phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-10-04 20:14:13 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into remove-batch-inference

Browse source
This commit is contained in:
Matthew Farrellee 2025-09-26 11:25:56 -04:00
commit 32b87bf88a
748 changed files with 127607 additions and 50032 deletions
Download patch file Download diff file Expand all files Collapse all files

40
llama_stack/apis/agents/agents.py
View file

@ -27,6 +27,7 @@ from llama_stack.apis.inference import (
)
from llama_stack.apis.safety import SafetyViolation
from llama_stack.apis.tools import ToolDef
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
from .openai_responses import (
@ -481,7 +482,7 @@ class Agents(Protocol):
- Agents can also use Memory to retrieve information from knowledge bases. See the RAG Tool and Vector IO APIs for more details.
"""
@webmethod(route="/agents", method="POST", descriptive_name="create_agent")
@webmethod(route="/agents", method="POST", descriptive_name="create_agent", level=LLAMA_STACK_API_V1)
async def create_agent(
self,
agent_config: AgentConfig,
@ -494,7 +495,10 @@ class Agents(Protocol):
...
@webmethod(
route="/agents/{agent_id}/session/{session_id}/turn", method="POST", descriptive_name="create_agent_turn"
route="/agents/{agent_id}/session/{session_id}/turn",
method="POST",
descriptive_name="create_agent_turn",
level=LLAMA_STACK_API_V1,
)
async def create_agent_turn(
self,
@ -524,6 +528,7 @@ class Agents(Protocol):
route="/agents/{agent_id}/session/{session_id}/turn/{turn_id}/resume",
method="POST",
descriptive_name="resume_agent_turn",
level=LLAMA_STACK_API_V1,
)
async def resume_agent_turn(
self,
@ -549,6 +554,7 @@ class Agents(Protocol):
@webmethod(
route="/agents/{agent_id}/session/{session_id}/turn/{turn_id}",
method="GET",
level=LLAMA_STACK_API_V1,
)
async def get_agents_turn(
self,
@ -568,6 +574,7 @@ class Agents(Protocol):
@webmethod(
route="/agents/{agent_id}/session/{session_id}/turn/{turn_id}/step/{step_id}",
method="GET",
level=LLAMA_STACK_API_V1,
)
async def get_agents_step(
self,
@ -586,7 +593,12 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}/session", method="POST", descriptive_name="create_agent_session")
@webmethod(
route="/agents/{agent_id}/session",
method="POST",
descriptive_name="create_agent_session",
level=LLAMA_STACK_API_V1,
)
async def create_agent_session(
self,
agent_id: str,
@ -600,7 +612,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}/session/{session_id}", method="GET")
@webmethod(route="/agents/{agent_id}/session/{session_id}", method="GET", level=LLAMA_STACK_API_V1)
async def get_agents_session(
self,
session_id: str,
@ -616,7 +628,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}/session/{session_id}", method="DELETE")
@webmethod(route="/agents/{agent_id}/session/{session_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def delete_agents_session(
self,
session_id: str,
@ -629,7 +641,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}", method="DELETE")
@webmethod(route="/agents/{agent_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def delete_agent(
self,
agent_id: str,
@ -640,7 +652,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents", method="GET")
@webmethod(route="/agents", method="GET", level=LLAMA_STACK_API_V1)
async def list_agents(self, start_index: int | None = None, limit: int | None = None) -> PaginatedResponse:
"""List all agents.
@ -650,7 +662,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}", method="GET")
@webmethod(route="/agents/{agent_id}", method="GET", level=LLAMA_STACK_API_V1)
async def get_agent(self, agent_id: str) -> Agent:
"""Describe an agent by its ID.
@ -659,7 +671,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/agents/{agent_id}/sessions", method="GET")
@webmethod(route="/agents/{agent_id}/sessions", method="GET", level=LLAMA_STACK_API_V1)
async def list_agent_sessions(
self,
agent_id: str,
@ -682,7 +694,7 @@ class Agents(Protocol):
#
# Both of these APIs are inherently stateful.
@webmethod(route="/openai/v1/responses/{response_id}", method="GET")
@webmethod(route="/openai/v1/responses/{response_id}", method="GET", level=LLAMA_STACK_API_V1)
async def get_openai_response(
self,
response_id: str,
@ -694,7 +706,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/openai/v1/responses", method="POST")
@webmethod(route="/openai/v1/responses", method="POST", level=LLAMA_STACK_API_V1)
async def create_openai_response(
self,
input: str | list[OpenAIResponseInput],
@ -719,7 +731,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/openai/v1/responses", method="GET")
@webmethod(route="/openai/v1/responses", method="GET", level=LLAMA_STACK_API_V1)
async def list_openai_responses(
self,
after: str | None = None,
@ -737,7 +749,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/openai/v1/responses/{response_id}/input_items", method="GET")
@webmethod(route="/openai/v1/responses/{response_id}/input_items", method="GET", level=LLAMA_STACK_API_V1)
async def list_openai_response_input_items(
self,
response_id: str,
@ -759,7 +771,7 @@ class Agents(Protocol):
"""
...
@webmethod(route="/openai/v1/responses/{response_id}", method="DELETE")
@webmethod(route="/openai/v1/responses/{response_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def delete_openai_response(self, response_id: str) -> OpenAIDeleteResponseObject:
"""Delete an OpenAI response by its ID.

9
llama_stack/apis/batches/batches.py
View file

@ -8,6 +8,7 @@ from typing import Literal, Protocol, runtime_checkable
from pydantic import BaseModel, Field
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, webmethod
try:
@ -42,7 +43,7 @@ class Batches(Protocol):
Note: This API is currently under active development and may undergo changes.
"""
@webmethod(route="/openai/v1/batches", method="POST")
@webmethod(route="/openai/v1/batches", method="POST", level=LLAMA_STACK_API_V1)
async def create_batch(
self,
input_file_id: str,
@ -62,7 +63,7 @@ class Batches(Protocol):
"""
...
@webmethod(route="/openai/v1/batches/{batch_id}", method="GET")
@webmethod(route="/openai/v1/batches/{batch_id}", method="GET", level=LLAMA_STACK_API_V1)
async def retrieve_batch(self, batch_id: str) -> BatchObject:
"""Retrieve information about a specific batch.
@ -71,7 +72,7 @@ class Batches(Protocol):
"""
...
@webmethod(route="/openai/v1/batches/{batch_id}/cancel", method="POST")
@webmethod(route="/openai/v1/batches/{batch_id}/cancel", method="POST", level=LLAMA_STACK_API_V1)
async def cancel_batch(self, batch_id: str) -> BatchObject:
"""Cancel a batch that is in progress.
@ -80,7 +81,7 @@ class Batches(Protocol):
"""
...
@webmethod(route="/openai/v1/batches", method="GET")
@webmethod(route="/openai/v1/batches", method="GET", level=LLAMA_STACK_API_V1)
async def list_batches(
self,
after: str | None = None,

19
llama_stack/apis/benchmarks/benchmarks.py
View file

@ -8,6 +8,7 @@ from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, Field
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1, LLAMA_STACK_API_V1ALPHA
from llama_stack.schema_utils import json_schema_type, webmethod
@ -53,7 +54,8 @@ class ListBenchmarksResponse(BaseModel):
@runtime_checkable
class Benchmarks(Protocol):
@webmethod(route="/eval/benchmarks", method="GET")
@webmethod(route="/eval/benchmarks", method="GET", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/eval/benchmarks", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def list_benchmarks(self) -> ListBenchmarksResponse:
"""List all benchmarks.
@ -61,7 +63,8 @@ class Benchmarks(Protocol):
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}", method="GET")
@webmethod(route="/eval/benchmarks/{benchmark_id}", method="GET", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/eval/benchmarks/{benchmark_id}", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def get_benchmark(
self,
benchmark_id: str,
@ -73,7 +76,8 @@ class Benchmarks(Protocol):
"""
...
@webmethod(route="/eval/benchmarks", method="POST")
@webmethod(route="/eval/benchmarks", method="POST", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/eval/benchmarks", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def register_benchmark(
self,
benchmark_id: str,
@ -93,3 +97,12 @@ class Benchmarks(Protocol):
:param metadata: The metadata to use for the benchmark.
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}", method="DELETE", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/eval/benchmarks/{benchmark_id}", method="DELETE", level=LLAMA_STACK_API_V1ALPHA)
async def unregister_benchmark(self, benchmark_id: str) -> None:
"""Unregister a benchmark.
:param benchmark_id: The ID of the benchmark to unregister.
"""
...

7
llama_stack/apis/common/errors.py
View file

@ -79,3 +79,10 @@ class ConflictError(ValueError):
def __init__(self, message: str) -> None:
super().__init__(message)
class TokenValidationError(ValueError):
"""raised when token validation fails during authentication"""
def __init__(self, message: str) -> None:
super().__init__(message)

5
llama_stack/apis/datasetio/datasetio.py
View file

@ -8,6 +8,7 @@ from typing import Any, Protocol, runtime_checkable
from llama_stack.apis.common.responses import PaginatedResponse
from llama_stack.apis.datasets import Dataset
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import webmethod
@ -20,7 +21,7 @@ class DatasetIO(Protocol):
# keeping for aligning with inference/safety, but this is not used
dataset_store: DatasetStore
@webmethod(route="/datasetio/iterrows/{dataset_id:path}", method="GET")
@webmethod(route="/datasetio/iterrows/{dataset_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def iterrows(
self,
dataset_id: str,
@ -44,7 +45,7 @@ class DatasetIO(Protocol):
"""
...
@webmethod(route="/datasetio/append-rows/{dataset_id:path}", method="POST")
@webmethod(route="/datasetio/append-rows/{dataset_id:path}", method="POST", level=LLAMA_STACK_API_V1)
async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None:
"""Append rows to a dataset.

9
llama_stack/apis/datasets/datasets.py
View file

@ -10,6 +10,7 @@ from typing import Annotated, Any, Literal, Protocol
from pydantic import BaseModel, Field
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -145,7 +146,7 @@ class ListDatasetsResponse(BaseModel):
class Datasets(Protocol):
@webmethod(route="/datasets", method="POST")
@webmethod(route="/datasets", method="POST", level=LLAMA_STACK_API_V1)
async def register_dataset(
self,
purpose: DatasetPurpose,
@ -214,7 +215,7 @@ class Datasets(Protocol):
"""
...
@webmethod(route="/datasets/{dataset_id:path}", method="GET")
@webmethod(route="/datasets/{dataset_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_dataset(
self,
dataset_id: str,
@ -226,7 +227,7 @@ class Datasets(Protocol):
"""
...
@webmethod(route="/datasets", method="GET")
@webmethod(route="/datasets", method="GET", level=LLAMA_STACK_API_V1)
async def list_datasets(self) -> ListDatasetsResponse:
"""List all datasets.
@ -234,7 +235,7 @@ class Datasets(Protocol):
"""
...
@webmethod(route="/datasets/{dataset_id:path}", method="DELETE")
@webmethod(route="/datasets/{dataset_id:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_dataset(
self,
dataset_id: str,

2
llama_stack/apis/datatypes.py
View file

@ -102,6 +102,7 @@ class Api(Enum, metaclass=DynamicApiMeta):
:cvar benchmarks: Benchmark suite management
:cvar tool_groups: Tool group organization
:cvar files: File storage and management
:cvar prompts: Prompt versions and management
:cvar inspect: Built-in system inspection and introspection
"""
@ -127,6 +128,7 @@ class Api(Enum, metaclass=DynamicApiMeta):
benchmarks = "benchmarks"
tool_groups = "tool_groups"
files = "files"
prompts = "prompts"
# built-in API
inspect = "inspect"

32
llama_stack/apis/eval/eval.py
View file

@ -13,6 +13,7 @@ from llama_stack.apis.common.job_types import Job
from llama_stack.apis.inference import SamplingParams, SystemMessage
from llama_stack.apis.scoring import ScoringResult
from llama_stack.apis.scoring_functions import ScoringFnParams
from llama_stack.apis.version import LLAMA_STACK_API_V1, LLAMA_STACK_API_V1ALPHA
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -83,7 +84,8 @@ class EvaluateResponse(BaseModel):
class Eval(Protocol):
"""Llama Stack Evaluation API for running evaluations on model and agent candidates."""
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs", method="POST")
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs", method="POST", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def run_eval(
self,
benchmark_id: str,
@ -97,7 +99,10 @@ class Eval(Protocol):
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}/evaluations", method="POST")
@webmethod(
route="/eval/benchmarks/{benchmark_id}/evaluations", method="POST", level=LLAMA_STACK_API_V1, deprecated=True
)
@webmethod(route="/eval/benchmarks/{benchmark_id}/evaluations", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def evaluate_rows(
self,
benchmark_id: str,
@ -115,7 +120,10 @@ class Eval(Protocol):
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="GET")
@webmethod(
route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="GET", level=LLAMA_STACK_API_V1, deprecated=True
)
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def job_status(self, benchmark_id: str, job_id: str) -> Job:
"""Get the status of a job.
@ -125,7 +133,13 @@ class Eval(Protocol):
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="DELETE")
@webmethod(
route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}",
method="DELETE",
level=LLAMA_STACK_API_V1,
deprecated=True,
)
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="DELETE", level=LLAMA_STACK_API_V1ALPHA)
async def job_cancel(self, benchmark_id: str, job_id: str) -> None:
"""Cancel a job.
@ -134,7 +148,15 @@ class Eval(Protocol):
"""
...
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}/result", method="GET")
@webmethod(
route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}/result",
method="GET",
level=LLAMA_STACK_API_V1,
deprecated=True,
)
@webmethod(
route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}/result", method="GET", level=LLAMA_STACK_API_V1ALPHA
)
async def job_result(self, benchmark_id: str, job_id: str) -> EvaluateResponse:
"""Get the result of a job.

36
llama_stack/apis/files/files.py
View file

@ -5,12 +5,13 @@
# the root directory of this source tree.
from enum import StrEnum
from typing import Annotated, Literal, Protocol, runtime_checkable
from typing import Annotated, ClassVar, Literal, Protocol, runtime_checkable
from fastapi import File, Form, Response, UploadFile
from pydantic import BaseModel
from pydantic import BaseModel, Field
from llama_stack.apis.common.responses import Order
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -49,6 +50,23 @@ class OpenAIFileObject(BaseModel):
purpose: OpenAIFilePurpose
@json_schema_type
class ExpiresAfter(BaseModel):
"""
Control expiration of uploaded files.
Params:
- anchor, must be "created_at"
- seconds, must be int between 3600 and 2592000 (1 hour to 30 days)
"""
MIN: ClassVar[int] = 3600 # 1 hour
MAX: ClassVar[int] = 2592000 # 30 days
anchor: Literal["created_at"]
seconds: int = Field(..., ge=3600, le=2592000)
@json_schema_type
class ListOpenAIFileResponse(BaseModel):
"""
@ -87,11 +105,14 @@ class OpenAIFileDeleteResponse(BaseModel):
@trace_protocol
class Files(Protocol):
# OpenAI Files API Endpoints
@webmethod(route="/openai/v1/files", method="POST")
@webmethod(route="/openai/v1/files", method="POST", level=LLAMA_STACK_API_V1)
async def openai_upload_file(
self,
file: Annotated[UploadFile, File()],
purpose: Annotated[OpenAIFilePurpose, Form()],
expires_after_anchor: Annotated[str | None, Form(alias="expires_after[anchor]")] = None,
expires_after_seconds: Annotated[int | None, Form(alias="expires_after[seconds]")] = None,
# TODO: expires_after is producing strange openapi spec, params are showing up as a required w/ oneOf being null
) -> OpenAIFileObject:
"""
Upload a file that can be used across various endpoints.
@ -99,6 +120,7 @@ class Files(Protocol):
The file upload should be a multipart form request with:
- file: The File object (not file name) to be uploaded.
- purpose: The intended purpose of the uploaded file.
- expires_after: Optional form values describing expiration for the file. Expected expires_after[anchor] = "created_at", expires_after[seconds] = {integer}. Seconds must be between 3600 and 2592000 (1 hour to 30 days).
:param file: The uploaded file object containing content and metadata (filename, content_type, etc.).
:param purpose: The intended purpose of the uploaded file (e.g., "assistants", "fine-tune").
@ -106,7 +128,7 @@ class Files(Protocol):
"""
...
@webmethod(route="/openai/v1/files", method="GET")
@webmethod(route="/openai/v1/files", method="GET", level=LLAMA_STACK_API_V1)
async def openai_list_files(
self,
after: str | None = None,
@ -125,7 +147,7 @@ class Files(Protocol):
"""
...
@webmethod(route="/openai/v1/files/{file_id}", method="GET")
@webmethod(route="/openai/v1/files/{file_id}", method="GET", level=LLAMA_STACK_API_V1)
async def openai_retrieve_file(
self,
file_id: str,
@ -138,7 +160,7 @@ class Files(Protocol):
"""
...
@webmethod(route="/openai/v1/files/{file_id}", method="DELETE")
@webmethod(route="/openai/v1/files/{file_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def openai_delete_file(
self,
file_id: str,
@ -151,7 +173,7 @@ class Files(Protocol):
"""
...
@webmethod(route="/openai/v1/files/{file_id}/content", method="GET")
@webmethod(route="/openai/v1/files/{file_id}/content", method="GET", level=LLAMA_STACK_API_V1)
async def openai_retrieve_file_content(
self,
file_id: str,

19
llama_stack/apis/inference/inference.py
View file

@ -21,6 +21,7 @@ from llama_stack.apis.common.content_types import ContentDelta, InterleavedConte
from llama_stack.apis.common.responses import Order
from llama_stack.apis.models import Model
from llama_stack.apis.telemetry import MetricResponseMixin
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.models.llama.datatypes import (
BuiltinTool,
StopReason,
@ -1006,7 +1007,7 @@ class InferenceProvider(Protocol):
model_store: ModelStore | None = None
@webmethod(route="/inference/completion", method="POST")
@webmethod(route="/inference/completion", method="POST", level=LLAMA_STACK_API_V1)
async def completion(
self,
model_id: str,
@ -1029,7 +1030,7 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/inference/chat-completion", method="POST")
@webmethod(route="/inference/chat-completion", method="POST", level=LLAMA_STACK_API_V1)
async def chat_completion(
self,
model_id: str,
@ -1069,7 +1070,7 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/inference/embeddings", method="POST")
@webmethod(route="/inference/embeddings", method="POST", level=LLAMA_STACK_API_V1)
async def embeddings(
self,
model_id: str,
@ -1089,7 +1090,7 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/inference/rerank", method="POST", experimental=True)
@webmethod(route="/inference/rerank", method="POST", experimental=True, level=LLAMA_STACK_API_V1)
async def rerank(
self,
model: str,
@ -1108,7 +1109,7 @@ class InferenceProvider(Protocol):
raise NotImplementedError("Reranking is not implemented")
return # this is so mypy's safe-super rule will consider the method concrete
@webmethod(route="/openai/v1/completions", method="POST")
@webmethod(route="/openai/v1/completions", method="POST", level=LLAMA_STACK_API_V1)
async def openai_completion(
self,
# Standard OpenAI completion parameters
@ -1159,7 +1160,7 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/openai/v1/chat/completions", method="POST")
@webmethod(route="/openai/v1/chat/completions", method="POST", level=LLAMA_STACK_API_V1)
async def openai_chat_completion(
self,
model: str,
@ -1215,7 +1216,7 @@ class InferenceProvider(Protocol):
"""
...
@webmethod(route="/openai/v1/embeddings", method="POST")
@webmethod(route="/openai/v1/embeddings", method="POST", level=LLAMA_STACK_API_V1)
async def openai_embeddings(
self,
model: str,
@ -1244,7 +1245,7 @@ class Inference(InferenceProvider):
- Embedding models: these models generate embeddings to be used for semantic search.
"""
@webmethod(route="/openai/v1/chat/completions", method="GET")
@webmethod(route="/openai/v1/chat/completions", method="GET", level=LLAMA_STACK_API_V1)
async def list_chat_completions(
self,
after: str | None = None,
@ -1262,7 +1263,7 @@ class Inference(InferenceProvider):
"""
raise NotImplementedError("List chat completions is not implemented")
@webmethod(route="/openai/v1/chat/completions/{completion_id}", method="GET")
@webmethod(route="/openai/v1/chat/completions/{completion_id}", method="GET", level=LLAMA_STACK_API_V1)
async def get_chat_completion(self, completion_id: str) -> OpenAICompletionWithInputMessages:
"""Describe a chat completion by its ID.

7
llama_stack/apis/inspect/inspect.py
View file

@ -8,6 +8,7 @@ from typing import Protocol, runtime_checkable
from pydantic import BaseModel
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.datatypes import HealthStatus
from llama_stack.schema_utils import json_schema_type, webmethod
@ -57,7 +58,7 @@ class ListRoutesResponse(BaseModel):
@runtime_checkable
class Inspect(Protocol):
@webmethod(route="/inspect/routes", method="GET")
@webmethod(route="/inspect/routes", method="GET", level=LLAMA_STACK_API_V1)
async def list_routes(self) -> ListRoutesResponse:
"""List all available API routes with their methods and implementing providers.
@ -65,7 +66,7 @@ class Inspect(Protocol):
"""
...
@webmethod(route="/health", method="GET")
@webmethod(route="/health", method="GET", level=LLAMA_STACK_API_V1)
async def health(self) -> HealthInfo:
"""Get the current health status of the service.
@ -73,7 +74,7 @@ class Inspect(Protocol):
"""
...
@webmethod(route="/version", method="GET")
@webmethod(route="/version", method="GET", level=LLAMA_STACK_API_V1)
async def version(self) -> VersionInfo:
"""Get the version of the service.

11
llama_stack/apis/models/models.py
View file

@ -10,6 +10,7 @@ from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, ConfigDict, Field, field_validator
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -102,7 +103,7 @@ class OpenAIListModelsResponse(BaseModel):
@runtime_checkable
@trace_protocol
class Models(Protocol):
@webmethod(route="/models", method="GET")
@webmethod(route="/models", method="GET", level=LLAMA_STACK_API_V1)
async def list_models(self) -> ListModelsResponse:
"""List all models.
@ -110,7 +111,7 @@ class Models(Protocol):
"""
...
@webmethod(route="/openai/v1/models", method="GET")
@webmethod(route="/openai/v1/models", method="GET", level=LLAMA_STACK_API_V1)
async def openai_list_models(self) -> OpenAIListModelsResponse:
"""List models using the OpenAI API.
@ -118,7 +119,7 @@ class Models(Protocol):
"""
...
@webmethod(route="/models/{model_id:path}", method="GET")
@webmethod(route="/models/{model_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_model(
self,
model_id: str,
@ -130,7 +131,7 @@ class Models(Protocol):
"""
...
@webmethod(route="/models", method="POST")
@webmethod(route="/models", method="POST", level=LLAMA_STACK_API_V1)
async def register_model(
self,
model_id: str,
@ -150,7 +151,7 @@ class Models(Protocol):
"""
...
@webmethod(route="/models/{model_id:path}", method="DELETE")
@webmethod(route="/models/{model_id:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_model(
self,
model_id: str,

19
llama_stack/apis/post_training/post_training.py
View file

@ -13,6 +13,7 @@ from pydantic import BaseModel, Field
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.job_types import JobStatus
from llama_stack.apis.common.training_types import Checkpoint
from llama_stack.apis.version import LLAMA_STACK_API_V1, LLAMA_STACK_API_V1ALPHA
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -283,7 +284,8 @@ class PostTrainingJobArtifactsResponse(BaseModel):
class PostTraining(Protocol):
@webmethod(route="/post-training/supervised-fine-tune", method="POST")
@webmethod(route="/post-training/supervised-fine-tune", method="POST", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/supervised-fine-tune", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def supervised_fine_tune(
self,
job_uuid: str,
@ -310,7 +312,8 @@ class PostTraining(Protocol):
"""
...
@webmethod(route="/post-training/preference-optimize", method="POST")
@webmethod(route="/post-training/preference-optimize", method="POST", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/preference-optimize", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def preference_optimize(
self,
job_uuid: str,
@ -332,7 +335,8 @@ class PostTraining(Protocol):
"""
...
@webmethod(route="/post-training/jobs", method="GET")
@webmethod(route="/post-training/jobs", method="GET", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/jobs", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def get_training_jobs(self) -> ListPostTrainingJobsResponse:
"""Get all training jobs.
@ -340,7 +344,8 @@ class PostTraining(Protocol):
"""
...
@webmethod(route="/post-training/job/status", method="GET")
@webmethod(route="/post-training/job/status", method="GET", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/job/status", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def get_training_job_status(self, job_uuid: str) -> PostTrainingJobStatusResponse:
"""Get the status of a training job.
@ -349,7 +354,8 @@ class PostTraining(Protocol):
"""
...
@webmethod(route="/post-training/job/cancel", method="POST")
@webmethod(route="/post-training/job/cancel", method="POST", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/job/cancel", method="POST", level=LLAMA_STACK_API_V1ALPHA)
async def cancel_training_job(self, job_uuid: str) -> None:
"""Cancel a training job.
@ -357,7 +363,8 @@ class PostTraining(Protocol):
"""
...
@webmethod(route="/post-training/job/artifacts", method="GET")
@webmethod(route="/post-training/job/artifacts", method="GET", level=LLAMA_STACK_API_V1, deprecated=True)
@webmethod(route="/post-training/job/artifacts", method="GET", level=LLAMA_STACK_API_V1ALPHA)
async def get_training_job_artifacts(self, job_uuid: str) -> PostTrainingJobArtifactsResponse:
"""Get the artifacts of a training job.

9
llama_stack/apis/prompts/__init__.py Normal file
View file

@ -0,0 +1,9 @@
# 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.
from .prompts import ListPromptsResponse, Prompt, Prompts
__all__ = ["Prompt", "Prompts", "ListPromptsResponse"]

190
llama_stack/apis/prompts/prompts.py Normal file
View file

@ -0,0 +1,190 @@
# 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 re
import secrets
from typing import Protocol, runtime_checkable
from pydantic import BaseModel, Field, field_validator, model_validator
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@json_schema_type
class Prompt(BaseModel):
"""A prompt resource representing a stored OpenAI Compatible prompt template in Llama Stack.
:param prompt: The system prompt text with variable placeholders. Variables are only supported when using the Responses API.
:param version: Version (integer starting at 1, incremented on save)
:param prompt_id: Unique identifier formatted as 'pmpt_<48-digit-hash>'
:param variables: List of prompt variable names that can be used in the prompt template
:param is_default: Boolean indicating whether this version is the default version for this prompt
"""
prompt: str | None = Field(default=None, description="The system prompt with variable placeholders")
version: int = Field(description="Version (integer starting at 1, incremented on save)", ge=1)
prompt_id: str = Field(description="Unique identifier in format 'pmpt_<48-digit-hash>'")
variables: list[str] = Field(
default_factory=list, description="List of variable names that can be used in the prompt template"
)
is_default: bool = Field(
default=False, description="Boolean indicating whether this version is the default version"
)
@field_validator("prompt_id")
@classmethod
def validate_prompt_id(cls, prompt_id: str) -> str:
if not isinstance(prompt_id, str):
raise TypeError("prompt_id must be a string in format 'pmpt_<48-digit-hash>'")
if not prompt_id.startswith("pmpt_"):
raise ValueError("prompt_id must start with 'pmpt_' prefix")
hex_part = prompt_id[5:]
if len(hex_part) != 48:
raise ValueError("prompt_id must be in format 'pmpt_<48-digit-hash>' (48 lowercase hex chars)")
for char in hex_part:
if char not in "0123456789abcdef":
raise ValueError("prompt_id hex part must contain only lowercase hex characters [0-9a-f]")
return prompt_id
@field_validator("version")
@classmethod
def validate_version(cls, prompt_version: int) -> int:
if prompt_version < 1:
raise ValueError("version must be >= 1")
return prompt_version
@model_validator(mode="after")
def validate_prompt_variables(self):
"""Validate that all variables used in the prompt are declared in the variables list."""
if not self.prompt:
return self
prompt_variables = set(re.findall(r"{{\s*(\w+)\s*}}", self.prompt))
declared_variables = set(self.variables)
undeclared = prompt_variables - declared_variables
if undeclared:
raise ValueError(f"Prompt contains undeclared variables: {sorted(undeclared)}")
return self
@classmethod
def generate_prompt_id(cls) -> str:
# Generate 48 hex characters (24 bytes)
random_bytes = secrets.token_bytes(24)
hex_string = random_bytes.hex()
return f"pmpt_{hex_string}"
class ListPromptsResponse(BaseModel):
"""Response model to list prompts."""
data: list[Prompt]
@runtime_checkable
@trace_protocol
class Prompts(Protocol):
"""Protocol for prompt management operations."""
@webmethod(route="/prompts", method="GET", level=LLAMA_STACK_API_V1)
async def list_prompts(self) -> ListPromptsResponse:
"""List all prompts.
:returns: A ListPromptsResponse containing all prompts.
"""
...
@webmethod(route="/prompts/{prompt_id}/versions", method="GET", level=LLAMA_STACK_API_V1)
async def list_prompt_versions(
self,
prompt_id: str,
) -> ListPromptsResponse:
"""List all versions of a specific prompt.
:param prompt_id: The identifier of the prompt to list versions for.
:returns: A ListPromptsResponse containing all versions of the prompt.
"""
...
@webmethod(route="/prompts/{prompt_id}", method="GET", level=LLAMA_STACK_API_V1)
async def get_prompt(
self,
prompt_id: str,
version: int | None = None,
) -> Prompt:
"""Get a prompt by its identifier and optional version.
:param prompt_id: The identifier of the prompt to get.
:param version: The version of the prompt to get (defaults to latest).
:returns: A Prompt resource.
"""
...
@webmethod(route="/prompts", method="POST", level=LLAMA_STACK_API_V1)
async def create_prompt(
self,
prompt: str,
variables: list[str] | None = None,
) -> Prompt:
"""Create a new prompt.
:param prompt: The prompt text content with variable placeholders.
:param variables: List of variable names that can be used in the prompt template.
:returns: The created Prompt resource.
"""
...
@webmethod(route="/prompts/{prompt_id}", method="PUT", level=LLAMA_STACK_API_V1)
async def update_prompt(
self,
prompt_id: str,
prompt: str,
version: int,
variables: list[str] | None = None,
set_as_default: bool = True,
) -> Prompt:
"""Update an existing prompt (increments version).
:param prompt_id: The identifier of the prompt to update.
:param prompt: The updated prompt text content.
:param version: The current version of the prompt being updated.
:param variables: Updated list of variable names that can be used in the prompt template.
:param set_as_default: Set the new version as the default (default=True).
:returns: The updated Prompt resource with incremented version.
"""
...
@webmethod(route="/prompts/{prompt_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def delete_prompt(
self,
prompt_id: str,
) -> None:
"""Delete a prompt.
:param prompt_id: The identifier of the prompt to delete.
"""
...
@webmethod(route="/prompts/{prompt_id}/set-default-version", method="PUT", level=LLAMA_STACK_API_V1)
async def set_default_version(
self,
prompt_id: str,
version: int,
) -> Prompt:
"""Set which version of a prompt should be the default in get_prompt (latest).
:param prompt_id: The identifier of the prompt.
:param version: The version to set as default.
:returns: The prompt with the specified version now set as default.
"""
...

5
llama_stack/apis/providers/providers.py
View file

@ -8,6 +8,7 @@ from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.datatypes import HealthResponse
from llama_stack.schema_utils import json_schema_type, webmethod
@ -45,7 +46,7 @@ class Providers(Protocol):
Providers API for inspecting, listing, and modifying providers and their configurations.
"""
@webmethod(route="/providers", method="GET")
@webmethod(route="/providers", method="GET", level=LLAMA_STACK_API_V1)
async def list_providers(self) -> ListProvidersResponse:
"""List all available providers.
@ -53,7 +54,7 @@ class Providers(Protocol):
"""
...
@webmethod(route="/providers/{provider_id}", method="GET")
@webmethod(route="/providers/{provider_id}", method="GET", level=LLAMA_STACK_API_V1)
async def inspect_provider(self, provider_id: str) -> ProviderInfo:
"""Get detailed information about a specific provider.

1
llama_stack/apis/resource.py
View file

@ -19,6 +19,7 @@ class ResourceType(StrEnum):
benchmark = "benchmark"
tool = "tool"
tool_group = "tool_group"
prompt = "prompt"
class Resource(BaseModel):

5
llama_stack/apis/safety/safety.py
View file

@ -11,6 +11,7 @@ from pydantic import BaseModel, Field
from llama_stack.apis.inference import Message
from llama_stack.apis.shields import Shield
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -97,7 +98,7 @@ class ShieldStore(Protocol):
class Safety(Protocol):
shield_store: ShieldStore
@webmethod(route="/safety/run-shield", method="POST")
@webmethod(route="/safety/run-shield", method="POST", level=LLAMA_STACK_API_V1)
async def run_shield(
self,
shield_id: str,
@ -113,7 +114,7 @@ class Safety(Protocol):
"""
...
@webmethod(route="/openai/v1/moderations", method="POST")
@webmethod(route="/openai/v1/moderations", method="POST", level=LLAMA_STACK_API_V1)
async def run_moderation(self, input: str | list[str], model: str) -> ModerationObject:
"""Classifies if text and/or image inputs are potentially harmful.
:param input: Input (or inputs) to classify.

5
llama_stack/apis/scoring/scoring.py
View file

@ -9,6 +9,7 @@ from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel
from llama_stack.apis.scoring_functions import ScoringFn, ScoringFnParams
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, webmethod
# mapping of metric to value
@ -61,7 +62,7 @@ class ScoringFunctionStore(Protocol):
class Scoring(Protocol):
scoring_function_store: ScoringFunctionStore
@webmethod(route="/scoring/score-batch", method="POST")
@webmethod(route="/scoring/score-batch", method="POST", level=LLAMA_STACK_API_V1)
async def score_batch(
self,
dataset_id: str,
@ -77,7 +78,7 @@ class Scoring(Protocol):
"""
...
@webmethod(route="/scoring/score", method="POST")
@webmethod(route="/scoring/score", method="POST", level=LLAMA_STACK_API_V1)
async def score(
self,
input_rows: list[dict[str, Any]],

15
llama_stack/apis/scoring_functions/scoring_functions.py
View file

@ -18,6 +18,7 @@ from pydantic import BaseModel, Field
from llama_stack.apis.common.type_system import ParamType
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -160,7 +161,7 @@ class ListScoringFunctionsResponse(BaseModel):
@runtime_checkable
class ScoringFunctions(Protocol):
@webmethod(route="/scoring-functions", method="GET")
@webmethod(route="/scoring-functions", method="GET", level=LLAMA_STACK_API_V1)
async def list_scoring_functions(self) -> ListScoringFunctionsResponse:
"""List all scoring functions.
@ -168,7 +169,7 @@ class ScoringFunctions(Protocol):
"""
...
@webmethod(route="/scoring-functions/{scoring_fn_id:path}", method="GET")
@webmethod(route="/scoring-functions/{scoring_fn_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_scoring_function(self, scoring_fn_id: str, /) -> ScoringFn:
"""Get a scoring function by its ID.
@ -177,7 +178,7 @@ class ScoringFunctions(Protocol):
"""
...
@webmethod(route="/scoring-functions", method="POST")
@webmethod(route="/scoring-functions", method="POST", level=LLAMA_STACK_API_V1)
async def register_scoring_function(
self,
scoring_fn_id: str,
@ -197,3 +198,11 @@ class ScoringFunctions(Protocol):
:param params: The parameters for the scoring function for benchmark eval, these can be overridden for app eval.
"""
...
@webmethod(route="/scoring-functions/{scoring_fn_id:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_scoring_function(self, scoring_fn_id: str) -> None:
"""Unregister a scoring function.
:param scoring_fn_id: The ID of the scoring function to unregister.
"""
...

9
llama_stack/apis/shields/shields.py
View file

@ -9,6 +9,7 @@ from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -49,7 +50,7 @@ class ListShieldsResponse(BaseModel):
@runtime_checkable
@trace_protocol
class Shields(Protocol):
@webmethod(route="/shields", method="GET")
@webmethod(route="/shields", method="GET", level=LLAMA_STACK_API_V1)
async def list_shields(self) -> ListShieldsResponse:
"""List all shields.
@ -57,7 +58,7 @@ class Shields(Protocol):
"""
...
@webmethod(route="/shields/{identifier:path}", method="GET")
@webmethod(route="/shields/{identifier:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_shield(self, identifier: str) -> Shield:
"""Get a shield by its identifier.
@ -66,7 +67,7 @@ class Shields(Protocol):
"""
...
@webmethod(route="/shields", method="POST")
@webmethod(route="/shields", method="POST", level=LLAMA_STACK_API_V1)
async def register_shield(
self,
shield_id: str,
@ -84,7 +85,7 @@ class Shields(Protocol):
"""
...
@webmethod(route="/shields/{identifier:path}", method="DELETE")
@webmethod(route="/shields/{identifier:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_shield(self, identifier: str) -> None:
"""Unregister a shield.

3
llama_stack/apis/synthetic_data_generation/synthetic_data_generation.py
View file

@ -10,6 +10,7 @@ from typing import Any, Protocol
from pydantic import BaseModel
from llama_stack.apis.inference import Message
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.schema_utils import json_schema_type, webmethod
@ -59,7 +60,7 @@ class SyntheticDataGenerationResponse(BaseModel):
class SyntheticDataGeneration(Protocol):
@webmethod(route="/synthetic-data-generation/generate")
@webmethod(route="/synthetic-data-generation/generate", level=LLAMA_STACK_API_V1)
def synthetic_data_generate(
self,
dialogs: list[Message],

29
llama_stack/apis/telemetry/telemetry.py
View file

@ -16,6 +16,7 @@ from typing import (
from pydantic import BaseModel, Field
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.models.llama.datatypes import Primitive
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -412,7 +413,7 @@ class QueryMetricsResponse(BaseModel):
@runtime_checkable
class Telemetry(Protocol):
@webmethod(route="/telemetry/events", method="POST")
@webmethod(route="/telemetry/events", method="POST", level=LLAMA_STACK_API_V1)
async def log_event(
self,
event: Event,
@ -425,7 +426,7 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/traces", method="POST", required_scope=REQUIRED_SCOPE)
@webmethod(route="/telemetry/traces", method="POST", required_scope=REQUIRED_SCOPE, level=LLAMA_STACK_API_V1)
async def query_traces(
self,
attribute_filters: list[QueryCondition] | None = None,
@ -443,7 +444,9 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/traces/{trace_id:path}", method="GET", required_scope=REQUIRED_SCOPE)
@webmethod(
route="/telemetry/traces/{trace_id:path}", method="GET", required_scope=REQUIRED_SCOPE, level=LLAMA_STACK_API_V1
)
async def get_trace(self, trace_id: str) -> Trace:
"""Get a trace by its ID.
@ -453,7 +456,10 @@ class Telemetry(Protocol):
...
@webmethod(
route="/telemetry/traces/{trace_id:path}/spans/{span_id:path}", method="GET", required_scope=REQUIRED_SCOPE
route="/telemetry/traces/{trace_id:path}/spans/{span_id:path}",
method="GET",
required_scope=REQUIRED_SCOPE,
level=LLAMA_STACK_API_V1,
)
async def get_span(self, trace_id: str, span_id: str) -> Span:
"""Get a span by its ID.
@ -464,7 +470,12 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/spans/{span_id:path}/tree", method="POST", required_scope=REQUIRED_SCOPE)
@webmethod(
route="/telemetry/spans/{span_id:path}/tree",
method="POST",
required_scope=REQUIRED_SCOPE,
level=LLAMA_STACK_API_V1,
)
async def get_span_tree(
self,
span_id: str,
@ -480,7 +491,7 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/spans", method="POST", required_scope=REQUIRED_SCOPE)
@webmethod(route="/telemetry/spans", method="POST", required_scope=REQUIRED_SCOPE, level=LLAMA_STACK_API_V1)
async def query_spans(
self,
attribute_filters: list[QueryCondition],
@ -496,7 +507,7 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/spans/export", method="POST")
@webmethod(route="/telemetry/spans/export", method="POST", level=LLAMA_STACK_API_V1)
async def save_spans_to_dataset(
self,
attribute_filters: list[QueryCondition],
@ -513,7 +524,9 @@ class Telemetry(Protocol):
"""
...
@webmethod(route="/telemetry/metrics/{metric_name}", method="POST", required_scope=REQUIRED_SCOPE)
@webmethod(
route="/telemetry/metrics/{metric_name}", method="POST", required_scope=REQUIRED_SCOPE, level=LLAMA_STACK_API_V1
)
async def query_metrics(
self,
metric_name: str,

5
llama_stack/apis/tools/rag_tool.py
View file

@ -11,6 +11,7 @@ from pydantic import BaseModel, Field, field_validator
from typing_extensions import runtime_checkable
from llama_stack.apis.common.content_types import URL, InterleavedContent
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -185,7 +186,7 @@ class RAGQueryConfig(BaseModel):
@runtime_checkable
@trace_protocol
class RAGToolRuntime(Protocol):
@webmethod(route="/tool-runtime/rag-tool/insert", method="POST")
@webmethod(route="/tool-runtime/rag-tool/insert", method="POST", level=LLAMA_STACK_API_V1)
async def insert(
self,
documents: list[RAGDocument],
@ -200,7 +201,7 @@ class RAGToolRuntime(Protocol):
"""
...
@webmethod(route="/tool-runtime/rag-tool/query", method="POST")
@webmethod(route="/tool-runtime/rag-tool/query", method="POST", level=LLAMA_STACK_API_V1)
async def query(
self,
content: InterleavedContent,

17
llama_stack/apis/tools/tools.py
View file

@ -12,6 +12,7 @@ from typing_extensions import runtime_checkable
from llama_stack.apis.common.content_types import URL, InterleavedContent
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -151,7 +152,7 @@ class ListToolDefsResponse(BaseModel):
@runtime_checkable
@trace_protocol
class ToolGroups(Protocol):
@webmethod(route="/toolgroups", method="POST")
@webmethod(route="/toolgroups", method="POST", level=LLAMA_STACK_API_V1)
async def register_tool_group(
self,
toolgroup_id: str,
@ -168,7 +169,7 @@ class ToolGroups(Protocol):
"""
...
@webmethod(route="/toolgroups/{toolgroup_id:path}", method="GET")
@webmethod(route="/toolgroups/{toolgroup_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_tool_group(
self,
toolgroup_id: str,
@ -180,7 +181,7 @@ class ToolGroups(Protocol):
"""
...
@webmethod(route="/toolgroups", method="GET")
@webmethod(route="/toolgroups", method="GET", level=LLAMA_STACK_API_V1)
async def list_tool_groups(self) -> ListToolGroupsResponse:
"""List tool groups with optional provider.
@ -188,7 +189,7 @@ class ToolGroups(Protocol):
"""
...
@webmethod(route="/tools", method="GET")
@webmethod(route="/tools", method="GET", level=LLAMA_STACK_API_V1)
async def list_tools(self, toolgroup_id: str | None = None) -> ListToolsResponse:
"""List tools with optional tool group.
@ -197,7 +198,7 @@ class ToolGroups(Protocol):
"""
...
@webmethod(route="/tools/{tool_name:path}", method="GET")
@webmethod(route="/tools/{tool_name:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_tool(
self,
tool_name: str,
@ -209,7 +210,7 @@ class ToolGroups(Protocol):
"""
...
@webmethod(route="/toolgroups/{toolgroup_id:path}", method="DELETE")
@webmethod(route="/toolgroups/{toolgroup_id:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_toolgroup(
self,
toolgroup_id: str,
@ -238,7 +239,7 @@ class ToolRuntime(Protocol):
rag_tool: RAGToolRuntime | None = None
# TODO: This needs to be renamed once OPEN API generator name conflict issue is fixed.
@webmethod(route="/tool-runtime/list-tools", method="GET")
@webmethod(route="/tool-runtime/list-tools", method="GET", level=LLAMA_STACK_API_V1)
async def list_runtime_tools(
self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
) -> ListToolDefsResponse:
@ -250,7 +251,7 @@ class ToolRuntime(Protocol):
"""
...
@webmethod(route="/tool-runtime/invoke", method="POST")
@webmethod(route="/tool-runtime/invoke", method="POST", level=LLAMA_STACK_API_V1)
async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> ToolInvocationResult:
"""Run a tool with the given arguments.

9
llama_stack/apis/vector_dbs/vector_dbs.py
View file

@ -9,6 +9,7 @@ from typing import Literal, Protocol, runtime_checkable
from pydantic import BaseModel
from llama_stack.apis.resource import Resource, ResourceType
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.schema_utils import json_schema_type, webmethod
@ -65,7 +66,7 @@ class ListVectorDBsResponse(BaseModel):
@runtime_checkable
@trace_protocol
class VectorDBs(Protocol):
@webmethod(route="/vector-dbs", method="GET")
@webmethod(route="/vector-dbs", method="GET", level=LLAMA_STACK_API_V1)
async def list_vector_dbs(self) -> ListVectorDBsResponse:
"""List all vector databases.
@ -73,7 +74,7 @@ class VectorDBs(Protocol):
"""
...
@webmethod(route="/vector-dbs/{vector_db_id:path}", method="GET")
@webmethod(route="/vector-dbs/{vector_db_id:path}", method="GET", level=LLAMA_STACK_API_V1)
async def get_vector_db(
self,
vector_db_id: str,
@ -85,7 +86,7 @@ class VectorDBs(Protocol):
"""
...
@webmethod(route="/vector-dbs", method="POST")
@webmethod(route="/vector-dbs", method="POST", level=LLAMA_STACK_API_V1)
async def register_vector_db(
self,
vector_db_id: str,
@ -107,7 +108,7 @@ class VectorDBs(Protocol):
"""
...
@webmethod(route="/vector-dbs/{vector_db_id:path}", method="DELETE")
@webmethod(route="/vector-dbs/{vector_db_id:path}", method="DELETE", level=LLAMA_STACK_API_V1)
async def unregister_vector_db(self, vector_db_id: str) -> None:
"""Unregister a vector database.

39
llama_stack/apis/vector_io/vector_io.py
View file

@ -15,6 +15,7 @@ from pydantic import BaseModel, Field
from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.version import LLAMA_STACK_API_V1
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
from llama_stack.providers.utils.vector_io.vector_utils import generate_chunk_id
from llama_stack.schema_utils import json_schema_type, webmethod
@ -437,7 +438,7 @@ class VectorIO(Protocol):
# this will just block now until chunks are inserted, but it should
# probably return a Job instance which can be polled for completion
@webmethod(route="/vector-io/insert", method="POST")
@webmethod(route="/vector-io/insert", method="POST", level=LLAMA_STACK_API_V1)
async def insert_chunks(
self,
vector_db_id: str,
@ -455,7 +456,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/vector-io/query", method="POST")
@webmethod(route="/vector-io/query", method="POST", level=LLAMA_STACK_API_V1)
async def query_chunks(
self,
vector_db_id: str,
@ -472,7 +473,7 @@ class VectorIO(Protocol):
...
# OpenAI Vector Stores API endpoints
@webmethod(route="/openai/v1/vector_stores", method="POST")
@webmethod(route="/openai/v1/vector_stores", method="POST", level=LLAMA_STACK_API_V1)
async def openai_create_vector_store(
self,
name: str | None = None,
@ -498,7 +499,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores", method="GET")
@webmethod(route="/openai/v1/vector_stores", method="GET", level=LLAMA_STACK_API_V1)
async def openai_list_vector_stores(
self,
limit: int | None = 20,
@ -516,7 +517,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="GET")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="GET", level=LLAMA_STACK_API_V1)
async def openai_retrieve_vector_store(
self,
vector_store_id: str,
@ -528,7 +529,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="POST")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="POST", level=LLAMA_STACK_API_V1)
async def openai_update_vector_store(
self,
vector_store_id: str,
@ -546,7 +547,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="DELETE")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}", method="DELETE", level=LLAMA_STACK_API_V1)
async def openai_delete_vector_store(
self,
vector_store_id: str,
@ -558,7 +559,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/search", method="POST")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/search", method="POST", level=LLAMA_STACK_API_V1)
async def openai_search_vector_store(
self,
vector_store_id: str,
@ -584,7 +585,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files", method="POST")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files", method="POST", level=LLAMA_STACK_API_V1)
async def openai_attach_file_to_vector_store(
self,
vector_store_id: str,
@ -602,7 +603,7 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files", method="GET")
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files", method="GET", level=LLAMA_STACK_API_V1)
async def openai_list_files_in_vector_store(
self,
vector_store_id: str,
@ -624,7 +625,9 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="GET")
@webmethod(
route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="GET", level=LLAMA_STACK_API_V1
)
async def openai_retrieve_vector_store_file(
self,
vector_store_id: str,
@ -638,7 +641,11 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}/content", method="GET")
@webmethod(
route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}/content",
method="GET",
level=LLAMA_STACK_API_V1,
)
async def openai_retrieve_vector_store_file_contents(
self,
vector_store_id: str,
@ -652,7 +659,9 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="POST")
@webmethod(
route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="POST", level=LLAMA_STACK_API_V1
)
async def openai_update_vector_store_file(
self,
vector_store_id: str,
@ -668,7 +677,9 @@ class VectorIO(Protocol):
"""
...
@webmethod(route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="DELETE")
@webmethod(
route="/openai/v1/vector_stores/{vector_store_id}/files/{file_id}", method="DELETE", level=LLAMA_STACK_API_V1
)
async def openai_delete_vector_store_file(
self,
vector_store_id: str,

4
llama_stack/apis/version.py
View file

@ -4,4 +4,6 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
LLAMA_STACK_API_VERSION = "v1"
LLAMA_STACK_API_V1 = "v1"
LLAMA_STACK_API_V1BETA = "v1beta"
LLAMA_STACK_API_V1ALPHA = "v1alpha"

7
llama_stack/cli/stack/_build.py
View file

@ -45,6 +45,7 @@ from llama_stack.core.utils.dynamic import instantiate_class_type
from llama_stack.core.utils.exec import formulate_run_args, run_command
from llama_stack.core.utils.image_types import LlamaStackImageType
from llama_stack.providers.datatypes import Api
from llama_stack.providers.utils.sqlstore.sqlstore import SqliteSqlStoreConfig
DISTRIBS_PATH = Path(__file__).parent.parent.parent / "distributions"
@ -294,6 +295,12 @@ def _generate_run_config(
if build_config.external_providers_dir
else EXTERNAL_PROVIDERS_DIR,
)
if not run_config.inference_store:
run_config.inference_store = SqliteSqlStoreConfig(
**SqliteSqlStoreConfig.sample_run_config(
__distro_dir__=(DISTRIBS_BASE_DIR / image_name).as_posix(), db_name="inference_store.db"
)
)
# build providers dict
provider_registry = get_provider_registry(build_config)
for api in apis:

17
llama_stack/cli/verify_download.py
View file

@ -48,15 +48,12 @@ def setup_verify_download_parser(parser: argparse.ArgumentParser) -> None:
parser.set_defaults(func=partial(run_verify_cmd, parser=parser))
def calculate_md5(filepath: Path, chunk_size: int = 8192) -> str:
# NOTE: MD5 is used here only for download integrity verification,
# not for security purposes
# TODO: switch to SHA256
md5_hash = hashlib.md5(usedforsecurity=False)
def calculate_sha256(filepath: Path, chunk_size: int = 8192) -> str:
sha256_hash = hashlib.sha256()
with open(filepath, "rb") as f:
for chunk in iter(lambda: f.read(chunk_size), b""):
md5_hash.update(chunk)
return md5_hash.hexdigest()
sha256_hash.update(chunk)
return sha256_hash.hexdigest()
def load_checksums(checklist_path: Path) -> dict[str, str]:
@ -64,10 +61,10 @@ def load_checksums(checklist_path: Path) -> dict[str, str]:
with open(checklist_path) as f:
for line in f:
if line.strip():
md5sum, filepath = line.strip().split(" ", 1)
sha256sum, filepath = line.strip().split(" ", 1)
# Remove leading './' if present
filepath = filepath.lstrip("./")
checksums[filepath] = md5sum
checksums[filepath] = sha256sum
return checksums
@ -88,7 +85,7 @@ def verify_files(model_dir: Path, checksums: dict[str, str], console: Console) -
matches = False
if exists:
actual_hash = calculate_md5(full_path)
actual_hash = calculate_sha256(full_path)
matches = actual_hash == expected_hash
results.append(

4
llama_stack/core/build_container.sh
View file

@ -147,7 +147,7 @@ WORKDIR /app
RUN dnf -y update && dnf install -y iputils git net-tools wget \
vim-minimal python3.12 python3.12-pip python3.12-wheel \
python3.12-setuptools python3.12-devel gcc make && \
python3.12-setuptools python3.12-devel gcc gcc-c++ make && \
ln -s /bin/pip3.12 /bin/pip && ln -s /bin/python3.12 /bin/python && dnf clean all
ENV UV_SYSTEM_PYTHON=1
@ -164,7 +164,7 @@ RUN apt-get update && apt-get install -y \
procps psmisc lsof \
traceroute \
bubblewrap \
gcc \
gcc g++ \
&& rm -rf /var/lib/apt/lists/*
ENV UV_SYSTEM_PYTHON=1

20
llama_stack/core/client.py
View file

@ -15,7 +15,6 @@ import httpx
from pydantic import BaseModel, parse_obj_as
from termcolor import cprint
from llama_stack.apis.version import LLAMA_STACK_API_VERSION
from llama_stack.providers.datatypes import RemoteProviderConfig
_CLIENT_CLASSES = {}
@ -114,7 +113,24 @@ def create_api_client_class(protocol) -> type:
break
kwargs[param.name] = args[i]
url = f"{self.base_url}/{LLAMA_STACK_API_VERSION}/{webmethod.route.lstrip('/')}"
# Get all webmethods for this method (supports multiple decorators)
webmethods = getattr(method, "__webmethods__", [])
if not webmethods:
raise RuntimeError(f"Method {method} has no webmethod decorators")
# Choose the preferred webmethod (non-deprecated if available)
preferred_webmethod = None
for wm in webmethods:
if not getattr(wm, "deprecated", False):
preferred_webmethod = wm
break
# If no non-deprecated found, use the first one
if preferred_webmethod is None:
preferred_webmethod = webmethods[0]
url = f"{self.base_url}/{preferred_webmethod.level}/{preferred_webmethod.route.lstrip('/')}"
def convert(value):
if isinstance(value, list):

58
llama_stack/core/datatypes.py
View file

@ -7,6 +7,7 @@
from enum import StrEnum
from pathlib import Path
from typing import Annotated, Any, Literal, Self
from urllib.parse import urlparse
from pydantic import BaseModel, Field, field_validator, model_validator
@ -120,10 +121,6 @@ class AutoRoutedProviderSpec(ProviderSpec):
default=None,
)
@property
def pip_packages(self) -> list[str]:
raise AssertionError("Should not be called on AutoRoutedProviderSpec")
# Example: /models, /shields
class RoutingTableProviderSpec(ProviderSpec):
@ -212,6 +209,7 @@ class AuthProviderType(StrEnum):
OAUTH2_TOKEN = "oauth2_token"
GITHUB_TOKEN = "github_token"
CUSTOM = "custom"
KUBERNETES = "kubernetes"
class OAuth2TokenAuthConfig(BaseModel):
@ -282,8 +280,45 @@ class GitHubTokenAuthConfig(BaseModel):
)
class KubernetesAuthProviderConfig(BaseModel):
"""Configuration for Kubernetes authentication provider."""
type: Literal[AuthProviderType.KUBERNETES] = AuthProviderType.KUBERNETES
api_server_url: str = Field(
default="https://kubernetes.default.svc",
description="Kubernetes API server URL (e.g., https://api.cluster.domain:6443)",
)
verify_tls: bool = Field(default=True, description="Whether to verify TLS certificates")
tls_cafile: Path | None = Field(default=None, description="Path to CA certificate file for TLS verification")
claims_mapping: dict[str, str] = Field(
default_factory=lambda: {
"username": "roles",
"groups": "roles",
},
description="Mapping of Kubernetes user claims to access attributes",
)
@field_validator("api_server_url")
@classmethod
def validate_api_server_url(cls, v):
parsed = urlparse(v)
if not parsed.scheme or not parsed.netloc:
raise ValueError(f"api_server_url must be a valid URL with scheme and host: {v}")
if parsed.scheme not in ["http", "https"]:
raise ValueError(f"api_server_url scheme must be http or https: {v}")
return v
@field_validator("claims_mapping")
@classmethod
def validate_claims_mapping(cls, v):
for key, value in v.items():
if not value:
raise ValueError(f"claims_mapping value cannot be empty: {key}")
return v
AuthProviderConfig = Annotated[
OAuth2TokenAuthConfig | GitHubTokenAuthConfig | CustomAuthConfig,
OAuth2TokenAuthConfig | GitHubTokenAuthConfig | CustomAuthConfig | KubernetesAuthProviderConfig,
Field(discriminator="type"),
]
@ -392,6 +427,12 @@ class ServerConfig(BaseModel):
)
class InferenceStoreConfig(BaseModel):
sql_store_config: SqlStoreConfig
max_write_queue_size: int = Field(default=10000, description="Max queued writes for inference store")
num_writers: int = Field(default=4, description="Number of concurrent background writers")
class StackRunConfig(BaseModel):
version: int = LLAMA_STACK_RUN_CONFIG_VERSION
@ -425,11 +466,12 @@ Configuration for the persistence store used by the distribution registry. If no
a default SQLite store will be used.""",
)
inference_store: SqlStoreConfig | None = Field(
inference_store: InferenceStoreConfig | SqlStoreConfig | None = Field(
default=None,
description="""
Configuration for the persistence store used by the inference API. If not specified,
a default SQLite store will be used.""",
Configuration for the persistence store used by the inference API. Can be either a
InferenceStoreConfig (with queue tuning parameters) or a SqlStoreConfig (deprecated).
If not specified, a default SQLite store will be used.""",
)
# registry of "resources" in the distribution

27
llama_stack/core/distribution.py
View file

@ -16,16 +16,18 @@ from llama_stack.core.datatypes import BuildConfig, DistributionSpec
from llama_stack.core.external import load_external_apis
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
logger = get_logger(name=__name__, category="core")
INTERNAL_APIS = {Api.inspect, Api.providers, Api.prompts}
def stack_apis() -> list[Api]:
return list(Api)
@ -70,31 +72,16 @@ def builtin_automatically_routed_apis() -> list[AutoRoutedApiInfo]:
def providable_apis() -> list[Api]:
routing_table_apis = {x.routing_table_api for x in builtin_automatically_routed_apis()}
return [api for api in Api if api not in routing_table_apis and api != Api.inspect and api != Api.providers]
return [api for api in Api if api not in routing_table_apis and api not in INTERNAL_APIS]
def _load_remote_provider_spec(spec_data: dict[str, Any], api: Api) -> ProviderSpec:
adapter = AdapterSpec(**spec_data["adapter"])
spec = remote_provider_spec(
api=api,
adapter=adapter,
api_dependencies=[Api(dep) for dep in spec_data.get("api_dependencies", [])],
)
spec = RemoteProviderSpec(api=api, provider_type=f"remote::{spec_data['adapter_type']}", **spec_data)
return spec
def _load_inline_provider_spec(spec_data: dict[str, Any], api: Api, provider_name: str) -> ProviderSpec:
spec = InlineProviderSpec(
api=api,
provider_type=f"inline::{provider_name}",
pip_packages=spec_data.get("pip_packages", []),
module=spec_data["module"],
config_class=spec_data["config_class"],
api_dependencies=[Api(dep) for dep in spec_data.get("api_dependencies", [])],
optional_api_dependencies=[Api(dep) for dep in spec_data.get("optional_api_dependencies", [])],
provider_data_validator=spec_data.get("provider_data_validator"),
container_image=spec_data.get("container_image"),
)
spec = InlineProviderSpec(api=api, provider_type=f"inline::{provider_name}", **spec_data)
return spec

10
llama_stack/core/library_client.py
View file

@ -10,7 +10,6 @@ import json
import logging # allow-direct-logging
import os
import sys
from concurrent.futures import ThreadPoolExecutor
from enum import Enum
from io import BytesIO
from pathlib import Path
@ -41,7 +40,7 @@ from llama_stack.core.request_headers import (
from llama_stack.core.resolver import ProviderRegistry
from llama_stack.core.server.routes import RouteImpls, find_matching_route, initialize_route_impls
from llama_stack.core.stack import (
construct_stack,
Stack,
get_stack_run_config_from_distro,
replace_env_vars,
)
@ -148,7 +147,6 @@ class LlamaStackAsLibraryClient(LlamaStackClient):
self.async_client = AsyncLlamaStackAsLibraryClient(
config_path_or_distro_name, custom_provider_registry, provider_data, skip_logger_removal
)
self.pool_executor = ThreadPoolExecutor(max_workers=4)
self.provider_data = provider_data
self.loop = asyncio.new_event_loop()
@ -254,7 +252,10 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
try:
self.route_impls = None
self.impls = await construct_stack(self.config, self.custom_provider_registry)
stack = Stack(self.config, self.custom_provider_registry)
await stack.initialize()
self.impls = stack.impls
except ModuleNotFoundError as _e:
cprint(_e.msg, color="red", file=sys.stderr)
cprint(
@ -291,6 +292,7 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
)
raise _e
assert self.impls is not None
if Api.telemetry in self.impls:
setup_logger(self.impls[Api.telemetry])

0
llama_stack/providers/inline/scoring/basic/utils/bfcl/__init__.py → llama_stack/core/prompts/__init__.py
View file

233
llama_stack/core/prompts/prompts.py Normal file
View file

@ -0,0 +1,233 @@
# 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 json
from typing import Any
from pydantic import BaseModel
from llama_stack.apis.prompts import ListPromptsResponse, Prompt, Prompts
from llama_stack.core.datatypes import StackRunConfig
from llama_stack.core.utils.config_dirs import DISTRIBS_BASE_DIR
from llama_stack.providers.utils.kvstore import KVStore, kvstore_impl
from llama_stack.providers.utils.kvstore.config import SqliteKVStoreConfig
class PromptServiceConfig(BaseModel):
"""Configuration for the built-in prompt service.
:param run_config: Stack run configuration containing distribution info
"""
run_config: StackRunConfig
async def get_provider_impl(config: PromptServiceConfig, deps: dict[Any, Any]):
"""Get the prompt service implementation."""
impl = PromptServiceImpl(config, deps)
await impl.initialize()
return impl
class PromptServiceImpl(Prompts):
"""Built-in prompt service implementation using KVStore."""
def __init__(self, config: PromptServiceConfig, deps: dict[Any, Any]):
self.config = config
self.deps = deps
self.kvstore: KVStore
async def initialize(self) -> None:
kvstore_config = SqliteKVStoreConfig(
db_path=(DISTRIBS_BASE_DIR / self.config.run_config.image_name / "prompts.db").as_posix()
)
self.kvstore = await kvstore_impl(kvstore_config)
def _get_default_key(self, prompt_id: str) -> str:
"""Get the KVStore key that stores the default version number."""
return f"prompts:v1:{prompt_id}:default"
async def _get_prompt_key(self, prompt_id: str, version: int | None = None) -> str:
"""Get the KVStore key for prompt data, returning default version if applicable."""
if version:
return self._get_version_key(prompt_id, str(version))
default_key = self._get_default_key(prompt_id)
resolved_version = await self.kvstore.get(default_key)
if resolved_version is None:
raise ValueError(f"Prompt {prompt_id}:default not found")
return self._get_version_key(prompt_id, resolved_version)
def _get_version_key(self, prompt_id: str, version: str) -> str:
"""Get the KVStore key for a specific prompt version."""
return f"prompts:v1:{prompt_id}:{version}"
def _get_list_key_prefix(self) -> str:
"""Get the key prefix for listing prompts."""
return "prompts:v1:"
def _serialize_prompt(self, prompt: Prompt) -> str:
"""Serialize a prompt to JSON string for storage."""
return json.dumps(
{
"prompt_id": prompt.prompt_id,
"prompt": prompt.prompt,
"version": prompt.version,
"variables": prompt.variables or [],
"is_default": prompt.is_default,
}
)
def _deserialize_prompt(self, data: str) -> Prompt:
"""Deserialize a prompt from JSON string."""
obj = json.loads(data)
return Prompt(
prompt_id=obj["prompt_id"],
prompt=obj["prompt"],
version=obj["version"],
variables=obj.get("variables", []),
is_default=obj.get("is_default", False),
)
async def list_prompts(self) -> ListPromptsResponse:
"""List all prompts (default versions only)."""
prefix = self._get_list_key_prefix()
keys = await self.kvstore.keys_in_range(prefix, prefix + "\xff")
prompts = []
for key in keys:
if key.endswith(":default"):
try:
default_version = await self.kvstore.get(key)
if default_version:
prompt_id = key.replace(prefix, "").replace(":default", "")
version_key = self._get_version_key(prompt_id, default_version)
data = await self.kvstore.get(version_key)
if data:
prompt = self._deserialize_prompt(data)
prompts.append(prompt)
except (json.JSONDecodeError, KeyError):
continue
prompts.sort(key=lambda p: p.prompt_id or "", reverse=True)
return ListPromptsResponse(data=prompts)
async def get_prompt(self, prompt_id: str, version: int | None = None) -> Prompt:
"""Get a prompt by its identifier and optional version."""
key = await self._get_prompt_key(prompt_id, version)
data = await self.kvstore.get(key)
if data is None:
raise ValueError(f"Prompt {prompt_id}:{version if version else 'default'} not found")
return self._deserialize_prompt(data)
async def create_prompt(
self,
prompt: str,
variables: list[str] | None = None,
) -> Prompt:
"""Create a new prompt."""
if variables is None:
variables = []
prompt_obj = Prompt(
prompt_id=Prompt.generate_prompt_id(),
prompt=prompt,
version=1,
variables=variables,
)
version_key = self._get_version_key(prompt_obj.prompt_id, str(prompt_obj.version))
data = self._serialize_prompt(prompt_obj)
await self.kvstore.set(version_key, data)
default_key = self._get_default_key(prompt_obj.prompt_id)
await self.kvstore.set(default_key, str(prompt_obj.version))
return prompt_obj
async def update_prompt(
self,
prompt_id: str,
prompt: str,
version: int,
variables: list[str] | None = None,
set_as_default: bool = True,
) -> Prompt:
"""Update an existing prompt (increments version)."""
if version < 1:
raise ValueError("Version must be >= 1")
if variables is None:
variables = []
prompt_versions = await self.list_prompt_versions(prompt_id)
latest_prompt = max(prompt_versions.data, key=lambda x: int(x.version))
if version and latest_prompt.version != version:
raise ValueError(
f"'{version}' is not the latest prompt version for prompt_id='{prompt_id}'. Use the latest version '{latest_prompt.version}' in request."
)
current_version = latest_prompt.version if version is None else version
new_version = current_version + 1
updated_prompt = Prompt(prompt_id=prompt_id, prompt=prompt, version=new_version, variables=variables)
version_key = self._get_version_key(prompt_id, str(new_version))
data = self._serialize_prompt(updated_prompt)
await self.kvstore.set(version_key, data)
if set_as_default:
await self.set_default_version(prompt_id, new_version)
return updated_prompt
async def delete_prompt(self, prompt_id: str) -> None:
"""Delete a prompt and all its versions."""
await self.get_prompt(prompt_id)
prefix = f"prompts:v1:{prompt_id}:"
keys = await self.kvstore.keys_in_range(prefix, prefix + "\xff")
for key in keys:
await self.kvstore.delete(key)
async def list_prompt_versions(self, prompt_id: str) -> ListPromptsResponse:
"""List all versions of a specific prompt."""
prefix = f"prompts:v1:{prompt_id}:"
keys = await self.kvstore.keys_in_range(prefix, prefix + "\xff")
default_version = None
prompts = []
for key in keys:
data = await self.kvstore.get(key)
if key.endswith(":default"):
default_version = data
else:
if data:
prompt_obj = self._deserialize_prompt(data)
prompts.append(prompt_obj)
if not prompts:
raise ValueError(f"Prompt {prompt_id} not found")
for prompt in prompts:
prompt.is_default = str(prompt.version) == default_version
prompts.sort(key=lambda x: x.version)
return ListPromptsResponse(data=prompts)
async def set_default_version(self, prompt_id: str, version: int) -> Prompt:
"""Set which version of a prompt should be the default, If not set. the default is the latest."""
version_key = self._get_version_key(prompt_id, str(version))
data = await self.kvstore.get(version_key)
if data is None:
raise ValueError(f"Prompt {prompt_id} version {version} not found")
default_key = self._get_default_key(prompt_id)
await self.kvstore.set(default_key, str(version))
return self._deserialize_prompt(data)

12
llama_stack/core/resolver.py
View file

@ -19,6 +19,7 @@ from llama_stack.apis.inference import Inference, InferenceProvider
from llama_stack.apis.inspect import Inspect
from llama_stack.apis.models import Models
from llama_stack.apis.post_training import PostTraining
from llama_stack.apis.prompts import Prompts
from llama_stack.apis.providers import Providers as ProvidersAPI
from llama_stack.apis.safety import Safety
from llama_stack.apis.scoring import Scoring
@ -93,6 +94,7 @@ def api_protocol_map(external_apis: dict[Api, ExternalApiSpec] | None = None) ->
Api.tool_groups: ToolGroups,
Api.tool_runtime: ToolRuntime,
Api.files: Files,
Api.prompts: Prompts,
}
if external_apis:
@ -284,7 +286,15 @@ async def instantiate_providers(
if provider.provider_id is None:
continue
deps = {a: impls[a] for a in provider.spec.api_dependencies}
try:
deps = {a: impls[a] for a in provider.spec.api_dependencies}
except KeyError as e:
missing_api = e.args[0]
raise RuntimeError(
f"Failed to resolve '{provider.spec.api.value}' provider '{provider.provider_id}' of type '{provider.spec.provider_type}': "
f"required dependency '{missing_api.value}' is not available. "
f"Please add a '{missing_api.value}' provider to your configuration or check if the provider is properly configured."
) from e
for a in provider.spec.optional_api_dependencies:
if a in impls:
deps[a] = impls[a]

5
llama_stack/core/routers/__init__.py
View file

@ -78,7 +78,10 @@ async def get_auto_router_impl(
# TODO: move pass configs to routers instead
if api == Api.inference and run_config.inference_store:
inference_store = InferenceStore(run_config.inference_store, policy)
inference_store = InferenceStore(
config=run_config.inference_store,
policy=policy,
)
await inference_store.initialize()
api_to_dep_impl["store"] = inference_store

25
llama_stack/core/routers/inference.py
View file

@ -61,7 +61,7 @@ from llama_stack.models.llama.llama3.chat_format import ChatFormat
from llama_stack.models.llama.llama3.tokenizer import Tokenizer
from llama_stack.providers.datatypes import HealthResponse, HealthStatus, RoutingTable
from llama_stack.providers.utils.inference.inference_store import InferenceStore
from llama_stack.providers.utils.telemetry.tracing import get_current_span
from llama_stack.providers.utils.telemetry.tracing import enqueue_event, get_current_span
logger = get_logger(name=__name__, category="core::routers")
@ -88,6 +88,11 @@ class InferenceRouter(Inference):
async def shutdown(self) -> None:
logger.debug("InferenceRouter.shutdown")
if self.store:
try:
await self.store.shutdown()
except Exception as e:
logger.warning(f"Error during InferenceStore shutdown: {e}")
async def register_model(
self,
@ -158,7 +163,7 @@ class InferenceRouter(Inference):
metrics = self._construct_metrics(prompt_tokens, completion_tokens, total_tokens, model)
if self.telemetry:
for metric in metrics:
await self.telemetry.log_event(metric)
enqueue_event(metric)
return [MetricInResponse(metric=metric.metric, value=metric.value) for metric in metrics]
async def _count_tokens(
@ -391,7 +396,7 @@ class InferenceRouter(Inference):
model=model_obj,
)
for metric in metrics:
await self.telemetry.log_event(metric)
enqueue_event(metric)
# these metrics will show up in the client response.
response.metrics = (
@ -487,7 +492,7 @@ class InferenceRouter(Inference):
# Store the response with the ID that will be returned to the client
if self.store:
await self.store.store_chat_completion(response, messages)
asyncio.create_task(self.store.store_chat_completion(response, messages))
if self.telemetry:
metrics = self._construct_metrics(
@ -497,7 +502,7 @@ class InferenceRouter(Inference):
model=model_obj,
)
for metric in metrics:
await self.telemetry.log_event(metric)
enqueue_event(metric)
# these metrics will show up in the client response.
response.metrics = (
metrics if not hasattr(response, "metrics") or response.metrics is None else response.metrics + metrics
@ -624,7 +629,7 @@ class InferenceRouter(Inference):
"completion_tokens",
"total_tokens",
]: # Only log completion and total tokens
await self.telemetry.log_event(metric)
enqueue_event(metric)
# Return metrics in response
async_metrics = [
@ -670,7 +675,7 @@ class InferenceRouter(Inference):
)
for metric in completion_metrics:
if metric.metric in ["completion_tokens", "total_tokens"]: # Only log completion and total tokens
await self.telemetry.log_event(metric)
enqueue_event(metric)
# Return metrics in response
return [MetricInResponse(metric=metric.metric, value=metric.value) for metric in completion_metrics]
@ -715,7 +720,7 @@ class InferenceRouter(Inference):
choices_data[idx] = {
"content_parts": [],
"tool_calls_builder": {},
"finish_reason": None,
"finish_reason": "stop",
"logprobs_content_parts": [],
}
current_choice_data = choices_data[idx]
@ -766,7 +771,7 @@ class InferenceRouter(Inference):
model=model,
)
for metric in metrics:
await self.telemetry.log_event(metric)
enqueue_event(metric)
yield chunk
finally:
@ -815,4 +820,4 @@ class InferenceRouter(Inference):
object="chat.completion",
)
logger.debug(f"InferenceRouter.completion_response: {final_response}")
await self.store.store_chat_completion(final_response, messages)
asyncio.create_task(self.store.store_chat_completion(final_response, messages))

4
llama_stack/core/routing_tables/benchmarks.py
View file

@ -56,3 +56,7 @@ class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
provider_resource_id=provider_benchmark_id,
)
await self.register_object(benchmark)
async def unregister_benchmark(self, benchmark_id: str) -> None:
existing_benchmark = await self.get_benchmark(benchmark_id)
await self.unregister_object(existing_benchmark)

4
llama_stack/core/routing_tables/common.py
View file

@ -64,6 +64,10 @@ async def unregister_object_from_provider(obj: RoutableObject, p: Any) -> None:
return await p.unregister_shield(obj.identifier)
elif api == Api.datasetio:
return await p.unregister_dataset(obj.identifier)
elif api == Api.eval:
return await p.unregister_benchmark(obj.identifier)
elif api == Api.scoring:
return await p.unregister_scoring_function(obj.identifier)
elif api == Api.tool_runtime:
return await p.unregister_toolgroup(obj.identifier)
else:

2
llama_stack/core/routing_tables/models.py
View file

@ -33,7 +33,7 @@ class ModelsRoutingTable(CommonRoutingTableImpl, Models):
try:
models = await provider.list_models()
except Exception as e:
logger.exception(f"Model refresh failed for provider {provider_id}: {e}")
logger.warning(f"Model refresh failed for provider {provider_id}: {e}")
continue
self.listed_providers.add(provider_id)

4
llama_stack/core/routing_tables/scoring_functions.py
View file

@ -60,3 +60,7 @@ class ScoringFunctionsRoutingTable(CommonRoutingTableImpl, ScoringFunctions):
)
scoring_fn.provider_id = provider_id
await self.register_object(scoring_fn)
async def unregister_scoring_function(self, scoring_fn_id: str) -> None:
existing_scoring_fn = await self.get_scoring_function(scoring_fn_id)
await self.unregister_object(existing_scoring_fn)

26
llama_stack/core/routing_tables/vector_dbs.py
View file

@ -52,7 +52,6 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
provider_vector_db_id: str | None = None,
vector_db_name: str | None = None,
) -> VectorDB:
provider_vector_db_id = provider_vector_db_id or vector_db_id
if provider_id is None:
if len(self.impls_by_provider_id) > 0:
provider_id = list(self.impls_by_provider_id.keys())[0]
@ -69,14 +68,33 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
raise ModelTypeError(embedding_model, model.model_type, ModelType.embedding)
if "embedding_dimension" not in model.metadata:
raise ValueError(f"Model {embedding_model} does not have an embedding dimension")
provider = self.impls_by_provider_id[provider_id]
logger.warning(
"VectorDB is being deprecated in future releases in favor of VectorStore. Please migrate your usage accordingly."
)
vector_store = await provider.openai_create_vector_store(
name=vector_db_name or vector_db_id,
embedding_model=embedding_model,
embedding_dimension=model.metadata["embedding_dimension"],
provider_id=provider_id,
provider_vector_db_id=provider_vector_db_id,
)
vector_store_id = vector_store.id
actual_provider_vector_db_id = provider_vector_db_id or vector_store_id
logger.warning(
f"Ignoring vector_db_id {vector_db_id} and using vector_store_id {vector_store_id} instead. Setting VectorDB {vector_db_id} to VectorDB.vector_db_name"
)
vector_db_data = {
"identifier": vector_db_id,
"identifier": vector_store_id,
"type": ResourceType.vector_db.value,
"provider_id": provider_id,
"provider_resource_id": provider_vector_db_id,
"provider_resource_id": actual_provider_vector_db_id,
"embedding_model": embedding_model,
"embedding_dimension": model.metadata["embedding_dimension"],
"vector_db_name": vector_db_name,
"vector_db_name": vector_store.name,
}
vector_db = TypeAdapter(VectorDBWithOwner).validate_python(vector_db_data)
await self.register_object(vector_db)

93
llama_stack/core/server/auth_providers.py
View file

@ -8,16 +8,18 @@ import ssl
import time
from abc import ABC, abstractmethod
from asyncio import Lock
from urllib.parse import parse_qs, urlparse
from urllib.parse import parse_qs, urljoin, urlparse
import httpx
from jose import jwt
from pydantic import BaseModel, Field
from llama_stack.apis.common.errors import TokenValidationError
from llama_stack.core.datatypes import (
AuthenticationConfig,
CustomAuthConfig,
GitHubTokenAuthConfig,
KubernetesAuthProviderConfig,
OAuth2TokenAuthConfig,
User,
)
@ -162,7 +164,7 @@ class OAuth2TokenAuthProvider(AuthProvider):
auth=auth,
timeout=10.0, # Add a reasonable timeout
)
if response.status_code != 200:
if response.status_code != httpx.codes.OK:
logger.warning(f"Token introspection failed with status code: {response.status_code}")
raise ValueError(f"Token introspection failed: {response.status_code}")
@ -272,7 +274,7 @@ class CustomAuthProvider(AuthProvider):
json=auth_request.model_dump(),
timeout=10.0, # Add a reasonable timeout
)
if response.status_code != 200:
if response.status_code != httpx.codes.OK:
logger.warning(f"Authentication failed with status code: {response.status_code}")
raise ValueError(f"Authentication failed: {response.status_code}")
@ -374,6 +376,89 @@ async def _get_github_user_info(access_token: str, github_api_base_url: str) ->
}
class KubernetesAuthProvider(AuthProvider):
"""
Kubernetes authentication provider that validates tokens using the Kubernetes SelfSubjectReview API.
This provider integrates with Kubernetes API server by using the
/apis/authentication.k8s.io/v1/selfsubjectreviews endpoint to validate tokens and extract user information.
"""
def __init__(self, config: KubernetesAuthProviderConfig):
self.config = config
def _httpx_verify_value(self) -> bool | str:
"""
Build the value for httpx's `verify` parameter.
- False disables verification.
- Path string points to a CA bundle.
- True uses system defaults.
"""
if not self.config.verify_tls:
return False
if self.config.tls_cafile:
return self.config.tls_cafile.as_posix()
return True
async def validate_token(self, token: str, scope: dict | None = None) -> User:
"""Validate a token using Kubernetes SelfSubjectReview API endpoint."""
# Build the Kubernetes SelfSubjectReview API endpoint URL
review_api_url = urljoin(self.config.api_server_url, "/apis/authentication.k8s.io/v1/selfsubjectreviews")
# Create SelfSubjectReview request body
review_request = {"apiVersion": "authentication.k8s.io/v1", "kind": "SelfSubjectReview"}
verify = self._httpx_verify_value()
try:
async with httpx.AsyncClient(verify=verify, timeout=10.0) as client:
response = await client.post(
review_api_url,
json=review_request,
headers={
"Authorization": f"Bearer {token}",
"Content-Type": "application/json",
},
)
if response.status_code == httpx.codes.UNAUTHORIZED:
raise TokenValidationError("Invalid token")
if response.status_code != httpx.codes.CREATED:
logger.warning(f"Kubernetes SelfSubjectReview API failed with status code: {response.status_code}")
raise TokenValidationError(f"Token validation failed: {response.status_code}")
review_response = response.json()
# Extract user information from SelfSubjectReview response
status = review_response.get("status", {})
if not status:
raise ValueError("No status found in SelfSubjectReview response")
user_info = status.get("userInfo", {})
if not user_info:
raise ValueError("No userInfo found in SelfSubjectReview response")
username = user_info.get("username")
if not username:
raise ValueError("No username found in SelfSubjectReview response")
# Build user attributes from Kubernetes user info
user_attributes = get_attributes_from_claims(user_info, self.config.claims_mapping)
return User(
principal=username,
attributes=user_attributes,
)
except httpx.TimeoutException:
logger.warning("Kubernetes SelfSubjectReview API request timed out")
raise ValueError("Token validation timeout") from None
except Exception as e:
logger.warning(f"Error during token validation: {str(e)}")
raise ValueError(f"Token validation error: {str(e)}") from e
async def close(self):
"""Close any resources."""
pass
def create_auth_provider(config: AuthenticationConfig) -> AuthProvider:
"""Factory function to create the appropriate auth provider."""
provider_config = config.provider_config
@ -384,5 +469,7 @@ def create_auth_provider(config: AuthenticationConfig) -> AuthProvider:
return OAuth2TokenAuthProvider(provider_config)
elif isinstance(provider_config, GitHubTokenAuthConfig):
return GitHubTokenAuthProvider(provider_config)
elif isinstance(provider_config, KubernetesAuthProviderConfig):
return KubernetesAuthProvider(provider_config)
else:
raise ValueError(f"Unknown authentication provider config type: {type(provider_config)}")

30
llama_stack/core/server/routes.py
View file

@ -14,7 +14,6 @@ from starlette.routing import Route
from llama_stack.apis.datatypes import Api, ExternalApiSpec
from llama_stack.apis.tools import RAGToolRuntime, SpecialToolGroup
from llama_stack.apis.version import LLAMA_STACK_API_VERSION
from llama_stack.core.resolver import api_protocol_map
from llama_stack.schema_utils import WebMethod
@ -54,22 +53,23 @@ def get_all_api_routes(
protocol_methods.append((f"{tool_group.value}.{name}", method))
for name, method in protocol_methods:
if not hasattr(method, "__webmethod__"):
# Get all webmethods for this method (supports multiple decorators)
webmethods = getattr(method, "__webmethods__", [])
if not webmethods:
continue
# The __webmethod__ attribute is dynamically added by the @webmethod decorator
# mypy doesn't know about this dynamic attribute, so we ignore the attr-defined error
webmethod = method.__webmethod__ # type: ignore[attr-defined]
path = f"/{LLAMA_STACK_API_VERSION}/{webmethod.route.lstrip('/')}"
if webmethod.method == hdrs.METH_GET:
http_method = hdrs.METH_GET
elif webmethod.method == hdrs.METH_DELETE:
http_method = hdrs.METH_DELETE
else:
http_method = hdrs.METH_POST
routes.append(
(Route(path=path, methods=[http_method], name=name, endpoint=None), webmethod)
) # setting endpoint to None since don't use a Router object
# Create routes for each webmethod decorator
for webmethod in webmethods:
path = f"/{webmethod.level}/{webmethod.route.lstrip('/')}"
if webmethod.method == hdrs.METH_GET:
http_method = hdrs.METH_GET
elif webmethod.method == hdrs.METH_DELETE:
http_method = hdrs.METH_DELETE
else:
http_method = hdrs.METH_POST
routes.append(
(Route(path=path, methods=[http_method], name=name, endpoint=None), webmethod)
) # setting endpoint to None since don't use a Router object
apis[api] = routes

230
llama_stack/core/server/server.py
View file

@ -6,6 +6,7 @@
import argparse
import asyncio
import concurrent.futures
import functools
import inspect
import json
@ -24,7 +25,6 @@ from typing import Annotated, Any, get_origin
import httpx
import rich.pretty
import yaml
from aiohttp import hdrs
from fastapi import Body, FastAPI, HTTPException, Request, Response
from fastapi import Path as FastapiPath
from fastapi.exceptions import RequestValidationError
@ -44,23 +44,17 @@ from llama_stack.core.datatypes import (
process_cors_config,
)
from llama_stack.core.distribution import builtin_automatically_routed_apis
from llama_stack.core.external import ExternalApiSpec, load_external_apis
from llama_stack.core.external import load_external_apis
from llama_stack.core.request_headers import (
PROVIDER_DATA_VAR,
request_provider_data_context,
user_from_scope,
)
from llama_stack.core.resolver import InvalidProviderError
from llama_stack.core.server.routes import (
find_matching_route,
get_all_api_routes,
initialize_route_impls,
)
from llama_stack.core.server.routes import get_all_api_routes
from llama_stack.core.stack import (
Stack,
cast_image_name_to_string,
construct_stack,
replace_env_vars,
shutdown_stack,
validate_env_pair,
)
from llama_stack.core.utils.config import redact_sensitive_fields
@ -74,13 +68,12 @@ from llama_stack.providers.inline.telemetry.meta_reference.telemetry import (
)
from llama_stack.providers.utils.telemetry.tracing import (
CURRENT_TRACE_CONTEXT,
end_trace,
setup_logger,
start_trace,
)
from .auth import AuthenticationMiddleware
from .quota import QuotaMiddleware
from .tracing import TracingMiddleware
REPO_ROOT = Path(__file__).parent.parent.parent.parent
@ -132,15 +125,17 @@ def translate_exception(exc: Exception) -> HTTPException | RequestValidationErro
},
)
elif isinstance(exc, ConflictError):
return HTTPException(status_code=409, detail=str(exc))
return HTTPException(status_code=httpx.codes.CONFLICT, detail=str(exc))
elif isinstance(exc, ResourceNotFoundError):
return HTTPException(status_code=404, detail=str(exc))
return HTTPException(status_code=httpx.codes.NOT_FOUND, detail=str(exc))
elif isinstance(exc, ValueError):
return HTTPException(status_code=httpx.codes.BAD_REQUEST, detail=f"Invalid value: {str(exc)}")
elif isinstance(exc, BadRequestError):
return HTTPException(status_code=httpx.codes.BAD_REQUEST, detail=str(exc))
elif isinstance(exc, PermissionError | AccessDeniedError):
return HTTPException(status_code=httpx.codes.FORBIDDEN, detail=f"Permission denied: {str(exc)}")
elif isinstance(exc, ConnectionError | httpx.ConnectError):
return HTTPException(status_code=httpx.codes.BAD_GATEWAY, detail=str(exc))
elif isinstance(exc, asyncio.TimeoutError | TimeoutError):
return HTTPException(status_code=httpx.codes.GATEWAY_TIMEOUT, detail=f"Operation timed out: {str(exc)}")
elif isinstance(exc, NotImplementedError):
@ -154,21 +149,34 @@ def translate_exception(exc: Exception) -> HTTPException | RequestValidationErro
)
async def shutdown(app):
"""Initiate a graceful shutdown of the application.
Handled by the lifespan context manager. The shutdown process involves
shutting down all implementations registered in the application.
class StackApp(FastAPI):
"""
await shutdown_stack(app.__llama_stack_impls__)
A wrapper around the FastAPI application to hold a reference to the Stack instance so that we can
start background tasks (e.g. refresh model registry periodically) from the lifespan context manager.
"""
def __init__(self, config: StackRunConfig, *args, **kwargs):
super().__init__(*args, **kwargs)
self.stack: Stack = Stack(config)
# This code is called from a running event loop managed by uvicorn so we cannot simply call
# asyncio.run() to initialize the stack. We cannot await either since this is not an async
# function.
# As a workaround, we use a thread pool executor to run the initialize() method
# in a separate thread.
with concurrent.futures.ThreadPoolExecutor() as executor:
future = executor.submit(asyncio.run, self.stack.initialize())
future.result()
@asynccontextmanager
async def lifespan(app: FastAPI):
async def lifespan(app: StackApp):
logger.info("Starting up")
assert app.stack is not None
app.stack.create_registry_refresh_task()
yield
logger.info("Shutting down")
await shutdown(app)
await app.stack.shutdown()
def is_streaming_request(func_name: str, request: Request, **kwargs):
@ -285,65 +293,6 @@ def create_dynamic_typed_route(func: Any, method: str, route: str) -> Callable:
return route_handler
class TracingMiddleware:
def __init__(self, app, impls, external_apis: dict[str, ExternalApiSpec]):
self.app = app
self.impls = impls
self.external_apis = external_apis
# FastAPI built-in paths that should bypass custom routing
self.fastapi_paths = ("/docs", "/redoc", "/openapi.json", "/favicon.ico", "/static")
async def __call__(self, scope, receive, send):
if scope.get("type") == "lifespan":
return await self.app(scope, receive, send)
path = scope.get("path", "")
# Check if the path is a FastAPI built-in path
if path.startswith(self.fastapi_paths):
# Pass through to FastAPI's built-in handlers
logger.debug(f"Bypassing custom routing for FastAPI built-in path: {path}")
return await self.app(scope, receive, send)
if not hasattr(self, "route_impls"):
self.route_impls = initialize_route_impls(self.impls, self.external_apis)
try:
_, _, route_path, webmethod = find_matching_route(
scope.get("method", hdrs.METH_GET), path, self.route_impls
)
except ValueError:
# If no matching endpoint is found, pass through to FastAPI
logger.debug(f"No matching route found for path: {path}, falling back to FastAPI")
return await self.app(scope, receive, send)
trace_attributes = {"__location__": "server", "raw_path": path}
# Extract W3C trace context headers and store as trace attributes
headers = dict(scope.get("headers", []))
traceparent = headers.get(b"traceparent", b"").decode()
if traceparent:
trace_attributes["traceparent"] = traceparent
tracestate = headers.get(b"tracestate", b"").decode()
if tracestate:
trace_attributes["tracestate"] = tracestate
trace_path = webmethod.descriptive_name or route_path
trace_context = await start_trace(trace_path, trace_attributes)
async def send_with_trace_id(message):
if message["type"] == "http.response.start":
headers = message.get("headers", [])
headers.append([b"x-trace-id", str(trace_context.trace_id).encode()])
message["headers"] = headers
await send(message)
try:
return await self.app(scope, receive, send_with_trace_id)
finally:
await end_trace()
class ClientVersionMiddleware:
def __init__(self, app):
self.app = app
@ -384,73 +333,61 @@ class ClientVersionMiddleware:
return await self.app(scope, receive, send)
def main(args: argparse.Namespace | None = None):
"""Start the LlamaStack server."""
parser = argparse.ArgumentParser(description="Start the LlamaStack server.")
def create_app(
config_file: str | None = None,
env_vars: list[str] | None = None,
) -> StackApp:
"""Create and configure the FastAPI application.
add_config_distro_args(parser)
parser.add_argument(
"--port",
type=int,
default=int(os.getenv("LLAMA_STACK_PORT", 8321)),
help="Port to listen on",
)
parser.add_argument(
"--env",
action="append",
help="Environment variables in KEY=value format. Can be specified multiple times.",
)
Args:
config_file: Path to config file. If None, uses LLAMA_STACK_CONFIG env var or default resolution.
env_vars: List of environment variables in KEY=value format.
disable_version_check: Whether to disable version checking. If None, uses LLAMA_STACK_DISABLE_VERSION_CHECK env var.
# Determine whether the server args are being passed by the "run" command, if this is the case
# the args will be passed as a Namespace object to the main function, otherwise they will be
# parsed from the command line
if args is None:
args = parser.parse_args()
Returns:
Configured StackApp instance.
"""
config_file = config_file or os.getenv("LLAMA_STACK_CONFIG")
if config_file is None:
raise ValueError("No config file provided and LLAMA_STACK_CONFIG env var is not set")
config_or_distro = get_config_from_args(args)
config_file = resolve_config_or_distro(config_or_distro, Mode.RUN)
config_file = resolve_config_or_distro(config_file, Mode.RUN)
# Load and process configuration
logger_config = None
with open(config_file) as fp:
config_contents = yaml.safe_load(fp)
if isinstance(config_contents, dict) and (cfg := config_contents.get("logging_config")):
logger_config = LoggingConfig(**cfg)
logger = get_logger(name=__name__, category="core::server", config=logger_config)
if args.env:
for env_pair in args.env:
if env_vars:
for env_pair in env_vars:
try:
key, value = validate_env_pair(env_pair)
logger.info(f"Setting CLI environment variable {key} => {value}")
logger.info(f"Setting environment variable {key} => {value}")
os.environ[key] = value
except ValueError as e:
logger.error(f"Error: {str(e)}")
sys.exit(1)
raise ValueError(f"Invalid environment variable format: {env_pair}") from e
config = replace_env_vars(config_contents)
config = StackRunConfig(**cast_image_name_to_string(config))
_log_run_config(run_config=config)
app = FastAPI(
app = StackApp(
lifespan=lifespan,
docs_url="/docs",
redoc_url="/redoc",
openapi_url="/openapi.json",
config=config,
)
if not os.environ.get("LLAMA_STACK_DISABLE_VERSION_CHECK"):
app.add_middleware(ClientVersionMiddleware)
try:
# Create and set the event loop that will be used for both construction and server runtime
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
# Construct the stack in the persistent event loop
impls = loop.run_until_complete(construct_stack(config))
except InvalidProviderError as e:
logger.error(f"Error: {str(e)}")
sys.exit(1)
impls = app.stack.impls
if config.server.auth:
logger.info(f"Enabling authentication with provider: {config.server.auth.provider_config.type.value}")
@ -513,6 +450,7 @@ def main(args: argparse.Namespace | None = None):
apis_to_serve.add("inspect")
apis_to_serve.add("providers")
apis_to_serve.add("prompts")
for api_str in apis_to_serve:
api = Api(api_str)
@ -550,9 +488,54 @@ def main(args: argparse.Namespace | None = None):
app.exception_handler(RequestValidationError)(global_exception_handler)
app.exception_handler(Exception)(global_exception_handler)
app.__llama_stack_impls__ = impls
app.add_middleware(TracingMiddleware, impls=impls, external_apis=external_apis)
return app
def main(args: argparse.Namespace | None = None):
"""Start the LlamaStack server."""
parser = argparse.ArgumentParser(description="Start the LlamaStack server.")
add_config_distro_args(parser)
parser.add_argument(
"--port",
type=int,
default=int(os.getenv("LLAMA_STACK_PORT", 8321)),
help="Port to listen on",
)
parser.add_argument(
"--env",
action="append",
help="Environment variables in KEY=value format. Can be specified multiple times.",
)
# Determine whether the server args are being passed by the "run" command, if this is the case
# the args will be passed as a Namespace object to the main function, otherwise they will be
# parsed from the command line
if args is None:
args = parser.parse_args()
config_or_distro = get_config_from_args(args)
try:
app = create_app(
config_file=config_or_distro,
env_vars=args.env,
)
except Exception as e:
logger.error(f"Error creating app: {str(e)}")
sys.exit(1)
config_file = resolve_config_or_distro(config_or_distro, Mode.RUN)
with open(config_file) as fp:
config_contents = yaml.safe_load(fp)
if isinstance(config_contents, dict) and (cfg := config_contents.get("logging_config")):
logger_config = LoggingConfig(**cfg)
else:
logger_config = None
config = StackRunConfig(**cast_image_name_to_string(replace_env_vars(config_contents)))
import uvicorn
# Configure SSL if certificates are provided
@ -590,7 +573,6 @@ def main(args: argparse.Namespace | None = None):
if ssl_config:
uvicorn_config.update(ssl_config)
# Run uvicorn in the existing event loop to preserve background tasks
# We need to catch KeyboardInterrupt because uvicorn's signal handling
# re-raises SIGINT signals using signal.raise_signal(), which Python
# converts to KeyboardInterrupt. Without this catch, we'd get a confusing
@ -601,13 +583,9 @@ def main(args: argparse.Namespace | None = None):
# Another approach would be to ignore SIGINT entirely - let uvicorn handle it through its own
# signal handling but this is quite intrusive and not worth the effort.
try:
loop.run_until_complete(uvicorn.Server(uvicorn.Config(**uvicorn_config)).serve())
asyncio.run(uvicorn.Server(uvicorn.Config(**uvicorn_config)).serve())
except (KeyboardInterrupt, SystemExit):
logger.info("Received interrupt signal, shutting down gracefully...")
finally:
if not loop.is_closed():
logger.debug("Closing event loop")
loop.close()
def _log_run_config(run_config: StackRunConfig):

80
llama_stack/core/server/tracing.py Normal file
View file

@ -0,0 +1,80 @@
# 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.
from aiohttp import hdrs
from llama_stack.core.external import ExternalApiSpec
from llama_stack.core.server.routes import find_matching_route, initialize_route_impls
from llama_stack.log import get_logger
from llama_stack.providers.utils.telemetry.tracing import end_trace, start_trace
logger = get_logger(name=__name__, category="core::server")
class TracingMiddleware:
def __init__(self, app, impls, external_apis: dict[str, ExternalApiSpec]):
self.app = app
self.impls = impls
self.external_apis = external_apis
# FastAPI built-in paths that should bypass custom routing
self.fastapi_paths = ("/docs", "/redoc", "/openapi.json", "/favicon.ico", "/static")
async def __call__(self, scope, receive, send):
if scope.get("type") == "lifespan":
return await self.app(scope, receive, send)
path = scope.get("path", "")
# Check if the path is a FastAPI built-in path
if path.startswith(self.fastapi_paths):
# Pass through to FastAPI's built-in handlers
logger.debug(f"Bypassing custom routing for FastAPI built-in path: {path}")
return await self.app(scope, receive, send)
if not hasattr(self, "route_impls"):
self.route_impls = initialize_route_impls(self.impls, self.external_apis)
try:
_, _, route_path, webmethod = find_matching_route(
scope.get("method", hdrs.METH_GET), path, self.route_impls
)
except ValueError:
# If no matching endpoint is found, pass through to FastAPI
logger.debug(f"No matching route found for path: {path}, falling back to FastAPI")
return await self.app(scope, receive, send)
# Log deprecation warning if route is deprecated
if getattr(webmethod, "deprecated", False):
logger.warning(
f"DEPRECATED ROUTE USED: {scope.get('method', 'GET')} {path} - "
f"This route is deprecated and may be removed in a future version. "
f"Please check the docs for the supported version."
)
trace_attributes = {"__location__": "server", "raw_path": path}
# Extract W3C trace context headers and store as trace attributes
headers = dict(scope.get("headers", []))
traceparent = headers.get(b"traceparent", b"").decode()
if traceparent:
trace_attributes["traceparent"] = traceparent
tracestate = headers.get(b"tracestate", b"").decode()
if tracestate:
trace_attributes["tracestate"] = tracestate
trace_path = webmethod.descriptive_name or route_path
trace_context = await start_trace(trace_path, trace_attributes)
async def send_with_trace_id(message):
if message["type"] == "http.response.start":
headers = message.get("headers", [])
headers.append([b"x-trace-id", str(trace_context.trace_id).encode()])
message["headers"] = headers
await send(message)
try:
return await self.app(scope, receive, send_with_trace_id)
finally:
await end_trace()

165
llama_stack/core/stack.py
View file

@ -23,6 +23,7 @@ from llama_stack.apis.inference import Inference
from llama_stack.apis.inspect import Inspect
from llama_stack.apis.models import Models
from llama_stack.apis.post_training import PostTraining
from llama_stack.apis.prompts import Prompts
from llama_stack.apis.providers import Providers
from llama_stack.apis.safety import Safety
from llama_stack.apis.scoring import Scoring
@ -36,6 +37,7 @@ from llama_stack.apis.vector_io import VectorIO
from llama_stack.core.datatypes import Provider, StackRunConfig
from llama_stack.core.distribution import get_provider_registry
from llama_stack.core.inspect import DistributionInspectConfig, DistributionInspectImpl
from llama_stack.core.prompts.prompts import PromptServiceConfig, PromptServiceImpl
from llama_stack.core.providers import ProviderImpl, ProviderImplConfig
from llama_stack.core.resolver import ProviderRegistry, resolve_impls
from llama_stack.core.routing_tables.common import CommonRoutingTableImpl
@ -70,6 +72,7 @@ class LlamaStack(
ToolRuntime,
RAGToolRuntime,
Files,
Prompts,
):
pass
@ -103,12 +106,12 @@ async def register_resources(run_config: StackRunConfig, impls: dict[Api, Any]):
method = getattr(impls[api], register_method)
for obj in objects:
logger.debug(f"registering {rsrc.capitalize()} {obj} for provider {obj.provider_id}")
# Do not register models on disabled providers
if hasattr(obj, "provider_id") and (not obj.provider_id or obj.provider_id == "__disabled__"):
logger.debug(f"Skipping {rsrc.capitalize()} registration for disabled provider.")
continue
if hasattr(obj, "provider_id"):
# Do not register models on disabled providers
if not obj.provider_id or obj.provider_id == "__disabled__":
logger.debug(f"Skipping {rsrc.capitalize()} registration for disabled provider.")
continue
logger.debug(f"registering {rsrc.capitalize()} {obj} for provider {obj.provider_id}")
# we want to maintain the type information in arguments to method.
# instead of method(**obj.model_dump()), which may convert a typed attr to a dict,
@ -223,7 +226,10 @@ def replace_env_vars(config: Any, path: str = "") -> Any:
try:
result = re.sub(pattern, get_env_var, config)
return _convert_string_to_proper_type(result)
# Only apply type conversion if substitution actually happened
if result != config:
return _convert_string_to_proper_type(result)
return result
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
@ -300,76 +306,91 @@ def add_internal_implementations(impls: dict[Api, Any], run_config: StackRunConf
)
impls[Api.providers] = providers_impl
prompts_impl = PromptServiceImpl(
PromptServiceConfig(run_config=run_config),
deps=impls,
)
impls[Api.prompts] = prompts_impl
# Produces a stack of providers for the given run config. Not all APIs may be
# asked for in the run config.
async def construct_stack(
run_config: StackRunConfig, provider_registry: ProviderRegistry | None = None
) -> dict[Api, Any]:
if "LLAMA_STACK_TEST_INFERENCE_MODE" in os.environ:
from llama_stack.testing.inference_recorder import setup_inference_recording
class Stack:
def __init__(self, run_config: StackRunConfig, provider_registry: ProviderRegistry | None = None):
self.run_config = run_config
self.provider_registry = provider_registry
self.impls = None
# Produces a stack of providers for the given run config. Not all APIs may be
# asked for in the run config.
async def initialize(self):
if "LLAMA_STACK_TEST_INFERENCE_MODE" in os.environ:
from llama_stack.testing.inference_recorder import setup_inference_recording
global TEST_RECORDING_CONTEXT
TEST_RECORDING_CONTEXT = setup_inference_recording()
if TEST_RECORDING_CONTEXT:
TEST_RECORDING_CONTEXT.__enter__()
logger.info(f"Inference recording enabled: mode={os.environ.get('LLAMA_STACK_TEST_INFERENCE_MODE')}")
dist_registry, _ = await create_dist_registry(self.run_config.metadata_store, self.run_config.image_name)
policy = self.run_config.server.auth.access_policy if self.run_config.server.auth else []
impls = await resolve_impls(
self.run_config, self.provider_registry or get_provider_registry(self.run_config), dist_registry, policy
)
# Add internal implementations after all other providers are resolved
add_internal_implementations(impls, self.run_config)
if Api.prompts in impls:
await impls[Api.prompts].initialize()
await register_resources(self.run_config, impls)
await refresh_registry_once(impls)
self.impls = impls
def create_registry_refresh_task(self):
assert self.impls is not None, "Must call initialize() before starting"
global REGISTRY_REFRESH_TASK
REGISTRY_REFRESH_TASK = asyncio.create_task(refresh_registry_task(self.impls))
def cb(task):
import traceback
if task.cancelled():
logger.error("Model refresh task cancelled")
elif task.exception():
logger.error(f"Model refresh task failed: {task.exception()}")
traceback.print_exception(task.exception())
else:
logger.debug("Model refresh task completed")
REGISTRY_REFRESH_TASK.add_done_callback(cb)
async def shutdown(self):
for impl in self.impls.values():
impl_name = impl.__class__.__name__
logger.info(f"Shutting down {impl_name}")
try:
if hasattr(impl, "shutdown"):
await asyncio.wait_for(impl.shutdown(), timeout=5)
else:
logger.warning(f"No shutdown method for {impl_name}")
except TimeoutError:
logger.exception(f"Shutdown timeout for {impl_name}")
except (Exception, asyncio.CancelledError) as e:
logger.exception(f"Failed to shutdown {impl_name}: {e}")
global TEST_RECORDING_CONTEXT
TEST_RECORDING_CONTEXT = setup_inference_recording()
if TEST_RECORDING_CONTEXT:
TEST_RECORDING_CONTEXT.__enter__()
logger.info(f"Inference recording enabled: mode={os.environ.get('LLAMA_STACK_TEST_INFERENCE_MODE')}")
try:
TEST_RECORDING_CONTEXT.__exit__(None, None, None)
except Exception as e:
logger.error(f"Error during inference recording cleanup: {e}")
dist_registry, _ = await create_dist_registry(run_config.metadata_store, run_config.image_name)
policy = run_config.server.auth.access_policy if run_config.server.auth else []
impls = await resolve_impls(
run_config, provider_registry or get_provider_registry(run_config), dist_registry, policy
)
# Add internal implementations after all other providers are resolved
add_internal_implementations(impls, run_config)
await register_resources(run_config, impls)
await refresh_registry_once(impls)
global REGISTRY_REFRESH_TASK
REGISTRY_REFRESH_TASK = asyncio.create_task(refresh_registry_task(impls))
def cb(task):
import traceback
if task.cancelled():
logger.error("Model refresh task cancelled")
elif task.exception():
logger.error(f"Model refresh task failed: {task.exception()}")
traceback.print_exception(task.exception())
else:
logger.debug("Model refresh task completed")
REGISTRY_REFRESH_TASK.add_done_callback(cb)
return impls
async def shutdown_stack(impls: dict[Api, Any]):
for impl in impls.values():
impl_name = impl.__class__.__name__
logger.info(f"Shutting down {impl_name}")
try:
if hasattr(impl, "shutdown"):
await asyncio.wait_for(impl.shutdown(), timeout=5)
else:
logger.warning(f"No shutdown method for {impl_name}")
except TimeoutError:
logger.exception(f"Shutdown timeout for {impl_name}")
except (Exception, asyncio.CancelledError) as e:
logger.exception(f"Failed to shutdown {impl_name}: {e}")
global TEST_RECORDING_CONTEXT
if TEST_RECORDING_CONTEXT:
try:
TEST_RECORDING_CONTEXT.__exit__(None, None, None)
except Exception as e:
logger.error(f"Error during inference recording cleanup: {e}")
global REGISTRY_REFRESH_TASK
if REGISTRY_REFRESH_TASK:
REGISTRY_REFRESH_TASK.cancel()
global REGISTRY_REFRESH_TASK
if REGISTRY_REFRESH_TASK:
REGISTRY_REFRESH_TASK.cancel()
async def refresh_registry_once(impls: dict[Api, Any]):

2
llama_stack/core/start_stack.sh
View file

@ -123,6 +123,6 @@ if [[ "$env_type" == "venv" ]]; then
$other_args
elif [[ "$env_type" == "container" ]]; then
echo -e "${RED}Warning: Llama Stack no longer supports running Containers via the 'llama stack run' command.${NC}"
echo -e "Please refer to the documentation for more information: https://llama-stack.readthedocs.io/en/latest/distributions/building_distro.html#llama-stack-build"
echo -e "Please refer to the documentation for more information: https://llamastack.github.io/latest/distributions/building_distro.html#llama-stack-build"
exit 1
fi

8
llama_stack/core/store/registry.py
View file

@ -96,9 +96,11 @@ class DiskDistributionRegistry(DistributionRegistry):
async def register(self, obj: RoutableObjectWithProvider) -> bool:
existing_obj = await self.get(obj.type, obj.identifier)
# dont register if the object's providerid already exists
if existing_obj and existing_obj.provider_id == obj.provider_id:
return False
# warn if the object's providerid is different but proceed with registration
if existing_obj and existing_obj.provider_id != obj.provider_id:
logger.warning(
f"Object {existing_obj.type}:{existing_obj.identifier}'s {existing_obj.provider_id} provider is being replaced with {obj.provider_id}"
)
await self.kvstore.set(
KEY_FORMAT.format(type=obj.type, identifier=obj.identifier),

2
llama_stack/core/ui/README.md
View file

@ -6,7 +6,7 @@
## Developer Setup
1. Start up Llama Stack API server. More details [here](https://llama-stack.readthedocs.io/en/latest/getting_started/index.html).
1. Start up Llama Stack API server. More details [here](https://llamastack.github.io/latest/getting_started/index.htmll).
```
llama stack build --distro together --image-type venv

1
llama_stack/distributions/ci-tests/build.yaml
View file

@ -17,6 +17,7 @@ distribution_spec:
- provider_type: remote::vertexai
- provider_type: remote::groq
- provider_type: remote::sambanova
- provider_type: remote::azure
- provider_type: inline::sentence-transformers
vector_io:
- provider_type: inline::faiss

4
llama_stack/distributions/ci-tests/ci_tests.py
View file

@ -11,9 +11,7 @@ from ..starter.starter import get_distribution_template as get_starter_distribut
def get_distribution_template() -> DistributionTemplate:
template = get_starter_distribution_template()
name = "ci-tests"
template.name = name
template = get_starter_distribution_template(name="ci-tests")
template.description = "CI tests for Llama Stack"
return template

25
llama_stack/distributions/ci-tests/run.yaml
View file

@ -81,6 +81,13 @@ providers:
config:
url: https://api.sambanova.ai/v1
api_key: ${env.SAMBANOVA_API_KEY:=}
- provider_id: ${env.AZURE_API_KEY:+azure}
provider_type: remote::azure
config:
api_key: ${env.AZURE_API_KEY:=}
api_base: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers
provider_type: inline::sentence-transformers
vector_io:
@ -89,28 +96,28 @@ providers:
config:
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/faiss_store.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/faiss_store.db
- provider_id: sqlite-vec
provider_type: inline::sqlite-vec
config:
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/sqlite_vec.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/sqlite_vec.db
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/sqlite_vec_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/sqlite_vec_registry.db
- provider_id: ${env.MILVUS_URL:+milvus}
provider_type: inline::milvus
config:
db_path: ${env.MILVUS_DB_PATH:=~/.llama/distributions/starter}/milvus.db
db_path: ${env.MILVUS_DB_PATH:=~/.llama/distributions/ci-tests}/milvus.db
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/milvus_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/milvus_registry.db
- provider_id: ${env.CHROMADB_URL:+chromadb}
provider_type: remote::chromadb
config:
url: ${env.CHROMADB_URL:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter/}/chroma_remote_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests/}/chroma_remote_registry.db
- provider_id: ${env.PGVECTOR_DB:+pgvector}
provider_type: remote::pgvector
config:
@ -121,15 +128,15 @@ providers:
password: ${env.PGVECTOR_PASSWORD:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/pgvector_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/pgvector_registry.db
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/starter/files}
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/ci-tests/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/files_metadata.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/ci-tests}/files_metadata.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard

4
llama_stack/distributions/meta-reference-gpu/doc_template.md
View file

@ -1,7 +1,7 @@
---
orphan: true
---
# Meta Reference Distribution
# Meta Reference GPU Distribution
```{toctree}
:maxdepth: 2
@ -29,7 +29,7 @@ The following environment variables can be configured:
## Prerequisite: Downloading Models
Please use `llama model list --downloaded` to check that you have llama model checkpoints downloaded in `~/.llama` before proceeding. See [installation guide](https://llama-stack.readthedocs.io/en/latest/references/llama_cli_reference/download_models.html) here to download the models. Run `llama model list` to see the available models to download, and `llama model download` to download the checkpoints.
Please use `llama model list --downloaded` to check that you have llama model checkpoints downloaded in `~/.llama` before proceeding. See [installation guide](../../references/llama_cli_reference/download_models.md) here to download the models. Run `llama model list` to see the available models to download, and `llama model download` to download the checkpoints.
```
$ llama model list --downloaded

2
llama_stack/distributions/nvidia/build.yaml
View file

@ -23,6 +23,8 @@ distribution_spec:
- provider_type: inline::basic
tool_runtime:
- provider_type: inline::rag-runtime
files:
- provider_type: inline::localfs
image_type: venv
additional_pip_packages:
- aiosqlite

12
llama_stack/distributions/nvidia/doc_template.md
View file

@ -49,22 +49,22 @@ The deployed platform includes the NIM Proxy microservice, which is the service
### Datasetio API: NeMo Data Store
The NeMo Data Store microservice serves as the default file storage solution for the NeMo microservices platform. It exposts APIs compatible with the Hugging Face Hub client (`HfApi`), so you can use the client to interact with Data Store. The `NVIDIA_DATASETS_URL` environment variable should point to your NeMo Data Store endpoint.
See the {repopath}`NVIDIA Datasetio docs::llama_stack/providers/remote/datasetio/nvidia/README.md` for supported features and example usage.
See the [NVIDIA Datasetio docs](https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/datasetio/nvidia/README.md) for supported features and example usage.
### Eval API: NeMo Evaluator
The NeMo Evaluator microservice supports evaluation of LLMs. Launching an Evaluation job with NeMo Evaluator requires an Evaluation Config (an object that contains metadata needed by the job). A Llama Stack Benchmark maps to an Evaluation Config, so registering a Benchmark creates an Evaluation Config in NeMo Evaluator. The `NVIDIA_EVALUATOR_URL` environment variable should point to your NeMo Microservices endpoint.
See the {repopath}`NVIDIA Eval docs::llama_stack/providers/remote/eval/nvidia/README.md` for supported features and example usage.
See the [NVIDIA Eval docs](https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/eval/nvidia/README.md) for supported features and example usage.
### Post-Training API: NeMo Customizer
The NeMo Customizer microservice supports fine-tuning models. You can reference {repopath}`this list of supported models::llama_stack/providers/remote/post_training/nvidia/models.py` that can be fine-tuned using Llama Stack. The `NVIDIA_CUSTOMIZER_URL` environment variable should point to your NeMo Microservices endpoint.
The NeMo Customizer microservice supports fine-tuning models. You can reference [this list of supported models](https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/post_training/nvidia/models.py) that can be fine-tuned using Llama Stack. The `NVIDIA_CUSTOMIZER_URL` environment variable should point to your NeMo Microservices endpoint.
See the {repopath}`NVIDIA Post-Training docs::llama_stack/providers/remote/post_training/nvidia/README.md` for supported features and example usage.
See the [NVIDIA Post-Training docs](https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/post_training/nvidia/README.md) for supported features and example usage.
### Safety API: NeMo Guardrails
The NeMo Guardrails microservice sits between your application and the LLM, and adds checks and content moderation to a model. The `GUARDRAILS_SERVICE_URL` environment variable should point to your NeMo Microservices endpoint.
See the {repopath}`NVIDIA Safety docs::llama_stack/providers/remote/safety/nvidia/README.md` for supported features and example usage.
See the [NVIDIA Safety docs](https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/safety/nvidia/README.md) for supported features and example usage.
## Deploying models
In order to use a registered model with the Llama Stack APIs, ensure the corresponding NIM is deployed to your environment. For example, you can use the NIM Proxy microservice to deploy `meta/llama-3.2-1b-instruct`.
@ -138,4 +138,4 @@ llama stack run ./run.yaml \
```
## Example Notebooks
For examples of how to use the NVIDIA Distribution to run inference, fine-tune, evaluate, and run safety checks on your LLMs, you can reference the example notebooks in {repopath}`docs/notebooks/nvidia`.
For examples of how to use the NVIDIA Distribution to run inference, fine-tune, evaluate, and run safety checks on your LLMs, you can reference the example notebooks in [docs/notebooks/nvidia](https://github.com/meta-llama/llama-stack/tree/main/docs/notebooks/nvidia).

22
llama_stack/distributions/nvidia/nvidia.py
View file

@ -7,15 +7,15 @@
from pathlib import Path
from llama_stack.core.datatypes import BuildProvider, ModelInput, Provider, ShieldInput, ToolGroupInput
from llama_stack.distributions.template import DistributionTemplate, RunConfigSettings, get_model_registry
from llama_stack.distributions.template import DistributionTemplate, RunConfigSettings
from llama_stack.providers.inline.files.localfs.config import LocalfsFilesImplConfig
from llama_stack.providers.remote.datasetio.nvidia import NvidiaDatasetIOConfig
from llama_stack.providers.remote.eval.nvidia import NVIDIAEvalConfig
from llama_stack.providers.remote.inference.nvidia import NVIDIAConfig
from llama_stack.providers.remote.inference.nvidia.models import MODEL_ENTRIES
from llama_stack.providers.remote.safety.nvidia import NVIDIASafetyConfig
def get_distribution_template() -> DistributionTemplate:
def get_distribution_template(name: str = "nvidia") -> DistributionTemplate:
providers = {
"inference": [BuildProvider(provider_type="remote::nvidia")],
"vector_io": [BuildProvider(provider_type="inline::faiss")],
@ -30,6 +30,7 @@ def get_distribution_template() -> DistributionTemplate:
],
"scoring": [BuildProvider(provider_type="inline::basic")],
"tool_runtime": [BuildProvider(provider_type="inline::rag-runtime")],
"files": [BuildProvider(provider_type="inline::localfs")],
}
inference_provider = Provider(
@ -52,6 +53,11 @@ def get_distribution_template() -> DistributionTemplate:
provider_type="remote::nvidia",
config=NVIDIAEvalConfig.sample_run_config(),
)
files_provider = Provider(
provider_id="meta-reference-files",
provider_type="inline::localfs",
config=LocalfsFilesImplConfig.sample_run_config(f"~/.llama/distributions/{name}"),
)
inference_model = ModelInput(
model_id="${env.INFERENCE_MODEL}",
provider_id="nvidia",
@ -61,9 +67,6 @@ def get_distribution_template() -> DistributionTemplate:
provider_id="nvidia",
)
available_models = {
"nvidia": MODEL_ENTRIES,
}
default_tool_groups = [
ToolGroupInput(
toolgroup_id="builtin::rag",
@ -71,23 +74,21 @@ def get_distribution_template() -> DistributionTemplate:
),
]
default_models, _ = get_model_registry(available_models)
return DistributionTemplate(
name="nvidia",
name=name,
distro_type="self_hosted",
description="Use NVIDIA NIM for running LLM inference, evaluation and safety",
container_image=None,
template_path=Path(__file__).parent / "doc_template.md",
providers=providers,
available_models_by_provider=available_models,
run_configs={
"run.yaml": RunConfigSettings(
provider_overrides={
"inference": [inference_provider],
"datasetio": [datasetio_provider],
"eval": [eval_provider],
"files": [files_provider],
},
default_models=default_models,
default_tool_groups=default_tool_groups,
),
"run-with-safety.yaml": RunConfigSettings(
@ -97,6 +98,7 @@ def get_distribution_template() -> DistributionTemplate:
safety_provider,
],
"eval": [eval_provider],
"files": [files_provider],
},
default_models=[inference_model, safety_model],
default_shields=[ShieldInput(shield_id="${env.SAFETY_MODEL}", provider_id="nvidia")],

9
llama_stack/distributions/nvidia/run-with-safety.yaml
View file

@ -4,6 +4,7 @@ apis:
- agents
- datasetio
- eval
- files
- inference
- post_training
- safety
@ -88,6 +89,14 @@ providers:
tool_runtime:
- provider_id: rag-runtime
provider_type: inline::rag-runtime
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/nvidia/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/nvidia}/files_metadata.db
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/nvidia}/registry.db

89
llama_stack/distributions/nvidia/run.yaml
View file

@ -4,6 +4,7 @@ apis:
- agents
- datasetio
- eval
- files
- inference
- post_training
- safety
@ -77,91 +78,21 @@ providers:
tool_runtime:
- provider_id: rag-runtime
provider_type: inline::rag-runtime
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/nvidia/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/nvidia}/files_metadata.db
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/nvidia}/registry.db
inference_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/nvidia}/inference_store.db
models:
- metadata: {}
model_id: meta/llama3-8b-instruct
provider_id: nvidia
provider_model_id: meta/llama3-8b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama3-70b-instruct
provider_id: nvidia
provider_model_id: meta/llama3-70b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.1-8b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.1-8b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.1-70b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.1-70b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.1-405b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.1-405b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.2-1b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.2-1b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.2-3b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.2-3b-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.2-11b-vision-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.2-11b-vision-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.2-90b-vision-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.2-90b-vision-instruct
model_type: llm
- metadata: {}
model_id: meta/llama-3.3-70b-instruct
provider_id: nvidia
provider_model_id: meta/llama-3.3-70b-instruct
model_type: llm
- metadata:
embedding_dimension: 2048
context_length: 8192
model_id: nvidia/llama-3.2-nv-embedqa-1b-v2
provider_id: nvidia
provider_model_id: nvidia/llama-3.2-nv-embedqa-1b-v2
model_type: embedding
- metadata:
embedding_dimension: 1024
context_length: 512
model_id: nvidia/nv-embedqa-e5-v5
provider_id: nvidia
provider_model_id: nvidia/nv-embedqa-e5-v5
model_type: embedding
- metadata:
embedding_dimension: 4096
context_length: 512
model_id: nvidia/nv-embedqa-mistral-7b-v2
provider_id: nvidia
provider_model_id: nvidia/nv-embedqa-mistral-7b-v2
model_type: embedding
- metadata:
embedding_dimension: 1024
context_length: 512
model_id: snowflake/arctic-embed-l
provider_id: nvidia
provider_model_id: snowflake/arctic-embed-l
model_type: embedding
models: []
shields: []
vector_dbs: []
datasets: []

16
llama_stack/distributions/open-benchmark/open_benchmark.py
View file

@ -43,7 +43,7 @@ def get_inference_providers() -> tuple[list[Provider], dict[str, list[ProviderMo
"openai",
[
ProviderModelEntry(
provider_model_id="openai/gpt-4o",
provider_model_id="gpt-4o",
model_type=ModelType.llm,
)
],
@ -53,7 +53,7 @@ def get_inference_providers() -> tuple[list[Provider], dict[str, list[ProviderMo
"anthropic",
[
ProviderModelEntry(
provider_model_id="anthropic/claude-3-5-sonnet-latest",
provider_model_id="claude-3-5-sonnet-latest",
model_type=ModelType.llm,
)
],
@ -206,13 +206,6 @@ def get_distribution_template() -> DistributionTemplate:
uri="huggingface://datasets/llamastack/math_500?split=test",
),
),
DatasetInput(
dataset_id="bfcl",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/bfcl_v3?split=train",
),
),
DatasetInput(
dataset_id="ifeval",
purpose=DatasetPurpose.eval_messages_answer,
@ -250,11 +243,6 @@ def get_distribution_template() -> DistributionTemplate:
dataset_id="math_500",
scoring_functions=["basic::regex_parser_math_response"],
),
BenchmarkInput(
benchmark_id="meta-reference-bfcl",
dataset_id="bfcl",
scoring_functions=["basic::bfcl"],
),
BenchmarkInput(
benchmark_id="meta-reference-ifeval",
dataset_id="ifeval",

19
llama_stack/distributions/open-benchmark/run.yaml
View file

@ -136,14 +136,14 @@ inference_store:
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/open-benchmark}/inference_store.db
models:
- metadata: {}
model_id: openai/gpt-4o
model_id: gpt-4o
provider_id: openai
provider_model_id: openai/gpt-4o
provider_model_id: gpt-4o
model_type: llm
- metadata: {}
model_id: anthropic/claude-3-5-sonnet-latest
model_id: claude-3-5-sonnet-latest
provider_id: anthropic
provider_model_id: anthropic/claude-3-5-sonnet-latest
provider_model_id: claude-3-5-sonnet-latest
model_type: llm
- metadata: {}
model_id: gemini/gemini-1.5-flash
@ -188,12 +188,6 @@ datasets:
uri: huggingface://datasets/llamastack/math_500?split=test
metadata: {}
dataset_id: math_500
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/bfcl_v3?split=train
metadata: {}
dataset_id: bfcl
- purpose: eval/messages-answer
source:
type: uri
@ -228,11 +222,6 @@ benchmarks:
- basic::regex_parser_math_response
metadata: {}
benchmark_id: meta-reference-math-500
- dataset_id: bfcl
scoring_functions:
- basic::bfcl
metadata: {}
benchmark_id: meta-reference-bfcl
- dataset_id: ifeval
scoring_functions:
- basic::ifeval

1
llama_stack/distributions/starter-gpu/build.yaml
View file

@ -18,6 +18,7 @@ distribution_spec:
- provider_type: remote::vertexai
- provider_type: remote::groq
- provider_type: remote::sambanova
- provider_type: remote::azure
- provider_type: inline::sentence-transformers
vector_io:
- provider_type: inline::faiss

25
llama_stack/distributions/starter-gpu/run.yaml
View file

@ -81,6 +81,13 @@ providers:
config:
url: https://api.sambanova.ai/v1
api_key: ${env.SAMBANOVA_API_KEY:=}
- provider_id: ${env.AZURE_API_KEY:+azure}
provider_type: remote::azure
config:
api_key: ${env.AZURE_API_KEY:=}
api_base: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers
provider_type: inline::sentence-transformers
vector_io:
@ -89,28 +96,28 @@ providers:
config:
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/faiss_store.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/faiss_store.db
- provider_id: sqlite-vec
provider_type: inline::sqlite-vec
config:
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/sqlite_vec.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/sqlite_vec.db
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/sqlite_vec_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/sqlite_vec_registry.db
- provider_id: ${env.MILVUS_URL:+milvus}
provider_type: inline::milvus
config:
db_path: ${env.MILVUS_DB_PATH:=~/.llama/distributions/starter}/milvus.db
db_path: ${env.MILVUS_DB_PATH:=~/.llama/distributions/starter-gpu}/milvus.db
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/milvus_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/milvus_registry.db
- provider_id: ${env.CHROMADB_URL:+chromadb}
provider_type: remote::chromadb
config:
url: ${env.CHROMADB_URL:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter/}/chroma_remote_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu/}/chroma_remote_registry.db
- provider_id: ${env.PGVECTOR_DB:+pgvector}
provider_type: remote::pgvector
config:
@ -121,15 +128,15 @@ providers:
password: ${env.PGVECTOR_PASSWORD:=}
kvstore:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/pgvector_registry.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/pgvector_registry.db
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/starter/files}
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/starter-gpu/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter}/files_metadata.db
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/starter-gpu}/files_metadata.db
safety:
- provider_id: llama-guard
provider_type: inline::llama-guard

4
llama_stack/distributions/starter-gpu/starter_gpu.py
View file

@ -11,9 +11,7 @@ from ..starter.starter import get_distribution_template as get_starter_distribut
def get_distribution_template() -> DistributionTemplate:
template = get_starter_distribution_template()
name = "starter-gpu"
template.name = name
template = get_starter_distribution_template(name="starter-gpu")
template.description = "Quick start template for running Llama Stack with several popular providers. This distribution is intended for GPU-enabled environments."
template.providers["post_training"] = [

1
llama_stack/distributions/starter/build.yaml
View file

@ -18,6 +18,7 @@ distribution_spec:
- provider_type: remote::vertexai
- provider_type: remote::groq
- provider_type: remote::sambanova
- provider_type: remote::azure
- provider_type: inline::sentence-transformers
vector_io:
- provider_type: inline::faiss

7
llama_stack/distributions/starter/run.yaml
View file

@ -81,6 +81,13 @@ providers:
config:
url: https://api.sambanova.ai/v1
api_key: ${env.SAMBANOVA_API_KEY:=}
- provider_id: ${env.AZURE_API_KEY:+azure}
provider_type: remote::azure
config:
api_key: ${env.AZURE_API_KEY:=}
api_base: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers
provider_type: inline::sentence-transformers
vector_io:

25
llama_stack/distributions/starter/starter.py
View file

@ -59,6 +59,7 @@ ENABLED_INFERENCE_PROVIDERS = [
"cerebras",
"nvidia",
"bedrock",
"azure",
]
INFERENCE_PROVIDER_IDS = {
@ -68,6 +69,7 @@ INFERENCE_PROVIDER_IDS = {
"cerebras": "${env.CEREBRAS_API_KEY:+cerebras}",
"nvidia": "${env.NVIDIA_API_KEY:+nvidia}",
"vertexai": "${env.VERTEX_AI_PROJECT:+vertexai}",
"azure": "${env.AZURE_API_KEY:+azure}",
}
@ -76,12 +78,12 @@ def get_remote_inference_providers() -> list[Provider]:
remote_providers = [
provider
for provider in available_providers()
if isinstance(provider, RemoteProviderSpec) and provider.adapter.adapter_type in ENABLED_INFERENCE_PROVIDERS
if isinstance(provider, RemoteProviderSpec) and provider.adapter_type in ENABLED_INFERENCE_PROVIDERS
]
inference_providers = []
for provider_spec in remote_providers:
provider_type = provider_spec.adapter.adapter_type
provider_type = provider_spec.adapter_type
if provider_type in INFERENCE_PROVIDER_IDS:
provider_id = INFERENCE_PROVIDER_IDS[provider_type]
@ -99,9 +101,8 @@ def get_remote_inference_providers() -> list[Provider]:
return inference_providers
def get_distribution_template() -> DistributionTemplate:
def get_distribution_template(name: str = "starter") -> DistributionTemplate:
remote_inference_providers = get_remote_inference_providers()
name = "starter"
providers = {
"inference": [BuildProvider(provider_type=p.provider_type, module=p.module) for p in remote_inference_providers]
@ -278,5 +279,21 @@ def get_distribution_template() -> DistributionTemplate:
"http://localhost:11434",
"Ollama URL",
),
"AZURE_API_KEY": (
"",
"Azure API Key",
),
"AZURE_API_BASE": (
"",
"Azure API Base",
),
"AZURE_API_VERSION": (
"",
"Azure API Version",
),
"AZURE_API_TYPE": (
"azure",
"Azure API Type",
),
},
)

9
llama_stack/distributions/watsonx/run.yaml
View file

@ -10,6 +10,7 @@ apis:
- telemetry
- tool_runtime
- vector_io
- files
providers:
inference:
- provider_id: watsonx
@ -94,6 +95,14 @@ providers:
provider_type: inline::rag-runtime
- provider_id: model-context-protocol
provider_type: remote::model-context-protocol
files:
- provider_id: meta-reference-files
provider_type: inline::localfs
config:
storage_dir: ${env.FILES_STORAGE_DIR:=~/.llama/distributions/watsonx/files}
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/watsonx}/files_metadata.db
metadata_store:
type: sqlite
db_path: ${env.SQLITE_STORE_DIR:=~/.llama/distributions/watsonx}/registry.db

12
llama_stack/distributions/watsonx/watsonx.py
View file

@ -9,6 +9,7 @@ from pathlib import Path
from llama_stack.apis.models import ModelType
from llama_stack.core.datatypes import BuildProvider, ModelInput, Provider, ToolGroupInput
from llama_stack.distributions.template import DistributionTemplate, RunConfigSettings, get_model_registry
from llama_stack.providers.inline.files.localfs.config import LocalfsFilesImplConfig
from llama_stack.providers.inline.inference.sentence_transformers import (
SentenceTransformersInferenceConfig,
)
@ -16,7 +17,7 @@ from llama_stack.providers.remote.inference.watsonx import WatsonXConfig
from llama_stack.providers.remote.inference.watsonx.models import MODEL_ENTRIES
def get_distribution_template() -> DistributionTemplate:
def get_distribution_template(name: str = "watsonx") -> DistributionTemplate:
providers = {
"inference": [
BuildProvider(provider_type="remote::watsonx"),
@ -42,6 +43,7 @@ def get_distribution_template() -> DistributionTemplate:
BuildProvider(provider_type="inline::rag-runtime"),
BuildProvider(provider_type="remote::model-context-protocol"),
],
"files": [BuildProvider(provider_type="inline::localfs")],
}
inference_provider = Provider(
@ -79,9 +81,14 @@ def get_distribution_template() -> DistributionTemplate:
},
)
files_provider = Provider(
provider_id="meta-reference-files",
provider_type="inline::localfs",
config=LocalfsFilesImplConfig.sample_run_config(f"~/.llama/distributions/{name}"),
)
default_models, _ = get_model_registry(available_models)
return DistributionTemplate(
name="watsonx",
name=name,
distro_type="remote_hosted",
description="Use watsonx for running LLM inference",
container_image=None,
@ -92,6 +99,7 @@ def get_distribution_template() -> DistributionTemplate:
"run.yaml": RunConfigSettings(
provider_overrides={
"inference": [inference_provider, embedding_provider],
"files": [files_provider],
},
default_models=default_models + [embedding_model],
default_tool_groups=default_tool_groups,

88
llama_stack/providers/datatypes.py
View file

@ -131,6 +131,15 @@ class ProviderSpec(BaseModel):
""",
)
pip_packages: list[str] = Field(
default_factory=list,
description="The pip dependencies needed for this implementation",
)
provider_data_validator: str | None = Field(
default=None,
)
is_external: bool = Field(default=False, description="Notes whether this provider is an external provider.")
# used internally by the resolver; this is a hack for now
@ -145,45 +154,8 @@ class RoutingTable(Protocol):
async def get_provider_impl(self, routing_key: str) -> Any: ...
# TODO: this can now be inlined into RemoteProviderSpec
@json_schema_type
class AdapterSpec(BaseModel):
adapter_type: str = Field(
...,
description="Unique identifier for this adapter",
)
module: str = Field(
default_factory=str,
description="""
Fully-qualified name of the module to import. The module is expected to have:
- `get_adapter_impl(config, deps)`: returns the adapter implementation
""",
)
pip_packages: list[str] = Field(
default_factory=list,
description="The pip dependencies needed for this implementation",
)
config_class: str = Field(
description="Fully-qualified classname of the config for this provider",
)
provider_data_validator: str | None = Field(
default=None,
)
description: str | None = Field(
default=None,
description="""
A description of the provider. This is used to display in the documentation.
""",
)
@json_schema_type
class InlineProviderSpec(ProviderSpec):
pip_packages: list[str] = Field(
default_factory=list,
description="The pip dependencies needed for this implementation",
)
container_image: str | None = Field(
default=None,
description="""
@ -191,10 +163,6 @@ The container image to use for this implementation. If one is provided, pip_pack
If a provider depends on other providers, the dependencies MUST NOT specify a container image.
""",
)
# module field is inherited from ProviderSpec
provider_data_validator: str | None = Field(
default=None,
)
description: str | None = Field(
default=None,
description="""
@ -223,10 +191,15 @@ class RemoteProviderConfig(BaseModel):
@json_schema_type
class RemoteProviderSpec(ProviderSpec):
adapter: AdapterSpec = Field(
adapter_type: str = Field(
...,
description="Unique identifier for this adapter",
)
description: str | None = Field(
default=None,
description="""
If some code is needed to convert the remote responses into Llama Stack compatible
API responses, specify the adapter here.
A description of the provider. This is used to display in the documentation.
""",
)
@ -234,33 +207,6 @@ API responses, specify the adapter here.
def container_image(self) -> str | None:
return None
# module field is inherited from ProviderSpec
@property
def pip_packages(self) -> list[str]:
return self.adapter.pip_packages
@property
def provider_data_validator(self) -> str | None:
return self.adapter.provider_data_validator
def remote_provider_spec(
api: Api,
adapter: AdapterSpec,
api_dependencies: list[Api] | None = None,
optional_api_dependencies: list[Api] | None = None,
) -> RemoteProviderSpec:
return RemoteProviderSpec(
api=api,
provider_type=f"remote::{adapter.adapter_type}",
config_class=adapter.config_class,
module=adapter.module,
adapter=adapter,
api_dependencies=api_dependencies or [],
optional_api_dependencies=optional_api_dependencies or [],
)
class HealthStatus(StrEnum):
OK = "OK"

69
llama_stack/providers/inline/batches/reference/batches.py
View file

@ -178,9 +178,9 @@ class ReferenceBatchesImpl(Batches):
# TODO: set expiration time for garbage collection
if endpoint not in ["/v1/chat/completions"]:
if endpoint not in ["/v1/chat/completions", "/v1/completions"]:
raise ValueError(
f"Invalid endpoint: {endpoint}. Supported values: /v1/chat/completions. Code: invalid_value. Param: endpoint",
f"Invalid endpoint: {endpoint}. Supported values: /v1/chat/completions, /v1/completions. Code: invalid_value. Param: endpoint",
)
if completion_window != "24h":
@ -424,13 +424,21 @@ class ReferenceBatchesImpl(Batches):
)
valid = False
for param, expected_type, type_string in [
("model", str, "a string"),
# messages is specific to /v1/chat/completions
# we could skip validating messages here and let inference fail. however,
# that would be a very expensive way to find out messages is wrong.
("messages", list, "an array"), # TODO: allow messages to be a string?
]:
if batch.endpoint == "/v1/chat/completions":
required_params = [
("model", str, "a string"),
# messages is specific to /v1/chat/completions
# we could skip validating messages here and let inference fail. however,
# that would be a very expensive way to find out messages is wrong.
("messages", list, "an array"), # TODO: allow messages to be a string?
]
else: # /v1/completions
required_params = [
("model", str, "a string"),
("prompt", str, "a string"), # TODO: allow prompt to be a list of strings??
]
for param, expected_type, type_string in required_params:
if param not in body:
errors.append(
BatchError(
@ -591,20 +599,37 @@ class ReferenceBatchesImpl(Batches):
try:
# TODO(SECURITY): review body for security issues
request.body["messages"] = [convert_to_openai_message_param(msg) for msg in request.body["messages"]]
chat_response = await self.inference_api.openai_chat_completion(**request.body)
if request.url == "/v1/chat/completions":
request.body["messages"] = [convert_to_openai_message_param(msg) for msg in request.body["messages"]]
chat_response = await self.inference_api.openai_chat_completion(**request.body)
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(chat_response, "model_dump_json"), "Chat response must have model_dump_json method"
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id, # TODO: should this be different?
"body": chat_response.model_dump_json(),
},
}
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(chat_response, "model_dump_json"), "Chat response must have model_dump_json method"
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id, # TODO: should this be different?
"body": chat_response.model_dump_json(),
},
}
else: # /v1/completions
completion_response = await self.inference_api.openai_completion(**request.body)
# this is for mypy, we don't allow streaming so we'll get the right type
assert hasattr(completion_response, "model_dump_json"), (
"Completion response must have model_dump_json method"
)
return {
"id": request_id,
"custom_id": request.custom_id,
"response": {
"status_code": 200,
"request_id": request_id,
"body": completion_response.model_dump_json(),
},
}
except Exception as e:
logger.info(f"Error processing request {request.custom_id} in batch {batch_id}: {e}")
return {

7
llama_stack/providers/inline/eval/meta_reference/eval.py
View file

@ -75,6 +75,13 @@ class MetaReferenceEvalImpl(
)
self.benchmarks[task_def.identifier] = task_def
async def unregister_benchmark(self, benchmark_id: str) -> None:
if benchmark_id in self.benchmarks:
del self.benchmarks[benchmark_id]
key = f"{EVAL_TASKS_PREFIX}{benchmark_id}"
await self.kvstore.delete(key)
async def run_eval(
self,
benchmark_id: str,

10
llama_stack/providers/inline/files/localfs/files.py
View file

@ -44,7 +44,7 @@ class LocalfsFilesImpl(Files):
storage_path.mkdir(parents=True, exist_ok=True)
# Initialize SQL store for metadata
self.sql_store = AuthorizedSqlStore(sqlstore_impl(self.config.metadata_store))
self.sql_store = AuthorizedSqlStore(sqlstore_impl(self.config.metadata_store), self.policy)
await self.sql_store.create_table(
"openai_files",
{
@ -74,7 +74,7 @@ class LocalfsFilesImpl(Files):
if not self.sql_store:
raise RuntimeError("Files provider not initialized")
row = await self.sql_store.fetch_one("openai_files", policy=self.policy, where={"id": file_id})
row = await self.sql_store.fetch_one("openai_files", where={"id": file_id})
if not row:
raise ResourceNotFoundError(file_id, "File", "client.files.list()")
@ -86,11 +86,16 @@ class LocalfsFilesImpl(Files):
self,
file: Annotated[UploadFile, File()],
purpose: Annotated[OpenAIFilePurpose, Form()],
expires_after_anchor: Annotated[str | None, Form(alias="expires_after[anchor]")] = None,
expires_after_seconds: Annotated[int | None, Form(alias="expires_after[seconds]")] = None,
) -> OpenAIFileObject:
"""Upload a file that can be used across various endpoints."""
if not self.sql_store:
raise RuntimeError("Files provider not initialized")
if expires_after_anchor is not None or expires_after_seconds is not None:
raise NotImplementedError("File expiration is not supported by this provider")
file_id = self._generate_file_id()
file_path = self._get_file_path(file_id)
@ -145,7 +150,6 @@ class LocalfsFilesImpl(Files):
paginated_result = await self.sql_store.fetch_all(
table="openai_files",
policy=self.policy,
where=where_conditions if where_conditions else None,
order_by=[("created_at", order.value)],
cursor=("id", after) if after else None,

2
llama_stack/providers/inline/scoring/basic/scoring.py
View file

@ -22,7 +22,6 @@ from llama_stack.providers.utils.common.data_schema_validator import (
)
from .config import BasicScoringConfig
from .scoring_fn.bfcl_scoring_fn import BFCLScoringFn
from .scoring_fn.docvqa_scoring_fn import DocVQAScoringFn
from .scoring_fn.equality_scoring_fn import EqualityScoringFn
from .scoring_fn.ifeval_scoring_fn import IfEvalScoringFn
@ -37,7 +36,6 @@ FIXED_FNS = [
SubsetOfScoringFn,
RegexParserScoringFn,
RegexParserMathResponseScoringFn,
BFCLScoringFn,
IfEvalScoringFn,
DocVQAScoringFn,
]

93
llama_stack/providers/inline/scoring/basic/scoring_fn/bfcl_scoring_fn.py
View file

@ -1,93 +0,0 @@
# 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 json
import re
from typing import Any
from llama_stack.apis.scoring import ScoringResultRow
from llama_stack.apis.scoring_functions import ScoringFnParams
from llama_stack.providers.utils.scoring.base_scoring_fn import RegisteredBaseScoringFn
from ..utils.bfcl.ast_parser import decode_ast
from ..utils.bfcl.checker import ast_checker, is_empty_output
from .fn_defs.bfcl import bfcl
def postprocess(x: dict[str, Any], test_category: str) -> dict[str, Any]:
contain_func_call = False
error = None
error_type = None
checker_result = {}
try:
prediction = decode_ast(x["generated_answer"], x["language"]) or ""
contain_func_call = True
# if not is_function_calling_format_output(prediction):
if is_empty_output(prediction):
contain_func_call = False
error = "Did not output in the specified format. Note: the model_result is wrapped in a string to ensure json serializability."
error_type = "ast_decoder:decoder_wrong_output_format"
else:
checker_result = ast_checker(
json.loads(x["function"]),
prediction,
json.loads(x["ground_truth"]),
x["language"],
test_category=test_category,
model_name="",
)
except Exception as e:
prediction = ""
error = f"Invalid syntax. Failed to decode AST. {str(e)}"
error_type = "ast_decoder:decoder_failed"
return {
"prediction": prediction,
"contain_func_call": contain_func_call,
"valid": checker_result.get("valid", False),
"error": error or checker_result.get("error", ""),
"error_type": error_type or checker_result.get("error_type", ""),
}
def gen_valid(x: dict[str, Any]) -> dict[str, float]:
return {"valid": x["valid"]}
def gen_relevance_acc(x: dict[str, Any]) -> dict[str, float]:
# This function serves for both relevance and irrelevance tests, which share the exact opposite logic.
# If `test_category` is "irrelevance", the model is expected to output no function call.
# No function call means either the AST decoding fails (a error message is generated) or the decoded AST does not contain any function call (such as a empty list, `[]`).
# If `test_category` is "relevance", the model is expected to output to a function call, and empty list doesn't count as a function call.
acc = not x["contain_func_call"] if "irrelevance" in x["id"] else x["contain_func_call"]
return {"valid": float(acc)}
class BFCLScoringFn(RegisteredBaseScoringFn):
"""
A scoring_fn for BFCL
"""
def __init__(self, *args, **kwargs) -> None:
super().__init__(*args, **kwargs)
self.supported_fn_defs_registry = {
bfcl.identifier: bfcl,
}
async def score_row(
self,
input_row: dict[str, Any],
scoring_fn_identifier: str | None = "bfcl",
scoring_params: ScoringFnParams | None = None,
) -> ScoringResultRow:
test_category = re.sub(r"_[0-9_-]+$", "", input_row["id"])
score_result = postprocess(input_row, test_category)
if test_category in {"irrelevance", "live_relevance", "live_irrelevance"}:
score = gen_relevance_acc(score_result)["valid"]
else:
score = gen_valid(score_result)["valid"]
return {
"score": float(score),
}

21
llama_stack/providers/inline/scoring/basic/scoring_fn/fn_defs/bfcl.py
View file

@ -1,21 +0,0 @@
# 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.
from llama_stack.apis.common.type_system import NumberType
from llama_stack.apis.scoring_functions import (
AggregationFunctionType,
BasicScoringFnParams,
ScoringFn,
)
bfcl = ScoringFn(
identifier="basic::bfcl",
description="BFCL complex scoring",
return_type=NumberType(),
provider_id="basic",
provider_resource_id="bfcl",
params=BasicScoringFnParams(aggregation_functions=[AggregationFunctionType.accuracy]),
)

296
llama_stack/providers/inline/scoring/basic/utils/bfcl/ast_parser.py
View file

@ -1,296 +0,0 @@
# ruff: noqa
# 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 ast
from .tree_sitter import get_parser
def parse_java_function_call(source_code):
if not source_code.endswith(";"):
source_code += ";" # Necessary for the parser not to register an error
parser = get_parser("java")
tree = parser.parse(bytes(source_code, "utf8"))
root_node = tree.root_node
if root_node.has_error:
raise Exception("Error parsing java the source code.")
def get_text(node):
"""Returns the text represented by the node."""
return source_code[node.start_byte : node.end_byte]
def traverse_node(node, nested=False):
if node.type == "string_literal":
if nested:
return get_text(node)
# Strip surrounding quotes from string literals
return get_text(node)[1:-1]
elif node.type == "character_literal":
if nested:
return get_text(node)
# Strip surrounding single quotes from character literals
return get_text(node)[1:-1]
"""Traverse the node to collect texts for complex structures."""
if node.type in [
"identifier",
"class_literal",
"type_identifier",
"method_invocation",
]:
return get_text(node)
elif node.type == "array_creation_expression":
# Handle array creation expression specifically
type_node = node.child_by_field_name("type")
value_node = node.child_by_field_name("value")
type_text = traverse_node(type_node, True)
value_text = traverse_node(value_node, True)
return f"new {type_text}[]{value_text}"
elif node.type == "object_creation_expression":
# Handle object creation expression specifically
type_node = node.child_by_field_name("type")
arguments_node = node.child_by_field_name("arguments")
type_text = traverse_node(type_node, True)
if arguments_node:
# Process each argument carefully, avoiding unnecessary punctuation
argument_texts = []
for child in arguments_node.children:
if child.type not in [
",",
"(",
")",
]: # Exclude commas and parentheses
argument_text = traverse_node(child, True)
argument_texts.append(argument_text)
arguments_text = ", ".join(argument_texts)
return f"new {type_text}({arguments_text})"
else:
return f"new {type_text}()"
elif node.type == "set":
# Handling sets specifically
items = [traverse_node(n, True) for n in node.children if n.type not in [",", "set"]]
return "{" + ", ".join(items) + "}"
elif node.child_count > 0:
return "".join(traverse_node(child, True) for child in node.children)
else:
return get_text(node)
def extract_arguments(args_node):
arguments = {}
for child in args_node.children:
if child.type == "assignment_expression":
# For named parameters
name_node, value_node = child.children[0], child.children[2]
name = get_text(name_node)
value = traverse_node(value_node)
if name in arguments:
if not isinstance(arguments[name], list):
arguments[name] = [arguments[name]]
arguments[name].append(value)
else:
arguments[name] = value
# arguments.append({'name': name, 'value': value})
elif child.type in ["identifier", "class_literal", "set"]:
# For unnamed parameters and handling sets
value = traverse_node(child)
if None in arguments:
if not isinstance(arguments[None], list):
arguments[None] = [arguments[None]]
arguments[None].append(value)
else:
arguments[None] = value
return arguments
def traverse(node):
if node.type == "method_invocation":
# Extract the function name and its arguments
method_name = get_text(node.child_by_field_name("name"))
class_name_node = node.child_by_field_name("object")
if class_name_node:
class_name = get_text(class_name_node)
function_name = f"{class_name}.{method_name}"
else:
function_name = method_name
arguments_node = node.child_by_field_name("arguments")
if arguments_node:
arguments = extract_arguments(arguments_node)
for key, value in arguments.items():
if isinstance(value, list):
raise Exception("Error: Multiple arguments with the same name are not supported.")
return [{function_name: arguments}]
else:
for child in node.children:
result = traverse(child)
if result:
return result
result = traverse(root_node)
return result if result else {}
def parse_javascript_function_call(source_code):
if not source_code.endswith(";"):
source_code += ";" # Necessary for the parser not to register an error
parser = get_parser("javascript")
# Parse the source code
tree = parser.parse(bytes(source_code, "utf8"))
root_node = tree.root_node
if root_node.has_error:
raise Exception("Error js parsing the source code.")
# Function to recursively extract argument details
def extract_arguments(node):
args = {}
for child in node.children:
if child.type == "assignment_expression":
# Extract left (name) and right (value) parts of the assignment
name = child.children[0].text.decode("utf-8")
value = child.children[2].text.decode("utf-8")
if (value.startswith('"') and value.endswith('"')) or (value.startswith("'") and value.endswith("'")):
value = value[1:-1] # Trim the quotation marks
if name in args:
if not isinstance(args[name], list):
args[name] = [args[name]]
args[name].append(value)
else:
args[name] = value
elif child.type == "identifier" or child.type == "true":
# Handle non-named arguments and boolean values
value = child.text.decode("utf-8")
if None in args:
if not isinstance(args[None], list):
args[None] = [args[None]]
args[None].append(value)
else:
args[None] = value
return args
# Find the function call and extract its name and arguments
if root_node.type == "program":
for child in root_node.children:
if child.type == "expression_statement":
for sub_child in child.children:
if sub_child.type == "call_expression":
function_name = sub_child.children[0].text.decode("utf8")
arguments_node = sub_child.children[1]
parameters = extract_arguments(arguments_node)
for key, value in parameters.items():
if isinstance(value, list):
raise Exception("Error: Multiple arguments with the same name are not supported.")
result = [{function_name: parameters}]
return result
def ast_parse(input_str, language="Python"):
if language == "Python":
cleaned_input = input_str.strip("[]'")
parsed = ast.parse(cleaned_input, mode="eval")
extracted = []
if isinstance(parsed.body, ast.Call):
extracted.append(resolve_ast_call(parsed.body))
else:
for elem in parsed.body.elts:
extracted.append(resolve_ast_call(elem))
return extracted
elif language == "Java":
return parse_java_function_call(input_str[1:-1]) # Remove the [ and ] from the string
elif language == "JavaScript":
return parse_javascript_function_call(input_str[1:-1])
else:
raise NotImplementedError(f"Unsupported language: {language}")
def resolve_ast_call(elem):
# Handle nested attributes for deeply nested module paths
func_parts = []
func_part = elem.func
while isinstance(func_part, ast.Attribute):
func_parts.append(func_part.attr)
func_part = func_part.value
if isinstance(func_part, ast.Name):
func_parts.append(func_part.id)
func_name = ".".join(reversed(func_parts))
args_dict = {}
# Parse when args are simply passed as an unnamed dictionary arg
for arg in elem.args:
if isinstance(arg, ast.Dict):
for key, value in zip(arg.keys, arg.values):
if isinstance(key, ast.Constant):
arg_name = key.value
output = resolve_ast_by_type(value)
args_dict[arg_name] = output
for arg in elem.keywords:
output = resolve_ast_by_type(arg.value)
args_dict[arg.arg] = output
return {func_name: args_dict}
def resolve_ast_by_type(value):
if isinstance(value, ast.Constant):
if value.value is Ellipsis:
output = "..."
else:
output = value.value
elif isinstance(value, ast.UnaryOp):
output = -value.operand.value
elif isinstance(value, ast.List):
output = [resolve_ast_by_type(v) for v in value.elts]
elif isinstance(value, ast.Dict):
output = {resolve_ast_by_type(k): resolve_ast_by_type(v) for k, v in zip(value.keys, value.values)}
elif isinstance(value, ast.NameConstant): # Added this condition to handle boolean values
output = value.value
elif isinstance(value, ast.BinOp): # Added this condition to handle function calls as arguments
output = eval(ast.unparse(value))
elif isinstance(value, ast.Name):
output = value.id
elif isinstance(value, ast.Call):
if len(value.keywords) == 0:
output = ast.unparse(value)
else:
output = resolve_ast_call(value)
elif isinstance(value, ast.Tuple):
output = tuple(resolve_ast_by_type(v) for v in value.elts)
elif isinstance(value, ast.Lambda):
output = eval(ast.unparse(value.body[0].value))
elif isinstance(value, ast.Ellipsis):
output = "..."
elif isinstance(value, ast.Subscript):
try:
output = ast.unparse(value.body[0].value)
except:
output = ast.unparse(value.value) + "[" + ast.unparse(value.slice) + "]"
else:
raise Exception(f"Unsupported AST type: {type(value)}")
return output
def decode_ast(result, language="Python"):
func = result
func = func.replace("\n", "") # remove new line characters
if not func.startswith("["):
func = "[" + func
if not func.endswith("]"):
func = func + "]"
decoded_output = ast_parse(func, language)
return decoded_output
def decode_execute(result):
func = result
func = func.replace("\n", "") # remove new line characters
if not func.startswith("["):
func = "[" + func
if not func.endswith("]"):
func = func + "]"
decode_output = ast_parse(func)
execution_list = []
for function_call in decode_output:
for key, value in function_call.items():
execution_list.append(f"{key}({','.join([f'{k}={repr(v)}' for k, v in value.items()])})")
return execution_list

989
llama_stack/providers/inline/scoring/basic/utils/bfcl/checker.py
View file

@ -1,989 +0,0 @@
# ruff: noqa
# 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 json
import re
import time
from typing import Any
# Comment out for now until we actually use the rest checker in evals
# import requests # Do not remove this import even though it seems to be unused. It's used in the executable_checker_rest function.
class NoAPIKeyError(Exception):
def __init__(self):
self.message = "❗️Please fill in the API keys in the function_credential_config.json file. If you do not provide the API keys, the executable test category results will be inaccurate."
super().__init__(self.message)
REAL_TIME_MATCH_ALLOWED_DIFFERENCE = 0.2
JAVA_TYPE_CONVERSION = {
"byte": int,
"short": int,
"integer": int,
"float": float,
"double": float,
"long": int,
"boolean": bool,
"char": str,
"Array": list,
"ArrayList": list,
"Set": set,
"HashMap": dict,
"Hashtable": dict,
"Queue": list, # this can be `queue.Queue` as well, for simplicity we check with list
"Stack": list,
"String": str,
"any": str,
}
JS_TYPE_CONVERSION = {
"String": str,
"integer": int,
"float": float,
"Bigint": int,
"Boolean": bool,
"dict": dict,
"array": list,
"any": str,
}
# We switch to conditional import for the following two imports to avoid unnecessary installations.
# User doesn't need to setup the tree-sitter packages if they are not running the test for that language.
# from js_type_converter import js_type_converter
# from java_type_converter import java_type_converter
PYTHON_TYPE_MAPPING = {
"string": str,
"integer": int,
"float": float,
"boolean": bool,
"array": list,
"tuple": list,
"dict": dict,
"any": str,
}
# This is the list of types that we need to recursively check its values
PYTHON_NESTED_TYPE_CHECK_LIST = ["array", "tuple"]
NESTED_CONVERSION_TYPE_LIST = ["Array", "ArrayList", "array"]
#### Helper functions for AST ####
def find_description(func_descriptions, name):
if type(func_descriptions) == list:
for func_description in func_descriptions:
if func_description["name"] == name:
return func_description
return None
else:
# it is a dict, there is only one function
return func_descriptions
def get_possible_answer_type(possible_answer: list):
for answer in possible_answer:
if answer != "": # Optional parameter
return type(answer)
return None
def type_checker(
param: str,
value,
possible_answer: list,
expected_type_description: str,
expected_type_converted,
nested_type_converted,
):
# NOTE: This type checker only supports nested type checking for one level deep.
# We didn't implement recursive type checking for nested types, as it's not needed for the current use case and it's very complex.
result: Any = {
"valid": True,
"error": [],
"is_variable": False,
"error_type": "type_error:simple",
}
is_variable = False
# check for the case where a variable is used instead of a actual value.
# use the type in possible_answer as the expected type
possible_answer_type = get_possible_answer_type(possible_answer)
# if possible_answer only contains optional parameters, we can't determine the type
if possible_answer_type != None:
# we are being precise here.
# in fact, possible_answer_type should always be string, as that's how we treat varibale in possible_answer
if possible_answer_type != expected_type_converted:
is_variable = True
# value is the same type as in function description
if type(value) == expected_type_converted:
# We don't need to do recursive check for simple types
if nested_type_converted == None:
result["is_variable"] = is_variable
return result
else:
for possible_answer_item in possible_answer:
flag = True # Each parameter should match to at least one possible answer type.
# Here, we assume that each item should be the same type. We could also relax it.
if type(possible_answer_item) == list:
for value_item in value:
checker_result = type_checker(
param,
value_item,
possible_answer_item,
str(nested_type_converted),
nested_type_converted,
None,
)
if not checker_result["valid"]:
flag = False
break
if flag:
return {"valid": True, "error": [], "is_variable": is_variable}
result["valid"] = False
result["error"] = [
f"Nested type checking failed for parameter {repr(param)}. Expected outer type {expected_type_description} with inner type {str(nested_type_converted)}. Parameter value: {repr(value)}."
]
result["error_type"] = "type_error:nested"
# value is not as expected, check for the case where a variable is used instead of a actual value
# use the type in possible_answer as the expected type
possible_answer_type = get_possible_answer_type(possible_answer)
# if possible_answer only contains optional parameters, we can't determine the type
if possible_answer_type != None:
# we are being precise here.
# in fact, possible_answer_type should always be string, as that's how we treat varibale in possible_answer
if type(value) == possible_answer_type:
result["is_variable"] = True
return result
result["valid"] = False
result["error"].append(
f"Incorrect type for parameter {repr(param)}. Expected type {expected_type_description}, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:simple"
return result
def standardize_string(input_string: str):
# This function standardizes the string by removing all the spaces, ",./-_*^" punctuation, and converting it to lowercase
# It will also convert all the single quotes to double quotes
# This is used to compare the model output with the possible answers
# We don't want to punish model for answer like April 1, 2024 vs April 1,2024, vs April 1 2024
regex_string = r"[ \,\.\/\-\_\*\^]"
return re.sub(regex_string, "", input_string).lower().replace("'", '"')
def string_checker(param: str, model_output: str, possible_answer: list):
standardize_possible_answer = []
standardize_model_output = standardize_string(model_output)
for i in range(len(possible_answer)):
if type(possible_answer[i]) == str:
standardize_possible_answer.append(standardize_string(possible_answer[i]))
if standardize_model_output not in standardize_possible_answer:
return {
"valid": False,
"error": [
f"Invalid value for parameter {repr(param)}: {repr(model_output)}. Expected one of {possible_answer}. Case insensitive."
],
"error_type": "value_error:string",
}
return {"valid": True, "error": []}
def list_checker(param: str, model_output: list, possible_answer: list):
# Convert the tuple to a list
standardize_model_output = list(model_output)
# If the element in the list is a string, we need to standardize it
for i in range(len(standardize_model_output)):
if type(standardize_model_output[i]) == str:
standardize_model_output[i] = standardize_string(model_output[i])
standardize_possible_answer: Any = []
# We also need to standardize the possible answers
for i in range(len(possible_answer)):
standardize_possible_answer.append([])
for j in range(len(possible_answer[i])):
if type(possible_answer[i][j]) == str:
standardize_possible_answer[i].append(standardize_string(possible_answer[i][j]))
else:
standardize_possible_answer[i].append(possible_answer[i][j])
if standardize_model_output not in standardize_possible_answer:
return {
"valid": False,
"error": [
f"Invalid value for parameter {repr(param)}: {repr(model_output)}. Expected one of {possible_answer}."
],
"error_type": "value_error:list/tuple",
}
return {"valid": True, "error": []}
def dict_checker(param: str, model_output: dict, possible_answers: list):
# This function works for simple dictionaries, but not dictionaries with nested dictionaries.
# The current dataset only contains simple dictionaries, so this is sufficient.
result = {"valid": False, "error": [], "error_type": "dict_checker:unclear"}
for i in range(len(possible_answers)):
if possible_answers[i] == "":
continue
result = {"valid": False, "error": [], "error_type": "dict_checker:unclear"}
flag = True
possible_answer = possible_answers[i]
# possible_anwer is a single dictionary
for key, value in model_output.items():
if key not in possible_answer:
result["valid"] = False
result["error"].append(f"Unexpected dict key parameter: '{key}'.") # type: ignore[attr-defined]
result["error_type"] = "value_error:dict_key"
flag = False
break
standardize_value = value
# If the value is a string, we need to standardize it
if type(value) == str:
standardize_value = standardize_string(value)
# We also need to standardize the possible answers if they are string
standardize_possible_answer = []
for i in range(len(possible_answer[key])):
if type(possible_answer[key][i]) == str:
standardize_possible_answer.append(standardize_string(possible_answer[key][i]))
else:
standardize_possible_answer.append(possible_answer[key][i])
if standardize_value not in standardize_possible_answer:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Invalid value for parameter {repr(key)}: {repr(value)}. Expected one of {standardize_possible_answer}."
)
result["error_type"] = "value_error:dict_value"
flag = False
break
for key, value in possible_answer.items():
if key not in model_output and "" not in value:
result["valid"] = False
result["error"].append(f"Missing dict key parameter: '{key}'.") # type: ignore[attr-defined]
result["error_type"] = "value_error:dict_key"
flag = False
break
if flag:
return {"valid": True, "error": []}
return result
def list_dict_checker(param: str, model_output: list, possible_answers: list):
# This function takes in a list of dictionaries and checks if each dictionary is valid
# The order of the dictionaries in the list must match the order of the possible answers
result = {"valid": False, "error": [], "error_type": "list_dict_checker:unclear"}
for answer_index in range(len(possible_answers)):
flag = True # True means so far, all dictionaries are valid
# Only proceed if the number of dictionaries in the list matches the number of dictionaries in the possible answers
if len(model_output) != len(possible_answers[answer_index]):
result["valid"] = False
result["error"] = ["Wrong number of dictionaries in the list."]
result["error_type"] = "value_error:list_dict_count"
flag = False
continue
for dict_index in range(len(model_output)):
result = dict_checker(
param,
model_output[dict_index],
[possible_answers[answer_index][dict_index]],
)
if not result["valid"]:
flag = False
break
if flag:
return {"valid": True, "error": []}
return result
def simple_function_checker(
func_description: dict,
model_output: dict,
possible_answer: dict,
language: str,
model_name: str,
):
possible_answer = list(possible_answer.values())[0]
# Extract function name and parameters details
func_name = func_description["name"]
param_details = func_description["parameters"]["properties"]
required_params = func_description["parameters"]["required"]
# Initialize a result dictionary
result = {
"valid": True,
"error": [],
"error_type": "simple_function_checker:unclear",
}
# Check if function name matches
if func_name not in model_output:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Function name {repr(func_name)} not found in model output."
)
result["error_type"] = "simple_function_checker:wrong_func_name"
return result
model_params = model_output[func_name]
# Check for required parameters in model output
for param in required_params:
if param not in model_params:
result["valid"] = False
result["error"].append(f"Missing required parameter: {repr(param)}.") # type: ignore[attr-defined]
result["error_type"] = "simple_function_checker:missing_required"
return result
# Validate types and values for each parameter in model output
for param, value in model_params.items():
if param not in param_details or param not in possible_answer:
result["valid"] = False
result["error"].append(f"Unexpected parameter: {repr(param)}.") # type: ignore[attr-defined]
result["error_type"] = "simple_function_checker:unexpected_param"
return result
full_param_details = param_details[param]
expected_type_description = full_param_details["type"] # This is a string
is_variable = False
nested_type_converted = None
if language == "Java":
from evals.utils.bfcl.java_type_converter import java_type_converter
expected_type_converted = JAVA_TYPE_CONVERSION[expected_type_description]
if expected_type_description in JAVA_TYPE_CONVERSION:
if type(value) != str:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Incorrect type for parameter {repr(param)}. Expected type String, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:java"
return result
if expected_type_description in NESTED_CONVERSION_TYPE_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = JAVA_TYPE_CONVERSION[nested_type]
value = java_type_converter(value, expected_type_description, nested_type)
else:
value = java_type_converter(value, expected_type_description)
elif language == "JavaScript":
from evals.utils.bfcl.js_type_converter import js_type_converter
expected_type_converted = JS_TYPE_CONVERSION[expected_type_description]
if expected_type_description in JS_TYPE_CONVERSION:
if type(value) != str:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Incorrect type for parameter {repr(param)}. Expected type String, got {type(value).__name__}. Parameter value: {repr(value)}."
)
result["error_type"] = "type_error:js"
return result
if expected_type_description in NESTED_CONVERSION_TYPE_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = JS_TYPE_CONVERSION[nested_type]
value = js_type_converter(value, expected_type_description, nested_type)
else:
value = js_type_converter(value, expected_type_description)
elif language == "Python":
expected_type_converted = PYTHON_TYPE_MAPPING[expected_type_description]
if expected_type_description in PYTHON_NESTED_TYPE_CHECK_LIST:
nested_type = param_details[param]["items"]["type"]
nested_type_converted = PYTHON_TYPE_MAPPING[nested_type]
# We convert all tuple value to list when the expected type is tuple.
# The conversion is necessary because any tuple in the possible answer would become a list after being processed through json.dump() and json.load().
# This does introduce some false positive (eg, when the model provides a list value instead of tuple). We hope to find a better solution in the future.
if expected_type_description == "tuple" and type(value) == tuple:
value = list(value)
# Allow python auto conversion from int to float
if language == "Python" and expected_type_description == "float" and type(value) == int:
value = float(value)
# Type checking
# In fact, we only check for Python here.
# Type check for other languages are handled by the type converter, and so their value (after conversion) is always correct.
type_check_result = type_checker(
param,
value,
possible_answer[param],
expected_type_description,
expected_type_converted,
nested_type_converted,
)
is_variable = type_check_result["is_variable"]
if not type_check_result["valid"]:
return type_check_result
# It doesn't make sense to special handle dictionaries and list of dictionaries if the value is a variable.
# We can just treat the variable as a string and use the normal flow.
if not is_variable:
# Special handle for dictionaries
if expected_type_converted == dict:
result = dict_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Special handle for list of dictionaries
elif expected_type_converted == list and nested_type_converted == dict:
result = list_dict_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Special handle for strings
elif expected_type_converted == str:
# We don't check for case sensitivity for string, as long as it's not a variable
result = string_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
elif expected_type_converted == list:
result = list_checker(param, value, possible_answer[param])
if not result["valid"]:
return result
continue
# Check if the value is within the possible answers
if value not in possible_answer[param]:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Invalid value for parameter {repr(param)}: {repr(value)}. Expected one of {possible_answer[param]}."
)
result["error_type"] = "value_error:others"
return result
# Check for optional parameters not provided but allowed
for param in possible_answer:
if param not in model_params and "" not in possible_answer[param]:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Optional parameter {repr(param)} not provided and not marked as optional."
)
result["error_type"] = "simple_function_checker:missing_optional"
return result
return result
def parallel_function_checker_enforce_order(
func_descriptions: list,
model_output: list,
possible_answers: dict,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "parallel_function_checker_enforce_order:wrong_count",
}
func_name_list = list(possible_answers.keys())
possible_answers_list = []
for key, value in possible_answers.items():
possible_answers_list.append({key: value})
for i in range(len(possible_answers_list)):
func_description = find_description(func_descriptions, func_name_list[i])
result = simple_function_checker(
func_description,
model_output[i],
possible_answers_list[i],
language,
model_name,
)
if not result["valid"]:
return result
return {"valid": True, "error": []}
def parallel_function_checker_no_order(
func_descriptions: list,
model_output: list,
possible_answers: list,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "parallel_function_checker_no_order:wrong_count",
}
matched_indices = []
# We go throught the possible answers one by one, and eliminate the model output that matches the possible answer
# It must be this way because we need ground truth to fetch the correct function description
for i in range(len(possible_answers)):
# possible_answers[i] is a dictionary with only one key
func_name_expected = list(possible_answers[i].keys())[0]
func_description = find_description(func_descriptions, func_name_expected)
all_errors = []
for index in range(len(model_output)):
if index in matched_indices:
continue
result = simple_function_checker(
func_description,
model_output[index],
possible_answers[i],
language,
model_name,
)
if result["valid"]:
matched_indices.append(index)
break
else:
all_errors.append(
{
f"Model Result Index {index}": {
"sub_error": result["error"],
"sub_error_type": result["error_type"],
"model_output_item": model_output[index],
"possible_answer_item": possible_answers[i],
}
}
)
if not result["valid"]:
considered_indices = [i for i in range(len(model_output)) if i not in matched_indices]
all_errors.insert(
0,
f"Could not find a matching function among index {considered_indices} of model output for index {i} of possible answers.", # type: ignore[arg-type]
)
return {
"valid": False,
"error": all_errors,
"error_type": "parallel_function_checker_no_order:cannot_find_match",
}
return {"valid": True, "error": []}
def multiple_function_checker(
func_descriptions: list,
model_output: list,
possible_answers: list,
language: str,
model_name: str,
):
if len(model_output) != len(possible_answers):
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "multiple_function_checker:wrong_count",
}
# possible_answers is a list of only one dictionary with only one key
func_name_expected = list(possible_answers[0].keys())[0]
func_description = find_description(func_descriptions, func_name_expected)
return simple_function_checker(
func_description,
model_output[0],
possible_answers[0],
language,
model_name,
)
def patten_matcher(exec_output, expected_result, function_call, is_sanity_check):
result = {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
if type(exec_output) != type(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result type for {repr(function_call)}. Expected type: {type(expected_result)}, but got: {type(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type",
"model_executed_output": exec_output,
}
if type(exec_output) == dict:
# We loose the requirement for the sanity check as the expected result used in the sanity check might not be the most up-to-date one.
# This happens when the key is a timestamp or a random number.
if is_sanity_check:
if len(exec_output) != len(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but wrong number of elements in the output. Expected length: {len(expected_result)}, but got: {len(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type:dict_length",
"model_executed_output": exec_output,
}
else:
return result
for key, value in expected_result.items():
if key not in exec_output:
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but key {repr(key)} not found in the model output."
],
"error_type": "executable_checker:wrong_result_type:dict_key_not_found",
"model_executed_output": exec_output,
}
for key, value in exec_output.items():
if key not in expected_result:
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type Dict, but key {repr(key)} not expected in the model output."
],
"error_type": "executable_checker:wrong_result_type:dict_extra_key",
"model_executed_output": exec_output,
}
if type(exec_output) == list:
if len(exec_output) != len(expected_result):
return {
"valid": False,
"error": [
f"Wrong execution result pattern for {repr(function_call)}. Expect type list, but wrong number of elements in the output. Expected length: {len(expected_result)}, but got: {len(exec_output)}."
],
"error_type": "executable_checker:wrong_result_type:list_length",
"model_executed_output": exec_output,
}
return result
#### Helper functions for Exec ####
def executable_checker_simple(
function_call: str,
expected_result,
expected_result_type: str,
is_sanity_check=False,
):
result = {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
exec_dict: Any = {}
try:
exec(
"from executable_python_function import *" + "\nresult=" + function_call,
exec_dict,
)
exec_output = exec_dict["result"]
except NoAPIKeyError as e:
raise e
except Exception as e:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Error in execution: {repr(function_call)}. Error: {str(e)}"
)
result["error_type"] = "executable_checker:execution_error"
return result
# We need to special handle the case where the execution result is a tuple and convert it to a list
# Because when json is stored, the tuple is converted to a list, and so the expected result is a list when loaded from json
if isinstance(exec_output, tuple):
exec_output = list(exec_output)
if expected_result_type == "exact_match":
if exec_output != expected_result:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}."
)
result["error_type"] = "executable_checker:wrong_result"
result["model_executed_output"] = exec_output
return result
elif expected_result_type == "real_time_match":
# Allow for 5% difference
if (type(expected_result) == float or type(expected_result) == int) and (
type(exec_output) == float or type(exec_output) == int
):
if not (
expected_result * (1 - REAL_TIME_MATCH_ALLOWED_DIFFERENCE)
<= exec_output
<= expected_result * (1 + REAL_TIME_MATCH_ALLOWED_DIFFERENCE)
):
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}. {REAL_TIME_MATCH_ALLOWED_DIFFERENCE * 100}% difference allowed."
)
result["error_type"] = "executable_checker:wrong_result_real_time"
result["model_executed_output"] = exec_output
return result
else:
result["valid"] = False
result["error"].append( # type: ignore[attr-defined]
f"Wrong execution result for {repr(function_call)}. Expected: {expected_result}, but got: {exec_output}. Type needs to be float or int for real time match criteria."
)
result["error_type"] = "executable_checker:wrong_result_real_time"
result["model_executed_output"] = exec_output
return result
else:
# structural match
pattern_match_result = patten_matcher(exec_output, expected_result, function_call, is_sanity_check)
if not pattern_match_result["valid"]:
return pattern_match_result
return result
def executable_checker_parallel_no_order(
decoded_result: list, expected_exec_result: list, expected_exec_result_type: list
):
if len(decoded_result) != len(expected_exec_result):
return {
"valid": False,
"error": [
f"Wrong number of functions provided. Expected {len(expected_exec_result)}, but got {len(decoded_result)}."
],
"error_type": "value_error:exec_result_count",
}
matched_indices = []
for i in range(len(expected_exec_result)):
all_errors = []
for index in range(len(decoded_result)):
if index in matched_indices:
continue
result = executable_checker_simple(
decoded_result[index],
expected_exec_result[i],
expected_exec_result_type[i],
False,
)
if result["valid"]:
matched_indices.append(index)
break
else:
all_errors.append(
{
f"Model Result Index {index}": {
"sub_error": result["error"],
"sub_error_type": result["error_type"],
"model_executed_output": (
result["model_executed_output"] if "model_executed_output" in result else None
),
}
}
)
if not result["valid"]:
considered_indices = [i for i in range(len(decoded_result)) if i not in matched_indices]
all_errors.insert(
0,
f"Could not find a matching function among index {considered_indices} of model output for index {i} of possible answers.", # type: ignore[arg-type]
)
return {
"valid": False,
"error": all_errors,
"error_type": "executable_checker:cannot_find_match",
}
return {"valid": True, "error": [], "error_type": "executable_checker:unclear"}
#### Main function ####
def executable_checker_rest(func_call, idx):
# Move this here for now to avoid needing to read this file / fix paths to be relative to dataset_dir. Fix when it's actually needed / used.
EVAL_GROUND_TRUTH_PATH = "/mnt/wsfuse/fair_llm_v2/datasets/eval/bfcl/rest-eval-response_v5.jsonl" # Ground truth file for v5 for rest execution
with open(EVAL_GROUND_TRUTH_PATH, "r") as f:
EVAL_GROUND_TRUTH = f.readlines()
if "https://geocode.maps.co" in func_call:
time.sleep(2)
if "requests_get" in func_call:
func_call = func_call.replace("requests_get", "requests.get")
try:
response = eval(func_call)
except Exception as e:
return {
"valid": False,
"error": [f"Execution failed. {str(e)}"],
"error_type": "executable_checker_rest:execution_error",
}
try:
if response.status_code == 200:
eval_GT_json = json.loads(EVAL_GROUND_TRUTH[idx])
try:
if isinstance(eval_GT_json, dict):
if isinstance(response.json(), dict):
if set(eval_GT_json.keys()) == set(response.json().keys()):
return {"valid": True, "error": [], "error_type": ""}
return {
"valid": False,
"error": ["Key inconsistency"],
"error_type": "executable_checker_rest:wrong_key",
}
return {
"valid": False,
"error": [f"Expected dictionary, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
elif isinstance(eval_GT_json, list):
if isinstance(response.json(), list):
if len(eval_GT_json) != len(response.json()):
return {
"valid": False,
"error": [f"Response list length inconsistency."],
"error_type": "value_error:exec_result_rest_count",
}
else:
for i in range(len(eval_GT_json)):
if set(eval_GT_json[i].keys()) != set(response.json()[i].keys()):
return {
"valid": False,
"error": [f"Key inconsistency"],
"error_type": "executable_checker_rest:wrong_key",
}
return {"valid": True, "error": []}
else:
return {
"valid": False,
"error": [f"Expected list, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
return {
"valid": False,
"error": [f"Expected dict or list, but got {type(response.json())}"],
"error_type": "executable_checker_rest:wrong_type",
}
except Exception as e:
return {
"valid": False,
"error": [
f"Error in execution and type checking. Status code: {response.status_code}. Error: {str(e)}"
],
"error_type": "executable_checker_rest:response_format_error",
}
else:
return {
"valid": False,
"error": [f"Execution result status code is not 200, got {response.status_code}"],
"error_type": "executable_checker_rest:wrong_status_code",
}
except Exception as e:
return {
"valid": False,
"error": [f"Cannot get status code of the response. Error: {str(e)}"],
"error_type": "executable_checker_rest:cannot_get_status_code",
}
def ast_checker(func_description, model_output, possible_answer, language, test_category, model_name):
if "parallel" in test_category:
return parallel_function_checker_no_order(func_description, model_output, possible_answer, language, model_name)
elif "multiple" in test_category:
return multiple_function_checker(func_description, model_output, possible_answer, language, model_name)
else:
if len(model_output) != 1:
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "simple_function_checker:wrong_count",
}
return simple_function_checker(
func_description[0],
model_output[0],
possible_answer[0],
language,
model_name,
)
def exec_checker(decoded_result: list, func_description: dict, test_category: str):
if "multiple" in test_category or "parallel" in test_category:
return executable_checker_parallel_no_order(
decoded_result,
func_description["execution_result"],
func_description["execution_result_type"],
)
else:
if len(decoded_result) != 1:
return {
"valid": False,
"error": ["Wrong number of functions."],
"error_type": "simple_exec_checker:wrong_count",
}
return executable_checker_simple(
decoded_result[0],
func_description["execution_result"][0],
func_description["execution_result_type"][0],
False,
)
def is_empty_output(decoded_output):
# This function is a patch to the ast decoder for relevance detection
# Sometimes the ast decoder will parse successfully, but the input doens't really have a function call
# [], [{}], and anything that is not in function calling format is considered empty (and thus should be marked as correct)
if not is_function_calling_format_output(decoded_output):
return True
if len(decoded_output) == 0:
return True
if len(decoded_output) == 1 and len(decoded_output[0]) == 0:
return True
def is_function_calling_format_output(decoded_output):
# Ensure the output is a list of dictionaries
if type(decoded_output) == list:
for item in decoded_output:
if type(item) != dict:
return False
return True
return False

40
llama_stack/providers/inline/scoring/basic/utils/bfcl/tree_sitter.py
View file

@ -1,40 +0,0 @@
# 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.
"""
Tree-sitter changes its API with unfortunate frequency. Modules that need it should
import it from here so that we can centrally manage things as necessary.
"""
# These currently work with tree-sitter 0.23.0
# NOTE: Don't import tree-sitter or any of the language modules in the main module
# because not all environments have them. Import lazily inside functions where needed.
import importlib
import typing
if typing.TYPE_CHECKING:
import tree_sitter
def get_language(language: str) -> "tree_sitter.Language":
import tree_sitter
language_module_name = f"tree_sitter_{language}"
try:
language_module = importlib.import_module(language_module_name)
except ModuleNotFoundError as exc:
raise ValueError(
f"Language {language} is not found. Please install the tree-sitter-{language} package."
) from exc
return tree_sitter.Language(language_module.language())
def get_parser(language: str, **kwargs) -> "tree_sitter.Parser":
import tree_sitter
lang = get_language(language)
return tree_sitter.Parser(lang, **kwargs)

3
llama_stack/providers/inline/scoring/llm_as_judge/scoring.py
View file

@ -63,6 +63,9 @@ class LlmAsJudgeScoringImpl(
async def register_scoring_function(self, function_def: ScoringFn) -> None:
self.llm_as_judge_fn.register_scoring_fn_def(function_def)
async def unregister_scoring_function(self, scoring_fn_id: str) -> None:
self.llm_as_judge_fn.unregister_scoring_fn_def(scoring_fn_id)
async def score_batch(
self,
dataset_id: str,

2
llama_stack/providers/inline/tool_runtime/rag/__init__.py
View file

@ -14,6 +14,6 @@ from .config import RagToolRuntimeConfig
async def get_provider_impl(config: RagToolRuntimeConfig, deps: dict[Api, Any]):
from .memory import MemoryToolRuntimeImpl
impl = MemoryToolRuntimeImpl(config, deps[Api.vector_io], deps[Api.inference])
impl = MemoryToolRuntimeImpl(config, deps[Api.vector_io], deps[Api.inference], deps[Api.files])
await impl.initialize()
return impl

12
llama_stack/providers/inline/tool_runtime/rag/context_retriever.py
View file

@ -8,7 +8,7 @@
from jinja2 import Template
from llama_stack.apis.common.content_types import InterleavedContent
from llama_stack.apis.inference import UserMessage
from llama_stack.apis.inference import OpenAIUserMessageParam
from llama_stack.apis.tools.rag_tool import (
DefaultRAGQueryGeneratorConfig,
LLMRAGQueryGeneratorConfig,
@ -61,16 +61,16 @@ async def llm_rag_query_generator(
messages = [interleaved_content_as_str(content)]
template = Template(config.template)
content = template.render({"messages": messages})
rendered_content: str = template.render({"messages": messages})
model = config.model
message = UserMessage(content=content)
response = await inference_api.chat_completion(
model_id=model,
message = OpenAIUserMessageParam(content=rendered_content)
response = await inference_api.openai_chat_completion(
model=model,
messages=[message],
stream=False,
)
query = response.completion_message.content
query = response.choices[0].message.content
return query

147
llama_stack/providers/inline/tool_runtime/rag/memory.py
View file

@ -5,10 +5,15 @@
# the root directory of this source tree.
import asyncio
import base64
import io
import mimetypes
import secrets
import string
from typing import Any
import httpx
from fastapi import UploadFile
from pydantic import TypeAdapter
from llama_stack.apis.common.content_types import (
@ -17,6 +22,7 @@ from llama_stack.apis.common.content_types import (
InterleavedContentItem,
TextContentItem,
)
from llama_stack.apis.files import Files, OpenAIFilePurpose
from llama_stack.apis.inference import Inference
from llama_stack.apis.tools import (
ListToolDefsResponse,
@ -30,14 +36,16 @@ from llama_stack.apis.tools import (
ToolParameter,
ToolRuntime,
)
from llama_stack.apis.vector_io import QueryChunksResponse, VectorIO
from llama_stack.apis.vector_io import (
QueryChunksResponse,
VectorIO,
VectorStoreChunkingStrategyStatic,
VectorStoreChunkingStrategyStaticConfig,
)
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import ToolGroupsProtocolPrivate
from llama_stack.providers.utils.inference.prompt_adapter import interleaved_content_as_str
from llama_stack.providers.utils.memory.vector_store import (
content_from_doc,
make_overlapped_chunks,
)
from llama_stack.providers.utils.memory.vector_store import parse_data_url
from .config import RagToolRuntimeConfig
from .context_retriever import generate_rag_query
@ -49,16 +57,59 @@ def make_random_string(length: int = 8):
return "".join(secrets.choice(string.ascii_letters + string.digits) for _ in range(length))
async def raw_data_from_doc(doc: RAGDocument) -> tuple[bytes, str]:
"""Get raw binary data and mime type from a RAGDocument for file upload."""
if isinstance(doc.content, URL):
if doc.content.uri.startswith("data:"):
parts = parse_data_url(doc.content.uri)
mime_type = parts["mimetype"]
data = parts["data"]
if parts["is_base64"]:
file_data = base64.b64decode(data)
else:
file_data = data.encode("utf-8")
return file_data, mime_type
else:
async with httpx.AsyncClient() as client:
r = await client.get(doc.content.uri)
r.raise_for_status()
mime_type = r.headers.get("content-type", "application/octet-stream")
return r.content, mime_type
else:
if isinstance(doc.content, str):
content_str = doc.content
else:
content_str = interleaved_content_as_str(doc.content)
if content_str.startswith("data:"):
parts = parse_data_url(content_str)
mime_type = parts["mimetype"]
data = parts["data"]
if parts["is_base64"]:
file_data = base64.b64decode(data)
else:
file_data = data.encode("utf-8")
return file_data, mime_type
else:
return content_str.encode("utf-8"), "text/plain"
class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRuntime):
def __init__(
self,
config: RagToolRuntimeConfig,
vector_io_api: VectorIO,
inference_api: Inference,
files_api: Files,
):
self.config = config
self.vector_io_api = vector_io_api
self.inference_api = inference_api
self.files_api = files_api
async def initialize(self):
pass
@ -78,27 +129,56 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
vector_db_id: str,
chunk_size_in_tokens: int = 512,
) -> None:
chunks = []
for doc in documents:
content = await content_from_doc(doc)
# TODO: we should add enrichment here as URLs won't be added to the metadata by default
chunks.extend(
make_overlapped_chunks(
doc.document_id,
content,
chunk_size_in_tokens,
chunk_size_in_tokens // 4,
doc.metadata,
)
)
if not chunks:
if not documents:
return
await self.vector_io_api.insert_chunks(
chunks=chunks,
vector_db_id=vector_db_id,
)
for doc in documents:
try:
try:
file_data, mime_type = await raw_data_from_doc(doc)
except Exception as e:
log.error(f"Failed to extract content from document {doc.document_id}: {e}")
continue
file_extension = mimetypes.guess_extension(mime_type) or ".txt"
filename = doc.metadata.get("filename", f"{doc.document_id}{file_extension}")
file_obj = io.BytesIO(file_data)
file_obj.name = filename
upload_file = UploadFile(file=file_obj, filename=filename)
try:
created_file = await self.files_api.openai_upload_file(
file=upload_file, purpose=OpenAIFilePurpose.ASSISTANTS
)
except Exception as e:
log.error(f"Failed to upload file for document {doc.document_id}: {e}")
continue
chunking_strategy = VectorStoreChunkingStrategyStatic(
static=VectorStoreChunkingStrategyStaticConfig(
max_chunk_size_tokens=chunk_size_in_tokens,
chunk_overlap_tokens=chunk_size_in_tokens // 4,
)
)
try:
await self.vector_io_api.openai_attach_file_to_vector_store(
vector_store_id=vector_db_id,
file_id=created_file.id,
attributes=doc.metadata,
chunking_strategy=chunking_strategy,
)
except Exception as e:
log.error(
f"Failed to attach file {created_file.id} to vector store {vector_db_id} for document {doc.document_id}: {e}"
)
continue
except Exception as e:
log.error(f"Unexpected error processing document {doc.document_id}: {e}")
continue
async def query(
self,
@ -131,8 +211,18 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
for vector_db_id in vector_db_ids
]
results: list[QueryChunksResponse] = await asyncio.gather(*tasks)
chunks = [c for r in results for c in r.chunks]
scores = [s for r in results for s in r.scores]
chunks = []
scores = []
for vector_db_id, result in zip(vector_db_ids, results, strict=False):
for chunk, score in zip(result.chunks, result.scores, strict=False):
if not hasattr(chunk, "metadata") or chunk.metadata is None:
chunk.metadata = {}
chunk.metadata["vector_db_id"] = vector_db_id
chunks.append(chunk)
scores.append(score)
if not chunks:
return RAGQueryResult(content=None)
@ -167,6 +257,7 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
metadata_keys_to_exclude_from_context = [
"token_count",
"metadata_token_count",
"vector_db_id",
]
metadata_for_context = {}
for k in chunk_metadata_keys_to_include_from_context:
@ -191,6 +282,7 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
"document_ids": [c.metadata["document_id"] for c in chunks[: len(picked)]],
"chunks": [c.content for c in chunks[: len(picked)]],
"scores": scores[: len(picked)],
"vector_db_ids": [c.metadata["vector_db_id"] for c in chunks[: len(picked)]],
},
)
@ -226,7 +318,6 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
if query_config:
query_config = TypeAdapter(RAGQueryConfig).validate_python(query_config)
else:
# handle someone passing an empty dict
query_config = RAGQueryConfig()
query = kwargs["query"]
@ -237,6 +328,6 @@ class MemoryToolRuntimeImpl(ToolGroupsProtocolPrivate, ToolRuntime, RAGToolRunti
)
return ToolInvocationResult(
content=result.content,
content=result.content or [],
metadata=result.metadata,
)

67
llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py
View file

@ -30,11 +30,11 @@ from llama_stack.providers.utils.kvstore.api import KVStore
from llama_stack.providers.utils.memory.openai_vector_store_mixin import OpenAIVectorStoreMixin
from llama_stack.providers.utils.memory.vector_store import (
RERANKER_TYPE_RRF,
RERANKER_TYPE_WEIGHTED,
ChunkForDeletion,
EmbeddingIndex,
VectorDBWithIndex,
)
from llama_stack.providers.utils.vector_io.vector_utils import WeightedInMemoryAggregator
logger = get_logger(name=__name__, category="vector_io")
@ -66,59 +66,6 @@ def _create_sqlite_connection(db_path):
return connection
def _normalize_scores(scores: dict[str, float]) -> dict[str, float]:
"""Normalize scores to [0,1] range using min-max normalization."""
if not scores:
return {}
min_score = min(scores.values())
max_score = max(scores.values())
score_range = max_score - min_score
if score_range > 0:
return {doc_id: (score - min_score) / score_range for doc_id, score in scores.items()}
return dict.fromkeys(scores, 1.0)
def _weighted_rerank(
vector_scores: dict[str, float],
keyword_scores: dict[str, float],
alpha: float = 0.5,
) -> dict[str, float]:
"""ReRanker that uses weighted average of scores."""
all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
normalized_vector_scores = _normalize_scores(vector_scores)
normalized_keyword_scores = _normalize_scores(keyword_scores)
return {
doc_id: (alpha * normalized_keyword_scores.get(doc_id, 0.0))
+ ((1 - alpha) * normalized_vector_scores.get(doc_id, 0.0))
for doc_id in all_ids
}
def _rrf_rerank(
vector_scores: dict[str, float],
keyword_scores: dict[str, float],
impact_factor: float = 60.0,
) -> dict[str, float]:
"""ReRanker that uses Reciprocal Rank Fusion."""
# Convert scores to ranks
vector_ranks = {
doc_id: i + 1 for i, (doc_id, _) in enumerate(sorted(vector_scores.items(), key=lambda x: x[1], reverse=True))
}
keyword_ranks = {
doc_id: i + 1 for i, (doc_id, _) in enumerate(sorted(keyword_scores.items(), key=lambda x: x[1], reverse=True))
}
all_ids = set(vector_scores.keys()) | set(keyword_scores.keys())
rrf_scores = {}
for doc_id in all_ids:
vector_rank = vector_ranks.get(doc_id, float("inf"))
keyword_rank = keyword_ranks.get(doc_id, float("inf"))
# RRF formula: score = 1/(k + r) where k is impact_factor and r is the rank
rrf_scores[doc_id] = (1.0 / (impact_factor + vector_rank)) + (1.0 / (impact_factor + keyword_rank))
return rrf_scores
def _make_sql_identifier(name: str) -> str:
return re.sub(r"[^a-zA-Z0-9_]", "_", name)
@ -398,14 +345,10 @@ class SQLiteVecIndex(EmbeddingIndex):
for chunk, score in zip(keyword_response.chunks, keyword_response.scores, strict=False)
}
# Combine scores using the specified reranker
if reranker_type == RERANKER_TYPE_WEIGHTED:
alpha = reranker_params.get("alpha", 0.5)
combined_scores = _weighted_rerank(vector_scores, keyword_scores, alpha)
else:
# Default to RRF for None, RRF, or any unknown types
impact_factor = reranker_params.get("impact_factor", 60.0)
combined_scores = _rrf_rerank(vector_scores, keyword_scores, impact_factor)
# Combine scores using the reranking utility
combined_scores = WeightedInMemoryAggregator.combine_search_results(
vector_scores, keyword_scores, reranker_type, reranker_params
)
# Sort by combined score and get top k results
sorted_items = sorted(combined_scores.items(), key=lambda x: x[1], reverse=True)

2
llama_stack/providers/registry/batches.py
View file

@ -13,7 +13,7 @@ def available_providers() -> list[ProviderSpec]:
InlineProviderSpec(
api=Api.batches,
provider_type="inline::reference",
pip_packages=["openai"],
pip_packages=[],
module="llama_stack.providers.inline.batches.reference",
config_class="llama_stack.providers.inline.batches.reference.config.ReferenceBatchesImplConfig",
api_dependencies=[

41
llama_stack/providers/registry/datasetio.py
View file

@ -6,11 +6,10 @@
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
@ -25,28 +24,26 @@ def available_providers() -> list[ProviderSpec]:
api_dependencies=[],
description="Local filesystem-based dataset I/O provider for reading and writing datasets to local storage.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.datasetio,
adapter=AdapterSpec(
adapter_type="huggingface",
pip_packages=[
"datasets",
],
module="llama_stack.providers.remote.datasetio.huggingface",
config_class="llama_stack.providers.remote.datasetio.huggingface.HuggingfaceDatasetIOConfig",
description="HuggingFace datasets provider for accessing and managing datasets from the HuggingFace Hub.",
),
adapter_type="huggingface",
provider_type="remote::huggingface",
pip_packages=[
"datasets>=4.0.0",
],
module="llama_stack.providers.remote.datasetio.huggingface",
config_class="llama_stack.providers.remote.datasetio.huggingface.HuggingfaceDatasetIOConfig",
description="HuggingFace datasets provider for accessing and managing datasets from the HuggingFace Hub.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.datasetio,
adapter=AdapterSpec(
adapter_type="nvidia",
pip_packages=[
"datasets",
],
module="llama_stack.providers.remote.datasetio.nvidia",
config_class="llama_stack.providers.remote.datasetio.nvidia.NvidiaDatasetIOConfig",
description="NVIDIA's dataset I/O provider for accessing datasets from NVIDIA's data platform.",
),
adapter_type="nvidia",
provider_type="remote::nvidia",
module="llama_stack.providers.remote.datasetio.nvidia",
config_class="llama_stack.providers.remote.datasetio.nvidia.NvidiaDatasetIOConfig",
pip_packages=[
"datasets>=4.0.0",
],
description="NVIDIA's dataset I/O provider for accessing datasets from NVIDIA's data platform.",
),
]

21
llama_stack/providers/registry/eval.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from llama_stack.providers.datatypes import AdapterSpec, Api, InlineProviderSpec, ProviderSpec, remote_provider_spec
from llama_stack.providers.datatypes import Api, InlineProviderSpec, ProviderSpec, RemoteProviderSpec
def available_providers() -> list[ProviderSpec]:
@ -25,17 +25,16 @@ def available_providers() -> list[ProviderSpec]:
],
description="Meta's reference implementation of evaluation tasks with support for multiple languages and evaluation metrics.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.eval,
adapter=AdapterSpec(
adapter_type="nvidia",
pip_packages=[
"requests",
],
module="llama_stack.providers.remote.eval.nvidia",
config_class="llama_stack.providers.remote.eval.nvidia.NVIDIAEvalConfig",
description="NVIDIA's evaluation provider for running evaluation tasks on NVIDIA's platform.",
),
adapter_type="nvidia",
pip_packages=[
"requests",
],
provider_type="remote::nvidia",
module="llama_stack.providers.remote.eval.nvidia",
config_class="llama_stack.providers.remote.eval.nvidia.NVIDIAEvalConfig",
description="NVIDIA's evaluation provider for running evaluation tasks on NVIDIA's platform.",
api_dependencies=[
Api.datasetio,
Api.datasets,

23
llama_stack/providers/registry/files.py
View file

@ -4,13 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
)
from llama_stack.providers.datatypes import Api, InlineProviderSpec, ProviderSpec, RemoteProviderSpec
from llama_stack.providers.utils.sqlstore.sqlstore import sql_store_pip_packages
@ -25,14 +19,13 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.files.localfs.config.LocalfsFilesImplConfig",
description="Local filesystem-based file storage provider for managing files and documents locally.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.files,
adapter=AdapterSpec(
adapter_type="s3",
pip_packages=["boto3"] + sql_store_pip_packages,
module="llama_stack.providers.remote.files.s3",
config_class="llama_stack.providers.remote.files.s3.config.S3FilesImplConfig",
description="AWS S3-based file storage provider for scalable cloud file management with metadata persistence.",
),
provider_type="remote::s3",
adapter_type="s3",
pip_packages=["boto3"] + sql_store_pip_packages,
module="llama_stack.providers.remote.files.s3",
config_class="llama_stack.providers.remote.files.s3.config.S3FilesImplConfig",
description="AWS S3-based file storage provider for scalable cloud file management with metadata persistence.",
),
]

380
llama_stack/providers/registry/inference.py
View file

@ -6,11 +6,10 @@
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
META_REFERENCE_DEPS = [
@ -49,180 +48,167 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.inference.sentence_transformers.config.SentenceTransformersInferenceConfig",
description="Sentence Transformers inference provider for text embeddings and similarity search.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="cerebras",
pip_packages=[
"cerebras_cloud_sdk",
],
module="llama_stack.providers.remote.inference.cerebras",
config_class="llama_stack.providers.remote.inference.cerebras.CerebrasImplConfig",
description="Cerebras inference provider for running models on Cerebras Cloud platform.",
),
adapter_type="cerebras",
provider_type="remote::cerebras",
pip_packages=[
"cerebras_cloud_sdk",
],
module="llama_stack.providers.remote.inference.cerebras",
config_class="llama_stack.providers.remote.inference.cerebras.CerebrasImplConfig",
description="Cerebras inference provider for running models on Cerebras Cloud platform.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="ollama",
pip_packages=["ollama", "aiohttp", "h11>=0.16.0"],
config_class="llama_stack.providers.remote.inference.ollama.OllamaImplConfig",
module="llama_stack.providers.remote.inference.ollama",
description="Ollama inference provider for running local models through the Ollama runtime.",
),
adapter_type="ollama",
provider_type="remote::ollama",
pip_packages=["ollama", "aiohttp", "h11>=0.16.0"],
config_class="llama_stack.providers.remote.inference.ollama.OllamaImplConfig",
module="llama_stack.providers.remote.inference.ollama",
description="Ollama inference provider for running local models through the Ollama runtime.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="vllm",
pip_packages=["openai"],
module="llama_stack.providers.remote.inference.vllm",
config_class="llama_stack.providers.remote.inference.vllm.VLLMInferenceAdapterConfig",
description="Remote vLLM inference provider for connecting to vLLM servers.",
),
adapter_type="vllm",
provider_type="remote::vllm",
pip_packages=[],
module="llama_stack.providers.remote.inference.vllm",
config_class="llama_stack.providers.remote.inference.vllm.VLLMInferenceAdapterConfig",
provider_data_validator="llama_stack.providers.remote.inference.vllm.VLLMProviderDataValidator",
description="Remote vLLM inference provider for connecting to vLLM servers.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="tgi",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.TGIImplConfig",
description="Text Generation Inference (TGI) provider for HuggingFace model serving.",
),
adapter_type="tgi",
provider_type="remote::tgi",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.TGIImplConfig",
description="Text Generation Inference (TGI) provider for HuggingFace model serving.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="hf::serverless",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.InferenceAPIImplConfig",
description="HuggingFace Inference API serverless provider for on-demand model inference.",
),
adapter_type="hf::serverless",
provider_type="remote::hf::serverless",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.InferenceAPIImplConfig",
description="HuggingFace Inference API serverless provider for on-demand model inference.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="hf::endpoint",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.InferenceEndpointImplConfig",
description="HuggingFace Inference Endpoints provider for dedicated model serving.",
),
provider_type="remote::hf::endpoint",
adapter_type="hf::endpoint",
pip_packages=["huggingface_hub", "aiohttp"],
module="llama_stack.providers.remote.inference.tgi",
config_class="llama_stack.providers.remote.inference.tgi.InferenceEndpointImplConfig",
description="HuggingFace Inference Endpoints provider for dedicated model serving.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="fireworks",
pip_packages=[
"fireworks-ai",
],
module="llama_stack.providers.remote.inference.fireworks",
config_class="llama_stack.providers.remote.inference.fireworks.FireworksImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.fireworks.FireworksProviderDataValidator",
description="Fireworks AI inference provider for Llama models and other AI models on the Fireworks platform.",
),
adapter_type="fireworks",
provider_type="remote::fireworks",
pip_packages=[
"fireworks-ai<=0.17.16",
],
module="llama_stack.providers.remote.inference.fireworks",
config_class="llama_stack.providers.remote.inference.fireworks.FireworksImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.fireworks.FireworksProviderDataValidator",
description="Fireworks AI inference provider for Llama models and other AI models on the Fireworks platform.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="together",
pip_packages=[
"together",
],
module="llama_stack.providers.remote.inference.together",
config_class="llama_stack.providers.remote.inference.together.TogetherImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.together.TogetherProviderDataValidator",
description="Together AI inference provider for open-source models and collaborative AI development.",
),
adapter_type="together",
provider_type="remote::together",
pip_packages=[
"together",
],
module="llama_stack.providers.remote.inference.together",
config_class="llama_stack.providers.remote.inference.together.TogetherImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.together.TogetherProviderDataValidator",
description="Together AI inference provider for open-source models and collaborative AI development.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="bedrock",
pip_packages=["boto3"],
module="llama_stack.providers.remote.inference.bedrock",
config_class="llama_stack.providers.remote.inference.bedrock.BedrockConfig",
description="AWS Bedrock inference provider for accessing various AI models through AWS's managed service.",
),
adapter_type="bedrock",
provider_type="remote::bedrock",
pip_packages=["boto3"],
module="llama_stack.providers.remote.inference.bedrock",
config_class="llama_stack.providers.remote.inference.bedrock.BedrockConfig",
description="AWS Bedrock inference provider for accessing various AI models through AWS's managed service.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="databricks",
pip_packages=[
"openai",
],
module="llama_stack.providers.remote.inference.databricks",
config_class="llama_stack.providers.remote.inference.databricks.DatabricksImplConfig",
description="Databricks inference provider for running models on Databricks' unified analytics platform.",
),
adapter_type="databricks",
provider_type="remote::databricks",
pip_packages=["databricks-sdk"],
module="llama_stack.providers.remote.inference.databricks",
config_class="llama_stack.providers.remote.inference.databricks.DatabricksImplConfig",
description="Databricks inference provider for running models on Databricks' unified analytics platform.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="nvidia",
pip_packages=[
"openai",
],
module="llama_stack.providers.remote.inference.nvidia",
config_class="llama_stack.providers.remote.inference.nvidia.NVIDIAConfig",
description="NVIDIA inference provider for accessing NVIDIA NIM models and AI services.",
),
adapter_type="nvidia",
provider_type="remote::nvidia",
pip_packages=[],
module="llama_stack.providers.remote.inference.nvidia",
config_class="llama_stack.providers.remote.inference.nvidia.NVIDIAConfig",
description="NVIDIA inference provider for accessing NVIDIA NIM models and AI services.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="runpod",
pip_packages=["openai"],
module="llama_stack.providers.remote.inference.runpod",
config_class="llama_stack.providers.remote.inference.runpod.RunpodImplConfig",
description="RunPod inference provider for running models on RunPod's cloud GPU platform.",
),
adapter_type="runpod",
provider_type="remote::runpod",
pip_packages=[],
module="llama_stack.providers.remote.inference.runpod",
config_class="llama_stack.providers.remote.inference.runpod.RunpodImplConfig",
description="RunPod inference provider for running models on RunPod's cloud GPU platform.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="openai",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.openai",
config_class="llama_stack.providers.remote.inference.openai.OpenAIConfig",
provider_data_validator="llama_stack.providers.remote.inference.openai.config.OpenAIProviderDataValidator",
description="OpenAI inference provider for accessing GPT models and other OpenAI services.",
),
adapter_type="openai",
provider_type="remote::openai",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.openai",
config_class="llama_stack.providers.remote.inference.openai.OpenAIConfig",
provider_data_validator="llama_stack.providers.remote.inference.openai.config.OpenAIProviderDataValidator",
description="OpenAI inference provider for accessing GPT models and other OpenAI services.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="anthropic",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.anthropic",
config_class="llama_stack.providers.remote.inference.anthropic.AnthropicConfig",
provider_data_validator="llama_stack.providers.remote.inference.anthropic.config.AnthropicProviderDataValidator",
description="Anthropic inference provider for accessing Claude models and Anthropic's AI services.",
),
adapter_type="anthropic",
provider_type="remote::anthropic",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.anthropic",
config_class="llama_stack.providers.remote.inference.anthropic.AnthropicConfig",
provider_data_validator="llama_stack.providers.remote.inference.anthropic.config.AnthropicProviderDataValidator",
description="Anthropic inference provider for accessing Claude models and Anthropic's AI services.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="gemini",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.gemini",
config_class="llama_stack.providers.remote.inference.gemini.GeminiConfig",
provider_data_validator="llama_stack.providers.remote.inference.gemini.config.GeminiProviderDataValidator",
description="Google Gemini inference provider for accessing Gemini models and Google's AI services.",
),
adapter_type="gemini",
provider_type="remote::gemini",
pip_packages=[
"litellm",
],
module="llama_stack.providers.remote.inference.gemini",
config_class="llama_stack.providers.remote.inference.gemini.GeminiConfig",
provider_data_validator="llama_stack.providers.remote.inference.gemini.config.GeminiProviderDataValidator",
description="Google Gemini inference provider for accessing Gemini models and Google's AI services.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="vertexai",
pip_packages=["litellm", "google-cloud-aiplatform"],
module="llama_stack.providers.remote.inference.vertexai",
config_class="llama_stack.providers.remote.inference.vertexai.VertexAIConfig",
provider_data_validator="llama_stack.providers.remote.inference.vertexai.config.VertexAIProviderDataValidator",
description="""Google Vertex AI inference provider enables you to use Google's Gemini models through Google Cloud's Vertex AI platform, providing several advantages:
adapter_type="vertexai",
provider_type="remote::vertexai",
pip_packages=[
"litellm",
"google-cloud-aiplatform",
],
module="llama_stack.providers.remote.inference.vertexai",
config_class="llama_stack.providers.remote.inference.vertexai.VertexAIConfig",
provider_data_validator="llama_stack.providers.remote.inference.vertexai.config.VertexAIProviderDataValidator",
description="""Google Vertex AI inference provider enables you to use Google's Gemini models through Google Cloud's Vertex AI platform, providing several advantages:
• Enterprise-grade security: Uses Google Cloud's security controls and IAM
• Better integration: Seamless integration with other Google Cloud services
@ -242,61 +228,73 @@ Available Models:
- vertex_ai/gemini-2.0-flash
- vertex_ai/gemini-2.5-flash
- vertex_ai/gemini-2.5-pro""",
),
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="groq",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.groq",
config_class="llama_stack.providers.remote.inference.groq.GroqConfig",
provider_data_validator="llama_stack.providers.remote.inference.groq.config.GroqProviderDataValidator",
description="Groq inference provider for ultra-fast inference using Groq's LPU technology.",
),
adapter_type="groq",
provider_type="remote::groq",
pip_packages=[
"litellm",
],
module="llama_stack.providers.remote.inference.groq",
config_class="llama_stack.providers.remote.inference.groq.GroqConfig",
provider_data_validator="llama_stack.providers.remote.inference.groq.config.GroqProviderDataValidator",
description="Groq inference provider for ultra-fast inference using Groq's LPU technology.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="llama-openai-compat",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.llama_openai_compat",
config_class="llama_stack.providers.remote.inference.llama_openai_compat.config.LlamaCompatConfig",
provider_data_validator="llama_stack.providers.remote.inference.llama_openai_compat.config.LlamaProviderDataValidator",
description="Llama OpenAI-compatible provider for using Llama models with OpenAI API format.",
),
adapter_type="llama-openai-compat",
provider_type="remote::llama-openai-compat",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.llama_openai_compat",
config_class="llama_stack.providers.remote.inference.llama_openai_compat.config.LlamaCompatConfig",
provider_data_validator="llama_stack.providers.remote.inference.llama_openai_compat.config.LlamaProviderDataValidator",
description="Llama OpenAI-compatible provider for using Llama models with OpenAI API format.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="sambanova",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.sambanova",
config_class="llama_stack.providers.remote.inference.sambanova.SambaNovaImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.sambanova.config.SambaNovaProviderDataValidator",
description="SambaNova inference provider for running models on SambaNova's dataflow architecture.",
),
adapter_type="sambanova",
provider_type="remote::sambanova",
pip_packages=[
"litellm",
],
module="llama_stack.providers.remote.inference.sambanova",
config_class="llama_stack.providers.remote.inference.sambanova.SambaNovaImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.sambanova.config.SambaNovaProviderDataValidator",
description="SambaNova inference provider for running models on SambaNova's dataflow architecture.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="passthrough",
pip_packages=[],
module="llama_stack.providers.remote.inference.passthrough",
config_class="llama_stack.providers.remote.inference.passthrough.PassthroughImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.passthrough.PassthroughProviderDataValidator",
description="Passthrough inference provider for connecting to any external inference service not directly supported.",
),
adapter_type="passthrough",
provider_type="remote::passthrough",
pip_packages=[],
module="llama_stack.providers.remote.inference.passthrough",
config_class="llama_stack.providers.remote.inference.passthrough.PassthroughImplConfig",
provider_data_validator="llama_stack.providers.remote.inference.passthrough.PassthroughProviderDataValidator",
description="Passthrough inference provider for connecting to any external inference service not directly supported.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.inference,
adapter=AdapterSpec(
adapter_type="watsonx",
pip_packages=["ibm_watson_machine_learning"],
module="llama_stack.providers.remote.inference.watsonx",
config_class="llama_stack.providers.remote.inference.watsonx.WatsonXConfig",
provider_data_validator="llama_stack.providers.remote.inference.watsonx.WatsonXProviderDataValidator",
description="IBM WatsonX inference provider for accessing AI models on IBM's WatsonX platform.",
),
adapter_type="watsonx",
provider_type="remote::watsonx",
pip_packages=["ibm_watsonx_ai"],
module="llama_stack.providers.remote.inference.watsonx",
config_class="llama_stack.providers.remote.inference.watsonx.WatsonXConfig",
provider_data_validator="llama_stack.providers.remote.inference.watsonx.WatsonXProviderDataValidator",
description="IBM WatsonX inference provider for accessing AI models on IBM's WatsonX platform.",
),
RemoteProviderSpec(
api=Api.inference,
provider_type="remote::azure",
adapter_type="azure",
pip_packages=["litellm"],
module="llama_stack.providers.remote.inference.azure",
config_class="llama_stack.providers.remote.inference.azure.AzureConfig",
provider_data_validator="llama_stack.providers.remote.inference.azure.config.AzureProviderDataValidator",
description="""
Azure OpenAI inference provider for accessing GPT models and other Azure services.
Provider documentation
https://learn.microsoft.com/en-us/azure/ai-foundry/openai/overview
""",
),
]

19
llama_stack/providers/registry/post_training.py
View file

@ -7,7 +7,7 @@
from typing import cast
from llama_stack.providers.datatypes import AdapterSpec, Api, InlineProviderSpec, ProviderSpec, remote_provider_spec
from llama_stack.providers.datatypes import Api, InlineProviderSpec, ProviderSpec, RemoteProviderSpec
# We provide two versions of these providers so that distributions can package the appropriate version of torch.
# The CPU version is used for distributions that don't have GPU support -- they result in smaller container images.
@ -48,7 +48,7 @@ def available_providers() -> list[ProviderSpec]:
InlineProviderSpec(
api=Api.post_training,
provider_type="inline::huggingface-gpu",
pip_packages=["trl", "transformers", "peft", "datasets", "torch"],
pip_packages=["trl", "transformers", "peft", "datasets>=4.0.0", "torch"],
module="llama_stack.providers.inline.post_training.huggingface",
config_class="llama_stack.providers.inline.post_training.huggingface.HuggingFacePostTrainingConfig",
api_dependencies=[
@ -57,14 +57,13 @@ def available_providers() -> list[ProviderSpec]:
],
description="HuggingFace-based post-training provider for fine-tuning models using the HuggingFace ecosystem.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.post_training,
adapter=AdapterSpec(
adapter_type="nvidia",
pip_packages=["requests", "aiohttp"],
module="llama_stack.providers.remote.post_training.nvidia",
config_class="llama_stack.providers.remote.post_training.nvidia.NvidiaPostTrainingConfig",
description="NVIDIA's post-training provider for fine-tuning models on NVIDIA's platform.",
),
adapter_type="nvidia",
provider_type="remote::nvidia",
pip_packages=["requests", "aiohttp"],
module="llama_stack.providers.remote.post_training.nvidia",
config_class="llama_stack.providers.remote.post_training.nvidia.NvidiaPostTrainingConfig",
description="NVIDIA's post-training provider for fine-tuning models on NVIDIA's platform.",
),
]

50
llama_stack/providers/registry/safety.py
View file

@ -6,11 +6,10 @@
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
@ -48,35 +47,32 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.safety.code_scanner.CodeScannerConfig",
description="Code Scanner safety provider for detecting security vulnerabilities and unsafe code patterns.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.safety,
adapter=AdapterSpec(
adapter_type="bedrock",
pip_packages=["boto3"],
module="llama_stack.providers.remote.safety.bedrock",
config_class="llama_stack.providers.remote.safety.bedrock.BedrockSafetyConfig",
description="AWS Bedrock safety provider for content moderation using AWS's safety services.",
),
adapter_type="bedrock",
provider_type="remote::bedrock",
pip_packages=["boto3"],
module="llama_stack.providers.remote.safety.bedrock",
config_class="llama_stack.providers.remote.safety.bedrock.BedrockSafetyConfig",
description="AWS Bedrock safety provider for content moderation using AWS's safety services.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.safety,
adapter=AdapterSpec(
adapter_type="nvidia",
pip_packages=["requests"],
module="llama_stack.providers.remote.safety.nvidia",
config_class="llama_stack.providers.remote.safety.nvidia.NVIDIASafetyConfig",
description="NVIDIA's safety provider for content moderation and safety filtering.",
),
adapter_type="nvidia",
provider_type="remote::nvidia",
pip_packages=["requests"],
module="llama_stack.providers.remote.safety.nvidia",
config_class="llama_stack.providers.remote.safety.nvidia.NVIDIASafetyConfig",
description="NVIDIA's safety provider for content moderation and safety filtering.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.safety,
adapter=AdapterSpec(
adapter_type="sambanova",
pip_packages=["litellm", "requests"],
module="llama_stack.providers.remote.safety.sambanova",
config_class="llama_stack.providers.remote.safety.sambanova.SambaNovaSafetyConfig",
provider_data_validator="llama_stack.providers.remote.safety.sambanova.config.SambaNovaProviderDataValidator",
description="SambaNova's safety provider for content moderation and safety filtering.",
),
adapter_type="sambanova",
provider_type="remote::sambanova",
pip_packages=["litellm", "requests"],
module="llama_stack.providers.remote.safety.sambanova",
config_class="llama_stack.providers.remote.safety.sambanova.SambaNovaSafetyConfig",
provider_data_validator="llama_stack.providers.remote.safety.sambanova.config.SambaNovaProviderDataValidator",
description="SambaNova's safety provider for content moderation and safety filtering.",
),
]

2
llama_stack/providers/registry/scoring.py
View file

@ -38,7 +38,7 @@ def available_providers() -> list[ProviderSpec]:
InlineProviderSpec(
api=Api.scoring,
provider_type="inline::braintrust",
pip_packages=["autoevals", "openai"],
pip_packages=["autoevals"],
module="llama_stack.providers.inline.scoring.braintrust",
config_class="llama_stack.providers.inline.scoring.braintrust.BraintrustScoringConfig",
api_dependencies=[

90
llama_stack/providers/registry/tool_runtime.py
View file

@ -6,11 +6,10 @@
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
@ -32,62 +31,57 @@ def available_providers() -> list[ProviderSpec]:
],
module="llama_stack.providers.inline.tool_runtime.rag",
config_class="llama_stack.providers.inline.tool_runtime.rag.config.RagToolRuntimeConfig",
api_dependencies=[Api.vector_io, Api.inference],
api_dependencies=[Api.vector_io, Api.inference, Api.files],
description="RAG (Retrieval-Augmented Generation) tool runtime for document ingestion, chunking, and semantic search.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.tool_runtime,
adapter=AdapterSpec(
adapter_type="brave-search",
module="llama_stack.providers.remote.tool_runtime.brave_search",
config_class="llama_stack.providers.remote.tool_runtime.brave_search.config.BraveSearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.brave_search.BraveSearchToolProviderDataValidator",
description="Brave Search tool for web search capabilities with privacy-focused results.",
),
adapter_type="brave-search",
provider_type="remote::brave-search",
module="llama_stack.providers.remote.tool_runtime.brave_search",
config_class="llama_stack.providers.remote.tool_runtime.brave_search.config.BraveSearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.brave_search.BraveSearchToolProviderDataValidator",
description="Brave Search tool for web search capabilities with privacy-focused results.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.tool_runtime,
adapter=AdapterSpec(
adapter_type="bing-search",
module="llama_stack.providers.remote.tool_runtime.bing_search",
config_class="llama_stack.providers.remote.tool_runtime.bing_search.config.BingSearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.bing_search.BingSearchToolProviderDataValidator",
description="Bing Search tool for web search capabilities using Microsoft's search engine.",
),
adapter_type="bing-search",
provider_type="remote::bing-search",
module="llama_stack.providers.remote.tool_runtime.bing_search",
config_class="llama_stack.providers.remote.tool_runtime.bing_search.config.BingSearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.bing_search.BingSearchToolProviderDataValidator",
description="Bing Search tool for web search capabilities using Microsoft's search engine.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.tool_runtime,
adapter=AdapterSpec(
adapter_type="tavily-search",
module="llama_stack.providers.remote.tool_runtime.tavily_search",
config_class="llama_stack.providers.remote.tool_runtime.tavily_search.config.TavilySearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.tavily_search.TavilySearchToolProviderDataValidator",
description="Tavily Search tool for AI-optimized web search with structured results.",
),
adapter_type="tavily-search",
provider_type="remote::tavily-search",
module="llama_stack.providers.remote.tool_runtime.tavily_search",
config_class="llama_stack.providers.remote.tool_runtime.tavily_search.config.TavilySearchToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.tavily_search.TavilySearchToolProviderDataValidator",
description="Tavily Search tool for AI-optimized web search with structured results.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.tool_runtime,
adapter=AdapterSpec(
adapter_type="wolfram-alpha",
module="llama_stack.providers.remote.tool_runtime.wolfram_alpha",
config_class="llama_stack.providers.remote.tool_runtime.wolfram_alpha.config.WolframAlphaToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.wolfram_alpha.WolframAlphaToolProviderDataValidator",
description="Wolfram Alpha tool for computational knowledge and mathematical calculations.",
),
adapter_type="wolfram-alpha",
provider_type="remote::wolfram-alpha",
module="llama_stack.providers.remote.tool_runtime.wolfram_alpha",
config_class="llama_stack.providers.remote.tool_runtime.wolfram_alpha.config.WolframAlphaToolConfig",
pip_packages=["requests"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.wolfram_alpha.WolframAlphaToolProviderDataValidator",
description="Wolfram Alpha tool for computational knowledge and mathematical calculations.",
),
remote_provider_spec(
RemoteProviderSpec(
api=Api.tool_runtime,
adapter=AdapterSpec(
adapter_type="model-context-protocol",
module="llama_stack.providers.remote.tool_runtime.model_context_protocol",
config_class="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderConfig",
pip_packages=["mcp>=1.8.1"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderDataValidator",
description="Model Context Protocol (MCP) tool for standardized tool calling and context management.",
),
adapter_type="model-context-protocol",
provider_type="remote::model-context-protocol",
module="llama_stack.providers.remote.tool_runtime.model_context_protocol",
config_class="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderConfig",
pip_packages=["mcp>=1.8.1"],
provider_data_validator="llama_stack.providers.remote.tool_runtime.model_context_protocol.config.MCPProviderDataValidator",
description="Model Context Protocol (MCP) tool for standardized tool calling and context management.",
),
]

192
llama_stack/providers/registry/vector_io.py
View file

@ -6,11 +6,10 @@
from llama_stack.providers.datatypes import (
AdapterSpec,
Api,
InlineProviderSpec,
ProviderSpec,
remote_provider_spec,
RemoteProviderSpec,
)
@ -300,14 +299,16 @@ See [sqlite-vec's GitHub repo](https://github.com/asg017/sqlite-vec/tree/main) f
Please refer to the sqlite-vec provider documentation.
""",
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(
adapter_type="chromadb",
pip_packages=["chromadb-client"],
module="llama_stack.providers.remote.vector_io.chroma",
config_class="llama_stack.providers.remote.vector_io.chroma.ChromaVectorIOConfig",
description="""
RemoteProviderSpec(
api=Api.vector_io,
adapter_type="chromadb",
provider_type="remote::chromadb",
pip_packages=["chromadb-client"],
module="llama_stack.providers.remote.vector_io.chroma",
config_class="llama_stack.providers.remote.vector_io.chroma.ChromaVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
description="""
[Chroma](https://www.trychroma.com/) is an inline and remote vector
database provider for Llama Stack. It allows you to store and query vectors directly within a Chroma database.
That means you're not limited to storing vectors in memory or in a separate service.
@ -340,9 +341,6 @@ pip install chromadb
## Documentation
See [Chroma's documentation](https://docs.trychroma.com/docs/overview/introduction) for more details about Chroma in general.
""",
),
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
InlineProviderSpec(
api=Api.vector_io,
@ -387,14 +385,16 @@ See [Chroma's documentation](https://docs.trychroma.com/docs/overview/introducti
""",
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(
adapter_type="pgvector",
pip_packages=["psycopg2-binary"],
module="llama_stack.providers.remote.vector_io.pgvector",
config_class="llama_stack.providers.remote.vector_io.pgvector.PGVectorVectorIOConfig",
description="""
RemoteProviderSpec(
api=Api.vector_io,
adapter_type="pgvector",
provider_type="remote::pgvector",
pip_packages=["psycopg2-binary"],
module="llama_stack.providers.remote.vector_io.pgvector",
config_class="llama_stack.providers.remote.vector_io.pgvector.PGVectorVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
description="""
[PGVector](https://github.com/pgvector/pgvector) is a remote vector database provider for Llama Stack. It
allows you to store and query vectors directly in memory.
That means you'll get fast and efficient vector retrieval.
@ -404,6 +404,60 @@ That means you'll get fast and efficient vector retrieval.
- Easy to use
- Fully integrated with Llama Stack
There are three implementations of search for PGVectoIndex available:
1. Vector Search:
- How it works:
- Uses PostgreSQL's vector extension (pgvector) to perform similarity search
- Compares query embeddings against stored embeddings using Cosine distance or other distance metrics
- Eg. SQL query: SELECT document, embedding &lt;=&gt; %s::vector AS distance FROM table ORDER BY distance
-Characteristics:
- Semantic understanding - finds documents similar in meaning even if they don't share keywords
- Works with high-dimensional vector embeddings (typically 768, 1024, or higher dimensions)
- Best for: Finding conceptually related content, handling synonyms, cross-language search
2. Keyword Search
- How it works:
- Uses PostgreSQL's full-text search capabilities with tsvector and ts_rank
- Converts text to searchable tokens using to_tsvector('english', text). Default language is English.
- Eg. SQL query: SELECT document, ts_rank(tokenized_content, plainto_tsquery('english', %s)) AS score
- Characteristics:
- Lexical matching - finds exact keyword matches and variations
- Uses GIN (Generalized Inverted Index) for fast text search performance
- Scoring: Uses PostgreSQL's ts_rank function for relevance scoring
- Best for: Exact term matching, proper names, technical terms, Boolean-style queries
3. Hybrid Search
- How it works:
- Combines both vector and keyword search results
- Runs both searches independently, then merges results using configurable reranking
- Two reranking strategies available:
- Reciprocal Rank Fusion (RRF) - (default: 60.0)
- Weighted Average - (default: 0.5)
- Characteristics:
- Best of both worlds: semantic understanding + exact matching
- Documents appearing in both searches get boosted scores
- Configurable balance between semantic and lexical matching
- Best for: General-purpose search where you want both precision and recall
4. Database Schema
The PGVector implementation stores data optimized for all three search types:
CREATE TABLE vector_store_xxx (
id TEXT PRIMARY KEY,
document JSONB, -- Original document
embedding vector(dimension), -- For vector search
content_text TEXT, -- Raw text content
tokenized_content TSVECTOR -- For keyword search
);
-- Indexes for performance
CREATE INDEX content_gin_idx ON table USING GIN(tokenized_content); -- Keyword search
-- Vector index created automatically by pgvector
## Usage
To use PGVector in your Llama Stack project, follow these steps:
@ -412,6 +466,25 @@ To use PGVector in your Llama Stack project, follow these steps:
2. Configure your Llama Stack project to use pgvector. (e.g. remote::pgvector).
3. Start storing and querying vectors.
## This is an example how you can set up your environment for using PGVector
1. Export env vars:
```bash
export ENABLE_PGVECTOR=true
export PGVECTOR_HOST=localhost
export PGVECTOR_PORT=5432
export PGVECTOR_DB=llamastack
export PGVECTOR_USER=llamastack
export PGVECTOR_PASSWORD=llamastack
```
2. Create DB:
```bash
psql -h localhost -U postgres -c "CREATE ROLE llamastack LOGIN PASSWORD 'llamastack';"
psql -h localhost -U postgres -c "CREATE DATABASE llamastack OWNER llamastack;"
psql -h localhost -U llamastack -d llamastack -c "CREATE EXTENSION IF NOT EXISTS vector;"
```
## Installation
You can install PGVector using docker:
@ -422,19 +495,18 @@ docker pull pgvector/pgvector:pg17
## Documentation
See [PGVector's documentation](https://github.com/pgvector/pgvector) for more details about PGVector in general.
""",
),
),
RemoteProviderSpec(
api=Api.vector_io,
adapter_type="weaviate",
provider_type="remote::weaviate",
pip_packages=["weaviate-client"],
module="llama_stack.providers.remote.vector_io.weaviate",
config_class="llama_stack.providers.remote.vector_io.weaviate.WeaviateVectorIOConfig",
provider_data_validator="llama_stack.providers.remote.vector_io.weaviate.WeaviateRequestProviderData",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(
adapter_type="weaviate",
pip_packages=["weaviate-client"],
module="llama_stack.providers.remote.vector_io.weaviate",
config_class="llama_stack.providers.remote.vector_io.weaviate.WeaviateVectorIOConfig",
provider_data_validator="llama_stack.providers.remote.vector_io.weaviate.WeaviateRequestProviderData",
description="""
description="""
[Weaviate](https://weaviate.io/) is a vector database provider for Llama Stack.
It allows you to store and query vectors directly within a Weaviate database.
That means you're not limited to storing vectors in memory or in a separate service.
@ -449,6 +521,7 @@ Weaviate supports:
- Metadata filtering
- Multi-modal retrieval
## Usage
To use Weaviate in your Llama Stack project, follow these steps:
@ -464,9 +537,6 @@ To install Weaviate see the [Weaviate quickstart documentation](https://weaviate
## Documentation
See [Weaviate's documentation](https://weaviate.io/developers/weaviate) for more details about Weaviate in general.
""",
),
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
InlineProviderSpec(
api=Api.vector_io,
@ -520,28 +590,29 @@ docker pull qdrant/qdrant
See the [Qdrant documentation](https://qdrant.tech/documentation/) for more details about Qdrant in general.
""",
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(
adapter_type="qdrant",
pip_packages=["qdrant-client"],
module="llama_stack.providers.remote.vector_io.qdrant",
config_class="llama_stack.providers.remote.vector_io.qdrant.QdrantVectorIOConfig",
description="""
Please refer to the inline provider documentation.
""",
),
RemoteProviderSpec(
api=Api.vector_io,
adapter_type="qdrant",
provider_type="remote::qdrant",
pip_packages=["qdrant-client"],
module="llama_stack.providers.remote.vector_io.qdrant",
config_class="llama_stack.providers.remote.vector_io.qdrant.QdrantVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
description="""
Please refer to the inline provider documentation.
""",
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(
adapter_type="milvus",
pip_packages=["pymilvus>=2.4.10"],
module="llama_stack.providers.remote.vector_io.milvus",
config_class="llama_stack.providers.remote.vector_io.milvus.MilvusVectorIOConfig",
description="""
RemoteProviderSpec(
api=Api.vector_io,
adapter_type="milvus",
provider_type="remote::milvus",
pip_packages=["pymilvus>=2.4.10"],
module="llama_stack.providers.remote.vector_io.milvus",
config_class="llama_stack.providers.remote.vector_io.milvus.MilvusVectorIOConfig",
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
description="""
[Milvus](https://milvus.io/) is an inline and remote vector database provider for Llama Stack. It
allows you to store and query vectors directly within a Milvus database.
That means you're not limited to storing vectors in memory or in a separate service.
@ -562,7 +633,13 @@ To use Milvus in your Llama Stack project, follow these steps:
## Installation
You can install Milvus using pymilvus:
If you want to use inline Milvus, you can install:
```bash
pip install pymilvus[milvus-lite]
```
If you want to use remote Milvus, you can install:
```bash
pip install pymilvus
@ -732,14 +809,11 @@ See the [Milvus documentation](https://milvus.io/docs/install-overview.md) for m
For more details on TLS configuration, refer to the [TLS setup guide](https://milvus.io/docs/tls.md).
""",
),
api_dependencies=[Api.inference],
optional_api_dependencies=[Api.files],
),
InlineProviderSpec(
api=Api.vector_io,
provider_type="inline::milvus",
pip_packages=["pymilvus>=2.4.10"],
pip_packages=["pymilvus[milvus-lite]>=2.4.10"],
module="llama_stack.providers.inline.vector_io.milvus",
config_class="llama_stack.providers.inline.vector_io.milvus.MilvusVectorIOConfig",
api_dependencies=[Api.inference],

16
llama_stack/providers/remote/eval/nvidia/eval.py
View file

@ -14,7 +14,6 @@ from llama_stack.apis.datasets import Datasets
from llama_stack.apis.inference import Inference
from llama_stack.apis.scoring import Scoring, ScoringResult
from llama_stack.providers.datatypes import BenchmarksProtocolPrivate
from llama_stack.providers.remote.inference.nvidia.models import MODEL_ENTRIES
from llama_stack.providers.utils.inference.model_registry import ModelRegistryHelper
from .....apis.common.job_types import Job, JobStatus
@ -45,24 +44,29 @@ class NVIDIAEvalImpl(
self.inference_api = inference_api
self.agents_api = agents_api
ModelRegistryHelper.__init__(self, model_entries=MODEL_ENTRIES)
ModelRegistryHelper.__init__(self)
async def initialize(self) -> None: ...
async def shutdown(self) -> None: ...
async def _evaluator_get(self, path):
async def _evaluator_get(self, path: str):
"""Helper for making GET requests to the evaluator service."""
response = requests.get(url=f"{self.config.evaluator_url}{path}")
response.raise_for_status()
return response.json()
async def _evaluator_post(self, path, data):
async def _evaluator_post(self, path: str, data: dict[str, Any]):
"""Helper for making POST requests to the evaluator service."""
response = requests.post(url=f"{self.config.evaluator_url}{path}", json=data)
response.raise_for_status()
return response.json()
async def _evaluator_delete(self, path: str) -> None:
"""Helper for making DELETE requests to the evaluator service."""
response = requests.delete(url=f"{self.config.evaluator_url}{path}")
response.raise_for_status()
async def register_benchmark(self, task_def: Benchmark) -> None:
"""Register a benchmark as an evaluation configuration."""
await self._evaluator_post(
@ -75,6 +79,10 @@ class NVIDIAEvalImpl(
},
)
async def unregister_benchmark(self, benchmark_id: str) -> None:
"""Unregister a benchmark evaluation configuration from NeMo Evaluator."""
await self._evaluator_delete(f"/v1/evaluation/configs/{DEFAULT_NAMESPACE}/{benchmark_id}")
async def run_eval(
self,
benchmark_id: str,

7
llama_stack/providers/remote/files/s3/__init__.py
View file

@ -6,15 +6,14 @@
from typing import Any
from llama_stack.core.datatypes import Api
from llama_stack.core.datatypes import AccessRule, Api
from .config import S3FilesImplConfig
async def get_adapter_impl(config: S3FilesImplConfig, deps: dict[Api, Any]):
async def get_adapter_impl(config: S3FilesImplConfig, deps: dict[Api, Any], policy: list[AccessRule] | None = None):
from .files import S3FilesImpl
# TODO: authorization policies and user separation
impl = S3FilesImpl(config)
impl = S3FilesImpl(config, policy or [])
await impl.initialize()
return impl

198
llama_stack/providers/remote/files/s3/files.py
View file

@ -4,9 +4,9 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import time
import uuid
from typing import Annotated
from datetime import UTC, datetime
from typing import Annotated, Any
import boto3
from botocore.exceptions import BotoCoreError, ClientError, NoCredentialsError
@ -15,14 +15,17 @@ from fastapi import File, Form, Response, UploadFile
from llama_stack.apis.common.errors import ResourceNotFoundError
from llama_stack.apis.common.responses import Order
from llama_stack.apis.files import (
ExpiresAfter,
Files,
ListOpenAIFileResponse,
OpenAIFileDeleteResponse,
OpenAIFileObject,
OpenAIFilePurpose,
)
from llama_stack.core.datatypes import AccessRule
from llama_stack.providers.utils.sqlstore.api import ColumnDefinition, ColumnType
from llama_stack.providers.utils.sqlstore.sqlstore import SqlStore, sqlstore_impl
from llama_stack.providers.utils.sqlstore.authorized_sqlstore import AuthorizedSqlStore
from llama_stack.providers.utils.sqlstore.sqlstore import sqlstore_impl
from .config import S3FilesImplConfig
@ -83,22 +86,85 @@ async def _create_bucket_if_not_exists(client: boto3.client, config: S3FilesImpl
raise RuntimeError(f"Failed to access S3 bucket '{config.bucket_name}': {e}") from e
def _make_file_object(
*,
id: str,
filename: str,
purpose: str,
bytes: int,
created_at: int,
expires_at: int,
**kwargs: Any, # here to ignore any additional fields, e.g. extra fields from AuthorizedSqlStore
) -> OpenAIFileObject:
"""
Construct an OpenAIFileObject and normalize expires_at.
If expires_at is greater than the max we treat it as no-expiration and
return None for expires_at.
The OpenAI spec says expires_at type is Integer, but the implementation
will return None for no expiration.
"""
obj = OpenAIFileObject(
id=id,
filename=filename,
purpose=OpenAIFilePurpose(purpose),
bytes=bytes,
created_at=created_at,
expires_at=expires_at,
)
if obj.expires_at is not None and obj.expires_at > (obj.created_at + ExpiresAfter.MAX):
obj.expires_at = None # type: ignore
return obj
class S3FilesImpl(Files):
"""S3-based implementation of the Files API."""
# TODO: implement expiration, for now a silly offset
_SILLY_EXPIRATION_OFFSET = 100 * 365 * 24 * 60 * 60
def __init__(self, config: S3FilesImplConfig) -> None:
def __init__(self, config: S3FilesImplConfig, policy: list[AccessRule]) -> None:
self._config = config
self.policy = policy
self._client: boto3.client | None = None
self._sql_store: SqlStore | None = None
self._sql_store: AuthorizedSqlStore | None = None
def _now(self) -> int:
"""Return current UTC timestamp as int seconds."""
return int(datetime.now(UTC).timestamp())
async def _get_file(self, file_id: str, return_expired: bool = False) -> dict[str, Any]:
where: dict[str, str | dict] = {"id": file_id}
if not return_expired:
where["expires_at"] = {">": self._now()}
if not (row := await self.sql_store.fetch_one("openai_files", where=where)):
raise ResourceNotFoundError(file_id, "File", "files.list()")
return row
async def _delete_file(self, file_id: str) -> None:
"""Delete a file from S3 and the database."""
try:
self.client.delete_object(
Bucket=self._config.bucket_name,
Key=file_id,
)
except ClientError as e:
if e.response["Error"]["Code"] != "NoSuchKey":
raise RuntimeError(f"Failed to delete file from S3: {e}") from e
await self.sql_store.delete("openai_files", where={"id": file_id})
async def _delete_if_expired(self, file_id: str) -> None:
"""If the file exists and is expired, delete it."""
if row := await self._get_file(file_id, return_expired=True):
if (expires_at := row.get("expires_at")) and expires_at <= self._now():
await self._delete_file(file_id)
async def initialize(self) -> None:
self._client = _create_s3_client(self._config)
await _create_bucket_if_not_exists(self._client, self._config)
self._sql_store = sqlstore_impl(self._config.metadata_store)
self._sql_store = AuthorizedSqlStore(sqlstore_impl(self._config.metadata_store), self.policy)
await self._sql_store.create_table(
"openai_files",
{
@ -121,7 +187,7 @@ class S3FilesImpl(Files):
return self._client
@property
def sql_store(self) -> SqlStore:
def sql_store(self) -> AuthorizedSqlStore:
assert self._sql_store is not None, "Provider not initialized"
return self._sql_store
@ -129,27 +195,47 @@ class S3FilesImpl(Files):
self,
file: Annotated[UploadFile, File()],
purpose: Annotated[OpenAIFilePurpose, Form()],
expires_after_anchor: Annotated[str | None, Form(alias="expires_after[anchor]")] = None,
expires_after_seconds: Annotated[int | None, Form(alias="expires_after[seconds]")] = None,
) -> OpenAIFileObject:
file_id = f"file-{uuid.uuid4().hex}"
filename = getattr(file, "filename", None) or "uploaded_file"
created_at = int(time.time())
expires_at = created_at + self._SILLY_EXPIRATION_OFFSET
created_at = self._now()
expires_after = None
if expires_after_anchor is not None or expires_after_seconds is not None:
# we use ExpiresAfter to validate input
expires_after = ExpiresAfter(
anchor=expires_after_anchor, # type: ignore[arg-type]
seconds=expires_after_seconds, # type: ignore[arg-type]
)
# the default is no expiration.
# to implement no expiration we set an expiration beyond the max.
# we'll hide this fact from users when returning the file object.
expires_at = created_at + ExpiresAfter.MAX * 42
# the default for BATCH files is 30 days, which happens to be the expiration max.
if purpose == OpenAIFilePurpose.BATCH:
expires_at = created_at + ExpiresAfter.MAX
if expires_after is not None:
expires_at = created_at + expires_after.seconds
content = await file.read()
file_size = len(content)
await self.sql_store.insert(
"openai_files",
{
"id": file_id,
"filename": filename,
"purpose": purpose.value,
"bytes": file_size,
"created_at": created_at,
"expires_at": expires_at,
},
)
entry: dict[str, Any] = {
"id": file_id,
"filename": filename,
"purpose": purpose.value,
"bytes": file_size,
"created_at": created_at,
"expires_at": expires_at,
}
await self.sql_store.insert("openai_files", entry)
try:
self.client.put_object(
@ -163,14 +249,7 @@ class S3FilesImpl(Files):
raise RuntimeError(f"Failed to upload file to S3: {e}") from e
return OpenAIFileObject(
id=file_id,
filename=filename,
purpose=purpose,
bytes=file_size,
created_at=created_at,
expires_at=expires_at,
)
return _make_file_object(**entry)
async def openai_list_files(
self,
@ -183,29 +262,19 @@ class S3FilesImpl(Files):
if not order:
order = Order.desc
where_conditions = {}
where_conditions: dict[str, Any] = {"expires_at": {">": self._now()}}
if purpose:
where_conditions["purpose"] = purpose.value
paginated_result = await self.sql_store.fetch_all(
table="openai_files",
where=where_conditions if where_conditions else None,
where=where_conditions,
order_by=[("created_at", order.value)],
cursor=("id", after) if after else None,
limit=limit,
)
files = [
OpenAIFileObject(
id=row["id"],
filename=row["filename"],
purpose=OpenAIFilePurpose(row["purpose"]),
bytes=row["bytes"],
created_at=row["created_at"],
expires_at=row["expires_at"],
)
for row in paginated_result.data
]
files = [_make_file_object(**row) for row in paginated_result.data]
return ListOpenAIFileResponse(
data=files,
@ -216,41 +285,20 @@ class S3FilesImpl(Files):
)
async def openai_retrieve_file(self, file_id: str) -> OpenAIFileObject:
row = await self.sql_store.fetch_one("openai_files", where={"id": file_id})
if not row:
raise ResourceNotFoundError(file_id, "File", "files.list()")
return OpenAIFileObject(
id=row["id"],
filename=row["filename"],
purpose=OpenAIFilePurpose(row["purpose"]),
bytes=row["bytes"],
created_at=row["created_at"],
expires_at=row["expires_at"],
)
await self._delete_if_expired(file_id)
row = await self._get_file(file_id)
return _make_file_object(**row)
async def openai_delete_file(self, file_id: str) -> OpenAIFileDeleteResponse:
row = await self.sql_store.fetch_one("openai_files", where={"id": file_id})
if not row:
raise ResourceNotFoundError(file_id, "File", "files.list()")
try:
self.client.delete_object(
Bucket=self._config.bucket_name,
Key=row["id"],
)
except ClientError as e:
if e.response["Error"]["Code"] != "NoSuchKey":
raise RuntimeError(f"Failed to delete file from S3: {e}") from e
await self.sql_store.delete("openai_files", where={"id": file_id})
await self._delete_if_expired(file_id)
_ = await self._get_file(file_id) # raises if not found
await self._delete_file(file_id)
return OpenAIFileDeleteResponse(id=file_id, deleted=True)
async def openai_retrieve_file_content(self, file_id: str) -> Response:
row = await self.sql_store.fetch_one("openai_files", where={"id": file_id})
if not row:
raise ResourceNotFoundError(file_id, "File", "files.list()")
await self._delete_if_expired(file_id)
row = await self._get_file(file_id)
try:
response = self.client.get_object(
@ -261,7 +309,7 @@ class S3FilesImpl(Files):
content = response["Body"].read()
except ClientError as e:
if e.response["Error"]["Code"] == "NoSuchKey":
await self.sql_store.delete("openai_files", where={"id": file_id})
await self._delete_file(file_id)
raise ResourceNotFoundError(file_id, "File", "files.list()") from e
raise RuntimeError(f"Failed to download file from S3: {e}") from e

Some files were not shown because too many files have changed in this diff Show more