Releases: VE-FORBRYDERNE/mtj-softtuner
Releases · VE-FORBRYDERNE/mtj-softtuner
v1.3.6
v1.3.5
v1.3.4
What's Changed
- Add embedding layernorm for BLOOM (007950d)
- Clip optax version to <= 0.1.2 to fix failing tests by @vfbd in #9
- Fix missing requirements by @vfbd in #10
- Fix loading of base OPT-125M model and finetuned OPT models by @vfbd in #11
- Fix aria2 model downloads by @vfbd in #12
Full Changelog: v1.3.3...v1.3.4
v1.3.3
What's Changed
- Add
|| true
in install.sh - Fix failing tests by @vfbd in #8
Full Changelog: v1.3.2...v1.3.3
v1.3.2
v1.3.1
What's Changed
- Change "SOFT-TUNING PROGRESS" to "CURRENT TRAINING STEP"
- Don't run tests in mtj_softtuner/kobold by @vfbd in #6
Full Changelog: v1.3.0...v1.3.1
v1.3.0
What's Changed
- Change "SOFT-TUNING PROGRESS" to "CURRENT TRAINING STEP"
- Add support for fractional number of epochs by @vfbd in #5
Full Changelog: v1.2.0...v1.3.0
v1.2.0
Model downloads are now significantly faster because they're now downloaded using aria2 instead of the Python requests module. OPT is now supported (up to 13B) and new prompt initialization methods were added.
What's Changed
- OPT support
- Kaiming initialization by @vfbd in #1
- Vocabulary sample initialization by @vfbd in #2
- Add
trainer.data.prompt_seed
and tests for prompt initialization by @vfbd in #3 - Use aria2 to download models if
aria2c
is in PATH by @vfbd in #4
Full Changelog: v1.1.4...v1.2.0
v1.1.4
Full Changelog: v1.1.3...v1.1.4
v1.1.3
Full Changelog: v1.1.2...v1.1.3