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

Revert "[1/2] sgl-kernel: Fuse routed scaling factor into select_experts" (#8706)

Browse Source
This commit is contained in:
Liangsheng Yin
2025-08-02 20:14:30 +08:00
committed by GitHub
parent ac6962ccd6
commit f9f0138f80
5 changed files with 12 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
sgl-kernel/tests/test_moe_fused_gate.py
View File

@@ -19,10 +19,7 @@ from sglang.srt.layers.moe.topk import biased_grouped_topk
],
)
@pytest.mark.parametrize("num_fused_shared_experts", [0, 1, 2])
@pytest.mark.parametrize("apply_routed_scaling_factor_on_output", [True, False])
def test_moe_fused_gate_combined(
seq_length, params, num_fused_shared_experts, apply_routed_scaling_factor_on_output
):
def test_moe_fused_gate_combined(seq_length, params, num_fused_shared_experts):
num_experts, num_expert_group, topk_group, topk = params
dtype = torch.float32
@@ -40,7 +37,6 @@ def test_moe_fused_gate_combined(
topk=topk,
num_fused_shared_experts=num_fused_shared_experts,
routed_scaling_factor=2.5,
apply_routed_scaling_factor_on_output=apply_routed_scaling_factor_on_output,
)
ref_output, ref_indices = biased_grouped_topk(
scores,
@@ -52,7 +48,6 @@ def test_moe_fused_gate_combined(
topk_group=topk_group,
num_fused_shared_experts=num_fused_shared_experts,
routed_scaling_factor=2.5,
apply_routed_scaling_factor_on_output=apply_routed_scaling_factor_on_output,
)
# When num_fused_shared_experts > 0, ignore the comparison of the last topk dimension