Remove pytest from runtime model import

[kstonekuan]

Summary

This removes an unintended runtime dependency on pytest from openpi.models_pytorch.gemma_pytorch.

The only usage was in a type annotation for past_key_values:

• before: list[torch.FloatTensor] | pytest.Cache | None
• after: list[torch.FloatTensor] | None

Why

pytest should not be required to import or serve runtime model code.

This showed up when using openpi in a production inference image that excludes dev dependencies.

In our serving environment, openpi.models_pytorch.gemma_pytorch was imported during model load, which caused the server to fail if pytest was not installed. The pytest reference was only used in a type annotation and was not part of runtime behavior.

Changes

• removed import pytest from src/openpi/models_pytorch/gemma_pytorch.py
• replaced the pytest.Cache annotation on past_key_values with a runtime-safe annotation

Validation

• verified that the only runtime use of pytest in this module was the type annotation
• verified the change is isolated to this import and annotation

Notes

I kept this patch intentionally narrow to avoid changing model behavior or cache semantics beyond removing the accidental test-only dependency.

[jimmyt857]

Thanks!