feat: Bring Your Own API (BYOA) (#2228)
Some checks failed
Coverage Badge / unit-tests (push) Failing after 1s
Integration Auth Tests / test-matrix (oauth2_token) (push) Failing after 1s
Installer CI / lint (push) Failing after 3s
Integration Tests / discover-tests (push) Successful in 3s
Installer CI / smoke-test-on-dev (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.12, inline::faiss) (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.12, inline::sqlite-vec) (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.12, inline::milvus) (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.12, remote::chromadb) (push) Failing after 6s
Python Package Build Test / build (3.12) (push) Failing after 3s
Python Package Build Test / build (3.13) (push) Failing after 2s
Test Llama Stack Build / generate-matrix (push) Successful in 3s
SqlStore Integration Tests / test-postgres (3.12) (push) Failing after 10s
Test Llama Stack Build / build-single-provider (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.13, inline::faiss) (push) Failing after 5s
SqlStore Integration Tests / test-postgres (3.13) (push) Failing after 13s
Unit Tests / unit-tests (3.13) (push) Failing after 6s
Test External API and Providers / test-external (venv) (push) Failing after 5s
Test Llama Stack Build / build-ubi9-container-distribution (push) Failing after 6s
Test Llama Stack Build / build-custom-container-distribution (push) Failing after 5s
Unit Tests / unit-tests (3.12) (push) Failing after 7s
Vector IO Integration Tests / test-matrix (3.13, remote::pgvector) (push) Failing after 6s
Update ReadTheDocs / update-readthedocs (push) Failing after 8s
Integration Tests / test-matrix (push) Failing after 8s
Vector IO Integration Tests / test-matrix (3.13, remote::chromadb) (push) Failing after 5s
Vector IO Integration Tests / test-matrix (3.12, remote::pgvector) (push) Failing after 6s
Vector IO Integration Tests / test-matrix (3.13, inline::sqlite-vec) (push) Failing after 6s
Vector IO Integration Tests / test-matrix (3.13, inline::milvus) (push) Failing after 5s
Test Llama Stack Build / build (push) Failing after 6s
Pre-commit / pre-commit (push) Successful in 57s

# What does this PR do?

Prototype on a new feature to allow new APIs to be plugged in Llama
Stack. Opened for early feedback on the approach and test appetite on
the functionality.

@ashwinb @raghotham open for early feedback, thanks!

---------

Signed-off-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
This commit is contained in:
Sébastien Han 2025-07-24 22:41:14 +02:00 committed by GitHub
parent 341504869e
commit 632cf9eb72
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
27 changed files with 828 additions and 229 deletions

View file

@ -14,6 +14,7 @@ from termcolor import cprint
from llama_stack.distribution.datatypes import BuildConfig
from llama_stack.distribution.distribution import get_provider_registry
from llama_stack.distribution.external import load_external_apis
from llama_stack.distribution.utils.exec import run_command
from llama_stack.distribution.utils.image_types import LlamaStackImageType
from llama_stack.providers.datatypes import Api
@ -105,6 +106,11 @@ def build_image(
normal_deps, special_deps = get_provider_dependencies(build_config)
normal_deps += SERVER_DEPENDENCIES
if build_config.external_apis_dir:
external_apis = load_external_apis(build_config)
if external_apis:
for _, api_spec in external_apis.items():
normal_deps.extend(api_spec.pip_packages)
if build_config.image_type == LlamaStackImageType.CONTAINER.value:
script = str(importlib.resources.files("llama_stack") / "distribution/build_container.sh")

View file

@ -387,6 +387,11 @@ a default SQLite store will be used.""",
description="Path to directory containing external provider implementations. The providers code and dependencies must be installed on the system.",
)
external_apis_dir: Path | None = Field(
default=None,
description="Path to directory containing external API implementations. The APIs code and dependencies must be installed on the system.",
)
@field_validator("external_providers_dir")
@classmethod
def validate_external_providers_dir(cls, v):
@ -418,6 +423,10 @@ class BuildConfig(BaseModel):
default_factory=list,
description="Additional pip packages to install in the distribution. These packages will be installed in the distribution environment.",
)
external_apis_dir: Path | None = Field(
default=None,
description="Path to directory containing external API implementations. The APIs code and dependencies must be installed on the system.",
)
@field_validator("external_providers_dir")
@classmethod

View file

@ -12,6 +12,7 @@ from typing import Any
import yaml
from pydantic import BaseModel
from llama_stack.distribution.external import load_external_apis
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import (
AdapterSpec,
@ -133,16 +134,34 @@ def get_provider_registry(
ValueError: If any provider spec is invalid
"""
ret: dict[Api, dict[str, ProviderSpec]] = {}
registry: dict[Api, dict[str, ProviderSpec]] = {}
for api in providable_apis():
name = api.name.lower()
logger.debug(f"Importing module {name}")
try:
module = importlib.import_module(f"llama_stack.providers.registry.{name}")
ret[api] = {a.provider_type: a for a in module.available_providers()}
registry[api] = {a.provider_type: a for a in module.available_providers()}
except ImportError as e:
logger.warning(f"Failed to import module {name}: {e}")
# Refresh providable APIs with external APIs if any
external_apis = load_external_apis(config)
for api, api_spec in external_apis.items():
name = api_spec.name.lower()
logger.info(f"Importing external API {name} module {api_spec.module}")
try:
module = importlib.import_module(api_spec.module)
registry[api] = {a.provider_type: a for a in module.available_providers()}
except (ImportError, AttributeError) as e:
# Populate the registry with an empty dict to avoid breaking the provider registry
# This assume that the in-tree provider(s) are not available for this API which means
# that users will need to use external providers for this API.
registry[api] = {}
logger.error(
f"Failed to import external API {name}: {e}. Could not populate the in-tree provider(s) registry for {api.name}. \n"
"Install the API package to load any in-tree providers for this API."
)
# Check if config has the external_providers_dir attribute
if config and hasattr(config, "external_providers_dir") and config.external_providers_dir:
external_providers_dir = os.path.abspath(os.path.expanduser(config.external_providers_dir))
@ -175,11 +194,9 @@ def get_provider_registry(
else:
spec = _load_inline_provider_spec(spec_data, api, provider_name)
provider_type_key = f"inline::{provider_name}"
logger.info(f"Loaded {provider_type} provider spec for {provider_type_key} from {spec_path}")
if provider_type_key in ret[api]:
if provider_type_key in registry[api]:
logger.warning(f"Overriding already registered provider {provider_type_key} for {api.name}")
ret[api][provider_type_key] = spec
registry[api][provider_type_key] = spec
logger.info(f"Successfully loaded external provider {provider_type_key}")
except yaml.YAMLError as yaml_err:
logger.error(f"Failed to parse YAML file {spec_path}: {yaml_err}")
@ -187,4 +204,4 @@ def get_provider_registry(
except Exception as e:
logger.error(f"Failed to load provider spec from {spec_path}: {e}")
raise e
return ret
return registry

View file

@ -0,0 +1,54 @@
# 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 yaml
from llama_stack.apis.datatypes import Api, ExternalApiSpec
from llama_stack.distribution.datatypes import BuildConfig, StackRunConfig
from llama_stack.log import get_logger
logger = get_logger(name=__name__, category="core")
def load_external_apis(config: StackRunConfig | BuildConfig | None) -> dict[Api, ExternalApiSpec]:
"""Load external API specifications from the configured directory.
Args:
config: StackRunConfig or BuildConfig containing the external APIs directory path
Returns:
A dictionary mapping API names to their specifications
"""
if not config or not config.external_apis_dir:
return {}
external_apis_dir = config.external_apis_dir.expanduser().resolve()
if not external_apis_dir.is_dir():
logger.error(f"External APIs directory is not a directory: {external_apis_dir}")
return {}
logger.info(f"Loading external APIs from {external_apis_dir}")
external_apis: dict[Api, ExternalApiSpec] = {}
# Look for YAML files in the external APIs directory
for yaml_path in external_apis_dir.glob("*.yaml"):
try:
with open(yaml_path) as f:
spec_data = yaml.safe_load(f)
spec = ExternalApiSpec(**spec_data)
api = Api.add(spec.name)
logger.info(f"Loaded external API spec for {spec.name} from {yaml_path}")
external_apis[api] = spec
except yaml.YAMLError as yaml_err:
logger.error(f"Failed to parse YAML file {yaml_path}: {yaml_err}")
raise
except Exception:
logger.exception(f"Failed to load external API spec from {yaml_path}")
raise
return external_apis

View file

@ -16,6 +16,7 @@ from llama_stack.apis.inspect import (
VersionInfo,
)
from llama_stack.distribution.datatypes import StackRunConfig
from llama_stack.distribution.external import load_external_apis
from llama_stack.distribution.server.routes import get_all_api_routes
from llama_stack.providers.datatypes import HealthStatus
@ -42,8 +43,9 @@ class DistributionInspectImpl(Inspect):
run_config: StackRunConfig = self.config.run_config
ret = []
api_to_routes = get_all_api_routes()
for api, endpoints in api_to_routes.items():
external_apis = load_external_apis(run_config)
all_endpoints = get_all_api_routes(external_apis)
for api, endpoints in all_endpoints.items():
# Always include provider and inspect APIs, filter others based on run config
if api.value in ["providers", "inspect"]:
ret.extend(

View file

@ -11,6 +11,7 @@ from llama_stack.apis.agents import Agents
from llama_stack.apis.benchmarks import Benchmarks
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.datatypes import ExternalApiSpec
from llama_stack.apis.eval import Eval
from llama_stack.apis.files import Files
from llama_stack.apis.inference import Inference, InferenceProvider
@ -35,6 +36,7 @@ from llama_stack.distribution.datatypes import (
StackRunConfig,
)
from llama_stack.distribution.distribution import builtin_automatically_routed_apis
from llama_stack.distribution.external import load_external_apis
from llama_stack.distribution.store import DistributionRegistry
from llama_stack.distribution.utils.dynamic import instantiate_class_type
from llama_stack.log import get_logger
@ -59,8 +61,16 @@ class InvalidProviderError(Exception):
pass
def api_protocol_map() -> dict[Api, Any]:
return {
def api_protocol_map(external_apis: dict[Api, ExternalApiSpec] | None = None) -> dict[Api, Any]:
"""Get a mapping of API types to their protocol classes.
Args:
external_apis: Optional dictionary of external API specifications
Returns:
Dictionary mapping API types to their protocol classes
"""
protocols = {
Api.providers: ProvidersAPI,
Api.agents: Agents,
Api.inference: Inference,
@ -83,10 +93,23 @@ def api_protocol_map() -> dict[Api, Any]:
Api.files: Files,
}
if external_apis:
for api, api_spec in external_apis.items():
try:
module = importlib.import_module(api_spec.module)
api_class = getattr(module, api_spec.protocol)
def api_protocol_map_for_compliance_check() -> dict[Api, Any]:
protocols[api] = api_class
except (ImportError, AttributeError):
logger.exception(f"Failed to load external API {api_spec.name}")
return protocols
def api_protocol_map_for_compliance_check(config: Any) -> dict[Api, Any]:
external_apis = load_external_apis(config)
return {
**api_protocol_map(),
**api_protocol_map(external_apis),
Api.inference: InferenceProvider,
}
@ -250,7 +273,7 @@ async def instantiate_providers(
dist_registry: DistributionRegistry,
run_config: StackRunConfig,
policy: list[AccessRule],
) -> dict:
) -> dict[Api, Any]:
"""Instantiates providers asynchronously while managing dependencies."""
impls: dict[Api, Any] = {}
inner_impls_by_provider_id: dict[str, dict[str, Any]] = {f"inner-{x.value}": {} for x in router_apis}
@ -360,7 +383,7 @@ async def instantiate_provider(
impl.__provider_spec__ = provider_spec
impl.__provider_config__ = config
protocols = api_protocol_map_for_compliance_check()
protocols = api_protocol_map_for_compliance_check(run_config)
additional_protocols = additional_protocols_map()
# TODO: check compliance for special tool groups
# the impl should be for Api.tool_runtime, the name should be the special tool group, the protocol should be the special tool group protocol

View file

@ -12,10 +12,10 @@ from typing import Any
from aiohttp import hdrs
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.distribution.resolver import api_protocol_map
from llama_stack.providers.datatypes import Api
from llama_stack.schema_utils import WebMethod
EndpointFunc = Callable[..., Any]
@ -32,10 +32,12 @@ def toolgroup_protocol_map():
}
def get_all_api_routes() -> dict[Api, list[tuple[Route, WebMethod]]]:
def get_all_api_routes(
external_apis: dict[Api, ExternalApiSpec] | None = None,
) -> dict[Api, list[tuple[Route, WebMethod]]]:
apis = {}
protocols = api_protocol_map()
protocols = api_protocol_map(external_apis)
toolgroup_protocols = toolgroup_protocol_map()
for api, protocol in protocols.items():
routes = []
@ -74,8 +76,8 @@ def get_all_api_routes() -> dict[Api, list[tuple[Route, WebMethod]]]:
return apis
def initialize_route_impls(impls: dict[Api, Any]) -> RouteImpls:
api_to_routes = get_all_api_routes()
def initialize_route_impls(impls, external_apis: dict[Api, ExternalApiSpec] | None = None) -> RouteImpls:
api_to_routes = get_all_api_routes(external_apis)
route_impls: RouteImpls = {}
def _convert_path_to_regex(path: str) -> str:

View file

@ -40,6 +40,7 @@ from llama_stack.distribution.datatypes import (
StackRunConfig,
)
from llama_stack.distribution.distribution import builtin_automatically_routed_apis
from llama_stack.distribution.external import ExternalApiSpec, load_external_apis
from llama_stack.distribution.request_headers import PROVIDER_DATA_VAR, User, request_provider_data_context
from llama_stack.distribution.resolver import InvalidProviderError
from llama_stack.distribution.server.routes import (
@ -282,9 +283,10 @@ def create_dynamic_typed_route(func: Any, method: str, route: str) -> Callable:
class TracingMiddleware:
def __init__(self, app, impls):
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")
@ -301,7 +303,7 @@ class TracingMiddleware:
return await self.app(scope, receive, send)
if not hasattr(self, "route_impls"):
self.route_impls = initialize_route_impls(self.impls)
self.route_impls = initialize_route_impls(self.impls, self.external_apis)
try:
_, _, route_path, webmethod = find_matching_route(
@ -486,7 +488,9 @@ def main(args: argparse.Namespace | None = None):
else:
setup_logger(TelemetryAdapter(TelemetryConfig(), {}))
all_routes = get_all_api_routes()
# Load external APIs if configured
external_apis = load_external_apis(config)
all_routes = get_all_api_routes(external_apis)
if config.apis:
apis_to_serve = set(config.apis)
@ -505,7 +509,10 @@ def main(args: argparse.Namespace | None = None):
api = Api(api_str)
routes = all_routes[api]
impl = impls[api]
try:
impl = impls[api]
except KeyError as e:
raise ValueError(f"Could not find provider implementation for {api} API") from e
for route, _ in routes:
if not hasattr(impl, route.name):
@ -536,7 +543,7 @@ def main(args: argparse.Namespace | None = None):
app.exception_handler(Exception)(global_exception_handler)
app.__llama_stack_impls__ = impls
app.add_middleware(TracingMiddleware, impls=impls)
app.add_middleware(TracingMiddleware, impls=impls, external_apis=external_apis)
import uvicorn