Module neuralops.tests.test_gptq_real

Numerical verification tests for production GPTQ.

Functions

def test_gptq_real_128x128()

Expand source code

def test_gptq_real_128x128():
    torch.manual_seed(42)
    layer = nn.Linear(128, 128, bias=False)
    orig_w = layer.weight.data.clone()
    quantizer = GPTQQuantizer(layer, bits=4, group_size=32)

    # Simulate realistic activations
    x = torch.randn(64, 128)
    quantizer.add_batch(x)
    w_q = quantizer.quantize(damp=0.01)

    assert w_q.shape == (128, 128)
    assert not torch.equal(layer.weight.data, orig_w)

    # Numerical error verification: output MSE should remain low
    with torch.no_grad():
        out_orig = torch.matmul(x, orig_w.t())
        out_q = layer(x)
        rel_err = torch.norm(out_orig - out_q) / torch.norm(out_orig)
        assert rel_err < 0.6  # High quality 4-bit representation

def test_quantize_model_gptq()

Expand source code

def test_quantize_model_gptq():
    model = nn.Sequential(nn.Linear(32, 32), nn.Linear(32, 16))
    calib = [torch.randn(8, 32) for _ in range(3)]
    quantize_model_gptq(model, calib, bits=4)
    out = model(torch.randn(2, 32))
    assert out.shape == (2, 16)