[Spec] Support logprobs with DFlash (#33459)

This commit is contained in:
Jason Mancuso
2026-08-05 12:37:12 -07:00
committed by GitHub
parent 1a045669e4
commit 36853b8ffc
4 changed files with 20 additions and 13 deletions
+2 -5
View File
@@ -12,7 +12,7 @@ from sglang.test.kits.radix_cache_server_kit import (
gen_radix_tree,
run_radix_attention_test,
)
from sglang.test.kits.spec_server_kits import SpecGrammarKit
from sglang.test.kits.spec_server_kits import SpecGrammarKit, SpecLogprobKit
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_DFLASH,
DEFAULT_TARGET_MODEL_DFLASH,
@@ -32,6 +32,7 @@ class TestDFlashServerBase(
GSM8KMixin,
JSONConstrainedMixin,
SpecGrammarKit,
SpecLogprobKit,
):
max_running_requests = 64
attention_backend = "triton" if is_hip() else "flashinfer"
@@ -132,10 +133,6 @@ class TestDFlashServerBase(
self.assertEqual(outputs[0], outputs[1])
assert self.process.poll() is None
@unittest.skip("DFLASH rejects return_logprob at admission")
def test_grammar_logprob_count_matches_completion_tokens(self):
pass
class TestDFlashServerPage256(TestDFlashServerBase):
page_size = 256