v0.3.0

Summary of Changes

• Switch to multistage dockerfile which greatly reduced the size of the image
• Refactor image scripts to remove launch_training and call sft_trainer directly.
  • Note that this affects the error codes returned from sft_trainer to user error code 1 and internal error code 203.
  • In addition, this affects the logging as parameter parsing logging is moved into sft_trainer which is harder to view.

What's Changed

• Switch to multistage dockerfile by @tharapalanivel in #154
• refactor: remove launch_training and call sft_trainer directly by @anhuong in #164
• docs: consolidate configs, add kfto config by @anhuong in #170
• fix: bloom model can't run with flash-attn by @anhuong in #173
• Update README.md for Lora modules by @Ssukriti in #174

Full Changelog: v0.2.0...v0.3.0