| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from functools import partial |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from vllm.model_executor.layers.activation import GeluAndMul, SiluAndMul |
| 11 | from vllm.platforms import CpuArchEnum, current_platform |
| 12 | |
| 13 | |
| 14 | @pytest.mark.skipif( |
| 15 | not current_platform.is_cpu() |
| 16 | or current_platform.get_cpu_architecture() == CpuArchEnum.POWERPC, |
| 17 | reason="Native CPU activations require a non-PowerPC CPU platform", |
| 18 | ) |
| 19 | @pytest.mark.parametrize( |
| 20 | "activation_factory,reference", |
| 21 | [ |
| 22 | (SiluAndMul, nn.SiLU()), |
| 23 | (GeluAndMul, nn.GELU()), |
| 24 | (partial(GeluAndMul, approximate="tanh"), nn.GELU(approximate="tanh")), |
| 25 | ], |
| 26 | ids=["silu", "gelu", "gelu_tanh"], |
| 27 | ) |
| 28 | def test_cpu_act_and_mul_without_compiled_ops( |
| 29 | default_vllm_config, monkeypatch, activation_factory, reference |
| 30 | ): |
| 31 | """CPU activations must construct and run without compiled wheel operators.""" |
| 32 | default_vllm_config.compilation_config.custom_ops = ["all"] |
| 33 | monkeypatch.setattr(torch.ops, "_C", object()) |
| 34 | x = torch.tensor([[-2.0, -1.0, 0.0, 1.0, 2.0, 3.0, 4.0, 5.0]], device="cpu") |
| 35 | |
| 36 | layer = activation_factory() |
| 37 | |
| 38 | torch.testing.assert_close(layer(x), reference(x[:, :4]) * x[:, 4:]) |