config: the per-instance families read the bags (#35026)

This commit is contained in:
Cheng Wan
2026-08-17 16:17:53 -07:00
committed by GitHub
parent a97bc8db32
commit cba3c5d5ac
45 changed files with 909 additions and 530 deletions
@@ -15,6 +15,7 @@ from sglang.srt.model_executor.runner.decode_cuda_graph_runner import (
from sglang.srt.model_executor.runner.prefill_cuda_graph_runner import (
PrefillCudaGraphRunner,
)
from sglang.srt.runtime_context import get_context
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
@@ -23,31 +24,29 @@ register_cpu_ci(est_time=1, suite="base-a-test-cpu")
class TestHiddenStateGraphRecapture(CustomTestCase):
def test_server_mode_sets_graph_capture_ceiling(self):
disabled = SimpleNamespace(
enable_return_hidden_states=False,
return_hidden_states_mode=None,
)
last = SimpleNamespace(
enable_return_hidden_states=True,
return_hidden_states_mode="last",
)
full = SimpleNamespace(
enable_return_hidden_states=True,
return_hidden_states_mode="full",
)
self.assertEqual(
get_server_return_hidden_states_mode(disabled),
CaptureHiddenMode.NULL,
)
self.assertEqual(
get_server_return_hidden_states_mode(last),
CaptureHiddenMode.LAST,
)
self.assertEqual(
get_server_return_hidden_states_mode(full),
CaptureHiddenMode.FULL,
cases = (
(dict(enable_return_hidden_states=False), CaptureHiddenMode.NULL),
(
dict(
enable_return_hidden_states=True, return_hidden_states_mode="last"
),
CaptureHiddenMode.LAST,
),
(
dict(
enable_return_hidden_states=True, return_hidden_states_mode="full"
),
CaptureHiddenMode.FULL,
),
)
for fields, expected in cases:
with self.subTest(**fields):
override = get_context().override_server_args(**fields)
override.install()
try:
self.assertEqual(get_server_return_hidden_states_mode(), expected)
finally:
override.restore()
@staticmethod
def _make_runner(runner_cls, capture_hidden_mode):
@@ -17,6 +17,7 @@ from sglang.srt.model_executor.runner.prefill_cuda_graph_runner import (
PrefillCudaGraphRunner,
)
from sglang.srt.model_executor.runner.shape_key import ShapeKey
from sglang.srt.runtime_context import get_context
from sglang.test.ci.ci_register import register_cpu_ci
from sglang.test.test_utils import CustomTestCase
@@ -111,13 +112,17 @@ class TestPrefillCudaGraphRunnerChunkedPrefix(CustomTestCase):
def test_eagle_target_tc_piecewise_skips_last_mode_capture(self):
eager_runner = object()
# The server-side hidden-state ceiling is a bag leaf.
override = get_context().override_server_args(
enable_return_hidden_states=True,
return_hidden_states_mode="last",
)
override.install()
self.addCleanup(override.restore)
model_runner = SimpleNamespace(
is_draft_worker=False,
spec_algorithm=SimpleNamespace(is_eagle=lambda: True),
server_args=SimpleNamespace(
enable_return_hidden_states=True,
return_hidden_states_mode="last",
),
server_args=SimpleNamespace(),
)
with patch.object(