Releases: mlcommons/logging
Releases · mlcommons/logging
2.0.0-rc1 release for training 1.1.0 submission
The logging infra for the 2.0 MLPerf Training submission is ready to use.
New features for the 2.0 round
- Support for the new SSD benchmark: Compliance rules and RCPs.
- Weight initialization compliance rule.
- Change open submission compile time to 4h. Closed submission set to 30 mins for good.
- RCP pruning to relax RCP checking rules.
- Support to include a scaling.json file in result directory to scale final scores.
- Repository checker for large and inadmissible files for git.
More releases will be available before the May 20th deadline as:
- RCPs may be added.
- Bugs may be fixed.
Instructions to run the logging infra on your submission package:
The whole thing (Package checker + Result summarizer + Repository checker)
./scripts/verify_for_v2.0_training.sh <submission dir>
Package checker only:
python3 -m mlperf_logging.package_checker <submission dir> training 2.0.0
Result summarizer only:
python3 -m mlperf_logging.result_summarizer <submission dir> training 2.0.0
Repository checker only:
python3 -m mlperf_logging.repo_checker <submission dir> training 2.0.0
1.1.0-rc4 release for training 1.1.0 submission
Contains recent bug fixes for 1.1.0 training submission.
1.1.0-rc3
1.1.0-rc2
hpc-1.0.0-rc2
Update version to hpc-1.0.0-rc2 (#182)
hpc-1.0.0-rc1
Update version to hpc-1.0.0-rc1 (#171)
1.1.0-rc1
1.0.0
1.0 Changes Summary:
Training - compliance checker
- Added Unet3D rules
- Added RNN-T rules
- Added checks for gradient_accumulation_steps
- BERT: updated target from 0.712 to 0.720
- Removed checks for obsolete models: GNMT, Transformer
Training - package checker
- Added seed checker
- Added calls to the RCP checker
- Added calls to the system description checker
- Added support for the Unet3d olympic scoring (reject top and bottom 4)
Training - reference convergence point checker
- Added initial version of RCP checker
Training - result summarizer
- Refactored olympic scoring calculation to be able to accommodate unet3d (reject top and bottom 4)
- Added calls to the RCP checker (disabled RCP checks for minigo)
Inference
- Added power parser script
System description
- Added initial list of system description entries
- Added commands fetching Nvidia system info
1.0.0-rc4
1.0 Changes Summary:
Training - compliance checker
- Added Unet3D rules
- Added RNN-T rules
- Added checks for gradient_accumulation_steps
- BERT: updated target from 0.712 to 0.720
- Removed checks for obsolete models: GNMT, Transformer
Training - package checker
- Added seed checker
- Added calls to the RCP checker
- Added calls to the system description checker
- Added support for the Unet3d olympic scoring (reject top and bottom 4)
Training - reference convergence point checker
- Added initial version of RCP checker
Training - result summarizer
- Refactored olympic scoring calculation to be able to accommodate unet3d (reject top and bottom 4)
- Added calls to the RCP checker (disabled RCP checks for minigo)
Inference
- Added power parser script
System description
- Added initial list of system description entries
- Added commands fetching Nvidia system info
1.0.0-rc3
1.0 Changes Summary:
Training - compliance checker
- Added Unet3D rules
- Added RNN-T rules
- Added checks for gradient_accumulation_steps
- BERT: updated target from 0.712 to 0.720
- Removed checks for obsolete models: GNMT, Transformer
Training - package checker
- Added seed checker
- Added calls to the RCP checker
- Added calls to the system description checker
- Added support for the Unet3d olympic scoring (reject top and bottom 4)
Training - reference convergence point checker
- Added initial version of RCP checker
Training - result summarizer
- Refactored olympic scoring calculation to be able to accommodate unet3d (reject top and bottom 4)
- Added calls to the RCP checker (disabled RCP checks for minigo)
Inference
- Added power parser script
System description
- Added initial list of system description entries
- Added commands fetching Nvidia system info