EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Set num_fused_shared_experts as num_shared_experts when shared_experts fusion is not disabled (#6736)

Browse Source
This commit is contained in:
Cheng Wan
2025-06-04 15:53:22 -07:00
committed by GitHub
parent f0f84975f4
commit 81964328b7
22 changed files with 381 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
sgl-kernel/tests/test_moe_fused_gate.py
View File

@@ -19,7 +19,7 @@ from sglang.srt.layers.moe.topk import biased_grouped_topk
(512, 16, 8, 16),
],
)
@pytest.mark.parametrize("num_fused_shared_experts", [0, 1])
@pytest.mark.parametrize("num_fused_shared_experts", [0, 1, 2])
def test_moe_fused_gate_combined(seq_length, dtype, params, num_fused_shared_experts):
num_experts, num_expert_group, topk_group, topk = params
@@ -27,7 +27,7 @@ def test_moe_fused_gate_combined(seq_length, dtype, params, num_fused_shared_exp
tensor = torch.rand((seq_length, num_experts)).to(dtype).cuda()
scores = tensor.clone()
bias = torch.rand(num_experts).to(dtype).cuda()
topk = topk + min(1, num_fused_shared_experts)
topk = topk + num_fused_shared_experts
output, indices = moe_fused_gate(
tensor,