Bug: Test suite fails on macOS Apple Silicon & Windows — does not exist

[prince-shakyaa]

Summary

All tests in gemma/gm/ crash at collection time with:

AttributeError: module 'numpy' has no attribute 'float128'. Did you mean: 'float16'?

This causes 162 errors during pytest collection and blocks virtually every unit test from running on macOS (Apple Silicon) and Windows.

Root Cause

np.float128 is a platform-specific alias for the C long double type. It is only available on Linux x86-64. It does not exist on:

• macOS Apple Silicon (ARM64) — where long double == double (64-bit)
• Windows — where long double == double (64-bit)

NumPy 2.x made this even stricter: np.float128 is no longer silently created on platforms where it isn't available — it simply raises AttributeError.

The issue originates in kauldron/ktyping/dtypes.py (line 136), which unconditionally accesses np.float128 at module load time:

# Current (broken on macOS ARM / Windows):
float128 = NpDType(np.float128)

Since every test in gemma/gm/ transitively imports from kauldron.ktyping, this causes all of them to fail immediately during pytest collection — before any test body is executed.

Reproduction

# On macOS Apple Silicon or Windows:
pip install -e '.[dev]'
pytest -vv -n auto --ignore=gemma/gm/tests/ --ignore=gemma/peft/_interceptors_test.py --ignore=gemma/sampler_test.py
# => 162 errors: AttributeError: module 'numpy' has no attribute 'float128'

Suggested Fix (in kauldron/ktyping/dtypes.py)

# Before:
float128 = NpDType(np.float128)

# After:
float128 = NpDType(np.float128) if hasattr(np, 'float128') else NpDType(np.float64)

This guards the import and gracefully falls back to float64 on platforms that do not have 128-bit long doubles.

Impact

• Severity: High — blocks 162 test cases from running on macOS ARM and Windows.
• Platforms affected: macOS Apple Silicon, Windows.
• Platforms not affected: Linux x86-64.

Environment

Python	3.12.13
Platform	macOS Apple Silicon (arm64)
NumPy	2.x
gemma	4.0.1 (editable install from HEAD)
kauldron	HEAD from GitHub