test_quantization.py

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
os.environ["JAX_PLATFORMS"] = "cpu"

import pytest
import jax
import jax.numpy as jnp
import numpy as np
import torch

from quantization import quantization_types


torch.autograd.grad_mode.set_grad_enabled(False)
backends = {"jax": jnp, "numpy": np, "torch": torch}
scales = (0.5, 1.0, 1.5)


@pytest.fixture(params=backends.values(), ids=backends.keys())
def backend(request):
    """Fixture to parameterize over all backends."""
    return request.param


def _uint8_range(backend):
    """Fixture to provide uint8 range for each backend."""
    return backend.arange(256, dtype=backend.uint8)


uint8_range = pytest.fixture(_uint8_range)


def _float_range(backend):
    """Fixture to provide float range for each backend."""
    return backend.linspace(-4, 4, 101, dtype=backend.float32)


float_range = pytest.fixture(_float_range)


@pytest.fixture(params=quantization_types)
def cls(request):
    """Fixture to parameterize over all quantization classes."""
    return request.param


@pytest.fixture
def default_instance(cls):
    """Fixture to parameterize over default instances of all quantization classes."""
    return cls()


@pytest.fixture(params=scales)
def scale(request):
    """Fixture to parameterize over different scale values."""
    return request.param


@pytest.fixture
def instance(cls, scale):
    """Fixture to parameterize over instances of all quantization classes, swept over all scale values."""
    return cls(scale=scale)


def test_interface(cls, default_instance):
    """Test that all quantization classes and instances have the expected methods."""
    assert isinstance(default_instance, cls)
    classattrs = ()
    classmethods = ()
    instanceattrs = ("scale",)
    instancemethods = ("quantize", "dequantize", "nonlinearity", "inv_nonlinearity")
    interface = {
        cls: (classattrs, classmethods), 
        default_instance: (instanceattrs, instancemethods)
    }
    for obj, (attrs, methods) in interface.items():
        for attr in attrs + methods:
            assert hasattr(obj, attr)
        for method in methods:
            assert callable(getattr(obj, method))


def test_quantize_returns_uint8(default_instance, float_range):
    """Test that quantize returns uint8 values."""
    quantized = default_instance.quantize(float_range)
    assert "uint8" in str(quantized.dtype).lower()


def test_dequantize_returns_float32(default_instance, uint8_range):
    """Test that dequantize returns float32 values."""
    dequantized = default_instance.dequantize(uint8_range)
    assert "float32" in str(dequantized.dtype).lower()


def test_roundtrip_consistency(instance, uint8_range):
    """Test that dequantize->quantize is lossless."""
    dequantized = instance.dequantize(uint8_range)
    requantized = instance.quantize(dequantized)
    np.testing.assert_array_equal(uint8_range, requantized)


def test_cross_backend_consistency(instance):
    """Test that dequantization gives identical results across all backends."""
    results = {
        name: instance.dequantize(_uint8_range(backend)) 
        for name, backend in backends.items()
    }
    baseline_name = list(backends.keys())[0]
    baseline_result = results[baseline_name]
    for name, result in results.items():
        if name == baseline_name:
            continue
        np.testing.assert_array_almost_equal(
            result,
            baseline_result, 
            err_msg=f"{name} result doesn't match {baseline_name}",
        )


def test_jax_jit(instance):
    """Test that the JAX backend works when the quantization methods are jitted."""
    data = _uint8_range(jnp)
    
    def roundtrip(x):
        return instance.quantize(instance.dequantize(x))
    
    nojit = roundtrip(data)
    yesjit = jax.jit(roundtrip)(data)
    np.testing.assert_array_equal(yesjit, nojit)


def test_torch_compile(instance):
    """Test that the PyTorch backend works when the quantization methods are compiled."""
    data = _uint8_range(torch)
    
    def roundtrip(x):
        return instance.quantize(instance.dequantize(x))
    
    nojit = roundtrip(data)
    yesjit = torch.compile(roundtrip)(data)
    np.testing.assert_array_equal(yesjit, nojit)