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

ci: tag-gated nightly migration — foundation + 40 whole-file moves (#24725)

Browse Source 
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
This commit is contained in:
Alison Shao
2026-05-14 16:11:15 -07:00
committed by GitHub
co-authored by hnyls2002 Liangsheng Yin
parent 67096f48bf
commit ba214ef3d3
78 changed files with 2254 additions and 2131 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/sessions/test_session_control.py
+1 -1
View File
@@ -24,7 +24,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=87, stage="stage-b", runner_config="1-gpu-large")
register_cuda_ci(est_time=87, stage="extra-a", runner_config="1-gpu-large")
def remove_prefix(text: str, prefix: str) -> str:
test/registered/sessions/test_session_latency.py
+1 -5
View File
@@ -31,11 +31,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(
est_time=122,
stage="stage-b",
runner_config="1-gpu-large",
)
register_cuda_ci(est_time=122, stage="extra-a", runner_config="1-gpu-large")
NUM_TURNS = 150
INPUT_LEN = 16
test/registered/sessions/test_streaming_session.py
+60 -1070
View File
File diff suppressed because it is too large Load Diff
test/registered/sessions/test_streaming_session_extra.py
+105
View File
@@ -0,0 +1,105 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.streaming_session_kit import StreamingSessionKitMixin
from sglang.test.server_fixtures.streaming_session_fixture import (
StreamingSessionServerBase,
)
from sglang.test.test_utils import (
DEFAULT_DRAFT_MODEL_EAGLE3,
DEFAULT_TARGET_MODEL_EAGLE3,
)
register_cuda_ci(est_time=691, stage="extra-a", runner_config="1-gpu-large")
class TestStreamingSessionRetractMixedChunk(
StreamingSessionServerBase, StreamingSessionKitMixin
):
"""Retract + --enable-mixed-chunk."""
extra_args = ["--chunked-prefill-size", "128", "--enable-mixed-chunk"]
env_overrides = [("SGLANG_TEST_RETRACT", True)]
class TestStreamingSessionRetractLargePage(
StreamingSessionServerBase, StreamingSessionKitMixin
):
"""Retract + page=256: exercises page-aligned `_free_tail`. Partial-page
free would corrupt pages still holding committed tokens."""
extra_args = ["--chunked-prefill-size", "4096", "--page-size", "256"]
env_overrides = [("SGLANG_TEST_RETRACT", True)]
# Common EAGLE3 spec args; reused by Eagle/EagleV2/EagleRetractLargePage variants.
_EAGLE3_SPEC_ARGS = [
"--dtype=float16",
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model",
DEFAULT_DRAFT_MODEL_EAGLE3,
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.7",
]
class TestStreamingSessionEagle(StreamingSessionServerBase, StreamingSessionKitMixin):
"""EAGLE3 spec v1 (overlap disabled); offset=-1 — see kit's note."""
kv_inherit_offset = -1
model = DEFAULT_TARGET_MODEL_EAGLE3
extra_args = [
"--disable-overlap-schedule",
"--chunked-prefill-size",
"512",
*_EAGLE3_SPEC_ARGS,
]
env_overrides = [("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", True)]
class TestStreamingSessionEagleV2(StreamingSessionServerBase, StreamingSessionKitMixin):
"""EAGLE3 spec v2 (overlap on)."""
model = DEFAULT_TARGET_MODEL_EAGLE3
extra_args = [
"--chunked-prefill-size",
"512",
*_EAGLE3_SPEC_ARGS,
]
env_overrides = [
("SGLANG_ENABLE_SPEC_V2", True),
("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", True),
]
class TestStreamingSessionEagleRetractLargePage(
StreamingSessionServerBase, StreamingSessionKitMixin
):
"""EAGLE3 spec v1 + retract + page=256: max-pressure on `_free_tail`
(spec tail + retract alloc-commit gap + page alignment)."""
kv_inherit_offset = -1
model = DEFAULT_TARGET_MODEL_EAGLE3
extra_args = [
"--disable-overlap-schedule",
"--chunked-prefill-size",
"4096",
*_EAGLE3_SPEC_ARGS,
"--page-size",
"256",
]
env_overrides = [
("SGLANG_TEST_RETRACT", True),
("SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN", True),
]
if __name__ == "__main__":
unittest.main()
test/registered/sessions/test_streaming_session_swa.py
+48 -122
View File
@@ -1,27 +1,15 @@
import os
import sys
import unittest
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
popen_launch_server,
from sglang.test.kits.streaming_session_kit import (
AbortLeakReproKitMixin,
StreamingSessionKitMixin,
)
# test/ has no __init__.py; add sibling dir so sibling module is importable
# when this file is run as a script via `python3 <path>`.
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from test_streaming_session import ( # noqa: E402
from sglang.test.server_fixtures.streaming_session_fixture import (
ABORT_REPRO_CHUNKED_PREFILL_SIZE,
ABORT_REPRO_CONTEXT_LEN,
ABORT_REPRO_PAGE_SIZE,
TestStreamingSession,
TestStreamingSessionAbortLeakRepro,
StreamingSessionServerBase,
)
register_cuda_ci(est_time=519, stage="stage-b", runner_config="1-gpu-large")
@@ -37,125 +25,63 @@ SWA_COMMON_ARGS = [
]
class TestStreamingSessionSWA(TestStreamingSession):
class TestStreamingSessionSWA(StreamingSessionServerBase, StreamingSessionKitMixin):
"""Baseline streaming session on a hybrid-SWA model."""
@classmethod
def setUpClass(cls):
cls.model = SWA_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--enable-streaming-session",
"--chunked-prefill-size",
"512",
*SWA_COMMON_ARGS,
],
)
cls.tokenizer = get_tokenizer(cls.model)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
model = SWA_MODEL
extra_args = ["--chunked-prefill-size", "512", *SWA_COMMON_ARGS]
class TestStreamingSessionSWARetractLargePage(TestStreamingSession):
class TestStreamingSessionSWARetractLargePage(
StreamingSessionServerBase, StreamingSessionKitMixin
):
"""SWA under retract decode with page=256."""
@classmethod
def setUpClass(cls):
cls.model = SWA_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
with (
envs.SGLANG_TEST_RETRACT.override(True),
envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2),
):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--enable-streaming-session",
"--chunked-prefill-size",
"4096",
"--page-size",
"256",
*SWA_COMMON_ARGS,
],
)
cls.tokenizer = get_tokenizer(cls.model)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
model = SWA_MODEL
extra_args = [
"--chunked-prefill-size",
"4096",
"--page-size",
"256",
*SWA_COMMON_ARGS,
]
env_overrides = [("SGLANG_TEST_RETRACT", True)]
class TestStreamingSessionSWARetractMixedChunk(TestStreamingSession):
class TestStreamingSessionSWARetractMixedChunk(
StreamingSessionServerBase, StreamingSessionKitMixin
):
"""SWA under retract decode with --enable-mixed-chunk."""
@classmethod
def setUpClass(cls):
cls.model = SWA_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
with (
envs.SGLANG_TEST_RETRACT.override(True),
envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2),
):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--enable-streaming-session",
"--chunked-prefill-size",
"128",
"--enable-mixed-chunk",
*SWA_COMMON_ARGS,
],
)
cls.tokenizer = get_tokenizer(cls.model)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
model = SWA_MODEL
extra_args = [
"--chunked-prefill-size",
"128",
"--enable-mixed-chunk",
*SWA_COMMON_ARGS,
]
env_overrides = [("SGLANG_TEST_RETRACT", True)]
class TestStreamingSessionSWAAbortLeakRepro(TestStreamingSessionAbortLeakRepro):
class TestStreamingSessionSWAAbortLeakRepro(
StreamingSessionServerBase, AbortLeakReproKitMixin
):
"""SWA abort-heavy chunked prefill leak repro."""
@classmethod
def setUpClass(cls):
cls.model = SWA_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
with envs.SGLANG_ENABLE_STRICT_MEM_CHECK_DURING_BUSY.override(2):
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--enable-streaming-session",
"--chunked-prefill-size",
str(ABORT_REPRO_CHUNKED_PREFILL_SIZE),
"--context-length",
str(ABORT_REPRO_CONTEXT_LEN),
"--page-size",
str(ABORT_REPRO_PAGE_SIZE),
"--max-running-requests",
"32",
"--log-level",
"info",
*SWA_COMMON_ARGS,
],
)
cls.tokenizer = get_tokenizer(cls.model)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
model = SWA_MODEL
extra_args = [
"--chunked-prefill-size",
str(ABORT_REPRO_CHUNKED_PREFILL_SIZE),
"--context-length",
str(ABORT_REPRO_CONTEXT_LEN),
"--page-size",
str(ABORT_REPRO_PAGE_SIZE),
"--max-running-requests",
"32",
"--log-level",
"info",
*SWA_COMMON_ARGS,
]
if __name__ == "__main__":