[CI] Fix FusedMoEConfig and input batch failure to recover CI (#1602)

Make CI happy

1.
c1909e7e8c
changed moeConfig init way
2.
48fb076cbc
changed input batch logic.

This PR address these change to vllm-ascend.

Closes: https://github.com/vllm-project/vllm-ascend/issues/1600

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
This commit is contained in:
wangxiyuan
2025-07-03 18:36:17 +08:00
committed by GitHub
parent d96da1f00c
commit a45dfde283
11 changed files with 173 additions and 134 deletions

View File

@@ -18,9 +18,12 @@
#
from typing import Optional
import pytest
import torch
from vllm.v1.sample.sampler import Sampler # noqa: F401
from vllm_ascend.utils import vllm_version_is
# Set tolerance to 1 for quant ops
DEFAULT_ATOL = 1e-3
DEFAULT_RTOL = 1e-3
@@ -118,6 +121,8 @@ def apply_top_k_top_p_new(
# test with leading dimension and merge seqlen and batch_size as num_tokens
@pytest.mark.skipif(not vllm_version_is("0.9.1"),
reason="apply_min_p has been removed after vllm 0.9.1")
@torch.inference_mode()
def test_apply_min_p() -> None:
logits = torch.randn((128, 7168)).npu()