| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | |
| 7 | from tests.kernels.utils import opcheck |
| 8 | from vllm import _custom_ops as ops # noqa: F401 |
| 9 | |
| 10 | |
| 11 | @pytest.mark.skipif( |
| 12 | not hasattr(torch.ops._C, "awq_dequantize"), |
| 13 | reason="AWQ is not supported on this GPU type.", |
| 14 | ) |
| 15 | def test_awq_dequantize_opcheck(monkeypatch: pytest.MonkeyPatch): |
| 16 | with monkeypatch.context() as m: |
| 17 | m.setenv("VLLM_USE_TRITON_AWQ", "0") |
| 18 | qweight = torch.randint( |
| 19 | -2000000000, 2000000000, (8192, 256), device="cuda", dtype=torch.int32 |
| 20 | ) |
| 21 | scales = torch.rand((64, 2048), device="cuda", dtype=torch.float16) |
| 22 | zeros = torch.empty((64, 256), device="cuda", dtype=torch.int32) |
| 23 | split_k_iters = 0 |
| 24 | thx = 0 |
| 25 | thy = 0 |
| 26 | opcheck( |
| 27 | torch.ops._C.awq_dequantize, |
| 28 | (qweight, scales, zeros, split_k_iters, thx, thy), |
| 29 | ) |