Summary

• Cast multimodal processor outputs to model.dtype before Gemma 4 generation.
• Prevents the T4 audio path from mixing Float audio features with Half model tensors.

Root cause

The notebook moved processor outputs to CUDA but did not normalize floating-point dtypes. On a T4, Gemma 4 runs in FP16 while audio features can remain FP32, causing masked_scatter_ to fail with a Float/Half mismatch during multimodal fusion.

Change

Updated the inference helper to use .to("cuda", dtype=model.dtype). Integer token IDs remain integer tensors.

Validation

• Notebook parses as valid JSON.
• git diff --check passes.
• Saved T4 notebook outputs show successful text, image, audio, and audio+image inference.
• Saved notebook contains zero error outputs.
• A fresh authenticated T4 rerun is still recommended for maintainer confirmation.

Note

This patch was prepared with assistance from Codex.