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

[Spec]: Make Triton standalone spec test deterministic (#25303)

Browse Source
This commit is contained in:
Ratish P
2026-05-18 17:36:10 -07:00
committed by GitHub
parent b79e4b1e68
commit dbac464726
2 changed files with 8 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/test/server_fixtures/standalone_fixture.py
+7 -2
View File
@@ -2,7 +2,8 @@
Variants combine this base with `CustomTestCase` and override class Variants combine this base with `CustomTestCase` and override class
attributes (`attention_backend`, plus optional `speculative_eagle_topk` / attributes (`attention_backend`, plus optional `speculative_eagle_topk` /
`speculative_num_draft_tokens` / `enable_spec_v2`) to select a backend `speculative_num_draft_tokens` / `enable_spec_v2` /
`enable_deterministic_inference`) to select a backend, deterministic mode,
and the V1 / V2 spec engine. and the V1 / V2 spec engine.
Pure mixin (does NOT inherit `TestCase`), so unittest does not collect Pure mixin (does NOT inherit `TestCase`), so unittest does not collect
@@ -41,11 +42,12 @@ class StandaloneServerBase:
speculative_eagle_topk: int = 1 speculative_eagle_topk: int = 1
speculative_num_draft_tokens: int = 5 speculative_num_draft_tokens: int = 5
enable_spec_v2: bool = True enable_spec_v2: bool = True
enable_deterministic_inference: bool = False
@classmethod @classmethod
def get_server_args(cls): def get_server_args(cls):
assert cls.attention_backend, f"{cls.__name__} must set `attention_backend`" assert cls.attention_backend, f"{cls.__name__} must set `attention_backend`"
return [ args = [
"--trust-remote-code", "--trust-remote-code",
"--cuda-graph-max-bs", "--cuda-graph-max-bs",
"8", "8",
@@ -64,6 +66,9 @@ class StandaloneServerBase:
"--attention-backend", "--attention-backend",
cls.attention_backend, cls.attention_backend,
] ]
if cls.enable_deterministic_inference:
args.append("--enable-deterministic-inference")
return args
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
test/registered/spec/test_spec_standalone_extra.py
+1
View File
@@ -21,6 +21,7 @@ class TestStandaloneSpeculativeDecodingTriton(StandaloneServerBase, CustomTestCa
speculative_eagle_topk = 2 speculative_eagle_topk = 2
speculative_num_draft_tokens = 7 speculative_num_draft_tokens = 7
enable_spec_v2 = False enable_spec_v2 = False
enable_deterministic_inference = True
class TestStandaloneSpeculativeDecodingFlashinfer(StandaloneServerBase, CustomTestCase): class TestStandaloneSpeculativeDecodingFlashinfer(StandaloneServerBase, CustomTestCase):