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

[6/N] MoE Refactor: Cleanup MoE-related configs (#8849)

Browse Source
This commit is contained in:
Cheng Wan
2025-08-14 21:14:53 -07:00
committed by GitHub
parent 584e1ab2d0
commit 295895120d
69 changed files with 956 additions and 1037 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
test/srt/test_fused_moe.py
View File

@@ -6,7 +6,7 @@ from tqdm import tqdm
from sglang.srt.layers.activation import SiluAndMul
from sglang.srt.layers.moe.fused_moe_triton.fused_moe import fused_moe
from sglang.srt.layers.moe.topk import select_experts
from sglang.srt.layers.moe.topk import TopKConfig, select_experts
from sglang.srt.layers.quantization.fp8_kernel import is_fp8_fnuz
from sglang.srt.layers.quantization.fp8_utils import normalize_e4m3fn_to_e4m3fnuz
from sglang.srt.utils import is_hip
@@ -136,19 +136,7 @@ class TestFusedMOE(CustomTestCase):
topk_output = select_experts(
hidden_states=a,
router_logits=score,
top_k=topk,
)
sglang_output = fused_moe(
a,
w1,
w2,
topk_output,
use_fp8_w8a8=True,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a1_scale,
a2_scale=a2_scale,
topk_config=TopKConfig(top_k=topk, renormalize=False),
)
torch_output = self.torch_naive_moe(
@@ -162,6 +150,18 @@ class TestFusedMOE(CustomTestCase):
a1_scale,
a2_scale,
)
sglang_output = fused_moe(
a,
w1,
w2,
topk_output,
use_fp8_w8a8=True,
w1_scale=w1_scale,
w2_scale=w2_scale,
a1_scale=a1_scale,
a2_scale=a2_scale,
)
torch.testing.assert_close(
sglang_output, torch_output, rtol=rtol, atol=atol
)
@@ -174,7 +174,7 @@ class TestFusedMOE(CustomTestCase):
topk_output = select_experts(
hidden_states=a,
router_logits=score,
top_k=topk,
topk_config=TopKConfig(top_k=topk, renormalize=False),
)
triton_output = fused_moe(a, w1, w2, topk_output)