v0.3.1

What's Changed

• feat: ck fp8 add hybrid format by @RuibinCheung in #279
• fix: use CUDA torch APIs in deep_ep to avoid hip_torch build failures by @zhenhuang12 in #281
• fix: amd-aiter uninstall and package check by @llying-001 in #289
• Dev/llying/update image v262 by @llying-001 in #287
• feat: attn support sbhd layout by @RuibinCheung in #275
• fix triton gemm determistic and correctness issue on gfx950 by @llying-001 in #286
• Allow Triton GEMM selectors to run without origami. by @kyle-256 in #296
• fix(typo): rename MXScalingRecipe, grouped gemm fp8 function name and remove scale args from quantize api by @RuibinCheung in #298
• refactor(moe): extensible EPBackend Protocol + EPBufferConfig & dispatcher test overhaul by @zhenhuang12 in #297
• fix: add input assertion for aiter attention by @RuibinCheung in #301
• fix: compile warning by @RuibinCheung in #302
• feat: clean unused env var and unify env var by @RuibinCheung in #305
• feat: attn add bhsd layout support by @RuibinCheung in #304
• fix(quantization): correct partial-tile indexing in reduce_row_kernel by @jasainio in #306
• fix: mxfp8 gemm WAR race between ds_read and buffer_load_lds by @xiaobochen-amd in #308
• fix: quantize mxfp8 int32 overflow by @xiaobochen-amd in #307
• [Attention] Torch.compile custom wrappers by @paulpak58 in #310
• update codeowners by @xiaobochen-amd in #315
• refact: ck gemm and grouped gemm directory structure by @RuibinCheung in #317
• feat(moe): add back deepep_use_comm_stream to force use single-stream by @zhenhuang12 in #314
• fix: add dispatch kwargs format print by @RuibinCheung in #316
• feat(grouped_gemm): switch default backend to Triton for BF16 and FP8 by @kyle-256 in #320
• fix: fix deepep combine hang by @zhenhuang12 in #327
• docs: add SECURITY.md for vulnerability reporting by @xiaobochen-amd in #331
• fix(grouped_gemm): bind per-stream hipBLASLt handles to fix bf16 multi-stream stall on gfx942 by @kyle-256 in #319
• [Attention] AITER varlen attention support by @paulpak58 in #332
• fix: hipblaslt_gemm correct workspace size for MI300X by @luiza-amd in #333
• fix: remove attn workaround by @RuibinCheung in #313
• Add HYBRID FP8 format support for Triton backend in gemm and grouped_gemm by @kyle-256 in #278
• feat(benchmark): support benchmark sharding for single case by @ChengYao-amd in #340
• feat(bench): add sharing support at benchmark yaml by @ChengYao-amd in #341
• feat: support hip permute kernel by @zhenhuang12 in #328
• feat(gemm): Drop input_precision="ieee" from FP8 tl.dot by @ChengYao-amd in #342
• feat(gemm): optimize blockwise FP8 GEMM Triton kernels (gfx950 / MI355X) by @ChengYao-amd in #339
• feat: add quantized tensor support by @RuibinCheung in #335
• docs: update example by @xiaobochen-amd in #346
• feat(jax): add 1g1p intranode and internode deepep for jax by @llying-001 in #344
• feat: refine rms norm ops by @RuibinCheung in #343
• refactor: moe_permute and DeepEPTokenDispatcher by @zhenhuang12 in #345
• [Op][Normalization] Triton BWD Norm Tuning + Persistent grid-stride by @paulpak58 in #352
• refactor(moe): rename num_permuted_tokens to permute_max_token_num in DeepEPTokenDispatcher by @zhenhuang12 in #354
• fix(moe): declare probs layout explicitly in moe_permute by @zhenhuang12 in #355
• feat: refactor origama and triton knobs by @RuibinCheung in #357
• fix(attention): match qkv_format strides in forward fake kernel by @jasainio in #362
• fix: add hybrid format test for float8linear by @RuibinCheung in #360
• feat(agent): Add Primus Turbo optimization agent skills by @ChengYao-amd in #338
• fix(agent): delete Chinese charactors by @ChengYao-amd in #368
• Add JAX vmap batching rules for MoE/grouped-GEMM primitives (pipeline-parallel support) by @yihuang-amd in #350
• build: pip-installable packaging + self-hosted PEP 503 release pipeline by @xiaobochen-amd in #239
• ci: build release sdist on the ROCm runner by @xiaobochen-amd in #370
• ci(release): drop submodule checkout in validate job by @xiaobochen-amd in #371
• opt(grouped_blockwise): MI300/MI355 grouped blockwise FP8 GEMM + HIP fused quant by @kyle-256 in #358
• fix: bring hipblaslt multi-stream grouped gemm back in autotune by @RuibinCheung in #367
• build(release): pin aiter by release tag and slim the sdist by @xiaobochen-amd in #372

New Contributors

• @luiza-amd made their first contribution in #333
• @yihuang-amd made their first contribution in #350

Full Changelog: v0.3.0...v0.3.1

v0.3.1-rc2

What's Changed

• opt(grouped_blockwise): MI300/MI355 grouped blockwise FP8 GEMM + HIP fused quant by @kyle-256 in #358
• fix: bring hipblaslt multi-stream grouped gemm back in autotune by @RuibinCheung in #367
• build(release): pin aiter by release tag and slim the sdist by @xiaobochen-amd in #372

Full Changelog: v0.3.1-rc1...v0.3.1-rc2

v0.3.1-rc1

What's Changed

• feat: ck fp8 add hybrid format by @RuibinCheung in #279
• fix: use CUDA torch APIs in deep_ep to avoid hip_torch build failures by @zhenhuang12 in #281
• fix: amd-aiter uninstall and package check by @llying-001 in #289
• Dev/llying/update image v262 by @llying-001 in #287
• feat: attn support sbhd layout by @RuibinCheung in #275
• fix triton gemm determistic and correctness issue on gfx950 by @llying-001 in #286
• Allow Triton GEMM selectors to run without origami. by @kyle-256 in #296
• fix(typo): rename MXScalingRecipe, grouped gemm fp8 function name and remove scale args from quantize api by @RuibinCheung in #298
• refactor(moe): extensible EPBackend Protocol + EPBufferConfig & dispatcher test overhaul by @zhenhuang12 in #297
• fix: add input assertion for aiter attention by @RuibinCheung in #301
• fix: compile warning by @RuibinCheung in #302
• feat: clean unused env var and unify env var by @RuibinCheung in #305
• feat: attn add bhsd layout support by @RuibinCheung in #304
• fix(quantization): correct partial-tile indexing in reduce_row_kernel by @jasainio in #306
• fix: mxfp8 gemm WAR race between ds_read and buffer_load_lds by @xiaobochen-amd in #308
• fix: quantize mxfp8 int32 overflow by @xiaobochen-amd in #307
• [Attention] Torch.compile custom wrappers by @paulpak58 in #310
• update codeowners by @xiaobochen-amd in #315
• refact: ck gemm and grouped gemm directory structure by @RuibinCheung in #317
• feat(moe): add back deepep_use_comm_stream to force use single-stream by @zhenhuang12 in #314
• fix: add dispatch kwargs format print by @RuibinCheung in #316
• feat(grouped_gemm): switch default backend to Triton for BF16 and FP8 by @kyle-256 in #320
• fix: fix deepep combine hang by @zhenhuang12 in #327
• docs: add SECURITY.md for vulnerability reporting by @xiaobochen-amd in #331
• fix(grouped_gemm): bind per-stream hipBLASLt handles to fix bf16 multi-stream stall on gfx942 by @kyle-256 in #319
• [Attention] AITER varlen attention support by @paulpak58 in #332
• fix: hipblaslt_gemm correct workspace size for MI300X by @luiza-amd in #333
• fix: remove attn workaround by @RuibinCheung in #313
• Add HYBRID FP8 format support for Triton backend in gemm and grouped_gemm by @kyle-256 in #278
• feat(benchmark): support benchmark sharding for single case by @ChengYao-amd in #340
• feat(bench): add sharing support at benchmark yaml by @ChengYao-amd in #341
• feat: support hip permute kernel by @zhenhuang12 in #328
• feat(gemm): Drop input_precision="ieee" from FP8 tl.dot by @ChengYao-amd in #342
• feat(gemm): optimize blockwise FP8 GEMM Triton kernels (gfx950 / MI355X) by @ChengYao-amd in #339
• feat: add quantized tensor support by @RuibinCheung in #335
• docs: update example by @xiaobochen-amd in #346
• feat(jax): add 1g1p intranode and internode deepep for jax by @llying-001 in #344
• feat: refine rms norm ops by @RuibinCheung in #343
• refactor: moe_permute and DeepEPTokenDispatcher by @zhenhuang12 in #345
• [Op][Normalization] Triton BWD Norm Tuning + Persistent grid-stride by @paulpak58 in #352
• refactor(moe): rename num_permuted_tokens to permute_max_token_num in DeepEPTokenDispatcher by @zhenhuang12 in #354
• fix(moe): declare probs layout explicitly in moe_permute by @zhenhuang12 in #355
• feat: refactor origama and triton knobs by @RuibinCheung in #357
• fix(attention): match qkv_format strides in forward fake kernel by @jasainio in #362
• fix: add hybrid format test for float8linear by @RuibinCheung in #360
• feat(agent): Add Primus Turbo optimization agent skills by @ChengYao-amd in #338
• fix(agent): delete Chinese charactors by @ChengYao-amd in #368
• Add JAX vmap batching rules for MoE/grouped-GEMM primitives (pipeline-parallel support) by @yihuang-amd in #350
• build: pip-installable packaging + self-hosted PEP 503 release pipeline by @xiaobochen-amd in #239
• ci: build release sdist on the ROCm runner by @xiaobochen-amd in #370
• ci(release): drop submodule checkout in validate job by @xiaobochen-amd in #371

New Contributors

• @jasainio made their first contribution in #306
• @paulpak58 made their first contribution in #310
• @luiza-amd made their first contribution in #333
• @yihuang-amd made their first contribution in #350

Full Changelog: v0.3.0...v0.3.1-rc1

v0.3.0

What's Changed

• feat(mxfp8): opt mxfp8 quantize kernel by @RuibinCheung in #165
• feat(deepep): support single-node moe_dispatch/combine on jax. by @zhenhuang12 in #162
• performance: jax group gemm align with pytorch version by @kyle-256 in #163
• Revert "performance: jax group gemm align with pytorch version" by @xiaobochen-amd in #168
• docs: update readme by @xiaobochen-amd in #169
• cicd: add benchamrk yaml by @xiaobochen-amd in #171
• feat: Backend Management and Auto-Tune Infrastructure by @xiaobochen-amd in #167
• feat: add hipblaslt grouped_gemm into Pytorch backend by @kyle-256 in #172
• Refactor the attention interface and add unified sequence parallel by @limou102 in #173
• chore: update readme & version number & cicd docker by @xiaobochen-amd in #174
• Fix import error of usp attention package by @limou102 in #177
• setup: if aiter already installed, skipping by @xiaobochen-amd in #178
• feat: add multi-process tests & clean unused code. by @xiaobochen-amd in #176
• refactor: hipblaslt grouped gemm c++ code. by @xiaobochen-amd in #181
• cicd : update docker image to v25.10 by @xiaobochen-amd in #183
• fix a autotune grouped gemm bug by @xiaobochen-amd in #184
• feat(mxfp4): Add mxfp4 gemm and fix scale calculate of mxfp8 by @RuibinCheung in #179
• feat: add compiler flag check for ROCm compatibility by @xiaobochen-amd in #185
• fix: add GPU sync to prevent auto-tune hang by @xiaobochen-amd in #186
• chore: set default build dir & rename grouped_gemm by @xiaobochen-amd in #187
• feat: turbo support ck based grouped gemm blockwise kernel by @kyle-256 in #182
• jax: add jax new feat and refactor ck_grouped_gemm/quant code by @xiaobochen-amd in #188
• feat(mxfp4/mxfp8): optimize mxfp4 performance and add 2D block quantization for mxfp8 by @RuibinCheung in #189
• setup: fix: redirect build dir to project root in editable install by @xiaobochen-amd in #191
• feat: update benchmark & cicd by @xiaobochen-amd in #192
• chore: update ck version to avoid rocm72 compile issue by @xiaobochen-amd in #193
• bench: update bench yaml by @xiaobochen-amd in #194
• feat: grouped gemm fp8 tensorwise support hipblaslt backend by @xiaobochen-amd in #196
• chore: add summarize_reults.py for benchmark by @xiaobochen-amd in #198
• benchmark: add more model config by @xiaobochen-amd in #199
• bench: add torch baseline by @xiaobochen-amd in #200
• feat(deepep): support third-party deepep as backend of TokenDispatcher. by @zhenhuang12 in #197
• feat(mxfp4): refactor gemm mxfp4 and mxfp8. fuse transpose, hadamard transform and quantization. by @RuibinCheung in #195
• feat(mxfp4/mxfp8): refine mxfp4 and mxfp8 by @RuibinCheung in #201
• bench: add h200 perf data by @xiaobochen-amd in #202
• bench: update csv format by @xiaobochen-amd in #205
• bench: remove slow backend and test case to avoid overtime by @xiaobochen-amd in #206
• feat(mxfp8/mxfp4): reduce quantize call in gemm mxfp8/mxfp4 and support fp4 turbo benchmark by @RuibinCheung in #203
• feat: add deepep intranode benchmark by @zhenhuang12 in #212
• opt: grouped_gemm_variable_k_postprocess perf by @xiaobochen-amd in #214
• fix: bench intranode deepep import error by @zhenhuang12 in #213
• [chore] fix multi zeros in group_lens issue by @kyle-256 in #216
• chore: add issue template by @xiaobochen-amd in #219
• feat(attn): support sink attention by @kyle-256 in #208
• chore: temporarily skip some hipBLASLt grouped GEMM test cases to ens… by @xiaobochen-amd in #222
• feat(mxfp4/mxfp8): refine quantization api and optimize kernels perf by @RuibinCheung in #207
• chore: update aiter & add attn deterministic benchmark by @xiaobochen-amd in #223
• cicd: update cicd runner by @xiaobochen-amd in #226
• [feat]: blockwise grouped_gemm_fp8 support hipgraph by @kyle-256 in #221
• feat(ci): build and install uccl-ep by @zhenhuang12 in #210
• bench: add TE fuser attention benchmark for AMD and NV by @vlluvia in #227
• [test] update blockwise grouped_gemm test case in hipgraph mode by @kyle-256 in #229
• feat(gg): hipblaslt backend support fp8 hybrid format in tensorwise gg by @RuibinCheung in #230
• feat: grouped gemm auto-tune on load-balance mode by @xiaobochen-amd in #231
• chore: update docker by @xiaobochen-amd in #232
• docs: update example.md by @xiaobochen-amd in #233
• fix: enable grouped gemm grad_out contiguous by @xiaobochen-amd in #234
• fix(grouped_gemm): use thread-local hipblaslt runner by @xiaobochen-amd in #238
• chore(aiter): update aiter to fix FA asm kernel bug by @RuibinCheung in #237
• test: add attn/gemm/gg deterministic test by @xiaobochen-amd in #224
• fix: scale may exist nan or inf by @RuibinCheung in #240
• [triton] re-write gemm/grouped_gemm triton backend for gfx942. by @kyle-256 in #235
• fix: force to use fgpu-rdc to compile by @zhenhuang12 in #236
• [feat] do not use triton autotune in gemm bf16 and fp8 tensorwise&rowwise kernel by @kyle-256 in #242
• ci: parallelize benchmarks across 8 GPUs via HIP_VISIBLE_DEVICES by @xiaobochen-amd in #246
• refactor(benchmark): config-driven GPU pool runner to replace inline … by @xiaobochen-amd in #247
• feat: attn 192/128 use aiter asm by @xiaobochen-amd in #248
• fix(typo): fp4 benchmark filename typo by @RuibinCheung in #245
• perf: update aiter for better attn-deterministic bwd perf. by @xiaobochen-amd in #250
• fix: use cheap fence to ensure memory order. by @zhenhuang12 in #251
• Fix HIPStreamMasqueradingAsCUDA build failure and setup.py issues by @t2h-fb in #253
• refactor(backend): introduce BackendEntry to support per-backend auto… by @xiaobochen-amd in #255
• update triton impl for mi300 & mi355 by @kyle-256 in #252
• reduce gemm blockwise test cases by @kyle-256 in #258
• feat: Make DeepEP CPU timeout configurable and increase default timeout by @zhenhuang12 in #256
• feat(mxfp4): integrate aiter fp4 gemm and support hip mxfp4 quant and shuffle kernel by @RuibinCheung in #249
• fix: refine get_fp8_dtype func by @RuibinCheung in #257
• fix sm_scale none bug by @xiaobochen-amd in #263
• ci: update turbo runner to tas node by @llying-001 in #273
• feat: refine mxfp8 quant kernel and optimzie mxfp4 quant kernel by @RuibinCheung in #261
• fix(JAX): correct moe dispatch backward gradients by @zhenhuang12 in #264
• opt: gemm mxfp8 kernel by @xiaobochen-amd in #274
• [feat] add slid...

v0.2.0

What's Changed

• fix(deepep): eliminate compile warning. by @zhenhuang12 in #123
• feat(deep_ep): support num_worst_token and use_defaulta_stream_as_comm_stream for internode. by @zhenhuang12 in #120
• feat(token_dispatcher): add DeepEPTokenDispatcher for MoE. by @zhenhuang12 in #114
• build: support multi-arch compilation (gfx942;gfx950) by @xiaobochen-amd in #124
• feat: attn add is_v3_atomic_fp32 env control by @xiaobochen-amd in #126
• chore: move router to moe dir by @xiaobochen-amd in #125
• [Sync-free MoE] feat: add swiglu, geglu and tokens_per_expert_to_mask api by @RuibinCheung in #122
• [HOTFIX] triton version requirement by @RuibinCheung in #130
• [Sync-free MoE] feat: refine act func by @RuibinCheung in #129
• fix(deepep): fix bug when use expert_capacity_factor by @zhenhuang12 in #127
• chore(docker): update default image to rocm/primus:v25.9_gfx942 by @xiaobochen-amd in #133
• [Aiter] Update aiter to fix pybind11 issue by @GeneDer in #132
• feat: gemm fp8 support cktile backend for both tensorwise and rowwise by @kyle-256 in #131
• [Fix] import activation module by @GeneDer in #137
• feat(deepep): move deep_ep header file to primus_turbo common header dir by @zhenhuang12 in #138
• feat(permute): permute op support to compute tokens_per_expert by @zhenhuang12 in #140
• feat: grouped gemm tensorwise impl update by @kyle-256 in #139
• chore: support jax=0.6.2 & jax cicd by @xiaobochen-amd in #136
• feat: add elementwise(unary/binary/quant/dequant) kernel by @xiaobochen-amd in #135
• chore: remove uselsee debug code by @kyle-256 in #141
• chore: refactor grouped gemm blockwise python code by @xiaobochen-amd in #142
• feat: add build ext and opt build efficiency by @xiaobochen-amd in #143
• feat: skip patch torch_extension when version >=2.8.0 by @zhenhuang12 in #144
• chore: refactor gemm fp8 api by @xiaobochen-amd in #145
• fix: skip disabled arch files in build by @xiaobochen-amd in #147
• chore: support quant gemm when m%128!=0 by @kyle-256 in #146
• feat: unify fp8 gemm API by @RuibinCheung in #148
• chore: update aiter version. by @xiaobochen-amd in #151
• add public primus-safe link in readme by @wenxie-amd in #152
• fix readme's typo error by @wenxie-amd in #153
• fix(deepep): fix internode_combine hang when set num_worst_token > 0 by @zhenhuang12 in #149
• [feat]: CK based block quant by @kyle-256 in #155
• feat(gemm): Add mxfp8 gemm and quantize kernel by @RuibinCheung in #154
• opt: grouped gemm perf when len(group_lens)==1 by @xiaobochen-amd in #158
• feat: add float4x2_e2m1 and float8_m8m0 data type by @xiaobochen-amd in #156
• feat(mxfp8): add k padding in bwd by @RuibinCheung in #160
• chore: Allow GEMM with k % 32 = 0 to participate in computation by @kyle-256 in #161
• feat: jax backend support grouped_gemm by @kyle-256 in #157

Full Changelog: v0.1.1...v0.2.0

v0.1.1

What's Changed

• feat(deepep): optimize gpu-cpu nosync. by @zhenhuang12 in #93
• feat: add std::numeric_limits specializations for fp8 (e4m3/e5m2) by @xiaobochen-amd in #98
• chore: set kPadN=true as default config by @kyle-256 in #99
• chore: fix groupedgemm fp8 tensorwise scale shape & add test cases by @xiaobochen-amd in #102
• feat(deepep): support internode for deepep. by @zhenhuang12 in #94
• chore: update variable_k implement by @kyle-256 in #104
• feat(deepep): improve intranode dispatch/combine performance. by @zhenhuang12 in #103
• chore: update readme & requirements.txt by @xiaobochen-amd in #105
• feat: use blockwise scaling for V in FP8 FA triton kernel by @hann-wang in #100
• feat: add python interfaces to create & destroy hip streams. by @yuankaichen-amd in #96
• chore: update gemm fp8 code. by @xiaobochen-amd in #78
• feat: refactor fp8 quant config by @RuibinCheung in #106
• feat: support rocm 7 by @xiaobochen-amd in #107
• fix: building on cpu nodes should not infer device arch by @GeneDer in #108
• chore: clean ck code by @xiaobochen-amd in #109
• doc: update readme by @xiaobochen-amd in #111
• doc(deepep): add DeepEP doc. by @zhenhuang12 in #110
• chore : clean unused cpp code & update benchmark code by @xiaobochen-amd in #112
• feat: build for arch==native by @xiaobochen-amd in #115
• doc: update primus product matrix. by @xiaobochen-amd in #116
• fix: gfx950 correctness issue by @kyle-256 in #117
• perf: update grouped gemm config on gfx355 to improve performance by @kyle-256 in #118
• build: remove -fgpu-rdc and --hip-link args by @wenxie-amd in #119

New Contributors

• @yuankaichen-amd made their first contribution in #96
• @GeneDer made their first contribution in #108
• @wenxie-amd made their first contribution in #119

Full Changelog: v0.1.0...v0.1.1

v0.1.0

What's Changed

• init: Project structure and basic operator setup by @xiaobochen-amd in #1
• ci: Add github CI by @xiaobochen-amd in #2
• feat: Add block-wise fp8gemm func by @xiaobochen-amd in #6
• ut: Numeric accuracy test for special functions and Gemm by @llying-001 in #8
• feat(async-tp): support fused_all_gather_matmul for async-tp. by @zhenhuang12 in #5
• benchmark: move numerical correctness to benchmark by @llying-001 in #9
• feat(attention): Add attention op. support bf16(ck/triton) and blockwise_fp8(Triton) by @kyle-256 in #7
• chore(format): integrate clang-format into pre-commit by @xiaobochen-amd in #10
• feat(gemm): Integrate ck gemm fp8 blockwise by @xiaobochen-amd in #13
• fix(gemm): ck fp8 gemm add meta func by @xiaobochen-amd in #14
• benchmark : add attention benchmark by @kyle-256 in #12
• fix(gemm): transA&transB use template in ck gemm fp8 blockwise by @xiaobochen-amd in #15
• feat: Add fp8 quant & dequant kernel and fp8 AlltoAll autograd function by @RuibinCheung in #17
• feat(ops): add blockwise fp8 grouped gemm by @xiaobochen-amd in #16
• feat(dtype): standard dtype naming and structure on cpp-side and python-side. by @xiaobochen-amd in #18
• feat(cpp): add cpp common code and float8 by @xiaobochen-amd in #19
• feat: Integrate hipblaslt with fp16/bf16 precision by @RuibinCheung in #20
• fix(attention): fix softmax_scale=None issue; rename TurboAttention by @kyle-256 in #22
• fix : fix attn bug when enable torch.compile by @xiaobochen-amd in #23
• ci : Use K8S by @haishuok0525 in #29
• feat(modules): Add MXLinear module. by @xiaobochen-amd in #25
• feat(moe): Add fused_moe_router function by @ChengYao-amd in #26
• refactor(gemm): refactor gemm ops/func and linear. by @xiaobochen-amd in #30
• feat(benchmark): add gemm_fp8 & grouped gemm benchmark; add flash_attn into attn benchmark by @kyle-256 in #32
• feat(async-tp): support fused_matmul_reduce_scatter(bf16) for async-tp. by @llying-001 in #28
• feat(async_tp): fused_all_gather_matmul support fp8. by @zhenhuang12 in #27
• feat(gemm): integrate hipblaslt fp8 by @RuibinCheung in #34
• refactor(ops): refactor gemm fp8 blockwise api by @xiaobochen-amd in #36
• feat(benchmark): add a torch pretrain e2e demo by @xiaobochen-amd in #38
• feat(attention): support attention context parallel with all2all communication type triton backend by @ChengYao-amd in #37
• feat(async-tp): remove triton-dist dependency. by @zhenhuang12 in #41
• feat(attention): update aiter, use hsa forward kernel when head_dim=128 by @kyle-256 in #43
• feat(rmsnorm): add rmsnorm hip kernel. by @xiaobochen-amd in #42
• feat(jax): add jax frontend & add rmsnorm lax by @xiaobochen-amd in #45
• feat(attention): add attn-cp all2all ck backend by @ChengYao-amd in #44
• chore: clean some benchmark code and modify grouped gemm code by @xiaobochen-amd in #40
• feat(deep_ep): primus_turbo integrate with deepep . by @zhenhuang12 in #48
• opt rmsnorm kernel perf by @xiaobochen-amd in #47
• feat(build): update build & remove aiter submodule by @xiaobochen-amd in #50
• feat(moe): add fused_moe_router_bkwd_triton by @ChengYao-amd in #51
• feat(attention): combine qkv all2all for cp-attn-a2a, refactor cp-att… by @ChengYao-amd in #53
• fix(deep_ep): fix intranode-dispatch bug and memory access fault on poolside 515B. by @zhenhuang12 in #54
• feat(grouped_gemm): add persistent grouped_gemm kernel. by @kyle-256 in #49
• feat(moe): fused router with scatter routing map & probs by @ChengYao-amd in #56
• feat: refine fp8 alltoall by @RuibinCheung in #57
• feat: update grouped gemm code & benchmark by @xiaobochen-amd in #58
• chore: add license header to source files by @xiaobochen-amd in #61
• feat(async-tp): support gemm_rs_overlap by pipeline method which splits inputs and use multi_stream copy by @llying-001 in #60
• feat(moe): fused router support arbitrary experts and selected groups by @ChengYao-amd in #62
• docs(readme): update readme.md by @xiaobochen-amd in #39
• fix(async_tp): make memory order safe. by @zhenhuang12 in #63
• fix(perf): fixed grouped_gemm_variable_k_postprocess perf issues by @xiaobochen-amd in #64
• feat(ck): update ck version and update ck grouped gemm code. by @xiaobochen-amd in #67
• refactor(attention): move files in utils dir to attention by @ChengYao-amd in #68
• fix: avoid aiter print warning in gfx950 by @xiaobochen-amd in #69
• feat: whl package name with git commit by @xiaobochen-amd in #70
• feat: add gfx950 arch to the FP8 FA triton kernel by @hann-wang in #52
• chore(attn): rename q_scale, k_scale, do_scale by @hann-wang in #71
• chore: convert transA/transB to trans_a/trans_b by @xiaobochen-amd in #74
• test(attention): numerical accuracy by @llying-001 in #72
• feat: gemm fp8 tensorwise support multi layouts and fp8-formats by @xiaobochen-amd in #76
• feat(deepep): dispatch return num_recv_tokens_per_expert of type tensor by @zhenhuang12 in #75
• feat(docs): add examples docs by @xiaobochen-amd in #77
• feat: grouped gemm support num_cu by @xiaobochen-amd in #79
• feat(attn): remove 192 128 head_size padding & refactor interface align with FA by @ChengYao-amd in #80
• chore: update setup and gitci by @xiaobochen-amd in #81
• feat(attn): refactor attn-cp ut by @ChengYao-amd in #82
• refactor(asynctp): refactor async-tp test cases. by @zhenhuang12 in #83
• fix: attention bug by @xiaobochen-amd in #84
• feat(grouped_gemm): grouped_gemm support fp8-tensorwise&fp8-rowwise. by @kyle-256 in #66
• chore: update README by @xiaobochen-amd in #85
• feat: add reduce kernel & opt quant tensorwise perf by @xiaobochen-amd in #86
• chore: update example and add codeowners cfg by @xiaobochen-amd in #87
• chore: add arch=native in setup & fix codeowners error path bug. by @xiaobochen-amd in #88
• feat(grouped_gemm): grouped_gemm_fp8 post-process optimization by @kyle-256 in #89
• 3rdparty: update aiter. by @xiaobochen-amd in #90
• docs: add source references and license headers for FlashAttention and TorchAO kernels by @xiaobochen-amd in #92
• chore: update ck & fix bf16 grouped gemm precision bug when n%256!=0 by @kyle-256 in #95
• chore: bump version to v0.1.0 by @xiaobochen-amd in #97

New Contributors

• @xiaobochen-amd made their first contribution in #1
• @llying-001 made their first contribution in #8
• @zhenhuang12 made their first contribution in https://github.com/AMD-AGI/...