v1.0.3

v1.0.3
Support torch 2.4 for flux.

v1.0.2

What's Changed

• add issue template by @liwenchangbdbz in #1
• add cutlass submodule and patches by @liwenchangbdbz in #2
• All gather and reduce scatter on SM80 by @zheng-ningxin in #3
• Reorganize and deduplicate files by @wenlei-bao in #4
• Add arXiv paper link by @wenlei-bao in #5
• Update BibTex by @wenlei-bao in #6
• Support IPC && SM90 version of AG-GEMM, GEMM-RS by @zheng-ningxin in #9
• fix the _allgather_base backend issue(issue11) by @zheng-ningxin in #12
• using c10::intrusive_ptrc10d::ProcessGroup as argument from python by @houqi in #13
• Add more device types for the time estimation. by @zheng-ningxin in #15
• Update README.md by @zheng-ningxin in #16
• zero out all the allocated shm buffer by @zheng-ningxin in #18
• feat: fix tuning for the all-gather gemm && move the reset-signal() to the forward critical path by @zheng-ningxin in #19
• Tune the AG performance for the llama-8b by @zheng-ningxin in #21
• Remove pynvshmem import in gemm_rs_80.py by @tlrmchlsmth in #22
• Support performance tunning for gemm-rs kernel on sm80 by @zheng-ningxin in #23
• add torch version to the whl name by @zheng-ningxin in #24

New Contributors

• @liwenchangbdbz made their first contribution in #1
• @zheng-ningxin made their first contribution in #3
• @wenlei-bao made their first contribution in #4
• @houqi made their first contribution in #13
• @tlrmchlsmth made their first contribution in #22

Full Changelog: https://github.com/bytedance/flux/commits/v1.0.2

v1.0.0

What's Changed

• add issue template by @liwenchangbdbz in #1
• add cutlass submodule and patches by @liwenchangbdbz in #2
• All gather and reduce scatter on SM80 by @zheng-ningxin in #3
• Reorganize and deduplicate files by @wenlei-bao in #4
• Add arXiv paper link by @wenlei-bao in #5
• Update BibTex by @wenlei-bao in #6
• Support IPC && SM90 version of AG-GEMM, GEMM-RS by @zheng-ningxin in #9
• fix the _allgather_base backend issue(issue11) by @zheng-ningxin in #12
• using c10::intrusive_ptrc10d::ProcessGroup as argument from python by @houqi in #13
• Add more device types for the time estimation. by @zheng-ningxin in #15
• Update README.md by @zheng-ningxin in #16
• zero out all the allocated shm buffer by @zheng-ningxin in #18
• feat: fix tuning for the all-gather gemm && move the reset-signal() to the forward critical path by @zheng-ningxin in #19
• Tune the AG performance for the llama-8b by @zheng-ningxin in #21
• Remove pynvshmem import in gemm_rs_80.py by @tlrmchlsmth in #22
• Support performance tunning for gemm-rs kernel on sm80 by @zheng-ningxin in #23
• add torch version to the whl name by @zheng-ningxin in #24

Full Changelog: https://github.com/bytedance/flux/commits/v1.0.0