[Bugfix] Try to handle older versions of pytorch (#9086)

This commit is contained in:
bnellnm
2024-10-08 14:28:12 -07:00
committed by GitHub
parent de24046fcd
commit bd37b9fbe2
3 changed files with 41 additions and 21 deletions
+5
View File
@@ -1,11 +1,14 @@
import os
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():
os.environ["VLLM_USE_TRITON_AWQ"] = "0"
qweight = torch.randint(-2000000000,
@@ -21,6 +24,8 @@ def test_awq_dequantize_opcheck():
(qweight, scales, zeros, split_k_iters, thx, thy))
@pytest.mark.skipif(not hasattr(torch.ops._C, "awq_gemm"),
reason="AWQ is not supported on this GPU type.")
def test_awq_gemm_opcheck():
os.environ["VLLM_USE_TRITON_AWQ"] = "0"
input = torch.rand((2, 8192), device='cuda', dtype=torch.float16)
+4
View File
@@ -7,6 +7,7 @@ import torch
from tests.kernels.utils import (compute_max_diff, stack_and_dev, torch_moe,
torch_moe_single)
from vllm import _custom_ops as ops
from vllm.model_executor.layers.fused_moe.fused_marlin_moe import (
fused_marlin_moe, single_marlin_moe)
from vllm.model_executor.layers.fused_moe.fused_moe import fused_topk
@@ -21,6 +22,9 @@ from vllm.scalar_type import scalar_types
@pytest.mark.parametrize("e", [8, 64])
@pytest.mark.parametrize("topk", [2, 6])
@pytest.mark.parametrize("group_size", [-1, 32, 64, 128])
@pytest.mark.skipif(not (ops.supports_moe_ops
and hasattr(torch.ops._moe_C, "marlin_gemm_moe")),
reason="Marlin is not supported on this GPU type.")
def test_fused_marlin_moe_awq(
m: int,
n: int,