Fix Gemma4 audio feature dtype mismatch in masked_scatter

[danielhanchen]

What does this PR do?

Aligns the dtype of Gemma 4 audio features to the text-embedding dtype before masked_scatter, so the audio merge behaves like the image and video merges.

In Gemma4Model.forward, the image and video branches cast their features to inputs_embeds.dtype before scattering them into the text embeddings:

image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
...
video_features = video_features.to(inputs_embeds.device, inputs_embeds.dtype)

but the audio branch only moved the features to the correct device, never the dtype:

inputs_embeds = inputs_embeds.masked_scatter(
    audio_mask.to(inputs_embeds.device), audio_features.to(inputs_embeds.device)  # dtype not aligned
)

When inputs_embeds and the audio encoder output end up in different dtypes (for example a float32 residual stream with a float16 audio tower, which happens under mixed-precision setups), this raises:

RuntimeError: masked_scatter_: expected self and source to have same dtypes but got Float and Half

The fix casts audio_features to inputs_embeds.dtype in the same call, matching the image and video paths. Text-only, image, and video inference are unaffected since they already cast. The change is made in modular_gemma4.py and mirrored in the generated modeling_gemma4.py.

Before submitting

• This PR fixes a bug (audio inference crash under mixed precision).
• Did you make sure to update the documentation with your changes? Not needed.
• Did you write any new necessary tests? Not needed; this aligns audio with the existing image/video dtype handling.

Who can review?

@ArthurZucker @Cyrilvallez

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gemma4

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 29929449179
Result: success | Grafana metrics are not available yet.

[Rocketknight1]

Yes, LGTM!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.