minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[AMD] Fix fused append+remap DeepEP equivalence test on aiter path (#29377)

Browse Source
This commit is contained in:
YC Yen-Ching Tseng
2026-06-26 16:04:20 -07:00
committed by GitHub
parent 72812db138
commit 5eebb4b2e8
1 changed files with 6 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/moe/test_fused_append_remap_deepep.py
+6 -5
View File
@@ -14,7 +14,7 @@ from sglang.srt.layers.moe.moe_runner.triton_utils.fused_moe_triton_kernels impo
fused_append_remap_shared_experts_deepep, fused_append_remap_shared_experts_deepep,
fused_append_shared_experts, fused_append_shared_experts,
) )
from sglang.srt.layers.moe.topk import TopKConfig, _remap_topk_for_deepep from sglang.srt.layers.moe.topk import TopKConfig, _remap_topk_for_deepep, _use_aiter
from sglang.srt.runtime_context import get_parallel from sglang.srt.runtime_context import get_parallel
from sglang.srt.utils import get_device from sglang.srt.utils import get_device
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
@@ -109,12 +109,13 @@ class TestFusedAppendRemapDeepEP(CustomTestCase):
def test_equivalence_with_eager_append_then_remap(self): def test_equivalence_with_eager_append_then_remap(self):
"""Fused kernel == fused_append_shared_experts() + _remap_topk_for_deepep(). """Fused kernel == fused_append_shared_experts() + _remap_topk_for_deepep().
The eager remap overwrites the shared weight with 1/routed_scaling_factor, The eager remap overwrites the shared weight: 1.0 on the aiter/HIP path
so the fused kernel is invoked with that same value to make the two paths (routed_scaling_factor is pre-folded into the routed topk weights), else
bit-identical (ids are identical regardless of the scaling factor). 1/routed_scaling_factor. The fused kernel is invoked with that same value
so the two paths stay bit-identical (ids match regardless of scaling).
""" """
rsf = 2.5 rsf = 2.5
scale_factor = 1.0 / rsf scale_factor = 1.0 if _use_aiter else 1.0 / rsf
for m, k, npr, ep_size, ep_rank, s in self.CASES: for m, k, npr, ep_size, ep_rank, s in self.CASES:
with self.subTest(m=m, k=k, npr=npr, ep_rank=ep_rank, s=s): with self.subTest(m=m, k=k, npr=npr, ep_rank=ep_rank, s=s):
shared_id_base, num_local_routed = self._shared_id_base( shared_id_base, num_local_routed = self._shared_id_base(