v0.6.3

Features

• support different ssh port on different nodes (#93) 6810245
• support multiple models in one training script (#113) 312bcc0 (#107) 0aec789

Fixes

• autotune service defaults with a fixed random seed (#117) a58c2de

Others

• sort q_adam variables for better performance (#102) f277549
• improve autotune speed metrics measurement for better accuracy (#86) e4ee5ee
• install.sh upgrades existing bagua package bc69890
• install.sh will not install Rust if already exist on the system 67e1efe

v0.6.2

Fixes

• fix QAdam gradient is not BaguaTensor during first stage 1d4dc82

v0.6.1

Features

• add QAdam algorithm (#92) 0dafd24
• broadcast model parameters on every algorithm reset e5b36dc
• wrap python op in communication stream context by default 51eb656
• add append op methods to python BaguaBucket class (#87) 84d8cbc

Fixes

• BaguaBucket.tensors should only contain original passed in tensors c4ff05f
• fix append python op callable reference 04019cc
• fix BaguaBacket.clear_ops() return value 8cb9f54

v0.6.0

⚠ BREAKING CHANGE

• Now end users should use model.with_bagua(...) API to use Bagua for communication. Algorithm developers can use bagua.torch_api.algorithms.Algorithm to easily develop new algorithms. Installation requires bagua-core >=0.3 now.

Features

• add algorithm import in bagua.torch_api ee73edc
• support reduction op and reduce ac8632c
• auto installation support centos (#50) 073a59e

Fixes

• fix algoirthm pre forward hook not returned e6c7c8d
• the environment variable LOCAL_SIZE has been renamed in LOCAL_WORLD_SIZE (#51) 801b25a

v0.5.0

⚠ BREAKING CHANGE

• contrib: load balancing dataloader and fused optimizer are now in bagua.torch_api.contrib module
• baguaelastic/distributed/launch.py now moved to bagua/distributed/run.py

Features

• add dependency installation script for ubuntu (#41) 4d820ab
• Elastic training (#31) 1a5964c
• add broadcast_buffer in bagua_init (#29) e761cc6
• support bagua-core 0.2 (#26) f1d2bfa

Fixes

• autotune: fix bucket size switch not effective (#48) 30b490a
• remove logging in load balancing dataloader to avoid deadlock (#35) e900383
• torch_api.distributed: cycle dependence (#16) 0314e24
• fix setup.py for low version setuptools (#14) 7d315c0
• fix baguaelastic launch script b069cd4=

v0.4.0

Initial public release of Bagua 🎆