mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2025-12-23 03:55:01 +08:00
e9add129ad
Signed-off-by: Matthias Gehre <matthias.gehre@amd.com> Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
50 lines
1.8 KiB
Python
50 lines
1.8 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import pytest
|
|
import torch
|
|
|
|
from tests.kernels.utils import opcheck
|
|
from vllm import _custom_ops as ops # noqa: F401
|
|
|
|
|
|
@pytest.mark.skipif(
|
|
not hasattr(torch.ops._C, "awq_dequantize"),
|
|
reason="AWQ is not supported on this GPU type.",
|
|
)
|
|
def test_awq_dequantize_opcheck(monkeypatch: pytest.MonkeyPatch):
|
|
with monkeypatch.context() as m:
|
|
m.setenv("VLLM_USE_TRITON_AWQ", "0")
|
|
qweight = torch.randint(
|
|
-2000000000, 2000000000, (8192, 256), device="cuda", dtype=torch.int32
|
|
)
|
|
scales = torch.rand((64, 2048), device="cuda", dtype=torch.float16)
|
|
zeros = torch.empty((64, 256), device="cuda", dtype=torch.int32)
|
|
split_k_iters = 0
|
|
thx = 0
|
|
thy = 0
|
|
opcheck(
|
|
torch.ops._C.awq_dequantize,
|
|
(qweight, scales, zeros, split_k_iters, thx, thy),
|
|
)
|
|
|
|
|
|
@pytest.mark.skip(reason="Not working; needs investigation.")
|
|
@pytest.mark.skipif(
|
|
not hasattr(torch.ops._C, "awq_gemm"),
|
|
reason="AWQ is not supported on this GPU type.",
|
|
)
|
|
def test_awq_gemm_opcheck(monkeypatch: pytest.MonkeyPatch):
|
|
with monkeypatch.context() as m:
|
|
m.setenv("VLLM_USE_TRITON_AWQ", "0")
|
|
input = torch.rand((2, 8192), device="cuda", dtype=torch.float16)
|
|
qweight = torch.randint(
|
|
-2000000000, 2000000000, (8192, 256), device="cuda", dtype=torch.int32
|
|
)
|
|
scales = torch.empty((64, 2048), device="cuda", dtype=torch.float16)
|
|
qzeros = torch.randint(
|
|
-2000000000, 2000000000, (64, 256), device="cuda", dtype=torch.int32
|
|
)
|
|
split_k_iters = 8
|
|
opcheck(torch.ops._C.awq_gemm, (input, qweight, scales, qzeros, split_k_iters))
|