- fixed bug on Windows with stopping to report status
- automatic fallback to AVX2 on linux
- DEPRECATED: custom QLI Protocol
- fixed bug with Qiner communication protocol
- fixed bug which decreases training performance over time
- Optimized Communication between Client and API
- Fixed blocking Bug
- Deprecated:
"AVX2":true
setting - Optimized Detection of AVX512 and added Fallback to AVX2
- Removed Deprecated Settings
- Added Setting
AllowOptimization
to allow the trainer to try to optimize environment. (Default: false) - Reduced Runner Footprint
- Bugfixing
- Automatic optimization of the environment
- Bugfixing
- Possibility to Overwrite Params like AVX512
- new argument
--version
to get current version of Client
- Bugfixing
- New Setting:
"serviceLock": true
to use qiner protocol for custom runner - New Log Format with display of average it/s
- Signaling if the trainer isn't starting properly
- Fixed AVX512 detection on Windows
- Fixed Thread Settings from Portal
- Automatic runner selection from Portal Settings
- Automatic thread selection from Portal Settings
- New Setting:
"useAvx": true
which forces the client to use AVX2 runner - New Setting:
"allowHwInfoCollect": true
which will allow the client to collect CPU Model, RAM Size and CPU Cache Size; with this information the client can better decide which runner should be used - Optimized Detection of AVX2/512
- Optimized Error Handling
- Added automated selection of AVX2/512 runner
- New Setting:
"customRunner": true
to enforce a custom runner