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

[SpecV2]: Reopen kl accuracy test for qwen3 + SpecV2 (#22104)

Browse Source
This commit is contained in:
Zhangheng
2026-04-05 23:26:40 +08:00
committed by GitHub
parent 51b276de74
commit f6c9072e42
1 changed files with 2 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
+2 -3
View File
@@ -68,11 +68,10 @@ class TestQwen3NextMTPTopk(
] ]
# TODO(hzh): After merging the PR that fixes specv2 to correctly return log probs, class TestQwen3NextMTPV2(GSM8KMixin, KLDivergenceMixin, DefaultServerBase):
# add KLDivergenceMixin back. https://github.com/sgl-project/sglang/pull/18645
class TestQwen3NextMTPV2(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL model = QWEN3_NEXT_MODEL
gsm8k_accuracy_thres = 0.93 gsm8k_accuracy_thres = 0.93
kl_div_thres = 0.0025
other_args = [ other_args = [
"--trust-remote-code", "--trust-remote-code",
"--speculative-algorithm", "--speculative-algorithm",