Merge branch 'main' of github.com:microsoft/agent-lightning into codex/add-tutorial-for-launching-workers

Author: ultmaster

.github/workflows/docs.yml

@@ -8,6 +8,10 @@ on:
       - 'v*'
   workflow_dispatch:
 
+concurrency:
+  group: docs-deploy
+  cancel-in-progress: false
+
 permissions:
   contents: write
   pages: write

agentlightning/trainer/trainer.py

@@ -114,6 +114,9 @@ class Trainer(TrainerLegacy):
     or a dictionary with the initialization parameters for the exporter.
     Deprecated. Use [`adapter`][agentlightning.Trainer.adapter] instead."""
 
+    port: Optional[int]
+    """Port forwarded to [`ClientServerExecutionStrategy`][agentlightning.ClientServerExecutionStrategy]."""
+
     def __init__(
         self,
         *,
@@ -126,6 +129,7 @@ def __init__(
         store: ComponentSpec[LightningStore] = None,
         runner: ComponentSpec[Runner[Any]] = None,
         strategy: ComponentSpec[ExecutionStrategy] = None,
+        port: Optional[int] = None,
         algorithm: ComponentSpec[Algorithm] = None,
         llm_proxy: ComponentSpec[LLMProxy] = None,
         n_workers: Optional[int] = None,
@@ -139,6 +143,10 @@ def __init__(
         Each keyword accepts either a concrete instance, a class, a callable factory, a
         registry string, or a lightweight configuration dictionary (see
         [`build_component()`][agentlightning.trainer.init_utils.build_component]).
+
+        When ``port`` is provided it is forwarded to
+        [`ClientServerExecutionStrategy`][agentlightning.ClientServerExecutionStrategy]
+        instances constructed (or supplied) for the trainer.
         """
         # Do not call super().__init__() here.
         # super().__init__() will call TrainerLegacy's initialization, which is not intended.
@@ -209,7 +217,13 @@ def __init__(
         self.store = self._make_store(store)
         self.runner = self._make_runner(runner)
 
-        self.strategy = self._make_strategy(strategy, n_runners=self.n_runners)
+        self.port = port
+
+        self.strategy = self._make_strategy(
+            strategy,
+            n_runners=self.n_runners,
+            port=port,
+        )
         if hasattr(self.strategy, "n_runners"):
             strategy_runners = getattr(self.strategy, "n_runners")
             if isinstance(strategy_runners, int) and strategy_runners > 0:
@@ -284,13 +298,20 @@ def _make_strategy(
         strategy: ComponentSpec[ExecutionStrategy],
         *,
         n_runners: int,
+        port: Optional[int] = None,
     ) -> ExecutionStrategy:
         """Resolve the execution strategy and seed defaults such as `n_runners`."""
         if isinstance(strategy, ExecutionStrategy):
+            if port is not None and isinstance(strategy, ClientServerExecutionStrategy):
+                strategy.server_port = port
             return strategy
         optional_defaults: Dict[str, Callable[[], Any]] = {"n_runners": lambda: n_runners}
+        if port is not None:
+            optional_defaults["server_port"] = lambda: port
 
         def default_factory() -> ExecutionStrategy:
+            if port is not None:
+                return ClientServerExecutionStrategy(n_runners=n_runners, server_port=port)
             return ClientServerExecutionStrategy(n_runners=n_runners)
 
         return build_component(

agentlightning/types/tracer.py

@@ -14,7 +14,7 @@
 from opentelemetry.sdk.trace import ReadableSpan
 from opentelemetry.sdk.trace.id_generator import RandomIdGenerator
 from opentelemetry.trace.status import Status as OtelStatus
-from pydantic import BaseModel
+from pydantic import BaseModel, ConfigDict
 
 __all__ = [
     "AttributeValue",
@@ -95,8 +95,7 @@ class SpanContext(BaseModel):
     trace_state: TraceState
     """Mapping from trace state key to its value."""
 
-    class Config:
-        allow_extra = True
+    model_config = ConfigDict(extra="allow")
 
     @classmethod
     def from_opentelemetry(cls, src: trace_api.SpanContext) -> "SpanContext":
@@ -119,8 +118,7 @@ class TraceStatus(BaseModel):
     description: Optional[str] = None
     """The description of the span. Same as OpenTelemetry `Status.description` type."""
 
-    class Config:
-        allow_extra = True
+    model_config = ConfigDict(extra="allow")
 
     @classmethod
     def from_opentelemetry(cls, src: OtelStatus) -> "TraceStatus":
@@ -143,8 +141,7 @@ class Event(BaseModel):
     timestamp: Optional[float] = None
     """The timestamp of the event. Same as OpenTelemetry `Event.timestamp` type."""
 
-    class Config:
-        allow_extra = True
+    model_config = ConfigDict(extra="allow")
 
     @classmethod
     def from_opentelemetry(cls, src: OtelEvent) -> "Event":
@@ -166,8 +163,7 @@ class Link(BaseModel):
     attributes: Optional[Attributes] = None
     """Optional attributes."""
 
-    class Config:
-        allow_extra = True
+    model_config = ConfigDict(extra="allow")
 
     @classmethod
     def from_opentelemetry(cls, src: trace_api.Link) -> "Link":
@@ -214,8 +210,7 @@ class Span(BaseModel):
     stable even as upstream OpenTelemetry types evolve.
     """
 
-    class Config:
-        allow_extra = True  # allow extra fields if needed
+    model_config = ConfigDict(extra="allow")
 
     rollout_id: str
     """The rollout which this span belongs to."""

docs/algorithm-zoo/verl.md

@@ -22,7 +22,7 @@ pip install agentlightning[verl]
 
 !!! note "Notes for Readers"
 
-    [VERL][agentlightning.algorithm.verl.VERL] in this article refers to a wrapper, provided by Agent-lightning, of the VERL framework. It's a subclass of [agentlightning.Algorithm][]. To differentiate it from the VERL framework, all references to the VERL framework shall use the term "VERL framework", and all references to the Agent-lightning wrapper shall be highlighted with a link.
+    [VERL][agentlightning.algorithm.verl.VERL] in this article refers to a wrapper, provided by Agent-lightning, of the [VERL framework](https://github.com/volcengine/verl). It's a subclass of [agentlightning.Algorithm][]. To differentiate it from the VERL framework, all references to the VERL framework shall use the term "VERL framework", and all references to the Agent-lightning wrapper shall be highlighted with a link.
 
 ## Resources
 

pyproject.toml

@@ -63,10 +63,12 @@ experiment = [
 ]
 
 core-legacy = [
-  "agentops<=0.4.18"
+  "agentops<=0.4.18",
+  "openai<2.0.0",
 ]
 core-stable = [
-  "agentops>=0.4.21"
+  "agentops>=0.4.21",
+  "openai>=2.0.0",
 ]
 
 # For PyTorch.
@@ -123,8 +125,8 @@ trl = [
   # Only work with PyTorch 2.8.0+.
   {include-group = "torch-stable"},
   # https://github.com/unslothai/unsloth/issues/3451
-  "unsloth<=2025.10.1",
-  "unsloth_zoo<=2025.10.1",
+  "unsloth>=2025.10.1,!=2025.10.2,!=2025.10.3,!=2025.10.4,!=2025.10.5,!=2025.10.6,!=2025.10.7,!=2025.10.8",
+  "unsloth_zoo>=2025.10.1,!=2025.10.2,!=2025.10.3,!=2025.10.4,!=2025.10.5,!=2025.10.6,!=2025.10.7,!=2025.10.8",
   "bitsandbytes",
   "peft",
   "datasets",
@@ -134,6 +136,13 @@ trl = [
   "vllm",
 ]
 
+# For Tinker integration.
+tinker = [
+  {include-group = "torch-stable"},
+  "tinker>=0.2.2",
+  "tinker_cookbook",
+]
+
 # Agent-related dependencies.
 autogen = [
   "autogen-agentchat",
@@ -160,6 +169,9 @@ sql = [
   "sqlparse",
   "nltk",
 ]
+crewai = [
+  "crewai[tools]>=1.2.0",
+]
 
 # Summarize into large installable groups.
 agents = [
@@ -168,9 +180,11 @@ agents = [
   {include-group = "langchain"},
   {include-group = "sql"},
   {include-group = "anthropic"},
+  {include-group = "crewai"},
 ]
 
 [tool.uv]
+required-version = ">=0.9.5"
 conflicts = [
   [
     { group = "core-legacy" },
@@ -189,15 +203,25 @@ environments = [
   "sys_platform == 'linux'",
 ]
 dependency-metadata = [
+  # Patch the dependencies of instructor to unpin "openai".
+  # This is a workaround for https://github.com/567-labs/instructor/issues/1852
+  { name = "instructor", version = "1.11.3", requires-dist = ["openai", "pydantic", "docstring-parser", "typer", "rich", "aiohttp", "tenacity", "pydantic-core", "jiter", "jinja2", "requests", "diskcache"] },
+]
+
+override-dependencies = [
+  # A conflict between the dependency of litellm[proxy] and crewai[tools]
+  "mcp>=1.19.0",
+  "uvicorn>=0.38.0",
   # Conflicts between packaging dependency of pyvers (dependency of tensordict) and agentops.
-  { name = "pyvers", version = "0.1.0", requires-dist = ["packaging>=24.0"] },
+  "packaging>=24.0",
 ]
 
 [tool.uv.sources]
 torch = [
   { index = "pytorch-cu128", group = "torch-cu128" },
   { index = "pytorch-cpu", group = "torch-cpu" },
 ]
+tinker_cookbook = { git = "https://github.com/thinking-machines-lab/tinker-cookbook", rev = "72ba5e6a1f52c0887e2674615e318ce21a39cc2a" }
 
 [[tool.uv.index]]
 name = "pypi"
@@ -216,6 +240,9 @@ flash-attn = [
   { requirement = "torch", match-runtime = true },
 ]
 
+[tool.uv.dependency-groups]
+tinker = {requires-python = ">=3.11"}
+
 [build-system]
 requires = ["hatchling"]
 build-backend = "hatchling.build"

tests/trainer/test_trainer_init.py

@@ -77,6 +77,46 @@ def test_trainer_with_client_server_strategy_dict() -> None:
     assert trainer.strategy.server_port == 9999
 
 
+def test_trainer_port_forwarded_to_client_server_strategy() -> None:
+    """Test that the top-level port argument configures the client-server strategy."""
+    trainer = agl.Trainer(
+        algorithm=agl.Baseline(),
+        n_runners=4,
+        port=8081,
+    )
+
+    assert isinstance(trainer.strategy, agl.ClientServerExecutionStrategy)
+    assert trainer.strategy.server_port == 8081
+
+
+def test_trainer_port_ignored_for_non_client_server_strategy() -> None:
+    """Test that port has no effect when using a non client-server strategy."""
+    trainer = agl.Trainer(
+        algorithm=agl.Baseline(),
+        n_runners=1,
+        port=8082,
+        strategy="shm",
+    )
+
+    assert isinstance(trainer.strategy, agl.SharedMemoryExecutionStrategy)
+    assert not hasattr(trainer.strategy, "server_port")
+
+
+def test_trainer_port_overrides_existing_client_server_strategy() -> None:
+    """Test that provided port overrides an initialized client-server strategy."""
+    strategy = agl.ClientServerExecutionStrategy(server_port=9000)
+
+    trainer = agl.Trainer(
+        algorithm=agl.Baseline(),
+        n_runners=1,
+        strategy=strategy,
+        port=9100,
+    )
+
+    assert trainer.strategy is strategy
+    assert trainer.strategy.server_port == 9100  # type: ignore
+
+
 def test_trainer_with_env_vars_for_execution_strategy(monkeypatch: pytest.MonkeyPatch) -> None:
     """Test that execution strategy supports environment variables to override values."""
     algorithm = agl.Baseline()