EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Feature] Support fine-grained shared expert overlap (#5482)

Browse Source 
Fine-grained control over shared expert overlap to prevent resource
contention.

- vLLM version: v0.13.0
- vLLM main:
5326c89803

---------

Signed-off-by: Jade Zheng <zheng.shoujian@outlook.com>
This commit is contained in:
Jade Zheng
2026-01-17 11:53:22 +08:00
committed by GitHub
parent 48e10de8c9
commit 22f253142a
9 changed files with 203 additions and 130 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
tests/ut/ops/test_moe_comm_method.py
View File

@@ -163,6 +163,7 @@ class TestMoECommMethod(TestBase):
"vllm_ascend.ops.fused_moe.moe_comm_method.TokenDispatcherWithAllGather"
)
@patch("vllm_ascend.ops.fused_moe.moe_comm_method.unified_apply_mlp")
@patch("torch.npu.current_stream", MagicMock())
def test_fused_experts_method(self, mock_unified_apply_mlp,
mock_token_dispatcher, mock_prepare_finalize,
mock_get_forward_context):