phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-03 09:53:45 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'upstream/main' into rm-build

Browse source
This commit is contained in:
Charlie Doern 2025-11-21 10:28:36 -05:00
commit c7afd497c8
122 changed files with 16310 additions and 785 deletions
Download patch file Download diff file Expand all files Collapse all files

23
src/llama_stack/distributions/ci-tests/run-with-postgres-store.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

23
src/llama_stack/distributions/ci-tests/run.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

3
src/llama_stack/distributions/nvidia/run-with-safety.yaml
View file

@ -16,9 +16,8 @@ providers:
- provider_id: nvidia
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: nvidia
provider_type: remote::nvidia
config:

3
src/llama_stack/distributions/nvidia/run.yaml
View file

@ -16,9 +16,8 @@ providers:
- provider_id: nvidia
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
vector_io:
- provider_id: faiss
provider_type: inline::faiss

4
src/llama_stack/distributions/open-benchmark/run.yaml
View file

@ -27,12 +27,12 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
vector_io:
- provider_id: sqlite-vec

2
src/llama_stack/distributions/postgres-demo/run.yaml
View file

@ -11,7 +11,7 @@ providers:
- provider_id: vllm-inference
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=http://localhost:8000/v1}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}

23
src/llama_stack/distributions/starter-gpu/run-with-postgres-store.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

23
src/llama_stack/distributions/starter-gpu/run.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

23
src/llama_stack/distributions/starter/run-with-postgres-store.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

23
src/llama_stack/distributions/starter/run.yaml
View file

@ -17,44 +17,43 @@ providers:
- provider_id: ${env.CEREBRAS_API_KEY:+cerebras}
provider_type: remote::cerebras
config:
base_url: https://api.cerebras.ai
base_url: https://api.cerebras.ai/v1
api_key: ${env.CEREBRAS_API_KEY:=}
- provider_id: ${env.OLLAMA_URL:+ollama}
provider_type: remote::ollama
config:
url: ${env.OLLAMA_URL:=http://localhost:11434}
base_url: ${env.OLLAMA_URL:=http://localhost:11434/v1}
- provider_id: ${env.VLLM_URL:+vllm}
provider_type: remote::vllm
config:
url: ${env.VLLM_URL:=}
base_url: ${env.VLLM_URL:=}
max_tokens: ${env.VLLM_MAX_TOKENS:=4096}
api_token: ${env.VLLM_API_TOKEN:=fake}
tls_verify: ${env.VLLM_TLS_VERIFY:=true}
- provider_id: ${env.TGI_URL:+tgi}
provider_type: remote::tgi
config:
url: ${env.TGI_URL:=}
base_url: ${env.TGI_URL:=}
- provider_id: fireworks
provider_type: remote::fireworks
config:
url: https://api.fireworks.ai/inference/v1
base_url: https://api.fireworks.ai/inference/v1
api_key: ${env.FIREWORKS_API_KEY:=}
- provider_id: together
provider_type: remote::together
config:
url: https://api.together.xyz/v1
base_url: https://api.together.xyz/v1
api_key: ${env.TOGETHER_API_KEY:=}
- provider_id: bedrock
provider_type: remote::bedrock
config:
api_key: ${env.AWS_BEDROCK_API_KEY:=}
api_key: ${env.AWS_BEARER_TOKEN_BEDROCK:=}
region_name: ${env.AWS_DEFAULT_REGION:=us-east-2}
- provider_id: ${env.NVIDIA_API_KEY:+nvidia}
provider_type: remote::nvidia
config:
url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}
base_url: ${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}
api_key: ${env.NVIDIA_API_KEY:=}
append_api_version: ${env.NVIDIA_APPEND_API_VERSION:=True}
- provider_id: openai
provider_type: remote::openai
config:
@ -76,18 +75,18 @@ providers:
- provider_id: groq
provider_type: remote::groq
config:
url: https://api.groq.com
base_url: https://api.groq.com/openai/v1
api_key: ${env.GROQ_API_KEY:=}
- provider_id: sambanova
provider_type: remote::sambanova
config:
url: https://api.sambanova.ai/v1
base_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:=}
base_url: ${env.AZURE_API_BASE:=}
api_version: ${env.AZURE_API_VERSION:=}
api_type: ${env.AZURE_API_TYPE:=}
- provider_id: sentence-transformers

2
src/llama_stack/distributions/watsonx/run.yaml
View file

@ -15,7 +15,7 @@ providers:
- provider_id: watsonx
provider_type: remote::watsonx
config:
url: ${env.WATSONX_BASE_URL:=https://us-south.ml.cloud.ibm.com}
base_url: ${env.WATSONX_BASE_URL:=https://us-south.ml.cloud.ibm.com}
api_key: ${env.WATSONX_API_KEY:=}
project_id: ${env.WATSONX_PROJECT_ID:=}
vector_io:

6
src/llama_stack/providers/inline/agents/meta_reference/__init__.py
View file

@ -23,12 +23,14 @@ async def get_provider_impl(
config,
deps[Api.inference],
deps[Api.vector_io],
deps[Api.safety],
deps.get(Api.safety),
deps[Api.tool_runtime],
deps[Api.tool_groups],
deps[Api.conversations],
policy,
deps[Api.prompts],
deps[Api.files],
telemetry_enabled,
policy,
)
await impl.initialize()
return impl

11
src/llama_stack/providers/inline/agents/meta_reference/agents.py
View file

@ -12,6 +12,7 @@ from llama_stack.providers.utils.responses.responses_store import ResponsesStore
from llama_stack_api import (
Agents,
Conversations,
Files,
Inference,
ListOpenAIResponseInputItem,
ListOpenAIResponseObject,
@ -22,6 +23,7 @@ from llama_stack_api import (
OpenAIResponsePrompt,
OpenAIResponseText,
Order,
Prompts,
ResponseGuardrail,
Safety,
ToolGroups,
@ -41,10 +43,12 @@ class MetaReferenceAgentsImpl(Agents):
config: MetaReferenceAgentsImplConfig,
inference_api: Inference,
vector_io_api: VectorIO,
safety_api: Safety,
safety_api: Safety | None,
tool_runtime_api: ToolRuntime,
tool_groups_api: ToolGroups,
conversations_api: Conversations,
prompts_api: Prompts,
files_api: Files,
policy: list[AccessRule],
telemetry_enabled: bool = False,
):
@ -56,7 +60,8 @@ class MetaReferenceAgentsImpl(Agents):
self.tool_groups_api = tool_groups_api
self.conversations_api = conversations_api
self.telemetry_enabled = telemetry_enabled
self.prompts_api = prompts_api
self.files_api = files_api
self.in_memory_store = InmemoryKVStoreImpl()
self.openai_responses_impl: OpenAIResponsesImpl | None = None
self.policy = policy
@ -73,6 +78,8 @@ class MetaReferenceAgentsImpl(Agents):
vector_io_api=self.vector_io_api,
safety_api=self.safety_api,
conversations_api=self.conversations_api,
prompts_api=self.prompts_api,
files_api=self.files_api,
)
async def shutdown(self) -> None:

108
src/llama_stack/providers/inline/agents/meta_reference/responses/openai_responses.py
View file

@ -4,6 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import re
import time
import uuid
from collections.abc import AsyncIterator
@ -18,13 +19,17 @@ from llama_stack.providers.utils.responses.responses_store import (
from llama_stack_api import (
ConversationItem,
Conversations,
Files,
Inference,
InvalidConversationIdError,
ListOpenAIResponseInputItem,
ListOpenAIResponseObject,
OpenAIChatCompletionContentPartParam,
OpenAIDeleteResponseObject,
OpenAIMessageParam,
OpenAIResponseInput,
OpenAIResponseInputMessageContentFile,
OpenAIResponseInputMessageContentImage,
OpenAIResponseInputMessageContentText,
OpenAIResponseInputTool,
OpenAIResponseMessage,
@ -34,7 +39,9 @@ from llama_stack_api import (
OpenAIResponseText,
OpenAIResponseTextFormat,
OpenAISystemMessageParam,
OpenAIUserMessageParam,
Order,
Prompts,
ResponseGuardrailSpec,
Safety,
ToolGroups,
@ -46,6 +53,7 @@ from .streaming import StreamingResponseOrchestrator
from .tool_executor import ToolExecutor
from .types import ChatCompletionContext, ToolContext
from .utils import (
convert_response_content_to_chat_content,
convert_response_input_to_chat_messages,
convert_response_text_to_chat_response_format,
extract_guardrail_ids,
@ -67,8 +75,10 @@ class OpenAIResponsesImpl:
tool_runtime_api: ToolRuntime,
responses_store: ResponsesStore,
vector_io_api: VectorIO, # VectorIO
safety_api: Safety,
safety_api: Safety | None,
conversations_api: Conversations,
prompts_api: Prompts,
files_api: Files,
):
self.inference_api = inference_api
self.tool_groups_api = tool_groups_api
@ -82,6 +92,8 @@ class OpenAIResponsesImpl:
tool_runtime_api=tool_runtime_api,
vector_io_api=vector_io_api,
)
self.prompts_api = prompts_api
self.files_api = files_api
async def _prepend_previous_response(
self,
@ -122,11 +134,13 @@ class OpenAIResponsesImpl:
# Use stored messages directly and convert only new input
message_adapter = TypeAdapter(list[OpenAIMessageParam])
messages = message_adapter.validate_python(previous_response.messages)
new_messages = await convert_response_input_to_chat_messages(input, previous_messages=messages)
new_messages = await convert_response_input_to_chat_messages(
input, previous_messages=messages, files_api=self.files_api
)
messages.extend(new_messages)
else:
# Backward compatibility: reconstruct from inputs
messages = await convert_response_input_to_chat_messages(all_input)
messages = await convert_response_input_to_chat_messages(all_input, files_api=self.files_api)
tool_context.recover_tools_from_previous_response(previous_response)
elif conversation is not None:
@ -138,7 +152,7 @@ class OpenAIResponsesImpl:
all_input = input
if not conversation_items.data:
# First turn - just convert the new input
messages = await convert_response_input_to_chat_messages(input)
messages = await convert_response_input_to_chat_messages(input, files_api=self.files_api)
else:
if not stored_messages:
all_input = conversation_items.data
@ -154,14 +168,82 @@ class OpenAIResponsesImpl:
all_input = input
messages = stored_messages or []
new_messages = await convert_response_input_to_chat_messages(all_input, previous_messages=messages)
new_messages = await convert_response_input_to_chat_messages(
all_input, previous_messages=messages, files_api=self.files_api
)
messages.extend(new_messages)
else:
all_input = input
messages = await convert_response_input_to_chat_messages(all_input)
messages = await convert_response_input_to_chat_messages(all_input, files_api=self.files_api)
return all_input, messages, tool_context
async def _prepend_prompt(
self,
messages: list[OpenAIMessageParam],
openai_response_prompt: OpenAIResponsePrompt | None,
) -> None:
"""Prepend prompt template to messages, resolving text/image/file variables.
:param messages: List of OpenAIMessageParam objects
:param openai_response_prompt: (Optional) OpenAIResponsePrompt object with variables
:returns: string of utf-8 characters
"""
if not openai_response_prompt or not openai_response_prompt.id:
return
prompt_version = int(openai_response_prompt.version) if openai_response_prompt.version else None
cur_prompt = await self.prompts_api.get_prompt(openai_response_prompt.id, prompt_version)
if not cur_prompt or not cur_prompt.prompt:
return
cur_prompt_text = cur_prompt.prompt
cur_prompt_variables = cur_prompt.variables
if not openai_response_prompt.variables:
messages.insert(0, OpenAISystemMessageParam(content=cur_prompt_text))
return
# Validate that all provided variables exist in the prompt
for name in openai_response_prompt.variables.keys():
if name not in cur_prompt_variables:
raise ValueError(f"Variable {name} not found in prompt {openai_response_prompt.id}")
# Separate text and media variables
text_substitutions = {}
media_content_parts: list[OpenAIChatCompletionContentPartParam] = []
for name, value in openai_response_prompt.variables.items():
# Text variable found
if isinstance(value, OpenAIResponseInputMessageContentText):
text_substitutions[name] = value.text
# Media variable found
elif isinstance(value, OpenAIResponseInputMessageContentImage | OpenAIResponseInputMessageContentFile):
converted_parts = await convert_response_content_to_chat_content([value], files_api=self.files_api)
if isinstance(converted_parts, list):
media_content_parts.extend(converted_parts)
# Eg: {{product_photo}} becomes "[Image: product_photo]"
# This gives the model textual context about what media exists in the prompt
var_type = value.type.replace("input_", "").replace("_", " ").title()
text_substitutions[name] = f"[{var_type}: {name}]"
def replace_variable(match: re.Match[str]) -> str:
var_name = match.group(1).strip()
return str(text_substitutions.get(var_name, match.group(0)))
pattern = r"\{\{\s*(\w+)\s*\}\}"
processed_prompt_text = re.sub(pattern, replace_variable, cur_prompt_text)
# Insert system message with resolved text
messages.insert(0, OpenAISystemMessageParam(content=processed_prompt_text))
# If we have media, create a new user message because allows to ingest images and files
if media_content_parts:
messages.append(OpenAIUserMessageParam(content=media_content_parts))
async def get_openai_response(
self,
response_id: str,
@ -273,6 +355,14 @@ class OpenAIResponsesImpl:
guardrail_ids = extract_guardrail_ids(guardrails) if guardrails else []
# Validate that Safety API is available if guardrails are requested
if guardrail_ids and self.safety_api is None:
raise ValueError(
"Cannot process guardrails: Safety API is not configured.\n\n"
"To use guardrails, ensure the Safety API is configured in your stack, or remove "
"the 'guardrails' parameter from your request."
)
if conversation is not None:
if previous_response_id is not None:
raise ValueError(
@ -289,6 +379,7 @@ class OpenAIResponsesImpl:
input=input,
conversation=conversation,
model=model,
prompt=prompt,
instructions=instructions,
previous_response_id=previous_response_id,
store=store,
@ -342,6 +433,7 @@ class OpenAIResponsesImpl:
instructions: str | None = None,
previous_response_id: str | None = None,
conversation: str | None = None,
prompt: OpenAIResponsePrompt | None = None,
store: bool | None = True,
temperature: float | None = None,
text: OpenAIResponseText | None = None,
@ -364,6 +456,9 @@ class OpenAIResponsesImpl:
if instructions:
messages.insert(0, OpenAISystemMessageParam(content=instructions))
# Prepend reusable prompt (if provided)
await self._prepend_prompt(messages, prompt)
# Structured outputs
response_format = await convert_response_text_to_chat_response_format(text)
@ -386,6 +481,7 @@ class OpenAIResponsesImpl:
ctx=ctx,
response_id=response_id,
created_at=created_at,
prompt=prompt,
text=text,
max_infer_iters=max_infer_iters,
parallel_tool_calls=parallel_tool_calls,

14
src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py
View file

@ -66,6 +66,8 @@ from llama_stack_api import (
OpenAIResponseUsage,
OpenAIResponseUsageInputTokensDetails,
OpenAIResponseUsageOutputTokensDetails,
OpenAIToolMessageParam,
Safety,
WebSearchToolTypes,
)
@ -111,7 +113,7 @@ class StreamingResponseOrchestrator:
max_infer_iters: int,
tool_executor, # Will be the tool execution logic from the main class
instructions: str | None,
safety_api,
safety_api: Safety | None,
guardrail_ids: list[str] | None = None,
prompt: OpenAIResponsePrompt | None = None,
parallel_tool_calls: bool | None = None,
@ -905,10 +907,16 @@ class StreamingResponseOrchestrator:
"""Coordinate execution of both function and non-function tool calls."""
# Execute non-function tool calls
for tool_call in non_function_tool_calls:
# Check if total calls made to built-in and mcp tools exceed max_tool_calls
# if total calls made to built-in and mcp tools exceed max_tool_calls
# then create a tool response message indicating the call was skipped
if self.max_tool_calls is not None and self.accumulated_builtin_tool_calls >= self.max_tool_calls:
logger.info(f"Ignoring built-in and mcp tool call since reached the limit of {self.max_tool_calls=}.")
break
skipped_call_message = OpenAIToolMessageParam(
content=f"Tool call skipped: maximum tool calls limit ({self.max_tool_calls}) reached.",
tool_call_id=tool_call.id,
)
next_turn_messages.append(skipped_call_message)
continue
# Find the item_id for this tool call
matching_item_id = None

129
src/llama_stack/providers/inline/agents/meta_reference/responses/utils.py
View file

@ -5,11 +5,14 @@
# the root directory of this source tree.
import asyncio
import base64
import mimetypes
import re
import uuid
from collections.abc import Sequence
from llama_stack_api import (
Files,
OpenAIAssistantMessageParam,
OpenAIChatCompletionContentPartImageParam,
OpenAIChatCompletionContentPartParam,
@ -18,6 +21,8 @@ from llama_stack_api import (
OpenAIChatCompletionToolCallFunction,
OpenAIChoice,
OpenAIDeveloperMessageParam,
OpenAIFile,
OpenAIFileFile,
OpenAIImageURL,
OpenAIJSONSchema,
OpenAIMessageParam,
@ -29,6 +34,7 @@ from llama_stack_api import (
OpenAIResponseInput,
OpenAIResponseInputFunctionToolCallOutput,
OpenAIResponseInputMessageContent,
OpenAIResponseInputMessageContentFile,
OpenAIResponseInputMessageContentImage,
OpenAIResponseInputMessageContentText,
OpenAIResponseInputTool,
@ -37,9 +43,11 @@ from llama_stack_api import (
OpenAIResponseMessage,
OpenAIResponseOutputMessageContent,
OpenAIResponseOutputMessageContentOutputText,
OpenAIResponseOutputMessageFileSearchToolCall,
OpenAIResponseOutputMessageFunctionToolCall,
OpenAIResponseOutputMessageMCPCall,
OpenAIResponseOutputMessageMCPListTools,
OpenAIResponseOutputMessageWebSearchToolCall,
OpenAIResponseText,
OpenAISystemMessageParam,
OpenAIToolMessageParam,
@ -49,6 +57,46 @@ from llama_stack_api import (
)
async def extract_bytes_from_file(file_id: str, files_api: Files) -> bytes:
"""
Extract raw bytes from file using the Files API.
:param file_id: The file identifier (e.g., "file-abc123")
:param files_api: Files API instance
:returns: Raw file content as bytes
:raises: ValueError if file cannot be retrieved
"""
try:
response = await files_api.openai_retrieve_file_content(file_id)
return bytes(response.body)
except Exception as e:
raise ValueError(f"Failed to retrieve file content for file_id '{file_id}': {str(e)}") from e
def generate_base64_ascii_text_from_bytes(raw_bytes: bytes) -> str:
"""
Converts raw binary bytes into a safe ASCII text representation for URLs
:param raw_bytes: the actual bytes that represents file content
:returns: string of utf-8 characters
"""
return base64.b64encode(raw_bytes).decode("utf-8")
def construct_data_url(ascii_text: str, mime_type: str | None) -> str:
"""
Construct data url with decoded data inside
:param ascii_text: ASCII content
:param mime_type: MIME type of file
:returns: data url string (eg. data:image/png,base64,%3Ch1%3EHello%2C%20World%21%3C%2Fh1%3E)
"""
if not mime_type:
mime_type = "application/octet-stream"
return f"data:{mime_type};base64,{ascii_text}"
async def convert_chat_choice_to_response_message(
choice: OpenAIChoice,
citation_files: dict[str, str] | None = None,
@ -78,11 +126,15 @@ async def convert_chat_choice_to_response_message(
async def convert_response_content_to_chat_content(
content: str | Sequence[OpenAIResponseInputMessageContent | OpenAIResponseOutputMessageContent],
files_api: Files | None,
) -> str | list[OpenAIChatCompletionContentPartParam]:
"""
Convert the content parts from an OpenAI Response API request into OpenAI Chat Completion content parts.
The content schemas of each API look similar, but are not exactly the same.
:param content: The content to convert
:param files_api: Files API for resolving file_id to raw file content (required if content contains files/images)
"""
if isinstance(content, str):
return content
@ -95,9 +147,68 @@ async def convert_response_content_to_chat_content(
elif isinstance(content_part, OpenAIResponseOutputMessageContentOutputText):
converted_parts.append(OpenAIChatCompletionContentPartTextParam(text=content_part.text))
elif isinstance(content_part, OpenAIResponseInputMessageContentImage):
detail = content_part.detail
image_mime_type = None
if content_part.image_url:
image_url = OpenAIImageURL(url=content_part.image_url, detail=content_part.detail)
image_url = OpenAIImageURL(url=content_part.image_url, detail=detail)
converted_parts.append(OpenAIChatCompletionContentPartImageParam(image_url=image_url))
elif content_part.file_id:
if files_api is None:
raise ValueError("file_ids are not supported by this implementation of the Stack")
image_file_response = await files_api.openai_retrieve_file(content_part.file_id)
if image_file_response.filename:
image_mime_type, _ = mimetypes.guess_type(image_file_response.filename)
raw_image_bytes = await extract_bytes_from_file(content_part.file_id, files_api)
ascii_text = generate_base64_ascii_text_from_bytes(raw_image_bytes)
image_data_url = construct_data_url(ascii_text, image_mime_type)
image_url = OpenAIImageURL(url=image_data_url, detail=detail)
converted_parts.append(OpenAIChatCompletionContentPartImageParam(image_url=image_url))
else:
raise ValueError(
f"Image content must have either 'image_url' or 'file_id'. "
f"Got image_url={content_part.image_url}, file_id={content_part.file_id}"
)
elif isinstance(content_part, OpenAIResponseInputMessageContentFile):
resolved_file_data = None
file_data = content_part.file_data
file_id = content_part.file_id
file_url = content_part.file_url
filename = content_part.filename
file_mime_type = None
if not any([file_data, file_id, file_url]):
raise ValueError(
f"File content must have at least one of 'file_data', 'file_id', or 'file_url'. "
f"Got file_data={file_data}, file_id={file_id}, file_url={file_url}"
)
if file_id:
if files_api is None:
raise ValueError("file_ids are not supported by this implementation of the Stack")
file_response = await files_api.openai_retrieve_file(file_id)
if not filename:
filename = file_response.filename
file_mime_type, _ = mimetypes.guess_type(file_response.filename)
raw_file_bytes = await extract_bytes_from_file(file_id, files_api)
ascii_text = generate_base64_ascii_text_from_bytes(raw_file_bytes)
resolved_file_data = construct_data_url(ascii_text, file_mime_type)
elif file_data:
if file_data.startswith("data:"):
resolved_file_data = file_data
else:
# Raw base64 data, wrap in data URL format
if filename:
file_mime_type, _ = mimetypes.guess_type(filename)
resolved_file_data = construct_data_url(file_data, file_mime_type)
elif file_url:
resolved_file_data = file_url
converted_parts.append(
OpenAIFile(
file=OpenAIFileFile(
file_data=resolved_file_data,
filename=filename,
)
)
)
elif isinstance(content_part, str):
converted_parts.append(OpenAIChatCompletionContentPartTextParam(text=content_part))
else:
@ -110,12 +221,14 @@ async def convert_response_content_to_chat_content(
async def convert_response_input_to_chat_messages(
input: str | list[OpenAIResponseInput],
previous_messages: list[OpenAIMessageParam] | None = None,
files_api: Files | None = None,
) -> list[OpenAIMessageParam]:
"""
Convert the input from an OpenAI Response API request into OpenAI Chat Completion messages.
:param input: The input to convert
:param previous_messages: Optional previous messages to check for function_call references
:param files_api: Files API for resolving file_id to raw file content (optional, required for file/image content)
"""
messages: list[OpenAIMessageParam] = []
if isinstance(input, list):
@ -169,6 +282,12 @@ async def convert_response_input_to_chat_messages(
elif isinstance(input_item, OpenAIResponseOutputMessageMCPListTools):
# the tool list will be handled separately
pass
elif isinstance(
input_item,
OpenAIResponseOutputMessageWebSearchToolCall | OpenAIResponseOutputMessageFileSearchToolCall,
):
# these tool calls are tracked internally but not converted to chat messages
pass
elif isinstance(input_item, OpenAIResponseMCPApprovalRequest) or isinstance(
input_item, OpenAIResponseMCPApprovalResponse
):
@ -176,7 +295,7 @@ async def convert_response_input_to_chat_messages(
pass
elif isinstance(input_item, OpenAIResponseMessage):
# Narrow type to OpenAIResponseMessage which has content and role attributes
content = await convert_response_content_to_chat_content(input_item.content)
content = await convert_response_content_to_chat_content(input_item.content, files_api)
message_type = await get_message_type_by_role(input_item.role)
if message_type is None:
raise ValueError(
@ -320,11 +439,15 @@ def is_function_tool_call(
return False
async def run_guardrails(safety_api: Safety, messages: str, guardrail_ids: list[str]) -> str | None:
async def run_guardrails(safety_api: Safety | None, messages: str, guardrail_ids: list[str]) -> str | None:
"""Run guardrails against messages and return violation message if blocked."""
if not messages:
return None
# If safety API is not available, skip guardrails
if safety_api is None:
return None
# Look up shields to get their provider_resource_id (actual model ID)
model_ids = []
# TODO: list_shields not in Safety interface but available at runtime via API routing

6
src/llama_stack/providers/registry/agents.py
View file

@ -30,11 +30,15 @@ def available_providers() -> list[ProviderSpec]:
config_class="llama_stack.providers.inline.agents.meta_reference.MetaReferenceAgentsImplConfig",
api_dependencies=[
Api.inference,
Api.safety,
Api.vector_io,
Api.tool_runtime,
Api.tool_groups,
Api.conversations,
Api.prompts,
Api.files,
],
optional_api_dependencies=[
Api.safety,
],
description="Meta's reference implementation of an agent system that can use tools, access vector databases, and perform complex reasoning tasks.",
),

4
src/llama_stack/providers/remote/inference/azure/azure.py
View file

@ -4,8 +4,6 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from urllib.parse import urljoin
from llama_stack.providers.utils.inference.openai_mixin import OpenAIMixin
from .config import AzureConfig
@ -22,4 +20,4 @@ class AzureInferenceAdapter(OpenAIMixin):
Returns the Azure API base URL from the configuration.
"""
return urljoin(str(self.config.api_base), "/openai/v1")
return str(self.config.base_url)

9
src/llama_stack/providers/remote/inference/azure/config.py
View file

@ -32,8 +32,9 @@ class AzureProviderDataValidator(BaseModel):
@json_schema_type
class AzureConfig(RemoteInferenceProviderConfig):
api_base: HttpUrl = Field(
description="Azure API base for Azure (e.g., https://your-resource-name.openai.azure.com)",
base_url: HttpUrl | None = Field(
default=None,
description="Azure API base for Azure (e.g., https://your-resource-name.openai.azure.com/openai/v1)",
)
api_version: str | None = Field(
default_factory=lambda: os.getenv("AZURE_API_VERSION"),
@ -48,14 +49,14 @@ class AzureConfig(RemoteInferenceProviderConfig):
def sample_run_config(
cls,
api_key: str = "${env.AZURE_API_KEY:=}",
api_base: str = "${env.AZURE_API_BASE:=}",
base_url: str = "${env.AZURE_API_BASE:=}",
api_version: str = "${env.AZURE_API_VERSION:=}",
api_type: str = "${env.AZURE_API_TYPE:=}",
**kwargs,
) -> dict[str, Any]:
return {
"api_key": api_key,
"api_base": api_base,
"base_url": base_url,
"api_version": api_version,
"api_type": api_type,
}

4
src/llama_stack/providers/remote/inference/bedrock/bedrock.py
View file

@ -37,7 +37,7 @@ class BedrockInferenceAdapter(OpenAIMixin):
"""
config: BedrockConfig
provider_data_api_key_field: str = "aws_bedrock_api_key"
provider_data_api_key_field: str = "aws_bearer_token_bedrock"
def get_base_url(self) -> str:
"""Get base URL for OpenAI client."""
@ -111,7 +111,7 @@ class BedrockInferenceAdapter(OpenAIMixin):
logger.error(f"AWS Bedrock authentication token expired: {error_msg}")
raise ValueError(
"AWS Bedrock authentication failed: Bearer token has expired. "
"The AWS_BEDROCK_API_KEY environment variable contains an expired pre-signed URL. "
"The AWS_BEARER_TOKEN_BEDROCK environment variable contains an expired pre-signed URL. "
"Please refresh your token by generating a new pre-signed URL with AWS credentials. "
"Refer to AWS Bedrock documentation for details on OpenAI-compatible endpoints."
) from e

6
src/llama_stack/providers/remote/inference/bedrock/config.py
View file

@ -12,9 +12,9 @@ from llama_stack.providers.utils.inference.model_registry import RemoteInference
class BedrockProviderDataValidator(BaseModel):
aws_bedrock_api_key: str | None = Field(
aws_bearer_token_bedrock: str | None = Field(
default=None,
description="API key for Amazon Bedrock",
description="API Key (Bearer token) for Amazon Bedrock",
)
@ -27,6 +27,6 @@ class BedrockConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(cls, **kwargs):
return {
"api_key": "${env.AWS_BEDROCK_API_KEY:=}",
"api_key": "${env.AWS_BEARER_TOKEN_BEDROCK:=}",
"region_name": "${env.AWS_DEFAULT_REGION:=us-east-2}",
}

4
src/llama_stack/providers/remote/inference/cerebras/cerebras.py
View file

@ -4,8 +4,6 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from urllib.parse import urljoin
from llama_stack.providers.utils.inference.openai_mixin import OpenAIMixin
from llama_stack_api import (
OpenAIEmbeddingsRequestWithExtraBody,
@ -21,7 +19,7 @@ class CerebrasInferenceAdapter(OpenAIMixin):
provider_data_api_key_field: str = "cerebras_api_key"
def get_base_url(self) -> str:
return urljoin(self.config.base_url, "v1")
return str(self.config.base_url)
async def openai_embeddings(
self,

8
src/llama_stack/providers/remote/inference/cerebras/config.py
View file

@ -7,12 +7,12 @@
import os
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
DEFAULT_BASE_URL = "https://api.cerebras.ai"
DEFAULT_BASE_URL = "https://api.cerebras.ai/v1"
class CerebrasProviderDataValidator(BaseModel):
@ -24,8 +24,8 @@ class CerebrasProviderDataValidator(BaseModel):
@json_schema_type
class CerebrasImplConfig(RemoteInferenceProviderConfig):
base_url: str = Field(
default=os.environ.get("CEREBRAS_BASE_URL", DEFAULT_BASE_URL),
base_url: HttpUrl | None = Field(
default=HttpUrl(os.environ.get("CEREBRAS_BASE_URL", DEFAULT_BASE_URL)),
description="Base URL for the Cerebras API",
)

10
src/llama_stack/providers/remote/inference/databricks/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field, SecretStr
from pydantic import BaseModel, Field, HttpUrl, SecretStr
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,9 +21,9 @@ class DatabricksProviderDataValidator(BaseModel):
@json_schema_type
class DatabricksImplConfig(RemoteInferenceProviderConfig):
url: str | None = Field(
base_url: HttpUrl | None = Field(
default=None,
description="The URL for the Databricks model serving endpoint",
description="The URL for the Databricks model serving endpoint (should include /serving-endpoints path)",
)
auth_credential: SecretStr | None = Field(
default=None,
@ -34,11 +34,11 @@ class DatabricksImplConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(
cls,
url: str = "${env.DATABRICKS_HOST:=}",
base_url: str = "${env.DATABRICKS_HOST:=}",
api_token: str = "${env.DATABRICKS_TOKEN:=}",
**kwargs: Any,
) -> dict[str, Any]:
return {
"url": url,
"base_url": base_url,
"api_token": api_token,
}

10
src/llama_stack/providers/remote/inference/databricks/databricks.py
View file

@ -29,15 +29,21 @@ class DatabricksInferenceAdapter(OpenAIMixin):
}
def get_base_url(self) -> str:
return f"{self.config.url}/serving-endpoints"
return str(self.config.base_url)
async def list_provider_model_ids(self) -> Iterable[str]:
# Filter out None values from endpoint names
api_token = self._get_api_key_from_config_or_provider_data()
# WorkspaceClient expects base host without /serving-endpoints suffix
base_url_str = str(self.config.base_url)
if base_url_str.endswith("/serving-endpoints"):
host = base_url_str[:-18] # Remove '/serving-endpoints'
else:
host = base_url_str
return [
endpoint.name # type: ignore[misc]
for endpoint in WorkspaceClient(
host=self.config.url, token=api_token
host=host, token=api_token
).serving_endpoints.list() # TODO: this is not async
]

8
src/llama_stack/providers/remote/inference/fireworks/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import Field
from pydantic import Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -14,14 +14,14 @@ from llama_stack_api import json_schema_type
@json_schema_type
class FireworksImplConfig(RemoteInferenceProviderConfig):
url: str = Field(
default="https://api.fireworks.ai/inference/v1",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.fireworks.ai/inference/v1"),
description="The URL for the Fireworks server",
)
@classmethod
def sample_run_config(cls, api_key: str = "${env.FIREWORKS_API_KEY:=}", **kwargs) -> dict[str, Any]:
return {
"url": "https://api.fireworks.ai/inference/v1",
"base_url": "https://api.fireworks.ai/inference/v1",
"api_key": api_key,
}

2
src/llama_stack/providers/remote/inference/fireworks/fireworks.py
View file

@ -24,4 +24,4 @@ class FireworksInferenceAdapter(OpenAIMixin):
provider_data_api_key_field: str = "fireworks_api_key"
def get_base_url(self) -> str:
return "https://api.fireworks.ai/inference/v1"
return str(self.config.base_url)

8
src/llama_stack/providers/remote/inference/groq/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,14 +21,14 @@ class GroqProviderDataValidator(BaseModel):
@json_schema_type
class GroqConfig(RemoteInferenceProviderConfig):
url: str = Field(
default="https://api.groq.com",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.groq.com/openai/v1"),
description="The URL for the Groq AI server",
)
@classmethod
def sample_run_config(cls, api_key: str = "${env.GROQ_API_KEY:=}", **kwargs) -> dict[str, Any]:
return {
"url": "https://api.groq.com",
"base_url": "https://api.groq.com/openai/v1",
"api_key": api_key,
}

2
src/llama_stack/providers/remote/inference/groq/groq.py
View file

@ -15,4 +15,4 @@ class GroqInferenceAdapter(OpenAIMixin):
provider_data_api_key_field: str = "groq_api_key"
def get_base_url(self) -> str:
return f"{self.config.url}/openai/v1"
return str(self.config.base_url)

8
src/llama_stack/providers/remote/inference/llama_openai_compat/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,14 +21,14 @@ class LlamaProviderDataValidator(BaseModel):
@json_schema_type
class LlamaCompatConfig(RemoteInferenceProviderConfig):
openai_compat_api_base: str = Field(
default="https://api.llama.com/compat/v1/",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.llama.com/compat/v1/"),
description="The URL for the Llama API server",
)
@classmethod
def sample_run_config(cls, api_key: str = "${env.LLAMA_API_KEY}", **kwargs) -> dict[str, Any]:
return {
"openai_compat_api_base": "https://api.llama.com/compat/v1/",
"base_url": "https://api.llama.com/compat/v1/",
"api_key": api_key,
}

2
src/llama_stack/providers/remote/inference/llama_openai_compat/llama.py
View file

@ -31,7 +31,7 @@ class LlamaCompatInferenceAdapter(OpenAIMixin):
:return: The Llama API base URL
"""
return self.config.openai_compat_api_base
return str(self.config.base_url)
async def openai_completion(
self,

16
src/llama_stack/providers/remote/inference/nvidia/config.py
View file

@ -7,7 +7,7 @@
import os
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -44,18 +44,14 @@ class NVIDIAConfig(RemoteInferenceProviderConfig):
URL of your running NVIDIA NIM and do not need to set the api_key.
"""
url: str = Field(
default_factory=lambda: os.getenv("NVIDIA_BASE_URL", "https://integrate.api.nvidia.com"),
base_url: HttpUrl | None = Field(
default_factory=lambda: os.getenv("NVIDIA_BASE_URL", "https://integrate.api.nvidia.com/v1"),
description="A base url for accessing the NVIDIA NIM",
)
timeout: int = Field(
default=60,
description="Timeout for the HTTP requests",
)
append_api_version: bool = Field(
default_factory=lambda: os.getenv("NVIDIA_APPEND_API_VERSION", "True").lower() != "false",
description="When set to false, the API version will not be appended to the base_url. By default, it is true.",
)
rerank_model_to_url: dict[str, str] = Field(
default_factory=lambda: {
"nv-rerank-qa-mistral-4b:1": "https://ai.api.nvidia.com/v1/retrieval/nvidia/reranking",
@ -68,13 +64,11 @@ class NVIDIAConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(
cls,
url: str = "${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com}",
base_url: HttpUrl | None = "${env.NVIDIA_BASE_URL:=https://integrate.api.nvidia.com/v1}",
api_key: str = "${env.NVIDIA_API_KEY:=}",
append_api_version: bool = "${env.NVIDIA_APPEND_API_VERSION:=True}",
**kwargs,
) -> dict[str, Any]:
return {
"url": url,
"base_url": base_url,
"api_key": api_key,
"append_api_version": append_api_version,
}

4
src/llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -44,7 +44,7 @@ class NVIDIAInferenceAdapter(OpenAIMixin):
}
async def initialize(self) -> None:
logger.info(f"Initializing NVIDIAInferenceAdapter({self.config.url})...")
logger.info(f"Initializing NVIDIAInferenceAdapter({self.config.base_url})...")
if _is_nvidia_hosted(self.config):
if not self.config.auth_credential:
@ -72,7 +72,7 @@ class NVIDIAInferenceAdapter(OpenAIMixin):
:return: The NVIDIA API base URL
"""
return f"{self.config.url}/v1" if self.config.append_api_version else self.config.url
return str(self.config.base_url)
async def list_provider_model_ids(self) -> Iterable[str]:
"""

2
src/llama_stack/providers/remote/inference/nvidia/utils.py
View file

@ -8,4 +8,4 @@ from . import NVIDIAConfig
def _is_nvidia_hosted(config: NVIDIAConfig) -> bool:
return "integrate.api.nvidia.com" in config.url
return "integrate.api.nvidia.com" in str(config.base_url)

12
src/llama_stack/providers/remote/inference/ollama/config.py
View file

@ -6,20 +6,22 @@
from typing import Any
from pydantic import Field, SecretStr
from pydantic import Field, HttpUrl, SecretStr
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
DEFAULT_OLLAMA_URL = "http://localhost:11434"
DEFAULT_OLLAMA_URL = "http://localhost:11434/v1"
class OllamaImplConfig(RemoteInferenceProviderConfig):
auth_credential: SecretStr | None = Field(default=None, exclude=True)
url: str = DEFAULT_OLLAMA_URL
base_url: HttpUrl | None = Field(default=HttpUrl(DEFAULT_OLLAMA_URL))
@classmethod
def sample_run_config(cls, url: str = "${env.OLLAMA_URL:=http://localhost:11434}", **kwargs) -> dict[str, Any]:
def sample_run_config(
cls, base_url: str = "${env.OLLAMA_URL:=http://localhost:11434/v1}", **kwargs
) -> dict[str, Any]:
return {
"url": url,
"base_url": base_url,
}

12
src/llama_stack/providers/remote/inference/ollama/ollama.py
View file

@ -55,17 +55,23 @@ class OllamaInferenceAdapter(OpenAIMixin):
# ollama client attaches itself to the current event loop (sadly?)
loop = asyncio.get_running_loop()
if loop not in self._clients:
self._clients[loop] = AsyncOllamaClient(host=self.config.url)
# Ollama client expects base URL without /v1 suffix
base_url_str = str(self.config.base_url)
if base_url_str.endswith("/v1"):
host = base_url_str[:-3]
else:
host = base_url_str
self._clients[loop] = AsyncOllamaClient(host=host)
return self._clients[loop]
def get_api_key(self):
return "NO KEY REQUIRED"
def get_base_url(self):
return self.config.url.rstrip("/") + "/v1"
return str(self.config.base_url)
async def initialize(self) -> None:
logger.info(f"checking connectivity to Ollama at `{self.config.url}`...")
logger.info(f"checking connectivity to Ollama at `{self.config.base_url}`...")
r = await self.health()
if r["status"] == HealthStatus.ERROR:
logger.warning(

6
src/llama_stack/providers/remote/inference/openai/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,8 +21,8 @@ class OpenAIProviderDataValidator(BaseModel):
@json_schema_type
class OpenAIConfig(RemoteInferenceProviderConfig):
base_url: str = Field(
default="https://api.openai.com/v1",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.openai.com/v1"),
description="Base URL for OpenAI API",
)

2
src/llama_stack/providers/remote/inference/openai/openai.py
View file

@ -35,4 +35,4 @@ class OpenAIInferenceAdapter(OpenAIMixin):
Returns the OpenAI API base URL from the configuration.
"""
return self.config.base_url
return str(self.config.base_url)

8
src/llama_stack/providers/remote/inference/passthrough/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import Field
from pydantic import Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -14,16 +14,16 @@ from llama_stack_api import json_schema_type
@json_schema_type
class PassthroughImplConfig(RemoteInferenceProviderConfig):
url: str = Field(
base_url: HttpUrl | None = Field(
default=None,
description="The URL for the passthrough endpoint",
)
@classmethod
def sample_run_config(
cls, url: str = "${env.PASSTHROUGH_URL}", api_key: str = "${env.PASSTHROUGH_API_KEY}", **kwargs
cls, base_url: HttpUrl | None = "${env.PASSTHROUGH_URL}", api_key: str = "${env.PASSTHROUGH_API_KEY}", **kwargs
) -> dict[str, Any]:
return {
"url": url,
"base_url": base_url,
"api_key": api_key,
}

4
src/llama_stack/providers/remote/inference/passthrough/passthrough.py
View file

@ -82,8 +82,8 @@ class PassthroughInferenceAdapter(NeedsRequestProviderData, Inference):
def _get_passthrough_url(self) -> str:
"""Get the passthrough URL from config or provider data."""
if self.config.url is not None:
return self.config.url
if self.config.base_url is not None:
return str(self.config.base_url)
provider_data = self.get_request_provider_data()
if provider_data is None:

6
src/llama_stack/providers/remote/inference/runpod/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field, SecretStr
from pydantic import BaseModel, Field, HttpUrl, SecretStr
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,7 +21,7 @@ class RunpodProviderDataValidator(BaseModel):
@json_schema_type
class RunpodImplConfig(RemoteInferenceProviderConfig):
url: str | None = Field(
base_url: HttpUrl | None = Field(
default=None,
description="The URL for the Runpod model serving endpoint",
)
@ -34,6 +34,6 @@ class RunpodImplConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(cls, **kwargs: Any) -> dict[str, Any]:
return {
"url": "${env.RUNPOD_URL:=}",
"base_url": "${env.RUNPOD_URL:=}",
"api_token": "${env.RUNPOD_API_TOKEN}",
}

2
src/llama_stack/providers/remote/inference/runpod/runpod.py
View file

@ -28,7 +28,7 @@ class RunpodInferenceAdapter(OpenAIMixin):
def get_base_url(self) -> str:
"""Get base URL for OpenAI client."""
return self.config.url
return str(self.config.base_url)
async def openai_chat_completion(
self,

8
src/llama_stack/providers/remote/inference/sambanova/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -21,14 +21,14 @@ class SambaNovaProviderDataValidator(BaseModel):
@json_schema_type
class SambaNovaImplConfig(RemoteInferenceProviderConfig):
url: str = Field(
default="https://api.sambanova.ai/v1",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.sambanova.ai/v1"),
description="The URL for the SambaNova AI server",
)
@classmethod
def sample_run_config(cls, api_key: str = "${env.SAMBANOVA_API_KEY:=}", **kwargs) -> dict[str, Any]:
return {
"url": "https://api.sambanova.ai/v1",
"base_url": "https://api.sambanova.ai/v1",
"api_key": api_key,
}

2
src/llama_stack/providers/remote/inference/sambanova/sambanova.py
View file

@ -25,4 +25,4 @@ class SambaNovaInferenceAdapter(OpenAIMixin):
:return: The SambaNova base URL
"""
return self.config.url
return str(self.config.base_url)

11
src/llama_stack/providers/remote/inference/tgi/config.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from pydantic import BaseModel, Field, SecretStr
from pydantic import BaseModel, Field, HttpUrl, SecretStr
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -15,18 +15,19 @@ from llama_stack_api import json_schema_type
class TGIImplConfig(RemoteInferenceProviderConfig):
auth_credential: SecretStr | None = Field(default=None, exclude=True)
url: str = Field(
description="The URL for the TGI serving endpoint",
base_url: HttpUrl | None = Field(
default=None,
description="The URL for the TGI serving endpoint (should include /v1 path)",
)
@classmethod
def sample_run_config(
cls,
url: str = "${env.TGI_URL:=}",
base_url: str = "${env.TGI_URL:=}",
**kwargs,
):
return {
"url": url,
"base_url": base_url,
}

20
src/llama_stack/providers/remote/inference/tgi/tgi.py
View file

@ -8,7 +8,7 @@
from collections.abc import Iterable
from huggingface_hub import AsyncInferenceClient, HfApi
from pydantic import SecretStr
from pydantic import HttpUrl, SecretStr
from llama_stack.log import get_logger
from llama_stack.providers.utils.inference.openai_mixin import OpenAIMixin
@ -23,7 +23,7 @@ log = get_logger(name=__name__, category="inference::tgi")
class _HfAdapter(OpenAIMixin):
url: str
base_url: HttpUrl
api_key: SecretStr
hf_client: AsyncInferenceClient
@ -36,7 +36,7 @@ class _HfAdapter(OpenAIMixin):
return "NO KEY REQUIRED"
def get_base_url(self):
return self.url
return self.base_url
async def list_provider_model_ids(self) -> Iterable[str]:
return [self.model_id]
@ -50,14 +50,20 @@ class _HfAdapter(OpenAIMixin):
class TGIAdapter(_HfAdapter):
async def initialize(self, config: TGIImplConfig) -> None:
if not config.url:
if not config.base_url:
raise ValueError("You must provide a URL in run.yaml (or via the TGI_URL environment variable) to use TGI.")
log.info(f"Initializing TGI client with url={config.url}")
self.hf_client = AsyncInferenceClient(model=config.url, provider="hf-inference")
log.info(f"Initializing TGI client with url={config.base_url}")
# Extract base URL without /v1 for HF client initialization
base_url_str = str(config.base_url).rstrip("/")
if base_url_str.endswith("/v1"):
base_url_for_client = base_url_str[:-3]
else:
base_url_for_client = base_url_str
self.hf_client = AsyncInferenceClient(model=base_url_for_client, provider="hf-inference")
endpoint_info = await self.hf_client.get_endpoint_info()
self.max_tokens = endpoint_info["max_total_tokens"]
self.model_id = endpoint_info["model_id"]
self.url = f"{config.url.rstrip('/')}/v1"
self.base_url = config.base_url
self.api_key = SecretStr("NO_KEY")

8
src/llama_stack/providers/remote/inference/together/config.py
View file

@ -6,7 +6,7 @@
from typing import Any
from pydantic import Field
from pydantic import Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -14,14 +14,14 @@ from llama_stack_api import json_schema_type
@json_schema_type
class TogetherImplConfig(RemoteInferenceProviderConfig):
url: str = Field(
default="https://api.together.xyz/v1",
base_url: HttpUrl | None = Field(
default=HttpUrl("https://api.together.xyz/v1"),
description="The URL for the Together AI server",
)
@classmethod
def sample_run_config(cls, **kwargs) -> dict[str, Any]:
return {
"url": "https://api.together.xyz/v1",
"base_url": "https://api.together.xyz/v1",
"api_key": "${env.TOGETHER_API_KEY:=}",
}

3
src/llama_stack/providers/remote/inference/together/together.py
View file

@ -9,7 +9,6 @@ from collections.abc import Iterable
from typing import Any, cast
from together import AsyncTogether # type: ignore[import-untyped]
from together.constants import BASE_URL # type: ignore[import-untyped]
from llama_stack.core.request_headers import NeedsRequestProviderData
from llama_stack.log import get_logger
@ -42,7 +41,7 @@ class TogetherInferenceAdapter(OpenAIMixin, NeedsRequestProviderData):
provider_data_api_key_field: str = "together_api_key"
def get_base_url(self):
return BASE_URL
return str(self.config.base_url)
def _get_client(self) -> AsyncTogether:
together_api_key = None

8
src/llama_stack/providers/remote/inference/vllm/config.py
View file

@ -6,7 +6,7 @@
from pathlib import Path
from pydantic import Field, SecretStr, field_validator
from pydantic import Field, HttpUrl, SecretStr, field_validator
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -14,7 +14,7 @@ from llama_stack_api import json_schema_type
@json_schema_type
class VLLMInferenceAdapterConfig(RemoteInferenceProviderConfig):
url: str | None = Field(
base_url: HttpUrl | None = Field(
default=None,
description="The URL for the vLLM model serving endpoint",
)
@ -48,11 +48,11 @@ class VLLMInferenceAdapterConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(
cls,
url: str = "${env.VLLM_URL:=}",
base_url: str = "${env.VLLM_URL:=}",
**kwargs,
):
return {
"url": url,
"base_url": base_url,
"max_tokens": "${env.VLLM_MAX_TOKENS:=4096}",
"api_token": "${env.VLLM_API_TOKEN:=fake}",
"tls_verify": "${env.VLLM_TLS_VERIFY:=true}",

6
src/llama_stack/providers/remote/inference/vllm/vllm.py
View file

@ -39,12 +39,12 @@ class VLLMInferenceAdapter(OpenAIMixin):
def get_base_url(self) -> str:
"""Get the base URL from config."""
if not self.config.url:
if not self.config.base_url:
raise ValueError("No base URL configured")
return self.config.url
return str(self.config.base_url)
async def initialize(self) -> None:
if not self.config.url:
if not self.config.base_url:
raise ValueError(
"You must provide a URL in run.yaml (or via the VLLM_URL environment variable) to use vLLM."
)

6
src/llama_stack/providers/remote/inference/watsonx/config.py
View file

@ -7,7 +7,7 @@
import os
from typing import Any
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, HttpUrl
from llama_stack.providers.utils.inference.model_registry import RemoteInferenceProviderConfig
from llama_stack_api import json_schema_type
@ -23,7 +23,7 @@ class WatsonXProviderDataValidator(BaseModel):
@json_schema_type
class WatsonXConfig(RemoteInferenceProviderConfig):
url: str = Field(
base_url: HttpUrl | None = Field(
default_factory=lambda: os.getenv("WATSONX_BASE_URL", "https://us-south.ml.cloud.ibm.com"),
description="A base url for accessing the watsonx.ai",
)
@ -39,7 +39,7 @@ class WatsonXConfig(RemoteInferenceProviderConfig):
@classmethod
def sample_run_config(cls, **kwargs) -> dict[str, Any]:
return {
"url": "${env.WATSONX_BASE_URL:=https://us-south.ml.cloud.ibm.com}",
"base_url": "${env.WATSONX_BASE_URL:=https://us-south.ml.cloud.ibm.com}",
"api_key": "${env.WATSONX_API_KEY:=}",
"project_id": "${env.WATSONX_PROJECT_ID:=}",
}

4
src/llama_stack/providers/remote/inference/watsonx/watsonx.py
View file

@ -255,7 +255,7 @@ class WatsonXInferenceAdapter(LiteLLMOpenAIMixin):
)
def get_base_url(self) -> str:
return self.config.url
return str(self.config.base_url)
# Copied from OpenAIMixin
async def check_model_availability(self, model: str) -> bool:
@ -316,7 +316,7 @@ class WatsonXInferenceAdapter(LiteLLMOpenAIMixin):
"""
Retrieves foundation model specifications from the watsonx.ai API.
"""
url = f"{self.config.url}/ml/v1/foundation_model_specs?version=2023-10-25"
url = f"{str(self.config.base_url)}/ml/v1/foundation_model_specs?version=2023-10-25"
headers = {
# Note that there is no authorization header. Listing models does not require authentication.
"Content-Type": "application/json",

209
src/llama_stack/providers/utils/inference/openai_compat.py
View file

@ -3,23 +3,10 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from collections.abc import Iterable
from typing import (
Any,
)
from openai.types.chat import (
ChatCompletionContentPartParam as OpenAIChatCompletionContentPartParam,
)
try:
from openai.types.chat import (
ChatCompletionMessageFunctionToolCall as OpenAIChatCompletionMessageFunctionToolCall,
)
except ImportError:
from openai.types.chat.chat_completion_message_tool_call import (
ChatCompletionMessageToolCall as OpenAIChatCompletionMessageFunctionToolCall,
)
from openai.types.chat import (
ChatCompletionMessageToolCall,
)
@ -32,18 +19,6 @@ from llama_stack.models.llama.datatypes import (
ToolCall,
ToolDefinition,
)
from llama_stack_api import (
URL,
GreedySamplingStrategy,
ImageContentItem,
JsonSchemaResponseFormat,
OpenAIResponseFormatParam,
SamplingParams,
TextContentItem,
TopKSamplingStrategy,
TopPSamplingStrategy,
_URLOrData,
)
logger = get_logger(name=__name__, category="providers::utils")
@ -73,42 +48,6 @@ class OpenAICompatCompletionResponse(BaseModel):
choices: list[OpenAICompatCompletionChoice]
def get_sampling_strategy_options(params: SamplingParams) -> dict:
options = {}
if isinstance(params.strategy, GreedySamplingStrategy):
options["temperature"] = 0.0
elif isinstance(params.strategy, TopPSamplingStrategy):
if params.strategy.temperature is not None:
options["temperature"] = params.strategy.temperature
if params.strategy.top_p is not None:
options["top_p"] = params.strategy.top_p
elif isinstance(params.strategy, TopKSamplingStrategy):
options["top_k"] = params.strategy.top_k
else:
raise ValueError(f"Unsupported sampling strategy: {params.strategy}")
return options
def get_sampling_options(params: SamplingParams | None) -> dict:
if not params:
return {}
options = {}
if params:
options.update(get_sampling_strategy_options(params))
if params.max_tokens:
options["max_tokens"] = params.max_tokens
if params.repetition_penalty is not None and params.repetition_penalty != 1.0:
options["repeat_penalty"] = params.repetition_penalty
if params.stop is not None:
options["stop"] = params.stop
return options
def text_from_choice(choice) -> str:
if hasattr(choice, "delta") and choice.delta:
return choice.delta.content # type: ignore[no-any-return] # external OpenAI types lack precise annotations
@ -253,154 +192,6 @@ def convert_tooldef_to_openai_tool(tool: ToolDefinition) -> dict:
return out
def _convert_stop_reason_to_openai_finish_reason(stop_reason: StopReason) -> str:
"""
Convert a StopReason to an OpenAI chat completion finish_reason.
"""
return {
StopReason.end_of_turn: "stop",
StopReason.end_of_message: "tool_calls",
StopReason.out_of_tokens: "length",
}.get(stop_reason, "stop")
def _convert_openai_finish_reason(finish_reason: str) -> StopReason:
"""
Convert an OpenAI chat completion finish_reason to a StopReason.
finish_reason: Literal["stop", "length", "tool_calls", ...]
- stop: model hit a natural stop point or a provided stop sequence
- length: maximum number of tokens specified in the request was reached
- tool_calls: model called a tool
->
class StopReason(Enum):
end_of_turn = "end_of_turn"
end_of_message = "end_of_message"
out_of_tokens = "out_of_tokens"
"""
# TODO(mf): are end_of_turn and end_of_message semantics correct?
return {
"stop": StopReason.end_of_turn,
"length": StopReason.out_of_tokens,
"tool_calls": StopReason.end_of_message,
}.get(finish_reason, StopReason.end_of_turn)
def _convert_openai_request_tools(tools: list[dict[str, Any]] | None = None) -> list[ToolDefinition]:
lls_tools: list[ToolDefinition] = []
if not tools:
return lls_tools
for tool in tools:
tool_fn = tool.get("function", {})
tool_name = tool_fn.get("name", None)
tool_desc = tool_fn.get("description", None)
tool_params = tool_fn.get("parameters", None)
lls_tool = ToolDefinition(
tool_name=tool_name,
description=tool_desc,
input_schema=tool_params, # Pass through entire JSON Schema
)
lls_tools.append(lls_tool)
return lls_tools
def _convert_openai_request_response_format(
response_format: OpenAIResponseFormatParam | None = None,
):
if not response_format:
return None
# response_format can be a dict or a pydantic model
response_format_dict = dict(response_format) # type: ignore[arg-type] # OpenAIResponseFormatParam union needs dict conversion
if response_format_dict.get("type", "") == "json_schema":
return JsonSchemaResponseFormat(
type="json_schema", # type: ignore[arg-type] # Literal["json_schema"] incompatible with expected type
json_schema=response_format_dict.get("json_schema", {}).get("schema", ""),
)
return None
def _convert_openai_tool_calls(
tool_calls: list[OpenAIChatCompletionMessageFunctionToolCall],
) -> list[ToolCall]:
"""
Convert an OpenAI ChatCompletionMessageToolCall list into a list of ToolCall.
OpenAI ChatCompletionMessageToolCall:
id: str
function: Function
type: Literal["function"]
OpenAI Function:
arguments: str
name: str
->
ToolCall:
call_id: str
tool_name: str
arguments: Dict[str, ...]
"""
if not tool_calls:
return [] # CompletionMessage tool_calls is not optional
return [
ToolCall(
call_id=call.id,
tool_name=call.function.name,
arguments=call.function.arguments,
)
for call in tool_calls
]
def _convert_openai_sampling_params(
max_tokens: int | None = None,
temperature: float | None = None,
top_p: float | None = None,
) -> SamplingParams:
sampling_params = SamplingParams()
if max_tokens:
sampling_params.max_tokens = max_tokens
# Map an explicit temperature of 0 to greedy sampling
if temperature == 0:
sampling_params.strategy = GreedySamplingStrategy()
else:
# OpenAI defaults to 1.0 for temperature and top_p if unset
if temperature is None:
temperature = 1.0
if top_p is None:
top_p = 1.0
sampling_params.strategy = TopPSamplingStrategy(temperature=temperature, top_p=top_p) # type: ignore[assignment] # SamplingParams.strategy union accepts this type
return sampling_params
def openai_content_to_content(content: str | Iterable[OpenAIChatCompletionContentPartParam] | None):
if content is None:
return ""
if isinstance(content, str):
return content
elif isinstance(content, list):
return [openai_content_to_content(c) for c in content]
elif hasattr(content, "type"):
if content.type == "text":
return TextContentItem(type="text", text=content.text) # type: ignore[attr-defined] # Iterable narrowed by hasattr check but mypy doesn't track
elif content.type == "image_url":
return ImageContentItem(type="image", image=_URLOrData(url=URL(uri=content.image_url.url))) # type: ignore[attr-defined] # Iterable narrowed by hasattr check but mypy doesn't track
else:
raise ValueError(f"Unknown content type: {content.type}")
else:
raise ValueError(f"Unknown content type: {content}")
async def prepare_openai_completion_params(**params):
async def _prepare_value(value: Any) -> Any:
new_value = value

28
src/llama_stack/providers/utils/inference/openai_mixin.py
View file

@ -213,6 +213,19 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
return api_key
def _validate_model_allowed(self, provider_model_id: str) -> None:
"""
Validate that the model is in the allowed_models list if configured.
:param provider_model_id: The provider-specific model ID to validate
:raises ValueError: If the model is not in the allowed_models list
"""
if self.config.allowed_models is not None and provider_model_id not in self.config.allowed_models:
raise ValueError(
f"Model '{provider_model_id}' is not in the allowed models list. "
f"Allowed models: {self.config.allowed_models}"
)
async def _get_provider_model_id(self, model: str) -> str:
"""
Get the provider-specific model ID from the model store.
@ -259,8 +272,11 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
Direct OpenAI completion API call.
"""
# TODO: fix openai_completion to return type compatible with OpenAI's API response
provider_model_id = await self._get_provider_model_id(params.model)
self._validate_model_allowed(provider_model_id)
completion_kwargs = await prepare_openai_completion_params(
model=await self._get_provider_model_id(params.model),
model=provider_model_id,
prompt=params.prompt,
best_of=params.best_of,
echo=params.echo,
@ -292,6 +308,9 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
"""
Direct OpenAI chat completion API call.
"""
provider_model_id = await self._get_provider_model_id(params.model)
self._validate_model_allowed(provider_model_id)
messages = params.messages
if self.download_images:
@ -313,7 +332,7 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
messages = [await _localize_image_url(m) for m in messages]
request_params = await prepare_openai_completion_params(
model=await self._get_provider_model_id(params.model),
model=provider_model_id,
messages=messages,
frequency_penalty=params.frequency_penalty,
function_call=params.function_call,
@ -351,10 +370,13 @@ class OpenAIMixin(NeedsRequestProviderData, ABC, BaseModel):
"""
Direct OpenAI embeddings API call.
"""
provider_model_id = await self._get_provider_model_id(params.model)
self._validate_model_allowed(provider_model_id)
# Build request params conditionally to avoid NotGiven/Omit type mismatch
# The OpenAI SDK uses Omit in signatures but NOT_GIVEN has type NotGiven
request_params: dict[str, Any] = {
"model": await self._get_provider_model_id(params.model),
"model": provider_model_id,
"input": params.input,
}
if params.encoding_format is not None:

69
src/llama_stack_api/vector_io.py
View file

@ -11,7 +11,7 @@
from typing import Annotated, Any, Literal, Protocol, runtime_checkable
from fastapi import Body, Query
from pydantic import BaseModel, Field
from pydantic import BaseModel, Field, field_validator
from llama_stack_api.common.tracing import telemetry_traceable
from llama_stack_api.inference import InterleavedContent
@ -372,6 +372,65 @@ VectorStoreFileStatus = Literal["completed"] | Literal["in_progress"] | Literal[
register_schema(VectorStoreFileStatus, name="VectorStoreFileStatus")
# VectorStoreFileAttributes type with OpenAPI constraints
VectorStoreFileAttributes = Annotated[
dict[str, Annotated[str, Field(max_length=512)] | float | bool],
Field(
max_length=16,
json_schema_extra={
"propertyNames": {"type": "string", "maxLength": 64},
"x-oaiTypeLabel": "map",
},
description=(
"Set of 16 key-value pairs that can be attached to an object. This can be "
"useful for storing additional information about the object in a structured "
"format, and querying for objects via API or the dashboard. Keys are strings "
"with a maximum length of 64 characters. Values are strings with a maximum "
"length of 512 characters, booleans, or numbers."
),
),
]
def _sanitize_vector_store_attributes(metadata: dict[str, Any] | None) -> dict[str, str | float | bool]:
"""
Sanitize metadata to VectorStoreFileAttributes spec (max 16 properties, primitives only).
Converts dict[str, Any] to dict[str, str | float | bool]:
- Preserves: str (truncated to 512 chars), bool, int/float (as float)
- Converts: list -> comma-separated string
- Filters: dict, None, other types
- Enforces: max 16 properties, max 64 char keys, max 512 char string values
"""
if not metadata:
return {}
sanitized: dict[str, str | float | bool] = {}
for key, value in metadata.items():
# Enforce max 16 properties
if len(sanitized) >= 16:
break
# Enforce max 64 char keys
if len(key) > 64:
continue
# Convert to supported primitive types
if isinstance(value, bool):
sanitized[key] = value
elif isinstance(value, int | float):
sanitized[key] = float(value)
elif isinstance(value, str):
# Enforce max 512 char string values
sanitized[key] = value[:512] if len(value) > 512 else value
elif isinstance(value, list):
# Convert lists to comma-separated strings (max 512 chars)
list_str = ", ".join(str(item) for item in value)
sanitized[key] = list_str[:512] if len(list_str) > 512 else list_str
return sanitized
@json_schema_type
class VectorStoreFileObject(BaseModel):
"""OpenAI Vector Store File object.
@ -389,7 +448,7 @@ class VectorStoreFileObject(BaseModel):
id: str
object: str = "vector_store.file"
attributes: dict[str, Any] = Field(default_factory=dict)
attributes: VectorStoreFileAttributes = Field(default_factory=dict)
chunking_strategy: VectorStoreChunkingStrategy
created_at: int
last_error: VectorStoreFileLastError | None = None
@ -397,6 +456,12 @@ class VectorStoreFileObject(BaseModel):
usage_bytes: int = 0
vector_store_id: str
@field_validator("attributes", mode="before")
@classmethod
def _validate_attributes(cls, v: dict[str, Any] | None) -> dict[str, str | float | bool]:
"""Sanitize attributes to match VectorStoreFileAttributes OpenAPI spec."""
return _sanitize_vector_store_attributes(v)
@json_schema_type
class VectorStoreListFilesResponse(BaseModel):