llama-stack 0.4.4__py3-none-any.whl → 0.5.0__py3-none-any.whl

llama_stack/providers/utils/safety.py

@@ -0,0 +1,114 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+import uuid
+from typing import TYPE_CHECKING
+
+from llama_stack_api import (
+    ModerationObject,
+    ModerationObjectResults,
+    OpenAIUserMessageParam,
+    RunModerationRequest,
+    RunShieldRequest,
+    RunShieldResponse,
+)
+
+if TYPE_CHECKING:
+    # Type stub for mypy - actual implementation provided by provider class
+    class _RunShieldProtocol:
+        async def run_shield(self, request: RunShieldRequest) -> RunShieldResponse: ...
+
+
+class ShieldToModerationMixin:
+    """
+    Mixin that provides run_moderation implementation by delegating to run_shield.
+
+    Providers must implement run_shield(request: RunShieldRequest) for this mixin to work.
+    Providers with custom run_moderation implementations will override this automatically.
+    """
+
+    if TYPE_CHECKING:
+        # Type hint for mypy - run_shield is provided by the mixed-in class
+        async def run_shield(self, request: RunShieldRequest) -> RunShieldResponse: ...
+
+    async def run_moderation(self, request: RunModerationRequest) -> ModerationObject:
+        """
+        Run moderation by converting input to messages and delegating to run_shield.
+
+        Args:
+            request: RunModerationRequest with input and model
+
+        Returns:
+            ModerationObject with results for each input
+
+        Raises:
+            ValueError: If model is None
+        """
+        if request.model is None:
+            raise ValueError(f"{self.__class__.__name__} moderation requires a model identifier")
+
+        inputs = request.input if isinstance(request.input, list) else [request.input]
+        results = []
+
+        for text_input in inputs:
+            # Convert string to OpenAI message format
+            message = OpenAIUserMessageParam(content=text_input)
+
+            # Call run_shield (must be implemented by the provider)
+            shield_request = RunShieldRequest(
+                shield_id=request.model,
+                messages=[message],
+            )
+            shield_response = await self.run_shield(shield_request)
+
+            # Convert RunShieldResponse to ModerationObjectResults
+            results.append(self._shield_response_to_moderation_result(shield_response))
+
+        return ModerationObject(
+            id=f"modr-{uuid.uuid4()}",
+            model=request.model,
+            results=results,
+        )
+
+    def _shield_response_to_moderation_result(
+        self,
+        shield_response: RunShieldResponse,
+    ) -> ModerationObjectResults:
+        """Convert RunShieldResponse to ModerationObjectResults.
+
+        Args:
+            shield_response: The response from run_shield
+
+        Returns:
+            ModerationObjectResults with appropriate fields set
+        """
+        if shield_response.violation is None:
+            # Safe content
+            return ModerationObjectResults(
+                flagged=False,
+                categories={},
+                category_scores={},
+                category_applied_input_types={},
+                user_message=None,
+                metadata={},
+            )
+
+        # Unsafe content - extract violation details
+        v = shield_response.violation
+        violation_type = v.metadata.get("violation_type", "unsafe")
+
+        # Ensure violation_type is a string (metadata values can be Any)
+        if not isinstance(violation_type, str):
+            violation_type = "unsafe"
+
+        return ModerationObjectResults(
+            flagged=True,
+            categories={violation_type: True},
+            category_scores={violation_type: 1.0},
+            category_applied_input_types={violation_type: ["text"]},
+            user_message=v.user_message,
+            metadata=v.metadata,
+        )

llama_stack/providers/utils/tools/mcp.py

@@ -8,6 +8,7 @@ import asyncio
 import hashlib
 from collections.abc import AsyncGenerator
 from contextlib import asynccontextmanager
+from dataclasses import dataclass
 from enum import Enum
 from typing import Any, cast
 
@@ -241,10 +242,12 @@ class MCPSessionManager:
             raise last_exception
         raise RuntimeError(f"Failed to create MCP session for {endpoint}")
 
-    async def close_all(self) -> None:
-        """Close all cached sessions.
+    async def __aenter__(self):
+        """Enter the async context manager."""
+        return self
 
-        Should be called at the end of a request to clean up resources.
+    async def __aexit__(self, exc_type, exc_val, exc_tb):
+        """Exit the async context manager and cleanup all sessions.
 
         Note: We catch BaseException (not just Exception) because:
         1. CancelledError is a BaseException and can occur during cleanup
@@ -275,6 +278,8 @@ class MCPSessionManager:
         if errors:
             logger.debug(f"Encountered {len(errors)} errors while closing MCP sessions (expected in streaming)")
 
+        return False
+
 
 @asynccontextmanager
 async def client_wrapper(endpoint: str, headers: dict[str, str]) -> AsyncGenerator[ClientSession, Any]:
@@ -470,3 +475,39 @@ async def invoke_mcp_tool(
     async with client_wrapper(endpoint, final_headers) as session:
         result = await session.call_tool(tool_name, kwargs)
         return _parse_mcp_result(result)
+
+
+@dataclass
+class MCPServerInfo:
+    """Server information from an MCP server."""
+
+    name: str
+    version: str
+    title: str | None = None
+    description: str | None = None
+
+
+async def get_mcp_server_info(
+    endpoint: str,
+    headers: dict[str, str] | None = None,
+    authorization: str | None = None,
+) -> MCPServerInfo:
+    """Get server info from an MCP server.
+    Args:
+        endpoint: MCP server endpoint URL
+        headers: Optional base headers to include
+        authorization: Optional OAuth access token (just the token, not "Bearer <token>")
+    Returns:
+        MCPServerInfo containing name, version, title, and description
+    """
+    final_headers = prepare_mcp_headers(headers, authorization)
+
+    async with client_wrapper(endpoint, final_headers) as session:
+        init_result = await session.initialize()
+
+        return MCPServerInfo(
+            name=init_result.serverInfo.name,
+            version=init_result.serverInfo.version,
+            title=init_result.serverInfo.title,
+            description=init_result.instructions,
+        )

llama_stack/testing/api_recorder.py

@@ -77,11 +77,14 @@ def _normalize_numeric_literal_strings(value: str) -> str:
     return _FLOAT_IN_STRING_PATTERN.sub(_replace, value)
 
 
-def _normalize_body_for_hash(value: Any) -> Any:
+def _normalize_body_for_hash(value: Any, exclude_stream_options: bool = False) -> Any:
     """Recursively normalize a JSON-like value to improve hash stability."""
 
     if isinstance(value, dict):
-        return {key: _normalize_body_for_hash(item) for key, item in value.items()}
+        normalized = {key: _normalize_body_for_hash(item) for key, item in value.items()}
+        if exclude_stream_options and "stream_options" in normalized:
+            del normalized["stream_options"]
+        return normalized
     if isinstance(value, list):
         return [_normalize_body_for_hash(item) for item in value]
     if isinstance(value, tuple):
@@ -146,7 +149,10 @@ def normalize_inference_request(method: str, url: str, headers: dict[str, Any],
 
     parsed = urlparse(url)
 
-    body_for_hash = _normalize_body_for_hash(body)
+    # Bedrock's OpenAI-compatible endpoint includes stream_options that vary between
+    # runs but don't affect the logical request. Exclude it for stable hashing.
+    is_bedrock = "bedrock" in parsed.netloc
+    body_for_hash = _normalize_body_for_hash(body, exclude_stream_options=is_bedrock)
 
     test_id = get_test_context()
     normalized: dict[str, Any] = {

{llama_stack-0.4.4.dist-info → llama_stack-0.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llama_stack
-Version: 0.4.4
+Version: 0.5.0
 Summary: Llama Stack
 Author-email: Meta Llama <llama-oss@meta.com>
 License: MIT
@@ -45,8 +45,12 @@ Requires-Dist: starlette>=0.49.1
 Requires-Dist: psycopg2-binary
 Requires-Dist: tornado>=6.5.3
 Requires-Dist: urllib3>=2.6.3
+Requires-Dist: oracledb>=3.4.1
+Requires-Dist: oci>=2.165.0
+Requires-Dist: numpy>=2.3.2
+Requires-Dist: mcp>=1.23.0
 Provides-Extra: client
-Requires-Dist: llama-stack-client==0.4.4; extra == "client"
+Requires-Dist: llama-stack-client==0.5.0; extra == "client"
 Dynamic: license-file
 
 # Llama Stack
@@ -158,6 +162,7 @@ Please checkout our [Documentation](https://llamastack.github.io/docs) page for
     * A [Zero-to-Hero Guide](https://github.com/meta-llama/llama-stack/tree/main/docs/zero_to_hero_guide) that guide you through all the key components of llama stack with code samples.
 * [Contributing](CONTRIBUTING.md)
     * [Adding a new API Provider](https://llamastack.github.io/docs/contributing/new_api_provider) to walk-through how to add a new API provider.
+    * [Release Process](RELEASE_PROCESS.md) for information about release schedules and versioning.
 
 ### Llama Stack Client SDKs
 
@@ -172,6 +177,13 @@ Check out our client SDKs for connecting to a Llama Stack server in your preferr
 
 You can find more example scripts with client SDKs to talk with the Llama Stack server in our [llama-stack-apps](https://github.com/meta-llama/llama-stack-apps/tree/main/examples) repo.
 
+## Community
+
+We hold regular community calls to discuss the latest developments and get feedback from the community.
+
+- Date: every Thursday
+- Time: 09:00 AM PST (check the [Community Event on Discord](https://discord.com/events/1257833999603335178/1413266296748900513) for the latest details)
+
 ## 🌟 GitHub Star History
 ## Star History