feat: convert provider config to a file path

currently provider.config is a dictionary. Introduce the ability to specify either a file path or the current in-file dictionary. Allowing users to specify a file path
enables more robust config management allowing stack administrators to swap in different provider configs seamlessly

Signed-off-by: Charlie Doern <cdoern@redhat.com>
This commit is contained in:
Charlie Doern 2025-06-23 11:20:54 -04:00
parent 4d0d2d685f
commit 31cc971503
14 changed files with 226 additions and 174 deletions

View file

@ -16,6 +16,7 @@ from llama_stack.distribution.datatypes import (
from llama_stack.distribution.distribution import (
builtin_automatically_routed_apis,
get_provider_registry,
resolve_config,
)
from llama_stack.distribution.utils.config_dirs import EXTERNAL_PROVIDERS_DIR
from llama_stack.distribution.utils.dynamic import instantiate_class_type
@ -30,7 +31,7 @@ def configure_single_provider(registry: dict[str, ProviderSpec], provider: Provi
config_type = instantiate_class_type(provider_spec.config_class)
try:
if provider.config:
existing = config_type(**provider.config)
existing = resolve_config(provider=provider, provider_spec=provider_spec)
else:
existing = None
except Exception:

View file

@ -150,7 +150,10 @@ class Provider(BaseModel):
# when the provider is enabled via a conditional environment variable
provider_id: str | None
provider_type: str
config: dict[str, Any]
config: str | dict[str, Any] | None = Field(
default=None,
description="Provider configuration dictionary or path to provider configuration file",
)
class LoggingConfig(BaseModel):

View file

@ -7,11 +7,15 @@
import glob
import importlib
import os
from pathlib import Path
from typing import Any
import yaml
from pydantic import BaseModel
from llama_stack.distribution.datatypes import Provider
from llama_stack.distribution.utils.dynamic import instantiate_class_type
from llama_stack.distribution.utils.env import replace_env_vars
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import (
AdapterSpec,
@ -188,3 +192,22 @@ def get_provider_registry(
logger.error(f"Failed to load provider spec from {spec_path}: {e}")
raise e
return ret
def resolve_config(provider: Provider, provider_spec: ProviderSpec | None = None, api: str | None = None):
if not provider_spec:
if not api:
raise ValueError("In order to get provider spec, must have API")
registry = get_provider_registry()
provider_spec = registry[Api(api)][provider.provider_type]
config_type = instantiate_class_type(provider_spec.config_class)
try:
if provider.config and isinstance(provider.config, str):
with open(Path(provider.config).expanduser().resolve()) as f:
config: dict[str, Any] = yaml.safe_load(f)
replaced = replace_env_vars(config)
return config_type(**replaced)
elif provider.config is not None:
return config_type(**provider.config)
except Exception as e:
raise ValueError("Error getting provider config") from e

View file

@ -10,6 +10,7 @@ from typing import Any
from pydantic import BaseModel
from llama_stack.apis.providers import ListProvidersResponse, ProviderInfo, Providers
from llama_stack.distribution.distribution import resolve_config
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import HealthResponse, HealthStatus
@ -51,12 +52,13 @@ class ProviderImpl(Providers):
# Skip providers that are not enabled
if p.provider_id is None:
continue
config = resolve_config(provider=p, api=api)
ret.append(
ProviderInfo(
api=api,
provider_id=p.provider_id,
provider_type=p.provider_type,
config=p.config,
config=dict(config),
health=providers_health.get(api, {}).get(
p.provider_id,
HealthResponse(

View file

@ -34,9 +34,8 @@ from llama_stack.distribution.datatypes import (
RoutingTableProviderSpec,
StackRunConfig,
)
from llama_stack.distribution.distribution import builtin_automatically_routed_apis
from llama_stack.distribution.distribution import builtin_automatically_routed_apis, resolve_config
from llama_stack.distribution.store import DistributionRegistry
from llama_stack.distribution.utils.dynamic import instantiate_class_type
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import (
Api,
@ -156,7 +155,7 @@ def specs_for_autorouted_apis(apis_to_serve: list[str] | set[str]) -> dict[str,
"__builtin__": ProviderWithSpec(
provider_id="__routing_table__",
provider_type="__routing_table__",
config={},
config=None,
spec=RoutingTableProviderSpec(
api=info.routing_table_api,
router_api=info.router_api,
@ -171,7 +170,7 @@ def specs_for_autorouted_apis(apis_to_serve: list[str] | set[str]) -> dict[str,
"__builtin__": ProviderWithSpec(
provider_id="__autorouted__",
provider_type="__autorouted__",
config={},
config=None,
spec=AutoRoutedProviderSpec(
api=info.router_api,
module="llama_stack.distribution.routers",
@ -329,8 +328,7 @@ async def instantiate_provider(
module = importlib.import_module(provider_spec.module)
args = []
if isinstance(provider_spec, RemoteProviderSpec):
config_type = instantiate_class_type(provider_spec.config_class)
config = config_type(**provider.config)
config = resolve_config(provider=provider, provider_spec=provider_spec)
method = "get_adapter_impl"
args = [config, deps]
@ -348,8 +346,7 @@ async def instantiate_provider(
else:
method = "get_provider_impl"
config_type = instantiate_class_type(provider_spec.config_class)
config = config_type(**provider.config)
config = resolve_config(provider=provider, provider_spec=provider_spec)
args = [config, deps]
if "policy" in inspect.signature(getattr(module, method)).parameters:
args.append(policy)

View file

@ -42,11 +42,10 @@ from llama_stack.distribution.server.routes import (
)
from llama_stack.distribution.stack import (
construct_stack,
replace_env_vars,
validate_env_pair,
)
from llama_stack.distribution.utils.config import redact_sensitive_fields
from llama_stack.distribution.utils.context import preserve_contexts_async_generator
from llama_stack.distribution.utils.env import replace_env_vars, validate_env_pair
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import Api
from llama_stack.providers.inline.telemetry.meta_reference.config import TelemetryConfig
@ -408,9 +407,10 @@ def main(args: argparse.Namespace | None = None):
log_line = f"Using config file: {config_file}"
elif args.template:
config_file = Path(REPO_ROOT) / "llama_stack" / "templates" / args.template / "run.yaml"
if not config_file.exists():
provider_configs = Path(REPO_ROOT) / "llama_stack" / "templates" / args.config / "provider_configs"
if not config_file.exists() or not provider_configs.exists():
raise ValueError(f"Template {args.template} does not exist")
log_line = f"Using template {args.template} config file: {config_file}"
log_line = f"Using template {args.template} config file: {config_file} and provider_config directory: {provider_configs}"
else:
raise ValueError("Either --config or --template must be provided")

View file

@ -5,8 +5,6 @@
# the root directory of this source tree.
import importlib.resources
import os
import re
import tempfile
from typing import Any
@ -40,6 +38,7 @@ from llama_stack.distribution.providers import ProviderImpl, ProviderImplConfig
from llama_stack.distribution.resolver import ProviderRegistry, resolve_impls
from llama_stack.distribution.store.registry import create_dist_registry
from llama_stack.distribution.utils.dynamic import instantiate_class_type
from llama_stack.distribution.utils.env import replace_env_vars
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import Api
@ -123,136 +122,6 @@ async def register_resources(run_config: StackRunConfig, impls: dict[Api, Any]):
)
class EnvVarError(Exception):
def __init__(self, var_name: str, path: str = ""):
self.var_name = var_name
self.path = path
super().__init__(
f"Environment variable '{var_name}' not set or empty {f'at {path}' if path else ''}. "
f"Use ${{env.{var_name}:=default_value}} to provide a default value, "
f"${{env.{var_name}:+value_if_set}} to make the field conditional, "
f"or ensure the environment variable is set."
)
def replace_env_vars(config: Any, path: str = "") -> Any:
if isinstance(config, dict):
result = {}
for k, v in config.items():
try:
result[k] = replace_env_vars(v, f"{path}.{k}" if path else k)
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return result
elif isinstance(config, list):
result = []
for i, v in enumerate(config):
try:
result.append(replace_env_vars(v, f"{path}[{i}]"))
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return result
elif isinstance(config, str):
# Pattern supports bash-like syntax: := for default and :+ for conditional and a optional value
pattern = r"\${env\.([A-Z0-9_]+)(?::([=+])([^}]*))?}"
def get_env_var(match: re.Match):
env_var = match.group(1)
operator = match.group(2) # '=' for default, '+' for conditional
value_expr = match.group(3)
env_value = os.environ.get(env_var)
if operator == "=": # Default value syntax: ${env.FOO:=default}
# If the env is set like ${env.FOO:=default} then use the env value when set
if env_value:
value = env_value
else:
# If the env is not set, look for a default value
# value_expr returns empty string (not None) when not matched
# This means ${env.FOO:=} and it's accepted and returns empty string - just like bash
if value_expr == "":
return ""
else:
value = value_expr
elif operator == "+": # Conditional value syntax: ${env.FOO:+value_if_set}
# If the env is set like ${env.FOO:+value_if_set} then use the value_if_set
if env_value:
if value_expr:
value = value_expr
# This means ${env.FOO:+}
else:
# Just like bash, this doesn't care whether the env is set or not and applies
# the value, in this case the empty string
return ""
else:
# Just like bash, this doesn't care whether the env is set or not, since it's not set
# we return an empty string
value = ""
else: # No operator case: ${env.FOO}
if not env_value:
raise EnvVarError(env_var, path)
value = env_value
# expand "~" from the values
return os.path.expanduser(value)
try:
result = re.sub(pattern, get_env_var, config)
return _convert_string_to_proper_type(result)
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return config
def _convert_string_to_proper_type(value: str) -> Any:
# This might be tricky depending on what the config type is, if 'str | None' we are
# good, if 'str' we need to keep the empty string... 'str | None' is more common and
# providers config should be typed this way.
# TODO: we could try to load the config class and see if the config has a field with type 'str | None'
# and then convert the empty string to None or not
if value == "":
return None
lowered = value.lower()
if lowered == "true":
return True
elif lowered == "false":
return False
try:
return int(value)
except ValueError:
pass
try:
return float(value)
except ValueError:
pass
return value
def validate_env_pair(env_pair: str) -> tuple[str, str]:
"""Validate and split an environment variable key-value pair."""
try:
key, value = env_pair.split("=", 1)
key = key.strip()
if not key:
raise ValueError(f"Empty key in environment variable pair: {env_pair}")
if not all(c.isalnum() or c == "_" for c in key):
raise ValueError(f"Key must contain only alphanumeric characters and underscores: {key}")
return key, value
except ValueError as e:
raise ValueError(
f"Invalid environment variable format '{env_pair}': {str(e)}. Expected format: KEY=value"
) from e
def add_internal_implementations(impls: dict[Api, Any], run_config: StackRunConfig) -> None:
"""Add internal implementations (inspect and providers) to the implementations dictionary.

View file

@ -0,0 +1,143 @@
# 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 os
import re
from typing import Any
from llama_stack.log import get_logger
logger = get_logger(name=__name__, category="core")
class EnvVarError(Exception):
def __init__(self, var_name: str, path: str = ""):
self.var_name = var_name
self.path = path
super().__init__(
f"Environment variable '{var_name}' not set or empty {f'at {path}' if path else ''}. "
f"Use ${{env.{var_name}:=default_value}} to provide a default value, "
f"${{env.{var_name}:+value_if_set}} to make the field conditional, "
f"or ensure the environment variable is set."
)
def replace_env_vars(config: Any, path: str = "") -> Any:
if isinstance(config, dict):
result_dict = {}
for k, v in config.items():
try:
result_dict[k] = replace_env_vars(v, f"{path}.{k}" if path else k)
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return result_dict
elif isinstance(config, list):
result_list = []
for i, v in enumerate(config):
try:
result_list.append(replace_env_vars(v, f"{path}[{i}]"))
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return result_list
elif isinstance(config, str):
# Pattern supports bash-like syntax: := for default and :+ for conditional and a optional value
pattern = r"\${env\.([A-Z0-9_]+)(?::([=+])([^}]*))?}"
def get_env_var(match: re.Match):
env_var = match.group(1)
operator = match.group(2) # '=' for default, '+' for conditional
value_expr = match.group(3)
env_value = os.environ.get(env_var)
if operator == "=": # Default value syntax: ${env.FOO:=default}
# If the env is set like ${env.FOO:=default} then use the env value when set
if env_value:
value = env_value
else:
# If the env is not set, look for a default value
# value_expr returns empty string (not None) when not matched
# This means ${env.FOO:=} and it's accepted and returns empty string - just like bash
if value_expr == "":
return ""
else:
value = value_expr
elif operator == "+": # Conditional value syntax: ${env.FOO:+value_if_set}
# If the env is set like ${env.FOO:+value_if_set} then use the value_if_set
if env_value:
if value_expr:
value = value_expr
# This means ${env.FOO:+}
else:
# Just like bash, this doesn't care whether the env is set or not and applies
# the value, in this case the empty string
return ""
else:
# Just like bash, this doesn't care whether the env is set or not, since it's not set
# we return an empty string
value = ""
else: # No operator case: ${env.FOO}
if not env_value:
raise EnvVarError(env_var, path)
value = env_value
# expand "~" from the values
return os.path.expanduser(value)
try:
result = re.sub(pattern, get_env_var, config)
return _convert_string_to_proper_type(result)
except EnvVarError as e:
raise EnvVarError(e.var_name, e.path) from None
return config
def _convert_string_to_proper_type(value: str) -> Any:
# This might be tricky depending on what the config type is, if 'str | None' we are
# good, if 'str' we need to keep the empty string... 'str | None' is more common and
# providers config should be typed this way.
# TODO: we could try to load the config class and see if the config has a field with type 'str | None'
# and then convert the empty string to None or not
if value == "":
return None
lowered = value.lower()
if lowered == "true":
return True
elif lowered == "false":
return False
try:
return int(value)
except ValueError:
pass
try:
return float(value)
except ValueError:
pass
return value
def validate_env_pair(env_pair: str) -> tuple[str, str]:
"""Validate and split an environment variable key-value pair."""
try:
key, value = env_pair.split("=", 1)
key = key.strip()
if not key:
raise ValueError(f"Empty key in environment variable pair: {env_pair}")
if not all(c.isalnum() or c == "_" for c in key):
raise ValueError(f"Key must contain only alphanumeric characters and underscores: {key}")
return key, value
except ValueError as e:
raise ValueError(
f"Invalid environment variable format '{env_pair}': {str(e)}. Expected format: KEY=value"
) from e