chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Use dsv3 optimized routing fused_topk_deepseek instead of moe_fused_gate (#15347)

Browse Source
This commit is contained in:
Lee Nau
2026-01-19 11:50:16 +08:00
committed by GitHub
parent ea879c7739
commit 84c8390514
3 changed files with 165 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/test_deepseek_v3_mtp.py
+2 -8
View File
@@ -82,10 +82,7 @@ class TestDeepseekV3MTP(CustomTestCase):
f"{avg_spec_accept_length=:.2f}\n"
)
self.assertGreater(metrics["accuracy"], 0.935)
if is_in_amd_ci():
self.assertGreater(avg_spec_accept_length, 2.8)
else:
self.assertGreater(avg_spec_accept_length, 2.9)
self.assertGreater(avg_spec_accept_length, 2.8)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
@@ -99,10 +96,7 @@ class TestDeepseekV3MTP(CustomTestCase):
f"{acc_length=:.2f}\n"
f"{speed=:.2f} token/s\n"
)
if is_in_amd_ci():
self.assertGreater(acc_length, 2.8)
else:
self.assertGreater(acc_length, 2.9)
self.assertGreater(acc_length, 2.8)
if is_in_amd_ci():
self.assertGreater(speed, 15)
else: