v0.4.0

What's Changed

• refactor: use a shared TCPStore in ParameterServer and create ProcessGroup using PrefixStore by @HubertZhang in #82
• feat: add StatelessProcessGroup to extend collective library by @kip-cxj in #66
• feat: release ipc buffers before calling update_weights_from_ipc's post_hook by @HubertZhang in #84

Full Changelog: v0.3.4...v0.4.0

v0.4.0-rc0

What's Changed

• refactor: use a shared TCPStore in ParameterServer and create ProcessGroup using PrefixStore by @HubertZhang in #82
• feat: add StatelessProcessGroup to extend collective library by @kip-cxj in #66
• feat: release ipc buffers before calling update_weights_from_ipc's post_hook by @HubertZhang in #84

Full Changelog: v0.3.4...v0.4.0-rc0

v0.3.4

What's Changed

• support mtp by @youzhedian in #81

New Contributors

• @youzhedian made their first contribution in #81

Full Changelog: v0.3.3...v0.3.4

v0.3.3

What's Changed

• fix: npu free host cache by @kip-cxj in #78
• bugfix: skip empty safetensors file in inplace_pin_memory by @HubertZhang in #79

Full Changelog: v0.3.2...v0.3.3

v0.3.2

What's Changed

• fix p2p update error when disable_h2d_buffer is true by @ruizhang1230 in #76
• fix: set current CUDA device in _inplace_pin_memory function by @SongXiaoXi in #77

Full Changelog: v0.3.1...v0.3.2

v0.3.1

Same as v0.3.1-rc0

v0.3.1-rc0

What's Changed

• Update use of environment variable in ps.py by @HubertZhang in #73
• misc: split ps.py file into multiple files by @specture724 in #64
• feat: cache device uuid in VllmWorkerExtension by @kip-cxj in #74

Full Changelog: v0.3.0-rc1...v0.3.1-rc0

v0.3.0

What's Changed

• feat: docs added for auto_pg, and auto_pg default set to True by @specture724 in #65
• hotfix: add a switch to disable inplace pinning of tensors by @specture724 in #68
• hotfix: inplace pin memory caused cudaErrorHostMemoryAlreadyRegistered by @specture724 in #69
• fix: CUDA OOM encountered with store based barrier by @specture724 in #70

Full Changelog: v0.3.0-rc0...v0.3.0-rc1

v0.2.3

• Disable "inplace pin memory" feature by default from 0.2.2, as it may cause issues

v0.3.0-rc0

What's Changed

• fix: use tcp store_based_barrier to control p2p update synchronization by @specture724 in #51

Full Changelog: v0.2.2...v0.3.0-rc0