MSCCL++ v0.9.0

What's Changed

• Fix lint.sh by @chhwang in #652
• Update the port channel tutorial doc by @chhwang in #653
• Fix test script by @Binyang2014 in #655
• Update EndpointConfig interfaces by @chhwang in #651
• New allreduce algo for small message size by @Binyang2014 in #647
• Fix docs by @chhwang in #656
• Fix docs version by @chhwang in #659
• Exclude irrelevant files from workflow triggers by @chhwang in #663
• Improving DSL documentation by @caiomcbr in #650
• Add exclude paths under pipeline triggers by @chhwang in #664
• Test peer accessibility after deployment by @chhwang in #661
• Rename nvls* files by @chhwang in #660
• Fix #651 by @chhwang in #662
• Add token pool for cuCreate API by @Binyang2014 in #628
• Auto-detect CUDA arch in CMake GPU check by @chhwang in #666
• FP8 support for Allreduce by @seagater in #646
• Resolve IBVerbs Loading Issues by @caiomcbr in #648
• Fixes for no-IB systems by @chhwang in #667
• Integrate MSCCL++ DSL to torch workload by @Binyang2014 in #620
• Add a new logger by @chhwang in #668
• upgrade codeql to v3 by @Binyang2014 in #676
• IB stack enhancements & bug fixes by @chhwang in #673
• Support Synchronous Initialization for Proxy Service by @caiomcbr in #679
• Supporting New Packet Kernel Operation at Executor by @caiomcbr in #677
• connect() APIs changed to return an instance instead of a shared_ptr by @chhwang in #680
• Fix Minor Issue Proxy Python Interface by @caiomcbr in #685
• Revise the mscclpp datatype by @seagater in #671
• Fix Error in Non IB Env at Executor by @caiomcbr in #686
• No IB Env CI Test by @caiomcbr in #687
• Fix Python bindings and tests by @chhwang in #690
• DSL Quick Start by @caiomcbr in #689
• Add CudaDeviceGuard by @chhwang in #691
• Optimized logger by @chhwang in #693
• Build fixes by @chhwang in #696
• Add an IB multi-node tutorial by @chhwang in #702
• Creating Documentation Section for MSCCL++ DSL by @caiomcbr in #706
• Make IB more configurable by @chhwang in #703
• Improve DSL Documentation by @caiomcbr in #707
• Add handle cache for AMD platform by @Binyang2014 in #698
• Add copilot-instructions.md by @chhwang in #602
• Use uncached memory on Rocm platform to avoid hang by @qishilu in #711
• Replace __HIP_PLATFORM_AMD__ to use internal macro by @Binyang2014 in #712
• Rename P2P log subsys into GPU by @chhwang in #716
• Minor fixes by @chhwang in #715
• Remove UB std:: declarations by @chhwang in #709
• Tune the nThreadsPerBlock for FP8 and Half datatype on MI300 by @seagater in #694
• Update container images for pipeline by @Binyang2014 in #717
• Add CUDA 13.0 Docker images by @chhwang in #720
• Bypassing SSCA alerts by @chhwang in #721
• Add GpuIpcMemHandle by @chhwang in #704
• Reduce CI build time by @chhwang in #723
• Use GpuIpcMem for NVLS connections by @chhwang in #719
• Fix ci issue by @Binyang2014 in #727
• Fix ci pipeline failure by @Binyang2014 in #729
• Torch integration by @Binyang2014 in #692
• fp8 nvls support (e5m2 and e4m3) by @mahdiehghazim in #730
• Support versioning for mscclpp document by @seagater in #724
• Revert "Support versioning for mscclpp document (#724)" by @seagater in #734
• Use native GPU architecture when NVIDIA GPU is detected; otherwise fall back to multi-arch build. by @mahdiehghazim in #732
• Update document versioning for PR #724 by @seagater in #735
• Fix the relative path extraction on github page by @seagater in #739
• Support multi-node in MemoryChannel tutorial by @chhwang in #726
• Address comments for PR #692 by @Binyang2014 in #733
• Refactor reduce kernel by @Binyang2014 in #738
• Fix cpplint error in main branch by @seagater in #740
• Update copilot-instructions.md by @chhwang in #722
• create CI pipeline for rocm by @Binyang2014 in #718
• Add a new IB stack impl that doesn't use RDMA atomics by @chhwang in #728
• Support Fusion for ReadPutPacket Operation at DSL by @caiomcbr in #742
• Refactor algo selection logic and introduce symmetric_memory env by @Binyang2014 in #741
• Support uint8 data type for Allreduce by @seagater in #736
• Add new CI pipeline for RCCL test by @Binyang2014 in #746
• Update dtype name by @Binyang2014 in #748
• address flagBuffer ownership issue by @Binyang2014 in #749
• Removing MPI Dependency by @caiomcbr in #743
• Address installation issue in some env by @Binyang2014 in #750
• Mahdieh/switchchannel test clean by @mahdiehghazim in #751
• Disabling Nanobind Memory Leak Warnings in Release Builds by @caiomcbr in #745
• Adjusting Communicator in Python API by @caiomcbr in #752
• Add CI pipeline for no-IB environment testing by @Binyang2014 in #755
• Add new algos for GB200 by @Binyang2014 in #747
• Add doc for perf tunning by @Binyang2014 in #756
• Adding Support to Setting Message Size Range in Native Algorithm API by @caiomcbr in #758
• Do threadInit/cudaSetDevice before other cuda calls by @wuxb45 in #757
• Fix NCCL fallback comm destroy and use latest NCCL release in CI by @Binyang2014 in #760
• Fix multicast handle leak, cuMemMap offset handling, and rename NVLS allreduce algorithms by @Binyang2014 in #759
• Fix use-after-free for fabric allocation handle in GpuIpcMemHandle by @Binyang2014 in #764
• Remove GTest dependency, add code coverage, and refactor unit tests and CI pipelines by @Copilot in #744
• Install default plans under MSCCLPP_CACHE_DIR/default by @ekwhoa in #769
• Use PTX red for D2D semaphore signal by @Binyang2014 in #768
• Add unit testing framework readme by @chhwang in #766
• Fix run-remote.sh to support multi-command scripts by @Binyang2014 in #770
• Fix CI/CD pipeline issues by @Binyang2014 in #773
• Support E4M3B15 datatype by @Binyang2014 in #765

New Contributors

• @qishilu made their first contribution in #711
• @mahdiehghazim made their first contribution in #730
• @wuxb45 made their first contribution in #757
• @ekwhoa made their first contribution in #769

Full Changelog: v0.8.0...v0.9.0