Repository navigation
What's Changed
- Adding Channel Type in ReduceSend Operation on DSL by Caio Rocha (@caiocbr) in #777
- Fix TBG on DSL Get Operation by Caio Rocha (@caiocbr) in #778
- IB
host-no-atomic: GDRCopy + mlx5dv Data Direct for memory-consistent low-latency signaling by Changho Hwang (@chhwang) in #753 - Adjusting Torch Integration Example by Caio Rocha (@caiocbr) in #779
- Add MSCCLPP_IB_GID_INDEX env by Binyang Li (@Binyang2014) in #780
- Updating NpKit by Caio Rocha (@caiocbr) in #785
- Adding CI Test to DSL Executor by Caio Rocha (@caiocbr) in #782
- Fix multi-node H100 CI: CUDA compat, deploy improvements by Binyang Li (@Binyang2014) in #781
- Fix nccl-test CI building for all GPU architectures by Binyang Li (@Binyang2014) in #786
- Support python wheel build by Binyang Li (@Binyang2014) in #787
- fix: isCuMemMapAllocated crashes on non-NVLS systems even with MSCCLPP_FORCE_DISABLE_NVLS=true by Binyang Li (@Binyang2014) with @Copilot in #790
- GDRCopy status message to string by Changho Hwang (@chhwang) in #793
- Fix FP8 ROCm build/test issues and dtype naming by Binyang Li (@Binyang2014) in #792
- Reset GPU tokens before reuse by Binyang Li (@Binyang2014) in #795
- Adding necessary macros for enabling mrc support by Mahdieh Ghazi (@mahdiehghazim) in #797
- Adding Support 4 Nodes AllReduce Small Message Size by Caio Rocha (@caiocbr) in #794
- Adding Support to bf16 Executor Tests by Caio Rocha (@caiocbr) in #801
- Enhance cross-node CudaIpc availability check by Changho Hwang (@chhwang) in #803
- Handle PortChannel flush asynchronously from the host proxy by Changho Hwang (@chhwang) in #802
- Clean up completed communicator receives by Binyang Li (@Binyang2014) in #804
- Include a static synchronization check in the DSL. by Caio Rocha (@caiocbr) in #806
- add nBlocks check for allreduce_allpair_packet algo by Binyang Li (@Binyang2014) in #807
- remove useless sync by Binyang Li (@Binyang2014) in #809
- Add check to filter invalid nblock/nthread candidates by Binyang Li (@Binyang2014) in #811
- Fix for allgather_fullmesh algo by Binyang Li (@Binyang2014) in #813
- Add collective benchmark and correctness check by Binyang Li (@Binyang2014) in #814
- Expose NVLS multicast granularity option for GpuBuffer by Binyang Li (@Binyang2014) in #815
- GB200 support: SendRecv DSL collective and per-channel executor connections by Binyang Li (@Binyang2014) in #810
- support rocm7.2 by Binyang Li (@Binyang2014) in #819
- Update port channel perf tests by Changho Hwang (@chhwang) in #820
- Migrate to C++20 and drop CUDA 11 support by @Empyreus in #822
- Fix unit test by Binyang Li (@Binyang2014) in #823
- Add IPC domain rank detection by Binyang Li (@Binyang2014) in #824
- Add support for allgather nvls algorithms by @Empyreus in #817
- Support FP8 NVLS multimem reduction by Binyang Li (@Binyang2014) in #826
- Fix the IB DMABUF path to be used from CUDA 12.8 by Changho Hwang (@chhwang) in #830
- Fix build with MSCCLPP_USE_IB=OFF when libibverbs headers are absent by @Empyreus in #833
- IB CI fixes: correct CQ-full warning bound, drain in-flight posts on stopProxy, expand multi-QP perf tests by Changho Hwang (@chhwang) in #808
- Increase
DefaultMaxCqPollNumfrom 1 to 256 by Changho Hwang (@chhwang) in #840 - Disable UT for Rocm7.2 and fix rccl test by Binyang Li (@Binyang2014) in #842
- Skip StopVMSS for A100 CI by Changho Hwang (@chhwang) in #845
- Update CI pipeline by Binyang Li (@Binyang2014) in #850
New Contributors
- @Empyreus made their first contribution in #822
Full Changelog: v0.9.0...v0.10.0