20251206

What's Changed

• add catlass ops demo by @ltcs11 in #200
• Add release package dependencies by @monkeyLoveding in #215
• bugfix: The rint interface causes ub to be unaligned. by @chenxu140 in #208
• add pybind11 by @monkeyLoveding in #217
• Upload wheel to release by @BourneSun0527 in #218
• Upload wheel to release (#218) by @BourneSun0527 in #220
• optimizer internode_dispatch hostbound by @zuje123 in #211
• Release workflow switch to cpu by @BourneSun0527 in #222
• switch release machine by @BourneSun0527 in #224
• Add cann path to LD_LIBRARY_PATH by @BourneSun0527 in #225
• [Feat] Lightning indexer op & GE helper engineering by @randgun in #203
• Fixing missing.so files by @monkeyLoveding in #226
• try Fixing missing.so files by @monkeyLoveding in #228
• add sinks_attenton for GPT-OSS by @Todobe in #216
• add zero_experts_compute_identity by @Todobe in #214
• Add performance testing section to the moe script by @goosj in #198
• normal_dispatch num_recv_tokens_per_expert_list support prefixSum by @zuje123 in #221
• A2 dispatch/combine layered operator adaptation for SGLang interface by @oagniqgnat in #209
• Add swiglu_oai for GPT-OSS by @Todobe in #233
• [DFX] Adaptable to multiple model validations for fused moe by @kaniel-outis in #229
• [Bugfix] add padding cases for causal_conv1d_update by @ltcs11 in #235
• [Feat] add chunk_gated_delta_rule triton support by @ltcs11 in #232
• Add two mixed-race tests: normal and low latency, normal and fused deep moe. by @goosj in #206
• debug deepep build by @BourneSun0527 in #231
• rework release build by @iforgetmyname in #237

New Contributors

• @ltcs11 made their first contribution in #200
• @monkeyLoveding made their first contribution in #215

Full Changelog: 2025112...2025120

20251128

What's Changed

• add test internode for deepep by @zuje123 in #193
• Support run normal mode deepep on a single A2 machine by @luanyundu in #201
• [Test] Testing the generalization of fused moe by @kaniel-outis in #167
• Add whl packages to Github Release by @BourneSun0527 in #204
• Add two scripts by @DubhepPan in #119
• support long cat on a3 by @luanyundu in #182
• calculate dispatch normal input parameters using npu instead of cpu by @lih827 in #177
• Add alloc_extend_kernel by @hw-csong in #196
• Modify deepep README_CN.md by @oagniqgnat in #187
• notify_dispatch kernel change magic from int32_t to uint64_t by @zuje123 in #202

Full Changelog: 2025112...2025112

20251120

What's Changed

• dispatch and combine batchsize support 4096 for A2 by @ruiqiangworking in #173
• remove redundant check by @ruiqiangworking in #175
• optimize deepep setup, package name with cann version by @zuje123 in #178
• deepep low_latency d&c support a2 single server by @zuje123 in #176
• Add README files for mlapo and batch_transpose_matmul by @randgun in #104
• Support device with different counts of AICore (FusedDeepMoe operator) by @wangqiankun13 in #180
• Add triton decode attention kernels by @RuixuanZhang06 in #184
• fix cann version check by @hustmf in #188
• update a verification of HCCL_BUFFSIZE for moe by @goosj in #183
• op transfer kv fixbug by @husf1130 in #194
• add_norm_bias and split_qkv_norm_rope for qwen3 by @chenxu140 in #157
• [Chore] Upgrade CANN to 8.3.RC1 by @iforgetmyname in #195

New Contributors

• @hustmf made their first contribution in #188
• @chenxu140 made their first contribution in #157

Full Changelog: 2025111...2025112

20251110

What's Changed

• Added custom low_latency operators for dispatch/combine in the A2 dec… by @oagniqgnat in #166
• deepep support internode api by @zuje123 in #169
• add layout to ops2 directory by @luanyundu in #171
• Modified the deep_ep README and add A2 operator performance data. by @oagniqgnat in #168
• feat: add verify_tree_greedy_kernel triton kernel by @ranjiewen in #165
• optimize a2 layered combine kernel code by @ruiqiangworking in #172
• feat:tiny bugfix&Performance Optimization by @Yael-X in #170

New Contributors

• @ruiqiangworking made their first contribution in #172

Full Changelog: 2025110...2025111

20251106

What's Changed

• Add dependency on the moe header file of CANN by @DubhepPan in #152
• support small bs = 1 or 2 by @wangyibo1005 in #150
• feat:adapt x86_64 compilation by @Yael-X in #143
• [DFX] Compatible with CAN 8.2 and CAN 8.3 by @kaniel-outis in #158
• add mla_preprocess test script by @LinyuanLi0046 in #153
• [DFX] adapt cann8.3 by @kaniel-outis in #159
• [bugfix] swiglu quant by @Liwansi in #162
• [New Ops] build tree efficient by @hw-csong in #161
• support shallow fused topk=-1 by @wangyibo1005 in #160
• support kvcacheio by @husf1130 in #163
• improve layout kernel on a2 by @luanyundu in #164

New Contributors

• @DubhepPan made their first contribution in #152
• @Liwansi made their first contribution in #162
• @hw-csong made their first contribution in #161

Full Changelog: 2025103...2025110

20251030

What's Changed

• add a2 dispatch layout and update its test by @luanyundu in #149
• support topk=-1 by @wangyibo1005 in #132
• add env to decide whether send out prefix sum or not by @luanyundu in #151
• refactor: make hiddenStateDim a class member in MlaTilingData, Follow up closed PR#82 by @LinyuanLi0046 in #133
• support cachemode int8_nzcache with bf16 in mla_preprocess by @LinyuanLi0046 in #135
• add op transfer_kv_dim_exchange by @husf1130 in #148
• impl fused_swiglu_quant with group_list for deepep-low-latency by @xiaobaicxy in #155
• [Kernel] add Flash-Linear-Attention/layernorm_gated Triton op by @iforgetmyname in #154

New Contributors

• @LinyuanLi0046 made their first contribution in #133
• @husf1130 made their first contribution in #148
• @xiaobaicxy made their first contribution in #155

Full Changelog: 2025102...2025103

20251023

What's Changed

• Change the padding generation from randperm back to arange by @oagniqgnat in #140
• LoRA: moving kernels from vllm-ascend repo by @vlserov in #128
• Update README.md of DeepEp by @goosj in #144

New Contributors

• @vlserov made their first contribution in #128
• @goosj made their first contribution in #144

Full Changelog: 2025102...2025102

20251022

What's Changed

• Update README.md: Add performace of normal and low latency dispatch/combine by @oagniqgnat in #106
• Support debug info for build by @jia-rundong in #99
• Update README by @oagniqgnat in #115
• Synchronous fusion moe by @kaniel-outis in #108
• Fix the severe performance degradation issue of the top9 dispatch in normal mode compared to top8. by @oagniqgnat in #117
• feat:add moe fused operator test draft by @Yael-X in #120
• mlapo fit different hidden state dim by @Todobe in #82
• Not use download.pytorch.org by @jia-rundong in #121
• EPLB for fused_deep_moe by @wangyibo1005 in #116
• [FusedDeepMoe] Support EPLB by @kaniel-outis in #118
• Support different token hidden sizes and gmm hidden sizes [FusedDeepMoe Operator] by @wangqiankun13 in #123
• Delete left useless code [FusedDeepMoe Operator] by @wangqiankun13 in #129
• update qwen3-next performance kernels by @iforgetmyname in #130
• [Bugfix] Remove unused code that causes split failure in Qwen3-Next by @iforgetmyname in #142

New Contributors

• @Todobe made their first contribution in #82
• @wangqiankun13 made their first contribution in #123

Full Changelog: 2025092...2025102

20250926

What's Changed

• Reapply fix for hccl buffer use and verify by @zuje123 in #91
• Compilation warnings pending cleanup by @oagniqgnat in #86
• Add CI to test args -a deepep by @jia-rundong in #84
• [Feature] Add diagnostic modules to dispatch and combine by @oagniqgnat in #95
• add FusedDeepMoe by @wangyibo1005 in #92
• fused_moe_for_sglang by @kaniel-outis in #94
• fix some bug for fused moe by @kaniel-outis in #102
• unfold layout expert limit and fix bug by @luanyundu in #107
• Added --pressure-test function in test_low_latency by @oagniqgnat in #101
• fix for sglang verl and readme by @lbk-sys in #98
• [feat] add batch_matmul_transpose op by @randgun in #77
• update fused moe readme by @kaniel-outis in #110
• Modify test_low_latency to support int8 quantization testing. by @oagniqgnat in #109
• feat:add env var to switch quant by @Yael-X in #112

New Contributors

• @oagniqgnat made their first contribution in #86
• @wangyibo1005 made their first contribution in #92
• @kaniel-outis made their first contribution in #94

Full Changelog: 2025091...2025092

20250913

What's Changed

• mlapo support bf16 KV Cache NZ format by @shengzhaotian in #79
• Fix the memory verification issue within intranode dispatch by @lih827 in #83
• [Feature] add fla and mamba kernels by @iforgetmyname in #87
• Revert "Fix the memory verification issue within intranode dispatch" by @iforgetmyname in #88
• Revert "Separate the buffers used by D/C and notify_dispatch to avoid conflicts" by @iforgetmyname in #89

Full Changelog: 2025090...2025091