[Refactor][CI] Reorganize test tree, and move slow e2e to weekly (#4951)
Signed-off-by: wangyu <410167048@qq.com>
This commit is contained in:
1 parent
925347332f
commit
7c17111ff5
158 files changed
+3479
-4130
No files matched your search
@@ -0,0 +1,110 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm_omni.experimental.fullduplex.engine.duplex_control_client import (
|
||||
DuplexControlClient,
|
||||
DuplexControlRequestError,
|
||||
)
|
||||
from vllm_omni.experimental.fullduplex.engine.lease import DuplexLeaseActivity
|
||||
from vllm_omni.experimental.fullduplex.engine.messages import (
|
||||
DuplexControlError,
|
||||
DuplexControlResultMessage,
|
||||
DuplexFence,
|
||||
ResumeDuplexSessionMessage,
|
||||
TouchDuplexSessionMessage,
|
||||
)
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
|
||||
|
||||
class _Transport:
|
||||
def __init__(self) -> None:
|
||||
self.calls = []
|
||||
|
||||
def execute(self, key, message, **kwargs):
|
||||
self.calls.append((key, message, kwargs))
|
||||
return DuplexControlResultMessage(
|
||||
control_id=message.control_id,
|
||||
fence=message.fence,
|
||||
operation="touch" if isinstance(message, TouchDuplexSessionMessage) else "resume",
|
||||
session_id=message.session_id,
|
||||
ok=True,
|
||||
stage_results=[],
|
||||
)
|
||||
|
||||
|
||||
class _UnsupportedTransport:
|
||||
def execute(self, key, message, **kwargs):
|
||||
del key, kwargs
|
||||
return DuplexControlResultMessage(
|
||||
control_id=message.control_id,
|
||||
fence=message.fence,
|
||||
operation="touch",
|
||||
session_id=message.session_id,
|
||||
ok=False,
|
||||
stage_results=[{"result": {"supported": False}}],
|
||||
unsupported_count=1,
|
||||
error=DuplexControlError(code="invalid_capability", message="unsupported", retryable=False),
|
||||
accepted_fence=DuplexFence(message.session_id, epoch=2),
|
||||
lease_generation=7,
|
||||
)
|
||||
|
||||
|
||||
def test_control_client_routes_touch_and_resume_by_control_id() -> None:
|
||||
transport = _Transport()
|
||||
control_ids = iter(("touch-id", "resume-id"))
|
||||
client = DuplexControlClient(transport, control_id_factory=lambda: next(control_ids))
|
||||
fence = DuplexFence("sid-client")
|
||||
|
||||
assert (
|
||||
client.touch(
|
||||
fence.session_id,
|
||||
fence=fence,
|
||||
activity=DuplexLeaseActivity.PLAYBACK_ACK,
|
||||
timeout=2.0,
|
||||
)["ok"]
|
||||
is True
|
||||
)
|
||||
assert (
|
||||
client.resume(
|
||||
fence.session_id,
|
||||
fence=fence,
|
||||
expected_lease_generation=7,
|
||||
timeout=3.0,
|
||||
)["ok"]
|
||||
is True
|
||||
)
|
||||
|
||||
touch_key, touch_message, touch_kwargs = transport.calls[0]
|
||||
assert touch_key == ("duplex", "touch-id")
|
||||
assert isinstance(touch_message, TouchDuplexSessionMessage)
|
||||
assert touch_message.activity == DuplexLeaseActivity.PLAYBACK_ACK.value
|
||||
assert touch_kwargs["timeout"] == 2.0
|
||||
|
||||
resume_key, resume_message, resume_kwargs = transport.calls[1]
|
||||
assert resume_key == ("duplex", "resume-id")
|
||||
assert isinstance(resume_message, ResumeDuplexSessionMessage)
|
||||
assert resume_message.expected_lease_generation == 7
|
||||
assert resume_kwargs["timeout"] == 3.0
|
||||
|
||||
|
||||
def test_control_client_treats_ok_false_as_authoritative_failure() -> None:
|
||||
client = DuplexControlClient(_UnsupportedTransport(), control_id_factory=lambda: "unsupported")
|
||||
fence = DuplexFence("sid-unsupported")
|
||||
|
||||
with pytest.raises(DuplexControlRequestError, match="duplex touch failed") as exc_info:
|
||||
client.touch(
|
||||
fence.session_id,
|
||||
fence=fence,
|
||||
activity=DuplexLeaseActivity.HEARTBEAT,
|
||||
timeout=1.0,
|
||||
)
|
||||
|
||||
assert exc_info.value.code == "invalid_capability"
|
||||
assert exc_info.value.retryable is False
|
||||
assert exc_info.value.accepted_fence == DuplexFence(fence.session_id, epoch=2)
|
||||
assert exc_info.value.lease_generation == 7
|
||||
File diff suppressed because it is too large.
Load diff
@@ -0,0 +1,68 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Focused tests for typed deployment-only runtime configuration."""
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import FrozenInstanceError
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm_omni.config.stage_config import (
|
||||
DuplexSessionRuntimeConfig,
|
||||
load_deploy_config,
|
||||
)
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
|
||||
|
||||
def test_duplex_session_runtime_defaults_are_typed_and_immutable(tmp_path) -> None:
|
||||
deploy_path = tmp_path / "duplex.yaml"
|
||||
deploy_path.write_text("duplex_session: {}\nstages: []\n", encoding="utf-8")
|
||||
|
||||
deploy = load_deploy_config(deploy_path)
|
||||
|
||||
assert isinstance(deploy.duplex_session, DuplexSessionRuntimeConfig)
|
||||
assert deploy.duplex_session.idle_ttl_s == 300.0
|
||||
assert deploy.duplex_session.disconnect_grace_s == 30.0
|
||||
assert deploy.duplex_session.reaper_interval_s == 5.0
|
||||
assert deploy.duplex_session.resume_replay_ttl_s == 60.0
|
||||
assert deploy.duplex_session.resume_replay_max_bytes_per_session == 8 * 1024 * 1024
|
||||
assert deploy.duplex_session.max_pending_input_bytes_per_session == 16 * 1024 * 1024
|
||||
assert deploy.duplex_session.max_pending_turns_per_session == 4
|
||||
assert deploy.duplex_session.max_sessions == 1
|
||||
assert deploy.duplex_session.completed_append_cache_size == 256
|
||||
with pytest.raises(FrozenInstanceError):
|
||||
deploy.duplex_session.idle_ttl_s = 1.0 # type: ignore[misc]
|
||||
|
||||
|
||||
def test_duplex_session_runtime_accepts_disabled_idle_expiry(tmp_path) -> None:
|
||||
deploy_path = tmp_path / "duplex.yaml"
|
||||
deploy_path.write_text("duplex_session:\n idle_ttl_s: null\nstages: []\n", encoding="utf-8")
|
||||
|
||||
deploy = load_deploy_config(deploy_path)
|
||||
|
||||
assert deploy.duplex_session.idle_ttl_s is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize(
|
||||
("name", "value"),
|
||||
[
|
||||
("idle_ttl_s", 0),
|
||||
("idle_ttl_s", -1),
|
||||
("disconnect_grace_s", 0),
|
||||
("reaper_interval_s", -1),
|
||||
("resume_replay_ttl_s", 0),
|
||||
("resume_replay_max_bytes_per_session", -1),
|
||||
("max_pending_input_bytes_per_session", 0),
|
||||
("max_pending_turns_per_session", -1),
|
||||
("max_sessions", 0),
|
||||
("completed_append_cache_size", 0),
|
||||
],
|
||||
)
|
||||
def test_duplex_session_runtime_rejects_non_positive_values(tmp_path, name: str, value: int) -> None:
|
||||
deploy_path = tmp_path / "duplex.yaml"
|
||||
deploy_path.write_text(f"duplex_session:\n {name}: {value}\nstages: []\n", encoding="utf-8")
|
||||
|
||||
with pytest.raises(ValueError, match=rf"duplex_session\.{name} must be positive"):
|
||||
load_deploy_config(deploy_path)
|
||||
@@ -0,0 +1,14 @@
|
||||
import pytest
|
||||
|
||||
from vllm_omni.experimental.fullduplex.engine.intermediate import get_stream_request_key
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
|
||||
|
||||
def test_get_stream_request_key_requires_stable_identifier():
|
||||
with pytest.raises(ValueError, match="stable request id"):
|
||||
get_stream_request_key({"ids": {"tts": [1, 2, 3]}})
|
||||
|
||||
|
||||
def test_get_stream_request_key_accepts_global_request_id():
|
||||
assert get_stream_request_key({"global_request_id": ["duplex-sid-stage0"]}) == "duplex-sid-stage0"
|
||||
@@ -0,0 +1,224 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from __future__ import annotations
|
||||
|
||||
from dataclasses import FrozenInstanceError
|
||||
|
||||
import pytest
|
||||
|
||||
from vllm_omni.experimental.fullduplex.engine.duplex_runtime import DuplexInputMode
|
||||
from vllm_omni.experimental.fullduplex.engine.duplex_session import DuplexSessionRuntimeManager
|
||||
from vllm_omni.experimental.fullduplex.engine.lease import (
|
||||
DuplexLeaseActivity,
|
||||
DuplexLeaseConfig,
|
||||
)
|
||||
from vllm_omni.experimental.fullduplex.engine.messages import DuplexFence
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
|
||||
|
||||
class FakeMonotonicClock:
|
||||
def __init__(self, value: float) -> None:
|
||||
self.value = value
|
||||
|
||||
def __call__(self) -> float:
|
||||
return self.value
|
||||
|
||||
def advance(self, seconds: float) -> None:
|
||||
self.value += seconds
|
||||
|
||||
|
||||
def _lease_config(*, idle_ttl_s: float | None = 300.0) -> DuplexLeaseConfig:
|
||||
return DuplexLeaseConfig(
|
||||
idle_ttl_s=idle_ttl_s,
|
||||
disconnect_grace_s=30.0,
|
||||
)
|
||||
|
||||
|
||||
def test_open_touch_detach_and_idle_expiry_use_monotonic_time() -> None:
|
||||
clock = FakeMonotonicClock(100.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence = DuplexFence("sid-expiry")
|
||||
session = manager.open_session(fence, lease_config=_lease_config())
|
||||
|
||||
assert session.lease.last_activity == 100.0
|
||||
clock.advance(10.0)
|
||||
session.touch(fence, DuplexLeaseActivity.HEARTBEAT)
|
||||
session.detach(fence)
|
||||
clock.advance(29.0)
|
||||
assert session.lease.disconnect_grace_expired(clock()) is False
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
clock.advance(272.0)
|
||||
expired = manager.collect_expired()
|
||||
|
||||
assert [item.session_id for item in expired] == ["sid-expiry"]
|
||||
assert expired[0].reason == "disconnect_grace_expired"
|
||||
assert manager.get("sid-expiry") is session
|
||||
manager.finalize_close_session(session)
|
||||
assert manager.get("sid-expiry") is None
|
||||
|
||||
|
||||
def test_detach_grace_and_resume_advance_token_independent_lease_generation() -> None:
|
||||
clock = FakeMonotonicClock(10.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence = DuplexFence("sid-resume")
|
||||
session = manager.open_session(fence, lease_config=_lease_config())
|
||||
|
||||
session.detach(fence)
|
||||
clock.advance(29.0)
|
||||
assert session.lease.disconnect_grace_expired(clock()) is False
|
||||
clock.advance(2.0)
|
||||
assert session.lease.disconnect_grace_expired(clock()) is True
|
||||
|
||||
generation = session.resume(fence, expected_lease_generation=0)
|
||||
|
||||
assert generation == 1
|
||||
assert session.lease.detached_at is None
|
||||
assert session.lease.disconnect_grace_expired(clock()) is False
|
||||
with pytest.raises(ValueError, match="lease generation mismatch"):
|
||||
session.resume(fence, expected_lease_generation=0)
|
||||
|
||||
|
||||
def test_active_operation_prevents_mid_transaction_expiry() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence = DuplexFence("sid-operation")
|
||||
session = manager.open_session(fence, lease_config=_lease_config(idle_ttl_s=5.0))
|
||||
|
||||
session.begin_operation(fence, "append-1")
|
||||
clock.advance(10.0)
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
session.end_operation(fence, "append-1")
|
||||
assert manager.collect_expired() == []
|
||||
clock.advance(6.0)
|
||||
assert [item.session_id for item in manager.collect_expired()] == ["sid-operation"]
|
||||
|
||||
|
||||
def test_sessions_have_independent_activity_deadlines_and_resources() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence_a = DuplexFence("sid-a")
|
||||
fence_b = DuplexFence("sid-b")
|
||||
session_a = manager.open_session(fence_a, lease_config=_lease_config(idle_ttl_s=10.0))
|
||||
session_b = manager.open_session(fence_b, lease_config=_lease_config(idle_ttl_s=10.0))
|
||||
session_a.reserve_stage_request(0, "req-a-reserved", fence=fence_a)
|
||||
session_a.bind_stage_request(1, "req-a-submitted", fence=fence_a)
|
||||
session_b.bind_stage_request(0, "req-b", fence=fence_b)
|
||||
|
||||
clock.advance(6.0)
|
||||
session_b.touch(fence_b, DuplexLeaseActivity.MODEL_OUTPUT)
|
||||
clock.advance(5.0)
|
||||
expired = manager.collect_expired()
|
||||
|
||||
assert len(expired) == 1
|
||||
assert expired[0].session_id == "sid-a"
|
||||
assert expired[0].reserved_request_ids == ("req-a-reserved",)
|
||||
assert expired[0].submitted_request_ids == ("req-a-submitted",)
|
||||
assert manager.get("sid-b") is session_b
|
||||
assert session_b.resource_request_ids() == ["req-b"]
|
||||
|
||||
|
||||
def test_close_and_reaper_select_exactly_one_terminal_transition() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence = DuplexFence("sid-race")
|
||||
manager.open_session(fence, lease_config=_lease_config(idle_ttl_s=1.0))
|
||||
clock.advance(2.0)
|
||||
|
||||
assert len(manager.collect_expired()) == 1
|
||||
assert manager.close_session(fence, reason="explicit_close") is None
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
fence_2 = DuplexFence("sid-race-2")
|
||||
manager.open_session(fence_2, lease_config=_lease_config(idle_ttl_s=1.0))
|
||||
assert manager.close_session(fence_2, reason="explicit_close") is not None
|
||||
clock.advance(2.0)
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
|
||||
def test_stale_fence_cannot_touch_detach_or_resume_lease() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
current = DuplexFence("sid-fence", epoch=1)
|
||||
stale = DuplexFence("sid-fence", epoch=0)
|
||||
session = manager.open_session(current, lease_config=_lease_config())
|
||||
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
session.touch(stale, DuplexLeaseActivity.APPEND)
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
session.detach(stale)
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
session.resume(stale, expected_lease_generation=0)
|
||||
|
||||
|
||||
def test_disabled_idle_expiry_never_collects_session() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
manager.open_session(DuplexFence("sid-no-expiry"), lease_config=_lease_config(idle_ttl_s=None))
|
||||
|
||||
clock.advance(1_000_000.0)
|
||||
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
|
||||
def test_detached_session_expires_at_disconnect_grace_when_idle_ttl_is_disabled() -> None:
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
fence = DuplexFence("sid-disconnect-grace")
|
||||
session = manager.open_session(fence, lease_config=_lease_config(idle_ttl_s=None))
|
||||
|
||||
session.detach(fence)
|
||||
clock.advance(29.0)
|
||||
assert manager.collect_expired() == []
|
||||
|
||||
clock.advance(1.0)
|
||||
expired = manager.collect_expired()
|
||||
|
||||
assert [item.session_id for item in expired] == [fence.session_id]
|
||||
assert expired[0].reason == "disconnect_grace_expired"
|
||||
|
||||
|
||||
def test_runtime_manager_enforces_server_owned_session_admission() -> None:
|
||||
manager = DuplexSessionRuntimeManager(max_sessions=2)
|
||||
manager.open_session(DuplexFence("sid-admission-a"))
|
||||
manager.open_session(DuplexFence("sid-admission-b"))
|
||||
|
||||
with pytest.raises(RuntimeError, match="duplex_session_capacity_exhausted"):
|
||||
manager.open_session(DuplexFence("sid-admission-c"))
|
||||
|
||||
|
||||
def test_completed_append_cache_is_bounded() -> None:
|
||||
manager = DuplexSessionRuntimeManager(completed_append_limit=2)
|
||||
fence = DuplexFence("sid-completed-cache")
|
||||
session = manager.open_session(fence)
|
||||
|
||||
for index in range(3):
|
||||
session.record_completed_append(
|
||||
f"operation-{index}",
|
||||
fence=fence,
|
||||
mode=DuplexInputMode.TURN_COMMIT_ONLY,
|
||||
final=True,
|
||||
stage_results=[{"index": index}],
|
||||
)
|
||||
|
||||
assert list(session.completed_appends) == ["operation-1", "operation-2"]
|
||||
|
||||
session.accept_fence(DuplexFence(fence.session_id, epoch=1))
|
||||
assert session.completed_appends == {}
|
||||
|
||||
|
||||
def test_lease_config_and_expiry_record_are_immutable() -> None:
|
||||
config = _lease_config()
|
||||
with pytest.raises(FrozenInstanceError):
|
||||
config.idle_ttl_s = 1.0 # type: ignore[misc]
|
||||
|
||||
clock = FakeMonotonicClock(0.0)
|
||||
manager = DuplexSessionRuntimeManager(clock=clock)
|
||||
manager.open_session(DuplexFence("sid-record"), lease_config=_lease_config(idle_ttl_s=1.0))
|
||||
clock.advance(2.0)
|
||||
record = manager.collect_expired()[0]
|
||||
with pytest.raises(FrozenInstanceError):
|
||||
record.reason = "changed" # type: ignore[misc]
|
||||
@@ -0,0 +1,480 @@
|
||||
from dataclasses import FrozenInstanceError
|
||||
from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
from vllm.sampling_params import SamplingParams
|
||||
|
||||
from vllm_omni.experimental.fullduplex.engine import duplex_runtime
|
||||
from vllm_omni.experimental.fullduplex.engine.contracts import (
|
||||
duplex_data_plane_request_info,
|
||||
duplex_resource_request_belongs_to_session,
|
||||
duplex_resource_request_id,
|
||||
)
|
||||
from vllm_omni.experimental.fullduplex.engine.duplex_runtime import (
|
||||
DuplexInputMode,
|
||||
DuplexOutputAction,
|
||||
DuplexOutputDecision,
|
||||
DuplexRuntimeCapabilities,
|
||||
DuplexSessionRuntimeManager,
|
||||
)
|
||||
from vllm_omni.experimental.fullduplex.engine.messages import DuplexFence
|
||||
from vllm_omni.experimental.fullduplex.minicpmo45.runtime import (
|
||||
MiniCPMO45DuplexRuntimeExtension,
|
||||
build_duplex_data_plane_prompt,
|
||||
duplex_scheduler_token_budget,
|
||||
)
|
||||
|
||||
pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
|
||||
|
||||
|
||||
def test_duplex_fence_is_immutable():
|
||||
fence = DuplexFence("session")
|
||||
|
||||
with pytest.raises(FrozenInstanceError):
|
||||
fence.epoch = 1 # type: ignore[misc]
|
||||
|
||||
assert not hasattr(fence, "__dict__")
|
||||
|
||||
|
||||
def test_duplex_runtime_tracks_stage_bindings_and_barge_in_epoch():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
session = manager.open_session(
|
||||
DuplexFence("sid-1"),
|
||||
capabilities=DuplexRuntimeCapabilities(
|
||||
input_modes={DuplexInputMode.APPEND_TOKENS},
|
||||
),
|
||||
)
|
||||
session.bind_stage_request(stage_id=0, request_id="req-stage0", fence=session.fence)
|
||||
session.bind_stage_request(stage_id=1, request_id="req-stage1", fence=session.fence)
|
||||
|
||||
update = session.append_input(mode=DuplexInputMode.APPEND_TOKENS, fence=session.fence)
|
||||
next_fence = DuplexFence("sid-1", epoch=1)
|
||||
stale_request_ids = session.release_fence(session.fence)
|
||||
session.accept_fence(next_fence)
|
||||
|
||||
assert update.seq == 1
|
||||
assert session.fence == next_fence
|
||||
assert stale_request_ids == ["req-stage0", "req-stage1"]
|
||||
assert session.stage_bindings == {}
|
||||
|
||||
|
||||
def test_duplex_runtime_tracks_same_request_id_for_each_pipeline_stage():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
fence = DuplexFence("sid-shared-pipeline-request")
|
||||
session = manager.open_session(fence)
|
||||
|
||||
session.reserve_stage_request(0, "req-shared", fence=fence)
|
||||
session.bind_stage_request(0, "req-shared", fence=fence)
|
||||
session.bind_stage_request(1, "req-shared", fence=fence)
|
||||
|
||||
assert session.stage_bindings[0].request_id == "req-shared"
|
||||
assert session.stage_bindings[1].request_id == "req-shared"
|
||||
assert session.resource_request_ids() == ["req-shared"]
|
||||
assert session.input_seq == 0
|
||||
|
||||
|
||||
def test_duplex_runtime_extension_validation_rejects_missing_methods():
|
||||
class IncompleteExtension:
|
||||
def configure_sampling_params(self, *, runtime_config, defaults):
|
||||
del runtime_config
|
||||
return defaults
|
||||
|
||||
with pytest.raises(TypeError, match="plan_append"):
|
||||
duplex_runtime.validate_duplex_runtime_extension(IncompleteExtension())
|
||||
|
||||
|
||||
def test_duplex_runtime_extension_validation_rejects_stage_count_mismatch():
|
||||
class WrongStageCountExtension(MiniCPMO45DuplexRuntimeExtension):
|
||||
def configure_sampling_params(self, *, runtime_config, defaults):
|
||||
del runtime_config
|
||||
return defaults[:1]
|
||||
|
||||
with pytest.raises(ValueError, match="one sampling parameter per stage"):
|
||||
duplex_runtime.validate_duplex_runtime_extension(
|
||||
WrongStageCountExtension(),
|
||||
sampling_defaults=(SamplingParams(), SamplingParams()),
|
||||
)
|
||||
|
||||
|
||||
def test_duplex_runtime_extension_validation_rejects_sampling_type_mismatch():
|
||||
class WrongSamplingTypeExtension(MiniCPMO45DuplexRuntimeExtension):
|
||||
def configure_sampling_params(self, *, runtime_config, defaults):
|
||||
del runtime_config
|
||||
return tuple(object() for _ in defaults)
|
||||
|
||||
with pytest.raises(TypeError, match="sampling parameter type"):
|
||||
duplex_runtime.validate_duplex_runtime_extension(
|
||||
WrongSamplingTypeExtension(),
|
||||
sampling_defaults=(SamplingParams(), SamplingParams()),
|
||||
)
|
||||
|
||||
|
||||
def test_duplex_output_decision_metadata_is_immutable():
|
||||
decision = DuplexOutputDecision(
|
||||
action=DuplexOutputAction.DIRECT_RESPONSE,
|
||||
metadata={"model_listen": True},
|
||||
)
|
||||
|
||||
with pytest.raises(TypeError):
|
||||
decision.metadata["model_listen"] = False
|
||||
|
||||
|
||||
def _decide_minicpmo_output(
|
||||
output: object,
|
||||
*,
|
||||
segment_token_ids: tuple[int, ...] = (),
|
||||
segment_output_metadata: dict | None = None,
|
||||
):
|
||||
return MiniCPMO45DuplexRuntimeExtension().decide_output(
|
||||
stage_id=0,
|
||||
final_stage_id=1,
|
||||
segment_finished=True,
|
||||
segment_token_ids=segment_token_ids,
|
||||
segment_output_metadata=segment_output_metadata or {},
|
||||
output=output,
|
||||
)
|
||||
|
||||
|
||||
def test_minicpmo_extension_owns_stage_sampling_overrides():
|
||||
defaults = (
|
||||
SamplingParams(max_tokens=4),
|
||||
SamplingParams(max_tokens=8),
|
||||
)
|
||||
|
||||
configured = MiniCPMO45DuplexRuntimeExtension().configure_sampling_params(
|
||||
runtime_config={
|
||||
"duplex_stage_max_tokens": {"0": 20},
|
||||
"duplex_stage_sampling_params": {"1": {"stop_token_ids": [151645]}},
|
||||
},
|
||||
defaults=defaults,
|
||||
)
|
||||
|
||||
assert configured[0].max_tokens == 20
|
||||
assert configured[1].stop_token_ids == [151645]
|
||||
assert defaults[0].max_tokens == 4
|
||||
assert 151645 not in (defaults[1].stop_token_ids or [])
|
||||
|
||||
|
||||
def test_minicpmo_output_decision_uses_raw_streaming_token_snapshot():
|
||||
decision = _decide_minicpmo_output(
|
||||
SimpleNamespace(outputs=[SimpleNamespace()]),
|
||||
segment_token_ids=(151705,),
|
||||
segment_output_metadata={"special_token_ids": {"listen_token_id": 151705}},
|
||||
)
|
||||
|
||||
assert decision is not None
|
||||
assert decision.action is DuplexOutputAction.DIRECT_RESPONSE
|
||||
assert decision.metadata["duplex_native_decision"] == "listen"
|
||||
assert decision.metadata["model_listen"] is True
|
||||
|
||||
|
||||
@pytest.mark.parametrize("attr", ["token_ids", "cumulative_token_ids"])
|
||||
def test_minicpmo_output_decision_ignores_output_level_token_history(attr):
|
||||
output = SimpleNamespace(
|
||||
multimodal_output={"special_token_ids": {"listen_token_id": 151705}},
|
||||
outputs=[SimpleNamespace()],
|
||||
**{attr: [42, 151705]},
|
||||
)
|
||||
|
||||
assert _decide_minicpmo_output(output) is None
|
||||
|
||||
|
||||
@pytest.mark.parametrize("attr", ["token_ids", "cumulative_token_ids"])
|
||||
def test_minicpmo_output_decision_uses_completion_token_ids(attr):
|
||||
output = SimpleNamespace(
|
||||
multimodal_output={"special_token_ids": {"listen_token_id": 151705}},
|
||||
outputs=[SimpleNamespace(**{attr: [42, 151705]})],
|
||||
)
|
||||
|
||||
assert _decide_minicpmo_output(output) is not None
|
||||
|
||||
|
||||
def test_minicpmo_output_decision_uses_completion_stop_reason():
|
||||
output = SimpleNamespace(
|
||||
multimodal_output={"special_token_ids": {"listen_token_id": 151705}},
|
||||
outputs=[SimpleNamespace(stop_reason=151705)],
|
||||
)
|
||||
|
||||
assert _decide_minicpmo_output(output) is not None
|
||||
|
||||
|
||||
def test_duplex_runtime_cancel_fence_rejects_late_append_and_accepts_next_epoch():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
cancelled_fence = DuplexFence("sid-cancel-race")
|
||||
next_fence = DuplexFence("sid-cancel-race", epoch=1)
|
||||
session = manager.open_session(
|
||||
cancelled_fence,
|
||||
capabilities=DuplexRuntimeCapabilities(
|
||||
input_modes={DuplexInputMode.APPEND_AUDIO_CHUNK},
|
||||
),
|
||||
)
|
||||
session.bind_stage_request(
|
||||
stage_id=0,
|
||||
request_id="req-cancelled",
|
||||
fence=cancelled_fence,
|
||||
)
|
||||
|
||||
stale_request_ids = session.cancel_fence(cancelled_fence, next_fence)
|
||||
|
||||
assert stale_request_ids == ["req-cancelled"]
|
||||
assert session.fence == next_fence
|
||||
assert session.stage_bindings == {}
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=cancelled_fence,
|
||||
)
|
||||
update = session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=next_fence,
|
||||
)
|
||||
assert update.seq == 1
|
||||
|
||||
|
||||
def test_duplex_runtime_stale_close_preserves_live_session_and_bindings():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
current_fence = DuplexFence("sid-stale-close", epoch=1)
|
||||
session = manager.open_session(
|
||||
current_fence,
|
||||
capabilities=DuplexRuntimeCapabilities(
|
||||
input_modes={DuplexInputMode.APPEND_AUDIO_CHUNK},
|
||||
),
|
||||
)
|
||||
session.bind_stage_request(0, "req-live", fence=current_fence)
|
||||
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
manager.close_session(DuplexFence("sid-stale-close"))
|
||||
|
||||
assert manager.get("sid-stale-close") is session
|
||||
assert session.stage_request_ids() == ["req-live"]
|
||||
|
||||
|
||||
def test_duplex_runtime_reopen_rejects_late_append_from_old_incarnation():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
old_fence = DuplexFence("sid-reopen", incarnation=0)
|
||||
old_session = manager.open_session(
|
||||
old_fence,
|
||||
capabilities=DuplexRuntimeCapabilities(
|
||||
input_modes={DuplexInputMode.APPEND_AUDIO_CHUNK},
|
||||
),
|
||||
)
|
||||
manager.close_session(old_fence)
|
||||
|
||||
new_fence = DuplexFence("sid-reopen", incarnation=1)
|
||||
new_session = manager.open_session(
|
||||
new_fence,
|
||||
capabilities=old_session.capabilities,
|
||||
)
|
||||
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
new_session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=old_fence,
|
||||
)
|
||||
assert (
|
||||
new_session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=new_fence,
|
||||
).seq
|
||||
== 1
|
||||
)
|
||||
|
||||
|
||||
def test_duplex_prompt_expands_incarnation_metadata():
|
||||
fence = DuplexFence("sid-incarnation", incarnation=3)
|
||||
|
||||
prompt = build_duplex_data_plane_prompt(
|
||||
request_id=duplex_resource_request_id(fence, "stage0"),
|
||||
fence=fence,
|
||||
session_config={},
|
||||
runtime_config={},
|
||||
seq=1,
|
||||
turn_seq=1,
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
payload={"is_speech": True},
|
||||
final=False,
|
||||
)
|
||||
|
||||
assert prompt["model_intermediate_buffer"]["duplex"]["incarnation"] == 3
|
||||
|
||||
|
||||
def test_duplex_runtime_tracks_turn_local_append_sequence():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
session = manager.open_session(
|
||||
DuplexFence("sid-turn-seq"),
|
||||
capabilities=DuplexRuntimeCapabilities(
|
||||
input_modes={DuplexInputMode.APPEND_AUDIO_CHUNK},
|
||||
),
|
||||
)
|
||||
|
||||
first = session.append_input(mode=DuplexInputMode.APPEND_AUDIO_CHUNK, fence=session.fence)
|
||||
second = session.append_input(mode=DuplexInputMode.APPEND_AUDIO_CHUNK, fence=session.fence)
|
||||
next_turn = DuplexFence("sid-turn-seq", turn_id=1, response_seq=1)
|
||||
third = session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=next_turn,
|
||||
)
|
||||
fourth = session.append_input(
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
fence=next_turn,
|
||||
)
|
||||
|
||||
assert [first.seq, second.seq, third.seq, fourth.seq] == [1, 2, 3, 4]
|
||||
assert [first.turn_seq, second.turn_seq, third.turn_seq, fourth.turn_seq] == [1, 2, 1, 2]
|
||||
assert [first.turn_id, second.turn_id, third.turn_id, fourth.turn_id] == [0, 0, 1, 1]
|
||||
|
||||
|
||||
def test_duplex_runtime_rejects_unsupported_append_mode():
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
session = manager.open_session(
|
||||
DuplexFence("sid-2"),
|
||||
capabilities=DuplexRuntimeCapabilities(input_modes={DuplexInputMode.TURN_COMMIT_ONLY}),
|
||||
)
|
||||
|
||||
with pytest.raises(ValueError, match="not supported"):
|
||||
session.append_input(mode=DuplexInputMode.APPEND_TOKENS, fence=session.fence)
|
||||
|
||||
|
||||
def test_duplex_data_plane_request_info_extracts_structured_stage_result():
|
||||
request_id, response_stage_id = duplex_data_plane_request_info(
|
||||
{
|
||||
"stage_results": [
|
||||
{"result": {"supported": True}},
|
||||
{
|
||||
"result": {
|
||||
"data_plane_append": True,
|
||||
"request_id": "duplex-sid-e0-stage0-s1",
|
||||
"response_stage_id": 1,
|
||||
}
|
||||
},
|
||||
]
|
||||
}
|
||||
)
|
||||
|
||||
assert request_id == "duplex-sid-e0-stage0-s1"
|
||||
assert response_stage_id == 1
|
||||
|
||||
|
||||
def test_duplex_data_plane_request_info_rejects_missing_request_id():
|
||||
assert duplex_data_plane_request_info(
|
||||
{
|
||||
"stage_results": [
|
||||
{
|
||||
"result": {
|
||||
"data_plane_append": True,
|
||||
"request_id": "",
|
||||
"response_stage_id": 1,
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
) == (None, None)
|
||||
|
||||
|
||||
def test_duplex_scheduler_token_budget_estimates_pcm_slots():
|
||||
assert (
|
||||
duplex_scheduler_token_budget(
|
||||
{
|
||||
"audio": "AAAAAA==",
|
||||
"format": "pcm_f32le",
|
||||
"sample_rate_hz": 16000,
|
||||
}
|
||||
)
|
||||
== 16
|
||||
)
|
||||
|
||||
|
||||
def test_duplex_scheduler_token_budget_ignores_client_budget_fields():
|
||||
assert (
|
||||
duplex_scheduler_token_budget(
|
||||
{
|
||||
"audio": "AAAAAA==",
|
||||
"format": "pcm_f32le",
|
||||
"duplex_num_input_tokens": 999,
|
||||
"num_input_tokens": 999,
|
||||
}
|
||||
)
|
||||
== 16
|
||||
)
|
||||
|
||||
|
||||
def test_resource_state_rejects_fence_regression_and_requires_explicit_fence():
|
||||
current = DuplexFence("sid", epoch=2, turn_id=3, response_seq=4)
|
||||
manager = DuplexSessionRuntimeManager()
|
||||
session = manager.open_session(
|
||||
current,
|
||||
capabilities=DuplexRuntimeCapabilities(input_modes={DuplexInputMode.APPEND_AUDIO_CHUNK}),
|
||||
)
|
||||
|
||||
for stale in (
|
||||
DuplexFence("sid", epoch=1, turn_id=99, response_seq=99),
|
||||
DuplexFence("sid", epoch=2, turn_id=2, response_seq=4),
|
||||
DuplexFence("sid", epoch=2, turn_id=3, response_seq=3),
|
||||
):
|
||||
with pytest.raises(RuntimeError, match="fence mismatch"):
|
||||
session.accept_fence(stale)
|
||||
assert session.fence == current
|
||||
|
||||
with pytest.raises(TypeError, match="fence"):
|
||||
session.bind_stage_request(0, "request")
|
||||
with pytest.raises(TypeError, match="fence"):
|
||||
session.append_input(mode=DuplexInputMode.APPEND_AUDIO_CHUNK)
|
||||
with pytest.raises(TypeError, match="DuplexFence"):
|
||||
manager.open_session("legacy-session")
|
||||
with pytest.raises(TypeError, match="DuplexFence"):
|
||||
manager.close_session("legacy-session")
|
||||
|
||||
|
||||
def test_resource_request_id_is_derived_from_fence_and_role():
|
||||
fence = DuplexFence("sid-with-dashes", epoch=7, turn_id=11, response_seq=13)
|
||||
|
||||
assert duplex_resource_request_id(fence, "stage0") == "duplex-s.c2lkLXdpdGgtZGFzaGVz.i.0.e.7.r.stage0"
|
||||
assert duplex_resource_request_id(fence, "stage1") == "duplex-s.c2lkLXdpdGgtZGFzaGVz.i.0.e.7.r.stage1"
|
||||
|
||||
|
||||
def test_resource_request_id_codec_separates_session_id_from_incarnation():
|
||||
embedded_incarnation = duplex_resource_request_id(
|
||||
DuplexFence("foo-i1", incarnation=0),
|
||||
"stage0",
|
||||
)
|
||||
actual_incarnation = duplex_resource_request_id(
|
||||
DuplexFence("foo", incarnation=1),
|
||||
"stage0",
|
||||
)
|
||||
|
||||
assert embedded_incarnation != actual_incarnation
|
||||
|
||||
|
||||
def test_resource_request_id_session_membership_uses_encoded_identity():
|
||||
request_id = duplex_resource_request_id(
|
||||
DuplexFence("sid-with-dashes", incarnation=2, epoch=7),
|
||||
"stage0",
|
||||
)
|
||||
|
||||
assert duplex_resource_request_belongs_to_session(request_id, "sid-with-dashes") is True
|
||||
assert duplex_resource_request_belongs_to_session(request_id, "sid") is False
|
||||
assert duplex_resource_request_belongs_to_session("duplex-s.invalid.i.x.e.7.r.stage0", "sid") is False
|
||||
|
||||
|
||||
def test_placeholder_budget_is_planned_inside_omni_engine_boundary():
|
||||
fence = DuplexFence("sid", turn_id=1, response_seq=1)
|
||||
prompt = build_duplex_data_plane_prompt(
|
||||
request_id=duplex_resource_request_id(fence, "stage0"),
|
||||
fence=fence,
|
||||
session_config={},
|
||||
runtime_config={},
|
||||
seq=2,
|
||||
turn_seq=1,
|
||||
mode=DuplexInputMode.APPEND_AUDIO_CHUNK,
|
||||
payload={
|
||||
"audio": "AAAAAA==",
|
||||
"format": "pcm_f32le",
|
||||
"duplex_num_input_tokens": 999,
|
||||
"num_input_tokens": 999,
|
||||
},
|
||||
final=False,
|
||||
)
|
||||
|
||||
assert len(prompt["prompt_token_ids"]) == 16
|
||||
assert prompt["model_intermediate_buffer"]["duplex"]["fence"] == fence
|
||||
assert prompt["model_intermediate_buffer"]["duplex"]["scheduler_token_budget"] == 16
|
||||
Reference in new issue
Block a user