v1.3.6

What's Changed

• chore: update cache-dit arch by @DefTruth in #932
• bc: deprecated serving module by @DefTruth in #933
• chore: suppress torch compile tuning logs by @DefTruth in #934
• compile: enabled descent_tuning by default by @DefTruth in #935
• docs: update quantization docs by @DefTruth in #937
• quant: add quantize backend enum by @DefTruth in #938
• kernel: refactor ops register by @DefTruth in #939
• chore: fix vllm-omni docs links by @DefTruth in #940
• examples: add cuda graph option by @DefTruth in #942
• chore: fix utils log info by @DefTruth in #943
• chore: add cuda graph usage docs by @DefTruth in #944
• chore: add cuda graph usage to overview by @DefTruth in #945
• CLI: add compile full-graph option by @DefTruth in #946
• chore: fix fullgraph param typo by @DefTruth in #947
• docs: add more cuda graph perf results by @DefTruth in #948
• docs: add more cuda graph perf results by @DefTruth in #949
• docs: update cuda graphs docs by @DefTruth in #950
• chore: allow cuda graph for dynamic compile by @DefTruth in #951
• feat: support cuda graph + fp8 rowwise by @DefTruth in #952
• chore: hotfix for mkdocs broken by @DefTruth in #953
• [1/N] feat: support svdquant w4a4 - kernels & skills by @DefTruth in #954
• pytest: fast_svd mode for testing by @DefTruth in #955
• [2/N] feat: streaming quantize for svdquant by @DefTruth in #956
• [3/N] feat: PTQ workflow for svdquant by @DefTruth in #957
• SKILL: add ptq-workflow-integration skill by @DefTruth in #958
• pytest: separate kernels and quantization tests by @DefTruth in #959
• chore: add docs strings to codebase by @DefTruth in #960
• chore: add svdq e2e example and format code by @DefTruth in #961
• SKILL: add Cute-DSL/CUDA/CUTLASS skills by @DefTruth in #962
• chore: update docs by @DefTruth in #963
• kernel: tune svdq w4a4 gemm stage/blk size for Ada by @DefTruth in #966
• kernel: unified ops register policy by @DefTruth in #967
• bench: refactor cache-dit bench by @DefTruth in #968
• svdquant: fast svd decompose, ~18x speedup by @DefTruth in #969
• [2/N] tune svdq w4a4 gemm for ada by @DefTruth in #970
• bc: refactor distributed codebase by @DefTruth in #971
• kernel: add cute-dsl based merge-attn-states kernel by @DefTruth in #973
• feat: extend SVDQ PTQ -> SVDQ DQ by @DefTruth in #974
• fix: support 3D input/output for W4A4 linear by @DefTruth in #975
• chore: support svdq-calib option in examples by @DefTruth in #976
• kernel: add cute-dsl based fp8 comm kernels by @DefTruth in #977
• [1/N] feat: support cute-dsl based svdquant w4a4 by @DefTruth in #978
• feat: support svdq-dq few shot by @DefTruth in #979
• chore: update svdq-dq few shot docs by @DefTruth in #980
• feat: support layerwise cpu offload by @DefTruth in #981
• [2/N] feat: support layerwise offload by @DefTruth in #982
• [3/N] feat: support layerwise offload by @DefTruth in #983
• [4/N] feat: support layerwise offload by @DefTruth in #984
• chore: unified all2all/ring comm api by @DefTruth in #985
• chore: refactor async ulysses codebase by @DefTruth in #986
• remove cutedsl based svdq kernels by @DefTruth in #987
• fix tensor parallel register import error by @DefTruth in #988
• feat: support sub cp_plan for context parallel by @DefTruth in #989
• chore: fix attention dispatch comments by @DefTruth in #990
• community: add tensorrt-llm x cache-dit link by @DefTruth in #991
• deps: use uv to install deps by @DefTruth in #992
• chore: update docs by @DefTruth in #993
• chore: add layerwise offload to overview by @DefTruth in #994
• chore: update layerwise offload cli quick start by @DefTruth in #995
• attention: fix sage-attn backend dispatch by @DefTruth in #996
• chore: add exclude-layers param to ptq example by @DefTruth in #997
• attention: separate attn backends by @DefTruth in #998
• svdq: support converter cli for dq workflow by @DefTruth in #999
• chore: fix typo by @DefTruth in #1000
• chore: revise quantization example in README by @DefTruth in #1001
• chore: update README by @DefTruth in #1002
• feat: support ray wrapper by @DefTruth in #1003

Full Changelog: v1.3.5...v1.3.6