the dec…

2026.01.19

What's Changed

• Add swiglu_oai_triton for GPTOSS by @Todobe in #270
• Optimize sinks attention for prefix cache by @Todobe in #260
• fix little batchsize and int8 quant on ci by @zhuyutong332 in #302
• fix bmm transpose in cann 8.5 by @randgun in #316
• Modify contribution guide by @BourneSun0527 in #315
• Integrate ccache for faster compilation by @randgun in #318
• add dfx for operator FusedDeepMoe by @wangyibo1005 in #317
• [Chore] CANN version bump to 8.5.0 by @iforgetmyname in #326

New Contributors

• @zhuyutong332 made their first contribution in #302

Full Changelog: 2026.01.12...2026.01.19

2026.01.12

What's Changed

• support add_gemma_rms_norm by @RuixuanZhang06 in #310
• feat:add performance compare by @Yael-X in #311
• [chore] version bump to 2026.01.12 by @iforgetmyname in #312

Full Changelog: 2026.01.09...2026.01.12

2026.01.09

What's Changed

• optimize gdn gating and fused_qkvzba_split_reshape_cat by @RuixuanZhang06 in #306

Full Changelog: 2026.01.07...2026.01.09

2026.01.07

What's Changed

• LoRA: Optimization LoRA kernels and refactoring by @vlserov in #284
• Support build with cann 8.5 by @BourneSun0527 in #283
• Added an environment variable to control whether to enable the Combine Ant Migration feature. by @oagniqgnat in #304
• Supplement A2 doc, software and hardware compatibility info by @zuje123 in #294
• fix layout numTokensPerExpertTensor partial Initialization bug by @zuje123 in #303

Full Changelog: 2025.12.31...2026.01.07