Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

优化Train相关的逻辑 (#3051)

DuTao85c510dda4fe
diff --git a/.gitignore b/.gitignore
index 316c6cd41..12d855a48 100644
--- a/.gitignore
+++ b/.gitignore
@@ -59,6 +59,7 @@ reports/
test-results.xml
test-reports/
report.html
+temp/
# Virtual environments
.env
diff --git a/openviking/session/train/components/dataset_service.py b/openviking/session/train/components/dataset_service.py
index 0016587dd..5ee77960e 100644
--- a/openviking/session/train/components/dataset_service.py
+++ b/openviking/session/train/components/dataset_service.py
@@ -640,13 +640,23 @@ def rollout_from_dict(data: dict[str, Any]) -> Rollout:
return Rollout(
case=case_from_dict(data["case"]),
- messages=[Message.from_dict(item) for item in data.get("messages", [])],
+ messages=[
+ Message.from_dict(_message_dict_with_defaults(item, index))
+ for index, item in enumerate(data.get("messages", []))
+ ],
policy_snapshot_id=data["policy_snapshot_id"],
evaluation=evaluation_from_dict(data.get("evaluation")),
metadata=dict(data.get("metadata") or {}),
)
+def _message_dict_with_defaults(data: dict[str, Any], index: int) -> dict[str, Any]:
+ """Accept lightweight remote message payloads that omit OpenViking-only ids."""
+ item = dict(data)
+ item.setdefault("id", f"remote_message_{index}")
+ return item
+
+
def evaluation_from_dict(data: dict[str, Any] | None) -> RubricEvaluation | None:
if data is None:
return None
diff --git a/openviking/session/train/components/remote.py b/openviking/session/train/components/remote.py
index bf2bba117..0b9555fe4 100644
--- a/openviking/session/train/components/remote.py
+++ b/openviking/session/train/components/remote.py
@@ -213,6 +213,7 @@ class RemoteRolloutExecutor:
transient_errors = 0
missing_execution_errors = 0
last_transient_error: BaseException | None = None
+ last_transient_response: httpx.Response | None = None
while True:
try:
response = await client.get(f"/v1/rollouts/executions/{execution_id}")
@@ -244,6 +245,17 @@ class RemoteRolloutExecutor:
min(self.poll_interval_seconds * missing_execution_errors, 10.0)
)
continue
+ if _is_retryable_poll_response(response):
+ transient_errors += 1
+ last_transient_response = response
+ if loop.time() >= deadline:
+ raise TimeoutError(
+ f"rollout execution {execution_id} polling timed out for case {case.name} "
+ f"after {self.execution_timeout_seconds}s; last polling response: "
+ f"{response.status_code} {_response_text(response)}"
+ )
+ await asyncio.sleep(min(self.poll_interval_seconds * transient_errors, 10.0))
+ continue
response.raise_for_status()
data = response.json()
status = data.get("status")
@@ -262,7 +274,12 @@ class RemoteRolloutExecutor:
f"; last polling error: {type(last_transient_error).__name__}: "
f"{last_transient_error}"
if last_transient_error is not None
- else ""
+ else (
+ f"; last polling response: {last_transient_response.status_code} "
+ f"{_response_text(last_transient_response)}"
+ if last_transient_response is not None
+ else ""
+ )
)
raise TimeoutError(
f"rollout execution {execution_id} timed out for case {case.name} "
@@ -303,6 +320,10 @@ def _require_execution_id(data: dict[str, Any], *, case: Case) -> str:
return execution_id
+def _is_retryable_poll_response(response: httpx.Response) -> bool:
+ return response.status_code == 408 or response.status_code == 429 or response.status_code >= 500
+
+
def _response_text(response: httpx.Response, *, max_chars: int = 500) -> str:
try:
text = response.text
diff --git a/openviking/session/train/pipeline.py b/openviking/session/train/pipeline.py
index f23c87fb0..a14e55b62 100644
--- a/openviking/session/train/pipeline.py
+++ b/openviking/session/train/pipeline.py
@@ -81,7 +81,8 @@ class OfflinePolicyOptimizationPipeline:
context: PipelineContext | Any,
) -> PipelineResult:
ctx = context if isinstance(context, PipelineContext) else PipelineContext()
- max_epochs = max(1, int(ctx.max_epochs or 1))
+ requested_epochs = ctx.max_epochs if ctx.max_epochs is not None else 1
+ max_epochs = max(0, int(requested_epochs))
case_loader = _train_case_loader(case_loader, ctx)
current_policy_set = policy_set
epoch_results: list[PipelineEpochResult] = []
diff --git a/openviking/session/train/run_batch_train_eval.py b/openviking/session/train/run_batch_train_eval.py
index 1f40cadb4..bb6c6fd41 100644
--- a/openviking/session/train/run_batch_train_eval.py
+++ b/openviking/session/train/run_batch_train_eval.py
@@ -36,6 +36,12 @@ def parse_args() -> argparse.Namespace:
default=200,
help="Concurrent OpenViking session.commit submissions during train (default: 200)",
)
+ parser.add_argument(
+ "--commit-timeout-seconds",
+ type=float,
+ default=None,
+ help="Max seconds to wait for each OpenViking session.commit task. Default waits indefinitely.",
+ )
parser.add_argument("--config", default=None, help="ov.conf path (optional)")
parser.add_argument("--server-url", default=None, help="OpenViking server URL. Defaults to ov.conf/ovcli.conf")
parser.add_argument("--api-key", default=None, help="OpenViking API key. Defaults to ov.conf/ovcli.conf")
@@ -188,6 +194,7 @@ async def main_async() -> int:
batch_size=args.batch_size,
concurrency=args.concurrency,
commit_concurrency=args.commit_concurrency,
+ commit_timeout_seconds=args.commit_timeout_seconds,
config_path=str(Path(args.config).expanduser()) if args.config else None,
server_url=args.server_url,
api_key=args.api_key,
diff --git a/tests/session/train/test_train_framework.py b/tests/session/train/test_train_framework.py
index 82685e196..e4586dbe4 100644
--- a/tests/session/train/test_train_framework.py
+++ b/tests/session/train/test_train_framework.py
@@ -351,6 +351,34 @@ async def test_default_policy_optimization_pipeline_runs_one_batch():
assert result.epochs[0].policy_snapshot_ids == ["snapshot-1"]
+async def test_policy_optimization_pipeline_allows_zero_epochs_without_training():
+ snapshotter = DummySnapshotter()
+ pipeline = OfflinePolicyOptimizationPipeline(
+ snapshotter=snapshotter,
+ rollout_executor=DummyExecutor(),
+ rollout_analyzer=DummyAnalyzer(),
+ gradient_estimator=DummyEstimator(),
+ policy_optimizer=DummyOptimizer(),
+ policy_updater=DummyUpdater(),
+ )
+
+ initial_policy_set = _policy_set()
+ result = await pipeline.train(
+ case_loader=ListCaseLoader([_case()]),
+ policy_set=initial_policy_set,
+ context=PipelineContext(max_epochs=0),
+ )
+
+ assert result.analyses == []
+ assert result.gradients == []
+ assert result.epochs == []
+ assert result.evaluation_passes == []
+ assert result.apply_result.updated_policy_set is initial_policy_set
+ assert result.metadata["max_epochs"] == 0
+ assert result.metadata["completed_epochs"] == 0
+
+
@pytest.mark.asyncio
async def test_offline_policy_optimization_pipeline_supports_train_and_eval():
snapshotter = DummySnapshotter()
diff --git a/tests/unit/test_remote_rollout_executor.py b/tests/unit/test_remote_rollout_executor.py
index e987e736d..5bd776238 100644
--- a/tests/unit/test_remote_rollout_executor.py
+++ b/tests/unit/test_remote_rollout_executor.py
@@ -128,3 +128,80 @@ def test_remote_rollout_executor_retries_transient_missing_execution(monkeypatch
assert len(rollouts) == 1
assert rollouts[0].case.name == "case-1"
assert calls.count(f"GET /v1/rollouts/executions/{execution_id}") == 2
+
+
+def test_remote_rollout_executor_retries_transient_poll_5xx(monkeypatch):
+ calls: list[str] = []
+ execution_id = "rollout_exec_gateway_blip"
+
+ def handler(request: httpx.Request) -> httpx.Response:
+ calls.append(f"{request.method} {request.url.path}")
+ if request.method == "POST" and request.url.path == "/v1/rollouts/execute":
+ return httpx.Response(200, json={"execution_id": execution_id, "status": "running"})
+ if request.method == "GET" and request.url.path.endswith(execution_id):
+ get_count = sum(1 for call in calls if call.startswith("GET "))
+ if get_count == 1:
+ return httpx.Response(502, text="Bad Gateway")
+ return httpx.Response(
+ 200,
+ json={
+ "execution_id": execution_id,
+ "status": "completed",
+ "rollout": {
+ "case": {
+ "name": "case-1",
+ "task_signature": "booking_duplicate",
+ "input": {"user_request": "cancel duplicate booking"},
+ "rubric": {
+ "name": "booking_rubric",
+ "description": "Cancel only the verified duplicate booking.",
+ "criteria": [
+ {
+ "name": "verify_duplicate",
+ "description": "Verify duplicate status first.",
+ "required": True,
+ "weight": 1.0,
+ "metadata": {},
+ }
+ ],
+ "metadata": {},
+ },
+ "metadata": {},
+ },
+ "messages": [],
+ "policy_snapshot_id": "snapshot-1",
+ "evaluation": None,
+ "metadata": {},
+ },
+ },
+ )
+ return httpx.Response(500, json={"error": "unexpected request"})
+
+ original_async_client = httpx.AsyncClient
+ monkeypatch.setattr(
+ httpx,
+ "AsyncClient",
+ lambda *args, **kwargs: original_async_client(
+ transport=httpx.MockTransport(handler),
+ base_url=kwargs.get("base_url"),
+ timeout=kwargs.get("timeout"),
+ ),
+ )
+
+ executor = RemoteRolloutExecutor(
+ service_url="http://rollout-service",
+ poll_interval_seconds=0.01,
+ execution_timeout_seconds=1.0,
+ )
+
+ rollouts = asyncio.run(
+ executor.execute(
+ [_case()],
+ _policy_set(),
+ ExecutionContext(policy_snapshot_id="snapshot-1"),
+ )
+ )
+
+ assert len(rollouts) == 1
+ assert rollouts[0].case.name == "case-1"
+ assert calls.count(f"GET /v1/rollouts/executions/{execution_id}") == 2