Gahow Wang
445e491123
third_party/vllm/ now tracked in git for direct patch management.
Based on vLLM v0.18.1 release with one patch applied:
vllm/v1/core/sched/scheduler.py:
Replace fatal assert with graceful skip when KV transfer callback
arrives for an already-aborted request during PD disaggregated serving.
Future vLLM modifications should be made directly in third_party/vllm/
and committed normally. The patches/ directory is kept as documentation
of what changed from upstream.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
36 lines
1.2 KiB
Python
36 lines
1.2 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
|
|
import torch
|
|
|
|
from tests.kernels.utils import opcheck
|
|
from vllm import _custom_ops as ops # noqa: F401
|
|
|
|
|
|
def test_gptq_shuffle_opcheck():
|
|
weight = torch.randint(
|
|
-2000000, 2000000, (1792, 4096), device="cuda", dtype=torch.int32
|
|
)
|
|
perm = torch.empty((0,), device="cuda", dtype=torch.int32)
|
|
bit = 4
|
|
opcheck(torch.ops._C.gptq_shuffle, (weight, perm, bit))
|
|
|
|
|
|
def test_gptq_gemm_opcheck():
|
|
a = torch.rand((240, 4096), device="cuda", dtype=torch.float16)
|
|
weight = torch.randint(
|
|
-2000000, 2000000, (512, 6144), device="cuda", dtype=torch.int32
|
|
)
|
|
zeros = torch.zeros((32, 768), device="cuda", dtype=torch.int32)
|
|
scales = torch.rand((32, 6144), device="cuda", dtype=torch.float16)
|
|
idx = torch.empty((0,), device="cuda", dtype=torch.int32)
|
|
use_exllama = True
|
|
bit = 4
|
|
# Test both GPTQv1 and GPTQv2 format
|
|
opcheck(
|
|
torch.ops._C.gptq_gemm, (a, weight, zeros, scales, idx, use_exllama, True, bit)
|
|
)
|
|
opcheck(
|
|
torch.ops._C.gptq_gemm, (a, weight, zeros, scales, idx, use_exllama, False, bit)
|
|
)
|