Starstruk Logo MarkStarstruk

v0.1.6

oumi-ai/oumiv0.1.6Feb 22, 2025by wizeng23

AI Summary

This release enhances VLM training by supporting multi-image inputs and fixes the launcher to work on RunPod and Lambda cloud providers.

Key Highlights

  • Support for multi-image VLM training
  • Added warnings for Qwen2.5-VL special requirements
  • Fixed Oumi launcher to run on RunPod and Lambda
  • Updated default training params for Qwen2-VL-2B-Instruct

New Features

  • RemoteParams no longer require an API URL
  • Qwen2-VL evaluation job fixes
  • Multi-image support in SGLang inference engine

Full Release Notes

## What's Changed
* Update RemoteParams to no longer require an API URL. by @taenin in https://github.com/oumi-ai/oumi/pull/1452
* [Tiny] Update default training params for Qwen2-VL-2B-Instruct by @optas in https://github.com/oumi-ai/oumi/pull/1454
* [Tiny] Add more warnings for "special" requirements of Qwen2.5-VL by @optas in https://github.com/oumi-ai/oumi/pull/1453
* Minor cleanup of oumi fetch by @taenin in https://github.com/oumi-ai/oumi/pull/1463
* Support for multi-image VLM training by @nikg4 in https://github.com/oumi-ai/oumi/pull/1448
* Remove a temp workaround in `pad_sequences` on the left side by @nikg4 in https://github.com/oumi-ai/oumi/pull/1464
* [tiny] Add warning that Oumi doesn't support Intel Macs by @wizeng23 in https://github.com/oumi-ai/oumi/pull/1467
* VLM-related logging improvements  by @nikg4 in https://github.com/oumi-ai/oumi/pull/1469
* Fix Oumi launcher to be able to run on RunPod and Lambda by @wizeng23 in https://github.com/oumi-ai/oumi/pull/1470
* Enable pre-release install for uv in pyproject.toml by @wizeng23 in https://github.com/oumi-ai/oumi/pull/1466


**Full Changelog**: https://github.com/oumi-ai/oumi/compare/v0.1.5...v0.1.6