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

2026-04-05 23:26:40 +08:00
parent 51b276de74
commit f6c9072e42
1 changed files with 2 additions and 3 deletions
@@ -68,11 +68,10 @@ class TestQwen3NextMTPTopk(
    ]


-# TODO(hzh): After merging the PR that fixes specv2 to correctly return log probs,
-# add KLDivergenceMixin back. https://github.com/sgl-project/sglang/pull/18645
-class TestQwen3NextMTPV2(GSM8KMixin, DefaultServerBase):
+class TestQwen3NextMTPV2(GSM8KMixin, KLDivergenceMixin, DefaultServerBase):
    model = QWEN3_NEXT_MODEL
    gsm8k_accuracy_thres = 0.93
+    kl_div_thres = 0.0025
    other_args = [
        "--trust-remote-code",
        "--speculative-algorithm",