verifiers 0.3.2.dev146__tar.gz → 0.3.2.dev148__tar.gz

{verifiers-0.3.2.dev146 → verifiers-0.3.2.dev148}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: verifiers
-Version: 0.3.2.dev146
+Version: 0.3.2.dev148
 Summary: Verifiers: Environments for LLM Reinforcement Learning
 Project-URL: Homepage, https://github.com/primeintellect-ai/verifiers
 Project-URL: Documentation, https://github.com/primeintellect-ai/verifiers

{verifiers-0.3.2.dev146 → verifiers-0.3.2.dev148}/verifiers/v1/agent.py

@@ -57,22 +57,13 @@ __all__ = ["Agent", "AgentConfig", "Agents", "TimeoutConfig", "make_agent"]
 logger = logging.getLogger(__name__)
 
 
-DEFAULT_ROLLOUT_TIMEOUT = 4 * 3600.0
-"""Agent solve-attempt budget when neither the eval config nor the task sets one."""
-
-
 def resolve_rollout_timeouts(timeout: TimeoutConfig, task: Task) -> RolloutTimeouts:
     """Apply an agent's stage-timeout precedence to one task."""
-    agent_timeout = (
-        timeout.rollout if timeout.rollout is not None else task.data.timeout.agent
-    )
-    if agent_timeout is None:
-        agent_timeout = DEFAULT_ROLLOUT_TIMEOUT
-    elif agent_timeout == 0:
-        agent_timeout = None  # explicit: unbounded
     return RolloutTimeouts(
         setup=timeout.setup if timeout.setup is not None else task.data.timeout.setup,
-        agent=agent_timeout,
+        agent=(
+            timeout.rollout if timeout.rollout is not None else task.data.timeout.agent
+        ),
         finalize=(
             timeout.finalize
             if timeout.finalize is not None

{verifiers-0.3.2.dev146 → verifiers-0.3.2.dev148}/verifiers/v1/configs/agent.py

@@ -2,7 +2,7 @@
 
 from typing import Any
 
-from pydantic import BaseModel, SerializeAsAny, model_validator
+from pydantic import BaseModel, Field, SerializeAsAny, model_validator
 from pydantic_config import BaseConfig
 
 from verifiers.v1.clients import ClientConfig
@@ -14,16 +14,17 @@ from verifiers.v1.utils.generic import deep_merge
 
 
 class TimeoutConfig(BaseConfig):
-    """Timeout (in seconds) for different phases of an agent's run."""
+    """Timeout (in seconds) for different phases of an agent's run. Each unset
+    stage takes the task's own timeout for it, else runs with no limit."""
 
-    setup: float | None = None  # one shared budget: task setup + provisioning
+    # one shared budget: task setup + provisioning
+    setup: float | None = Field(None, gt=0)
     """Timeout (in seconds) for task/harness setup through session preparation."""
-    rollout: float | None = None
-    """Timeout (in seconds) for the agent's solve attempt. Unset: the task's own
-    timeout, else 4 hours. `0` disables the timeout."""
-    finalize: float | None = None
+    rollout: float | None = Field(None, gt=0)
+    """Timeout (in seconds) for the agent's solve attempt."""
+    finalize: float | None = Field(None, gt=0)
     """Timeout (in seconds) for the task + harness finalize hooks."""
-    scoring: float | None = None
+    scoring: float | None = Field(None, gt=0)
     """Timeout (in seconds) for the task + harness metrics + scoring hooks."""
 
 

{verifiers-0.3.2.dev146 → verifiers-0.3.2.dev148}/verifiers/v1/rollout.py

@@ -209,7 +209,7 @@ class Rollout:
             self.runtime_config.type,
         )
         loop = asyncio.get_running_loop()
-        setup_deadline: float | None = None
+        setup_timeout: asyncio.Timeout | None = None
         try:
             runtime_env = dict(self.task.runtime_env())
             if self._borrowed_runtime is None:
@@ -237,12 +237,12 @@ class Rollout:
                 else loop.time() + self._timeouts.setup
             )
             async with (
-                asyncio.timeout_at(setup_deadline),
+                asyncio.timeout_at(setup_deadline) as setup_timeout,
                 boundary(TaskError, "task setup"),
             ):
                 await invoke(self.task.setup, {"trace": self.trace, "runtime": runtime})
             async with (
-                asyncio.timeout_at(setup_deadline),
+                asyncio.timeout_at(setup_deadline) as setup_timeout,
                 boundary(HarnessError, "harness setup"),
             ):
                 await self.harness.setup(runtime)
@@ -281,7 +281,7 @@ class Rollout:
             # execution policy while preserving the framework routes the agent uses.
             await runtime.prepare_execution([self._endpoint, *self._urls.values()])
             async with (
-                asyncio.timeout_at(setup_deadline),
+                asyncio.timeout_at(setup_deadline) as setup_timeout,
                 boundary(HarnessError, "opening harness session"),
             ):
                 harness_data = self.trace.task.data
@@ -349,9 +349,9 @@ class Rollout:
                         **session_kwargs,
                     )
         except TimeoutError as e:
-            # The setup deadline's own expiry is a timeout; a TimeoutError with no
-            # expired deadline (a runtime's own I/O) stays the raw failure.
-            if setup_deadline is not None and loop.time() >= setup_deadline:
+            # Only an expired setup context owns the timeout; I/O between these
+            # contexts can time out independently after the deadline passes.
+            if setup_timeout is not None and setup_timeout.expired():
                 self.timeout("setup")
             else:
                 self.fail(e)

{verifiers-0.3.2.dev146 → verifiers-0.3.2.dev148}/verifiers/v1/tasksets/harbor/taskset.py

@@ -77,10 +77,10 @@ class HarborConfig(TasksetConfig):
     tasks: list[str] | None = None
     """Optional subset of task names to load (None = all)."""
     ignore_timeouts: bool = True
-    """Drop each task's declared agent and verifier timeouts so rollouts use the
-    run-level `--timeout.*` limits or their defaults (4 h for the agent; `--timeout.rollout 0`
-    runs unbounded). Task timeouts are authored against Harbor's runtime and confound
-    model capability with inference speed; set False to apply them anyway."""
+    """Drop each task's declared agent and verifier timeouts so rollouts use only the
+    run-level `--timeout.*` limits (none by default). Task timeouts are authored
+    against Harbor's runtime and confound model capability with inference speed; set
+    False to apply them anyway."""
     timeout_multiplier: float = Field(1.0, gt=0)
     """Scale each task's agent and verifier timeouts. Only applies with
     `ignore_timeouts=False`."""