Eitan Porat

## Summary When using `nnx.List` (which stores items with integer dict keys) with Orbax checkpointing, integer keys are converted to strings during save but not converted back during restore. This...

# Background Nonogram is a logic puzzle. You are given a grid you need to color in black and white and "hints" - which indicate the run lengths. For example,...

I am getting 5.7M SPS in c implementation vs 2.7k in c implementation. Didn't tinker too much with the training parameters. It seems pretty good after 100 million steps.

# Description - Add audio encoder support for qwen3 * classes such as `AudioEncoder` and `Qwen3OmniAudioEncoder` * added the `use_audio` flag to the configuration. * various audio specific flags -...

# Description Add deepstack support from this paper https://arxiv.org/abs/2406.04334 idea is to inject intermediate representations of the vision encoder into the intermediate layers of the llm. # Tests I looked...

# Description Checkpoint conversion mapping from hf # Tests Checked that the weights are correct by comparing forward passes of the model. # Checklist Before submitting this PR, please make...

# Description This PR implements the pre-processing required for qwen3 omni it should be merged with @hengtaoguo's PR https://github.com/AI-Hypercomputer/maxtext/pull/2613 # Tests I ran inference successfully! # Checklist Before submitting this...

# Description Implemented multi-modal rope and integrated in training and inference # Tests Compared to the pytorch implementation `tests/multimodal_rope_check.py` # Checklist Before submitting this PR, please make sure (put X...