| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from tests.kernels.utils import opcheck |
| 7 | from vllm import _custom_ops as ops # noqa: F401 |
| 8 | |
| 9 | |
| 10 | def test_gptq_shuffle_opcheck(): |
| 11 | weight = torch.randint( |
| 12 | -2000000, 2000000, (1792, 4096), device="cuda", dtype=torch.int32 |
| 13 | ) |
| 14 | bit = 4 |
| 15 | opcheck(torch.ops._C.gptq_shuffle, (weight, bit)) |
| 16 | |
| 17 | |
| 18 | def test_gptq_gemm_opcheck(): |
| 19 | a = torch.rand((240, 4096), device="cuda", dtype=torch.float16) |
| 20 | weight = torch.randint( |
| 21 | -2000000, 2000000, (512, 6144), device="cuda", dtype=torch.int32 |
| 22 | ) |
| 23 | zeros = torch.zeros((32, 768), device="cuda", dtype=torch.int32) |
| 24 | scales = torch.rand((32, 6144), device="cuda", dtype=torch.float16) |
| 25 | use_exllama = True |
| 26 | bit = 4 |
| 27 | # Test both GPTQv1 and GPTQv2 format |
| 28 | opcheck(torch.ops._C.gptq_gemm, (a, weight, zeros, scales, use_exllama, True, bit)) |
| 29 | opcheck(torch.ops._C.gptq_gemm, (a, weight, zeros, scales, use_exllama, False, bit)) |