phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-07-20 03:40:05 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into nvidia-e2e-notebook

Browse source
This commit is contained in:
Jash Gulabrai 2025-05-06 11:12:34 -04:00
commit b1d941e1f0
447 changed files with 6462 additions and 64778 deletions
Download patch file Download diff file Expand all files Collapse all files

26
.github/actions/setup-ollama/action.yml vendored Normal file
View file

@ -0,0 +1,26 @@
name: Setup Ollama
description: Start Ollama and cache model
inputs:
models:
description: Comma-separated list of models to pull
default: "llama3.2:3b-instruct-fp16,all-minilm:latest"
runs:
using: "composite"
steps:
- name: Install and start Ollama
shell: bash
run: |
# the ollama installer also starts the ollama service
curl -fsSL https://ollama.com/install.sh | sh
# Do NOT cache models - pulling the cache is actually slower than just pulling the model.
# It takes ~45 seconds to pull the models from the cache and unpack it, but only 30 seconds to
# pull them directly.
# Maybe this is because the cache is being pulled at the same time by all the matrix jobs?
- name: Pull requested models
if: inputs.models != ''
shell: bash
run: |
for model in $(echo "${{ inputs.models }}" | tr ',' ' '); do
ollama pull "$model"
done

15
.github/workflows/integration-tests.yml vendored
View file

@ -38,19 +38,8 @@ jobs:
python-version: "3.10"
activate-environment: true
- name: Install and start Ollama
run: |
# the ollama installer also starts the ollama service
curl -fsSL https://ollama.com/install.sh | sh
# Do NOT cache models - pulling the cache is actually slower than just pulling the model.
# It takes ~45 seconds to pull the models from the cache and unpack it, but only 30 seconds to
# pull them directly.
# Maybe this is because the cache is being pulled at the same time by all the matrix jobs?
- name: Pull Ollama models (instruct and embed)
run: |
ollama pull llama3.2:3b-instruct-fp16
ollama pull all-minilm:latest
- name: Setup ollama
uses: ./.github/actions/setup-ollama
- name: Set Up Environment and Install Dependencies
run: |

2
.github/workflows/pre-commit.yml vendored
View file

@ -27,6 +27,8 @@ jobs:
.pre-commit-config.yaml
- uses: pre-commit/action@2c7b3805fd2a0fd8c1884dcaebf91fc102a13ecd # v3.0.1
env:
SKIP: no-commit-to-branch
- name: Verify if there are any diff files after pre-commit
run: |

2
.github/workflows/providers-build.yml vendored
View file

@ -153,7 +153,7 @@ jobs:
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Set up Python
uses: actions/setup-python@8d9ed9ac5c53483de85588cdf95a591a75ab9f55 # v5.5.0
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: '3.10'

12
.pre-commit-config.yaml
View file

@ -15,6 +15,18 @@ repos:
args: ['--maxkb=1000']
- id: end-of-file-fixer
exclude: '^(.*\.svg)$'
- id: no-commit-to-branch
- id: check-yaml
args: ["--unsafe"]
- id: detect-private-key
- id: requirements-txt-fixer
- id: mixed-line-ending
args: [--fix=lf] # Forces to replace line ending by LF (line feed)
- id: check-executables-have-shebangs
- id: check-json
- id: check-shebang-scripts-are-executable
- id: check-symlinks
- id: check-toml
- repo: https://github.com/Lucas-C/pre-commit-hooks
rev: v1.5.4

594
CHANGELOG.md
View file

@ -3,28 +3,28 @@
# v0.2.3
Published on: 2025-04-25T22:46:21Z
## Highlights
* OpenAI compatible inference endpoints and client-SDK support. `client.chat.completions.create()` now works.
* significant improvements and functionality added to the nVIDIA distribution
* many improvements to the test verification suite.
* new inference providers: Ramalama, IBM WatsonX
* many improvements to the Playground UI
## Highlights
* OpenAI compatible inference endpoints and client-SDK support. `client.chat.completions.create()` now works.
* significant improvements and functionality added to the nVIDIA distribution
* many improvements to the test verification suite.
* new inference providers: Ramalama, IBM WatsonX
* many improvements to the Playground UI
---
# v0.2.2
Published on: 2025-04-13T01:19:49Z
## Main changes
- Bring Your Own Provider (@leseb) - use out-of-tree provider code to execute the distribution server
- OpenAI compatible inference API in progress (@bbrowning)
- Provider verifications (@ehhuang)
- Many updates and fixes to playground
- Several llama4 related fixes
## Main changes
- Bring Your Own Provider (@leseb) - use out-of-tree provider code to execute the distribution server
- OpenAI compatible inference API in progress (@bbrowning)
- Provider verifications (@ehhuang)
- Many updates and fixes to playground
- Several llama4 related fixes
---
@ -38,10 +38,10 @@ Published on: 2025-04-05T23:13:00Z
# v0.2.0
Published on: 2025-04-05T19:04:29Z
## Llama 4 Support
Checkout more at https://www.llama.com
## Llama 4 Support
Checkout more at https://www.llama.com
---
@ -49,58 +49,58 @@ Checkout more at https://www.llama.com
# v0.1.9
Published on: 2025-03-29T00:52:23Z
### Build and Test Agents
* Agents: Entire document context with attachments
* RAG: Documentation with sqlite-vec faiss comparison
* Getting started: Fixes to getting started notebook.
### Agent Evals and Model Customization
* (**New**) Post-training: Add nemo customizer
### Better Engineering
* Moved sqlite-vec to non-blocking calls
* Don't return a payload on file delete
### Build and Test Agents
* Agents: Entire document context with attachments
* RAG: Documentation with sqlite-vec faiss comparison
* Getting started: Fixes to getting started notebook.
### Agent Evals and Model Customization
* (**New**) Post-training: Add nemo customizer
### Better Engineering
* Moved sqlite-vec to non-blocking calls
* Don't return a payload on file delete
---
# v0.1.8
Published on: 2025-03-24T01:28:50Z
# v0.1.8 Release Notes
### Build and Test Agents
* Safety: Integrated NVIDIA as a safety provider.
* VectorDB: Added Qdrant as an inline provider.
* Agents: Added support for multiple tool groups in agents.
* Agents: Simplified imports for Agents in client package
### Agent Evals and Model Customization
* Introduced DocVQA and IfEval benchmarks.
### Deploying and Monitoring Agents
* Introduced a Containerfile and image workflow for the Playground.
* Implemented support for Bearer (API Key) authentication.
* Added attribute-based access control for resources.
* Fixes on docker deployments: use --pull always and standardized the default port to 8321
* Deprecated: /v1/inspect/providers use /v1/providers/ instead
### Better Engineering
* Consolidated scripts under the ./scripts directory.
* Addressed mypy violations in various modules.
* Added Dependabot scans for Python dependencies.
* Implemented a scheduled workflow to update the changelog automatically.
* Enforced concurrency to reduce CI loads.
### New Contributors
* @cmodi-meta made their first contribution in https://github.com/meta-llama/llama-stack/pull/1650
* @jeffmaury made their first contribution in https://github.com/meta-llama/llama-stack/pull/1671
* @derekhiggins made their first contribution in https://github.com/meta-llama/llama-stack/pull/1698
* @Bobbins228 made their first contribution in https://github.com/meta-llama/llama-stack/pull/1745
# v0.1.8 Release Notes
### Build and Test Agents
* Safety: Integrated NVIDIA as a safety provider.
* VectorDB: Added Qdrant as an inline provider.
* Agents: Added support for multiple tool groups in agents.
* Agents: Simplified imports for Agents in client package
### Agent Evals and Model Customization
* Introduced DocVQA and IfEval benchmarks.
### Deploying and Monitoring Agents
* Introduced a Containerfile and image workflow for the Playground.
* Implemented support for Bearer (API Key) authentication.
* Added attribute-based access control for resources.
* Fixes on docker deployments: use --pull always and standardized the default port to 8321
* Deprecated: /v1/inspect/providers use /v1/providers/ instead
### Better Engineering
* Consolidated scripts under the ./scripts directory.
* Addressed mypy violations in various modules.
* Added Dependabot scans for Python dependencies.
* Implemented a scheduled workflow to update the changelog automatically.
* Enforced concurrency to reduce CI loads.
### New Contributors
* @cmodi-meta made their first contribution in https://github.com/meta-llama/llama-stack/pull/1650
* @jeffmaury made their first contribution in https://github.com/meta-llama/llama-stack/pull/1671
* @derekhiggins made their first contribution in https://github.com/meta-llama/llama-stack/pull/1698
* @Bobbins228 made their first contribution in https://github.com/meta-llama/llama-stack/pull/1745
**Full Changelog**: https://github.com/meta-llama/llama-stack/compare/v0.1.7...v0.1.8
---
@ -108,73 +108,73 @@ Published on: 2025-03-24T01:28:50Z
# v0.1.7
Published on: 2025-03-14T22:30:51Z
## 0.1.7 Release Notes
### Build and Test Agents
* Inference: ImageType is now refactored to LlamaStackImageType
* Inference: Added tests to measure TTFT
* Inference: Bring back usage metrics
* Agents: Added endpoint for get agent, list agents and list sessions
* Agents: Automated conversion of type hints in client tool for lite llm format
* Agents: Deprecated ToolResponseMessage in agent.resume API
* Added Provider API for listing and inspecting provider info
### Agent Evals and Model Customization
* Eval: Added new eval benchmarks Math 500 and BFCL v3
* Deploy and Monitoring of Agents
* Telemetry: Fix tracing to work across coroutines
### Better Engineering
* Display code coverage for unit tests
* Updated call sites (inference, tool calls, agents) to move to async non blocking calls
* Unit tests also run on Python 3.11, 3.12, and 3.13
* Added ollama inference to Integration tests CI
* Improved documentation across examples, testing, CLI, updated providers table )
## 0.1.7 Release Notes
### Build and Test Agents
* Inference: ImageType is now refactored to LlamaStackImageType
* Inference: Added tests to measure TTFT
* Inference: Bring back usage metrics
* Agents: Added endpoint for get agent, list agents and list sessions
* Agents: Automated conversion of type hints in client tool for lite llm format
* Agents: Deprecated ToolResponseMessage in agent.resume API
* Added Provider API for listing and inspecting provider info
### Agent Evals and Model Customization
* Eval: Added new eval benchmarks Math 500 and BFCL v3
* Deploy and Monitoring of Agents
* Telemetry: Fix tracing to work across coroutines
### Better Engineering
* Display code coverage for unit tests
* Updated call sites (inference, tool calls, agents) to move to async non blocking calls
* Unit tests also run on Python 3.11, 3.12, and 3.13
* Added ollama inference to Integration tests CI
* Improved documentation across examples, testing, CLI, updated providers table )
---
# v0.1.6
Published on: 2025-03-08T04:35:08Z
## 0.1.6 Release Notes
### Build and Test Agents
* Inference: Fixed support for inline vllm provider
* (**New**) Agent: Build & Monitor Agent Workflows with Llama Stack + Anthropic's Best Practice [Notebook](https://github.com/meta-llama/llama-stack/blob/main/docs/notebooks/Llama_Stack_Agent_Workflows.ipynb)
* (**New**) Agent: Revamped agent [documentation](https://llama-stack.readthedocs.io/en/latest/building_applications/agent.html) with more details and examples
* Agent: Unify tools and Python SDK Agents API
* Agent: AsyncAgent Python SDK wrapper supporting async client tool calls
* Agent: Support python functions without @client_tool decorator as client tools
* Agent: deprecation for allow_resume_turn flag, and remove need to specify tool_prompt_format
* VectorIO: MilvusDB support added
### Agent Evals and Model Customization
* (**New**) Agent: Llama Stack RAG Lifecycle [Notebook](https://github.com/meta-llama/llama-stack/blob/main/docs/notebooks/Llama_Stack_RAG_Lifecycle.ipynb)
* Eval: Documentation for eval, scoring, adding new benchmarks
* Eval: Distribution template to run benchmarks on llama & non-llama models
* Eval: Ability to register new custom LLM-as-judge scoring functions
* (**New**) Looking for contributors for open benchmarks. See [documentation](https://llama-stack.readthedocs.io/en/latest/references/evals_reference/index.html#open-benchmark-contributing-guide) for details.
### Deploy and Monitoring of Agents
* Better support for different log levels across all components for better monitoring
### Better Engineering
* Enhance OpenAPI spec to include Error types across all APIs
* Moved all tests to /tests and created unit tests to run on each PR
* Removed all dependencies on llama-models repo
## 0.1.6 Release Notes
### Build and Test Agents
* Inference: Fixed support for inline vllm provider
* (**New**) Agent: Build & Monitor Agent Workflows with Llama Stack + Anthropic's Best Practice [Notebook](https://github.com/meta-llama/llama-stack/blob/main/docs/notebooks/Llama_Stack_Agent_Workflows.ipynb)
* (**New**) Agent: Revamped agent [documentation](https://llama-stack.readthedocs.io/en/latest/building_applications/agent.html) with more details and examples
* Agent: Unify tools and Python SDK Agents API
* Agent: AsyncAgent Python SDK wrapper supporting async client tool calls
* Agent: Support python functions without @client_tool decorator as client tools
* Agent: deprecation for allow_resume_turn flag, and remove need to specify tool_prompt_format
* VectorIO: MilvusDB support added
### Agent Evals and Model Customization
* (**New**) Agent: Llama Stack RAG Lifecycle [Notebook](https://github.com/meta-llama/llama-stack/blob/main/docs/notebooks/Llama_Stack_RAG_Lifecycle.ipynb)
* Eval: Documentation for eval, scoring, adding new benchmarks
* Eval: Distribution template to run benchmarks on llama & non-llama models
* Eval: Ability to register new custom LLM-as-judge scoring functions
* (**New**) Looking for contributors for open benchmarks. See [documentation](https://llama-stack.readthedocs.io/en/latest/references/evals_reference/index.html#open-benchmark-contributing-guide) for details.
### Deploy and Monitoring of Agents
* Better support for different log levels across all components for better monitoring
### Better Engineering
* Enhance OpenAPI spec to include Error types across all APIs
* Moved all tests to /tests and created unit tests to run on each PR
* Removed all dependencies on llama-models repo
---
# v0.1.5.1
Published on: 2025-02-28T22:37:44Z
## 0.1.5.1 Release Notes
* Fixes for security risk in https://github.com/meta-llama/llama-stack/pull/1327 and https://github.com/meta-llama/llama-stack/pull/1328
## 0.1.5.1 Release Notes
* Fixes for security risk in https://github.com/meta-llama/llama-stack/pull/1327 and https://github.com/meta-llama/llama-stack/pull/1328
**Full Changelog**: https://github.com/meta-llama/llama-stack/compare/v0.1.5...v0.1.5.1
---
@ -182,176 +182,176 @@ Published on: 2025-02-28T22:37:44Z
# v0.1.5
Published on: 2025-02-28T18:14:01Z
## 0.1.5 Release Notes
### Build Agents
* Inference: Support more non-llama models (openai, anthropic, gemini)
* Inference: Can use the provider's model name in addition to the HF alias
* Inference: Fixed issues with calling tools that weren't specified in the prompt
* RAG: Improved system prompt for RAG and no more need for hard-coded rag-tool calling
* Embeddings: Added support for Nemo retriever embedding models
* Tools: Added support for MCP tools in Ollama Distribution
* Distributions: Added new Groq distribution
### Customize Models
* Save post-trained checkpoint in SafeTensor format to allow Ollama inference provider to use the post-trained model
### Monitor agents
* More comprehensive logging of agent steps including client tools
* Telemetry inputs/outputs are now structured and queryable
* Ability to retrieve agents session, turn, step by ids
### Better Engineering
* Moved executorch Swift code out of this repo into the llama-stack-client-swift repo, similar to kotlin
* Move most logging to use logger instead of prints
* Completed text /chat-completion and /completion tests
## 0.1.5 Release Notes
### Build Agents
* Inference: Support more non-llama models (openai, anthropic, gemini)
* Inference: Can use the provider's model name in addition to the HF alias
* Inference: Fixed issues with calling tools that weren't specified in the prompt
* RAG: Improved system prompt for RAG and no more need for hard-coded rag-tool calling
* Embeddings: Added support for Nemo retriever embedding models
* Tools: Added support for MCP tools in Ollama Distribution
* Distributions: Added new Groq distribution
### Customize Models
* Save post-trained checkpoint in SafeTensor format to allow Ollama inference provider to use the post-trained model
### Monitor agents
* More comprehensive logging of agent steps including client tools
* Telemetry inputs/outputs are now structured and queryable
* Ability to retrieve agents session, turn, step by ids
### Better Engineering
* Moved executorch Swift code out of this repo into the llama-stack-client-swift repo, similar to kotlin
* Move most logging to use logger instead of prints
* Completed text /chat-completion and /completion tests
---
# v0.1.4
Published on: 2025-02-25T00:02:43Z
## v0.1.4 Release Notes
Here are the key changes coming as part of this release:
### Build and Test Agents
* Inference: Added support for non-llama models
* Inference: Added option to list all downloaded models and remove models
* Agent: Introduce new api agents.resume_turn to include client side tool execution in the same turn
* Agent: AgentConfig introduces new variable “tool_config” that allows for better tool configuration and system prompt overrides
* Agent: Added logging for agent step start and completion times
* Agent: Added support for logging for tool execution metadata
* Embedding: Updated /inference/embeddings to support asymmetric models, truncation and variable sized outputs
* Embedding: Updated embedding models for Ollama, Together, and Fireworks with available defaults
* VectorIO: Improved performance of sqlite-vec using chunked writes
### Agent Evals and Model Customization
* Deprecated api /eval-tasks. Use /eval/benchmark instead
* Added CPU training support for TorchTune
### Deploy and Monitoring of Agents
* Consistent view of client and server tool calls in telemetry
### Better Engineering
* Made tests more data-driven for consistent evaluation
* Fixed documentation links and improved API reference generation
* Various small fixes for build scripts and system reliability
## v0.1.4 Release Notes
Here are the key changes coming as part of this release:
### Build and Test Agents
* Inference: Added support for non-llama models
* Inference: Added option to list all downloaded models and remove models
* Agent: Introduce new api agents.resume_turn to include client side tool execution in the same turn
* Agent: AgentConfig introduces new variable “tool_config” that allows for better tool configuration and system prompt overrides
* Agent: Added logging for agent step start and completion times
* Agent: Added support for logging for tool execution metadata
* Embedding: Updated /inference/embeddings to support asymmetric models, truncation and variable sized outputs
* Embedding: Updated embedding models for Ollama, Together, and Fireworks with available defaults
* VectorIO: Improved performance of sqlite-vec using chunked writes
### Agent Evals and Model Customization
* Deprecated api /eval-tasks. Use /eval/benchmark instead
* Added CPU training support for TorchTune
### Deploy and Monitoring of Agents
* Consistent view of client and server tool calls in telemetry
### Better Engineering
* Made tests more data-driven for consistent evaluation
* Fixed documentation links and improved API reference generation
* Various small fixes for build scripts and system reliability
---
# v0.1.3
Published on: 2025-02-14T20:24:32Z
## v0.1.3 Release
Here are some key changes that are coming as part of this release.
### Build and Test Agents
Streamlined the initial development experience
- Added support for llama stack run --image-type venv
- Enhanced vector store options with new sqlite-vec provider and improved Qdrant integration
- vLLM improvements for tool calling and logprobs
- Better handling of sporadic code_interpreter tool calls
### Agent Evals
Better benchmarking and Agent performance assessment
- Renamed eval API /eval-task to /benchmarks
- Improved documentation and notebooks for RAG and evals
### Deploy and Monitoring of Agents
Improved production readiness
- Added usage metrics collection for chat completions
- CLI improvements for provider information
- Improved error handling and system reliability
- Better model endpoint handling and accessibility
- Improved signal handling on distro server
### Better Engineering
Infrastructure and code quality improvements
- Faster text-based chat completion tests
- Improved testing for non-streaming agent apis
- Standardized import formatting with ruff linter
- Added conventional commits standard
- Fixed documentation parsing issues
## v0.1.3 Release
Here are some key changes that are coming as part of this release.
### Build and Test Agents
Streamlined the initial development experience
- Added support for llama stack run --image-type venv
- Enhanced vector store options with new sqlite-vec provider and improved Qdrant integration
- vLLM improvements for tool calling and logprobs
- Better handling of sporadic code_interpreter tool calls
### Agent Evals
Better benchmarking and Agent performance assessment
- Renamed eval API /eval-task to /benchmarks
- Improved documentation and notebooks for RAG and evals
### Deploy and Monitoring of Agents
Improved production readiness
- Added usage metrics collection for chat completions
- CLI improvements for provider information
- Improved error handling and system reliability
- Better model endpoint handling and accessibility
- Improved signal handling on distro server
### Better Engineering
Infrastructure and code quality improvements
- Faster text-based chat completion tests
- Improved testing for non-streaming agent apis
- Standardized import formatting with ruff linter
- Added conventional commits standard
- Fixed documentation parsing issues
---
# v0.1.2
Published on: 2025-02-07T22:06:49Z
# TL;DR
- Several stabilizations to development flows after the switch to `uv`
- Migrated CI workflows to new OSS repo - [llama-stack-ops](https://github.com/meta-llama/llama-stack-ops)
- Added automated rebuilds for ReadTheDocs
- Llama Stack server supports HTTPS
- Added system prompt overrides support
- Several bug fixes and improvements to documentation (check out Kubernetes deployment guide by @terrytangyuan )
# TL;DR
- Several stabilizations to development flows after the switch to `uv`
- Migrated CI workflows to new OSS repo - [llama-stack-ops](https://github.com/meta-llama/llama-stack-ops)
- Added automated rebuilds for ReadTheDocs
- Llama Stack server supports HTTPS
- Added system prompt overrides support
- Several bug fixes and improvements to documentation (check out Kubernetes deployment guide by @terrytangyuan )
---
# v0.1.1
Published on: 2025-02-02T02:29:24Z
A bunch of small / big improvements everywhere including support for Windows, switching to `uv` and many provider improvements.
A bunch of small / big improvements everywhere including support for Windows, switching to `uv` and many provider improvements.
---
# v0.1.0
Published on: 2025-01-24T17:47:47Z
We are excited to announce a stable API release of Llama Stack, which enables developers to build RAG applications and Agents using tools and safety shields, monitor and those agents with telemetry, and evaluate the agent with scoring functions.
## Context
GenAI application developers need more than just an LLM - they need to integrate tools, connect with their data sources, establish guardrails, and ground the LLM responses effectively. Currently, developers must piece together various tools and APIs, complicating the development lifecycle and increasing costs. The result is that developers are spending more time on these integrations rather than focusing on the application logic itself. The bespoke coupling of components also makes it challenging to adopt state-of-the-art solutions in the rapidly evolving GenAI space. This is particularly difficult for open models like Llama, as best practices are not widely established in the open.
Llama Stack was created to provide developers with a comprehensive and coherent interface that simplifies AI application development and codifies best practices across the Llama ecosystem. Since our launch in September 2024, we have seen a huge uptick in interest in Llama Stack APIs by both AI developers and from partners building AI services with Llama models. Partners like Nvidia, Fireworks, and Ollama have collaborated with us to develop implementations across various APIs, including inference, memory, and safety.
With Llama Stack, you can easily build a RAG agent which can also search the web, do complex math, and custom tool calling. You can use telemetry to inspect those traces, and convert telemetry into evals datasets. And with Llama Stack’s plugin architecture and prepackage distributions, you choose to run your agent anywhere - in the cloud with our partners, deploy your own environment using virtualenv, conda, or Docker, operate locally with Ollama, or even run on mobile devices with our SDKs. Llama Stack offers unprecedented flexibility while also simplifying the developer experience.
## Release
After iterating on the APIs for the last 3 months, today we’re launching a stable release (V1) of the Llama Stack APIs and the corresponding llama-stack server and client packages(v0.1.0). We now have automated tests for providers. These tests make sure that all provider implementations are verified. Developers can now easily and reliably select distributions or providers based on their specific requirements.
There are example standalone apps in llama-stack-apps.
## Key Features of this release
- **Unified API Layer**
- Inference: Run LLM models
- RAG: Store and retrieve knowledge for RAG
- Agents: Build multi-step agentic workflows
- Tools: Register tools that can be called by the agent
- Safety: Apply content filtering and safety policies
- Evaluation: Test model and agent quality
- Telemetry: Collect and analyze usage data and complex agentic traces
- Post Training ( Coming Soon ): Fine tune models for specific use cases
- **Rich Provider Ecosystem**
- Local Development: Meta's Reference, Ollama
- Cloud: Fireworks, Together, Nvidia, AWS Bedrock, Groq, Cerebras
- On-premises: Nvidia NIM, vLLM, TGI, Dell-TGI
- On-device: iOS and Android support
- **Built for Production**
- Pre-packaged distributions for common deployment scenarios
- Backwards compatibility across model versions
- Comprehensive evaluation capabilities
- Full observability and monitoring
- **Multiple developer interfaces**
- CLI: Command line interface
- Python SDK
- Swift iOS SDK
- Kotlin Android SDK
- **Sample llama stack applications**
- Python
- iOS
- Android
We are excited to announce a stable API release of Llama Stack, which enables developers to build RAG applications and Agents using tools and safety shields, monitor and those agents with telemetry, and evaluate the agent with scoring functions.
## Context
GenAI application developers need more than just an LLM - they need to integrate tools, connect with their data sources, establish guardrails, and ground the LLM responses effectively. Currently, developers must piece together various tools and APIs, complicating the development lifecycle and increasing costs. The result is that developers are spending more time on these integrations rather than focusing on the application logic itself. The bespoke coupling of components also makes it challenging to adopt state-of-the-art solutions in the rapidly evolving GenAI space. This is particularly difficult for open models like Llama, as best practices are not widely established in the open.
Llama Stack was created to provide developers with a comprehensive and coherent interface that simplifies AI application development and codifies best practices across the Llama ecosystem. Since our launch in September 2024, we have seen a huge uptick in interest in Llama Stack APIs by both AI developers and from partners building AI services with Llama models. Partners like Nvidia, Fireworks, and Ollama have collaborated with us to develop implementations across various APIs, including inference, memory, and safety.
With Llama Stack, you can easily build a RAG agent which can also search the web, do complex math, and custom tool calling. You can use telemetry to inspect those traces, and convert telemetry into evals datasets. And with Llama Stack’s plugin architecture and prepackage distributions, you choose to run your agent anywhere - in the cloud with our partners, deploy your own environment using virtualenv, conda, or Docker, operate locally with Ollama, or even run on mobile devices with our SDKs. Llama Stack offers unprecedented flexibility while also simplifying the developer experience.
## Release
After iterating on the APIs for the last 3 months, today we’re launching a stable release (V1) of the Llama Stack APIs and the corresponding llama-stack server and client packages(v0.1.0). We now have automated tests for providers. These tests make sure that all provider implementations are verified. Developers can now easily and reliably select distributions or providers based on their specific requirements.
There are example standalone apps in llama-stack-apps.
## Key Features of this release
- **Unified API Layer**
- Inference: Run LLM models
- RAG: Store and retrieve knowledge for RAG
- Agents: Build multi-step agentic workflows
- Tools: Register tools that can be called by the agent
- Safety: Apply content filtering and safety policies
- Evaluation: Test model and agent quality
- Telemetry: Collect and analyze usage data and complex agentic traces
- Post Training ( Coming Soon ): Fine tune models for specific use cases
- **Rich Provider Ecosystem**
- Local Development: Meta's Reference, Ollama
- Cloud: Fireworks, Together, Nvidia, AWS Bedrock, Groq, Cerebras
- On-premises: Nvidia NIM, vLLM, TGI, Dell-TGI
- On-device: iOS and Android support
- **Built for Production**
- Pre-packaged distributions for common deployment scenarios
- Backwards compatibility across model versions
- Comprehensive evaluation capabilities
- Full observability and monitoring
- **Multiple developer interfaces**
- CLI: Command line interface
- Python SDK
- Swift iOS SDK
- Kotlin Android SDK
- **Sample llama stack applications**
- Python
- iOS
- Android
---
@ -365,8 +365,8 @@ Published on: 2025-01-22T22:24:01Z
# v0.0.63
Published on: 2024-12-18T07:17:43Z
A small but important bug-fix release to update the URL datatype for the client-SDKs. The issue affected multimodal agentic turns especially.
A small but important bug-fix release to update the URL datatype for the client-SDKs. The issue affected multimodal agentic turns especially.
**Full Changelog**: https://github.com/meta-llama/llama-stack/compare/v0.0.62...v0.0.63
---
@ -402,39 +402,39 @@ Published on: 2024-11-22T00:36:09Z
# v0.0.53
Published on: 2024-11-20T22:18:00Z
🚀 Initial Release Notes for Llama Stack!
### Added
- Resource-oriented design for models, shields, memory banks, datasets and eval tasks
- Persistence for registered objects with distribution
- Ability to persist memory banks created for FAISS
- PostgreSQL KVStore implementation
- Environment variable placeholder support in run.yaml files
- Comprehensive Zero-to-Hero notebooks and quickstart guides
- Support for quantized models in Ollama
- Vision models support for Together, Fireworks, Meta-Reference, and Ollama, and vLLM
- Bedrock distribution with safety shields support
- Evals API with task registration and scoring functions
- MMLU and SimpleQA benchmark scoring functions
- Huggingface dataset provider integration for benchmarks
- Support for custom dataset registration from local paths
- Benchmark evaluation CLI tools with visualization tables
- RAG evaluation scoring functions and metrics
- Local persistence for datasets and eval tasks
### Changed
- Split safety into distinct providers (llama-guard, prompt-guard, code-scanner)
- Changed provider naming convention (`impls` → `inline`, `adapters` → `remote`)
- Updated API signatures for dataset and eval task registration
- Restructured folder organization for providers
- Enhanced Docker build configuration
- Added version prefixing for REST API routes
- Enhanced evaluation task registration workflow
- Improved benchmark evaluation output formatting
- Restructured evals folder organization for better modularity
### Removed
- `llama stack configure` command
🚀 Initial Release Notes for Llama Stack!
### Added
- Resource-oriented design for models, shields, memory banks, datasets and eval tasks
- Persistence for registered objects with distribution
- Ability to persist memory banks created for FAISS
- PostgreSQL KVStore implementation
- Environment variable placeholder support in run.yaml files
- Comprehensive Zero-to-Hero notebooks and quickstart guides
- Support for quantized models in Ollama
- Vision models support for Together, Fireworks, Meta-Reference, and Ollama, and vLLM
- Bedrock distribution with safety shields support
- Evals API with task registration and scoring functions
- MMLU and SimpleQA benchmark scoring functions
- Huggingface dataset provider integration for benchmarks
- Support for custom dataset registration from local paths
- Benchmark evaluation CLI tools with visualization tables
- RAG evaluation scoring functions and metrics
- Local persistence for datasets and eval tasks
### Changed
- Split safety into distinct providers (llama-guard, prompt-guard, code-scanner)
- Changed provider naming convention (`impls` → `inline`, `adapters` → `remote`)
- Updated API signatures for dataset and eval task registration
- Restructured folder organization for providers
- Enhanced Docker build configuration
- Added version prefixing for REST API routes
- Enhanced evaluation task registration workflow
- Improved benchmark evaluation output formatting
- Restructured evals folder organization for better modularity
### Removed
- `llama stack configure` command
---

15
CONTRIBUTING.md
View file

@ -141,11 +141,18 @@ uv sync
## Coding Style
* Comments should provide meaningful insights into the code. Avoid filler comments that simply describe the next step, as they create unnecessary clutter, same goes for docstrings.
* Prefer comments to clarify surprising behavior and/or relationships between parts of the code rather than explain what the next line of code does.
* Catching exceptions, prefer using a specific exception type rather than a broad catch-all like `Exception`.
* Comments should provide meaningful insights into the code. Avoid filler comments that simply
describe the next step, as they create unnecessary clutter, same goes for docstrings.
* Prefer comments to clarify surprising behavior and/or relationships between parts of the code
rather than explain what the next line of code does.
* Catching exceptions, prefer using a specific exception type rather than a broad catch-all like
`Exception`.
* Error messages should be prefixed with "Failed to ..."
* 4 spaces for indentation rather than tabs
* 4 spaces for indentation rather than tab
* When using `# noqa` to suppress a style or linter warning, include a comment explaining the
justification for bypassing the check.
* When using `# type: ignore` to suppress a mypy warning, include a comment explaining the
justification for bypassing the check.
## Common Tasks

6
docs/_static/css/my_theme.css vendored
View file

@ -27,3 +27,9 @@ pre {
white-space: pre-wrap !important;
word-break: break-all;
}
[data-theme="dark"] .mermaid {
background-color: #f4f4f6 !important;
border-radius: 6px;
padding: 0.5em;
}

3
docs/_static/llama-stack-spec.html vendored
View file

@ -6462,6 +6462,9 @@
"stream": {
"type": "boolean"
},
"temperature": {
"type": "number"
},
"tools": {
"type": "array",
"items": {

2
docs/_static/llama-stack-spec.yaml vendored
View file

@ -4506,6 +4506,8 @@ components:
type: boolean
stream:
type: boolean
temperature:
type: number
tools:
type: array
items:

70
docs/make.bat
View file

@ -1,35 +1,35 @@
@ECHO OFF
pushd %~dp0
REM Command file for Sphinx documentation
if "%SPHINXBUILD%" == "" (
set SPHINXBUILD=sphinx-build
)
set SOURCEDIR=.
set BUILDDIR=_build
%SPHINXBUILD% >NUL 2>NUL
if errorlevel 9009 (
echo.
echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
echo.installed, then set the SPHINXBUILD environment variable to point
echo.to the full path of the 'sphinx-build' executable. Alternatively you
echo.may add the Sphinx directory to PATH.
echo.
echo.If you don't have Sphinx installed, grab it from
echo.https://www.sphinx-doc.org/
exit /b 1
)
if "%1" == "" goto help
%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
goto end
:help
%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
:end
popd
@ECHO OFF
pushd %~dp0
REM Command file for Sphinx documentation
if "%SPHINXBUILD%" == "" (
set SPHINXBUILD=sphinx-build
)
set SOURCEDIR=.
set BUILDDIR=_build
%SPHINXBUILD% >NUL 2>NUL
if errorlevel 9009 (
echo.
echo.The 'sphinx-build' command was not found. Make sure you have Sphinx
echo.installed, then set the SPHINXBUILD environment variable to point
echo.to the full path of the 'sphinx-build' executable. Alternatively you
echo.may add the Sphinx directory to PATH.
echo.
echo.If you don't have Sphinx installed, grab it from
echo.https://www.sphinx-doc.org/
exit /b 1
)
if "%1" == "" goto help
%SPHINXBUILD% -M %1 %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
goto end
:help
%SPHINXBUILD% -M help %SOURCEDIR% %BUILDDIR% %SPHINXOPTS% %O%
:end
popd

3
docs/openapi_generator/pyopenapi/generator.py
View file

@ -6,6 +6,7 @@
import hashlib
import ipaddress
import types
import typing
from dataclasses import make_dataclass
from typing import Any, Dict, Set, Union
@ -189,7 +190,7 @@ class ContentBuilder:
else:
return "application/json"
if typing.get_origin(payload_type) is typing.Union:
if typing.get_origin(payload_type) in (typing.Union, types.UnionType):
media_types = []
item_types = []
for x in typing.get_args(payload_type):

12
docs/requirements.txt
View file

@ -1,16 +1,16 @@
sphinx==8.1.3
myst-parser
linkify
myst-parser
-e git+https://github.com/pytorch/pytorch_sphinx_theme.git#egg=pytorch_sphinx_theme
sphinx-rtd-theme>=1.0.0
sphinx_autobuild
sphinx==8.1.3
sphinx-copybutton
sphinx-design
sphinx-pdj-theme
sphinx_rtd_dark_mode
sphinx-rtd-theme>=1.0.0
sphinx-tabs
sphinx_autobuild
sphinx_rtd_dark_mode
sphinxcontrib-mermaid
sphinxcontrib-openapi
sphinxcontrib-redoc
sphinxcontrib-mermaid
sphinxcontrib-video
tomli

23
docs/source/building_applications/tools.md
View file

@ -43,27 +43,6 @@ The tool requires an API key which can be provided either in the configuration o
> **NOTE:** When using Tavily Search and Bing Search, the inference output will still display "Brave Search." This is because Llama models have been trained with Brave Search as a built-in tool. Tavily and bing is just being used in lieu of Brave search.
#### Code Interpreter
The Code Interpreter allows execution of Python code within a controlled environment.
```python
# Register Code Interpreter tool group
client.toolgroups.register(
toolgroup_id="builtin::code_interpreter", provider_id="code_interpreter"
)
```
Features:
- Secure execution environment using `bwrap` sandboxing
- Matplotlib support for generating plots
- Disabled dangerous system operations
- Configurable execution timeouts
> ⚠️ Important: The code interpreter tool can operate in a controlled environment locally or on Podman containers. To ensure proper functionality in containerized environments:
> - The container requires privileged access (e.g., --privileged).
> - Users without sufficient permissions may encounter permission errors. (`bwrap: Can't mount devpts on /newroot/dev/pts: Permission denied`)
> - 🔒 Security Warning: Privileged mode grants elevated access and bypasses security restrictions. Use only in local, isolated, or controlled environments.
#### WolframAlpha
@ -102,7 +81,7 @@ Features:
- Context retrieval with token limits
> **Note:** By default, llama stack run.yaml defines toolgroups for web search, code interpreter and rag, that are provided by tavily-search, code-interpreter and rag providers.
> **Note:** By default, llama stack run.yaml defines toolgroups for web search, wolfram alpha and rag, that are provided by tavily-search, wolfram-alpha and rag providers.
## Model Context Protocol (MCP) Tools

2
docs/source/distributions/remote_hosted_distro/watsonx.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-watsonx` distribution consists of the following pro
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss` |

2
docs/source/distributions/self_hosted_distro/bedrock.md
View file

@ -19,7 +19,7 @@ The `llamastack/distribution-bedrock` distribution consists of the following pro
| safety | `remote::bedrock` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/cerebras.md
View file

@ -12,7 +12,7 @@ The `llamastack/distribution-cerebras` distribution consists of the following pr
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/fireworks.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-fireworks` distribution consists of the following p
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `remote::wolfram-alpha`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `remote::wolfram-alpha`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/groq.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-groq` distribution consists of the following provid
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime` |
| vector_io | `inline::faiss` |

2
docs/source/distributions/self_hosted_distro/meta-reference-gpu.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-meta-reference-gpu` distribution consists of the fo
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/ollama.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-ollama` distribution consists of the following prov
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/passthrough.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-passthrough` distribution consists of the following
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `remote::wolfram-alpha`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `remote::wolfram-alpha`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/remote-vllm.md
View file

@ -21,7 +21,7 @@ The `llamastack/distribution-remote-vllm` distribution consists of the following
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/sambanova.md
View file

@ -19,7 +19,7 @@ The `llamastack/distribution-sambanova` distribution consists of the following p
| inference | `remote::sambanova` |
| safety | `inline::llama-guard` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/tgi.md
View file

@ -23,7 +23,7 @@ The `llamastack/distribution-tgi` distribution consists of the following provide
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/distributions/self_hosted_distro/together.md
View file

@ -22,7 +22,7 @@ The `llamastack/distribution-together` distribution consists of the following pr
| safety | `inline::llama-guard` |
| scoring | `inline::basic`, `inline::llm-as-judge`, `inline::braintrust` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::code-interpreter`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| tool_runtime | `remote::brave-search`, `remote::tavily-search`, `inline::rag-runtime`, `remote::model-context-protocol`, `remote::wolfram-alpha` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |

2
docs/source/providers/external.md
View file

@ -53,7 +53,9 @@ Here's a list of known external providers that you can use with Llama Stack:
| Name | Description | API | Type | Repository |
|------|-------------|-----|------|------------|
| KubeFlow Training | Train models with KubeFlow | Post Training | Remote | [llama-stack-provider-kft](https://github.com/opendatahub-io/llama-stack-provider-kft) |
| KubeFlow Pipelines | Train models with KubeFlow Pipelines | Post Training | Remote | [llama-stack-provider-kfp-trainer](https://github.com/opendatahub-io/llama-stack-provider-kfp-trainer) |
| RamaLama | Inference models with RamaLama | Inference | Remote | [ramalama-stack](https://github.com/containers/ramalama-stack) |
| TrustyAI LM-Eval | Evaluate models with TrustyAI LM-Eval | Eval | Remote | [llama-stack-provider-lmeval](https://github.com/trustyai-explainability/llama-stack-provider-lmeval) |
### Remote Provider Specification

107
docs/source/providers/vector_io/milvus.md Normal file
View file

@ -0,0 +1,107 @@
---
orphan: true
---
# Milvus
[Milvus](https://milvus.io/) is an inline and remote vector database provider for Llama Stack. It
allows you to store and query vectors directly within a Milvus database.
That means you're not limited to storing vectors in memory or in a separate service.
## Features
- Easy to use
- Fully integrated with Llama Stack
## Usage
To use Milvus in your Llama Stack project, follow these steps:
1. Install the necessary dependencies.
2. Configure your Llama Stack project to use Milvus.
3. Start storing and querying vectors.
## Installation
You can install Milvus using pymilvus:
```bash
pip install pymilvus
```
## Configuration
In Llama Stack, Milvus can be configured in two ways:
- **Inline (Local) Configuration** - Uses Milvus-Lite for local storage
- **Remote Configuration** - Connects to a remote Milvus server
### Inline (Local) Configuration
The simplest method is local configuration, which requires setting `db_path`, a path for locally storing Milvus-Lite files:
```yaml
vector_io:
- provider_id: milvus
provider_type: inline::milvus
config:
db_path: ~/.llama/distributions/together/milvus_store.db
```
### Remote Configuration
Remote configuration is suitable for larger data storage requirements:
#### Standard Remote Connection
```yaml
vector_io:
- provider_id: milvus
provider_type: remote::milvus
config:
uri: "http://<host>:<port>"
token: "<user>:<password>"
```
#### TLS-Enabled Remote Connection (One-way TLS)
For connections to Milvus instances with one-way TLS enabled:
```yaml
vector_io:
- provider_id: milvus
provider_type: remote::milvus
config:
uri: "https://<host>:<port>"
token: "<user>:<password>"
secure: True
server_pem_path: "/path/to/server.pem"
```
#### Mutual TLS (mTLS) Remote Connection
For connections to Milvus instances with mutual TLS (mTLS) enabled:
```yaml
vector_io:
- provider_id: milvus
provider_type: remote::milvus
config:
uri: "https://<host>:<port>"
token: "<user>:<password>"
secure: True
ca_pem_path: "/path/to/ca.pem"
client_pem_path: "/path/to/client.pem"
client_key_path: "/path/to/client.key"
```
#### Key Parameters for TLS Configuration
- **`secure`**: Enables TLS encryption when set to `true`. Defaults to `false`.
- **`server_pem_path`**: Path to the **server certificate** for verifying the server’s identity (used in one-way TLS).
- **`ca_pem_path`**: Path to the **Certificate Authority (CA) certificate** for validating the server certificate (required in mTLS).
- **`client_pem_path`**: Path to the **client certificate** file (required for mTLS).
- **`client_key_path`**: Path to the **client private key** file (required for mTLS).
## Documentation
See the [Milvus documentation](https://milvus.io/docs/install-overview.md) for more details about Milvus in general.
For more details on TLS configuration, refer to the [TLS setup guide](https://milvus.io/docs/tls.md).

31
docs/source/providers/vector_io/mivus.md
View file

@ -1,31 +0,0 @@
---
orphan: true
---
# Milvus
[Milvus](https://milvus.io/) is an inline and remote vector database provider for Llama Stack. It
allows you to store and query vectors directly within a Milvus database.
That means you're not limited to storing vectors in memory or in a separate service.
## Features
- Easy to use
- Fully integrated with Llama Stack
## Usage
To use Milvus in your Llama Stack project, follow these steps:
1. Install the necessary dependencies.
2. Configure your Llama Stack project to use Milvus.
3. Start storing and querying vectors.
## Installation
You can install Milvus using pymilvus:
```bash
pip install pymilvus
```
## Documentation
See the [Milvus documentation](https://milvus.io/docs/install-overview.md) for more details about Milvus in general.

8
docs/zero_to_hero_guide/README.md
View file

@ -86,11 +86,11 @@ If you're looking for more specific topics, we have a [Zero to Hero Guide](#next
llama stack build --template ollama --image-type conda
```
**Expected Output:**
```
```bash
...
Build Successful! Next steps:
1. Set the environment variables: LLAMA_STACK_PORT, OLLAMA_URL, INFERENCE_MODEL, SAFETY_MODEL
2. `llama stack run /Users/<username>/.llama/distributions/llamastack-ollama/ollama-run.yaml
Build Successful!
You can find the newly-built template here: ~/.llama/distributions/ollama/ollama-run.yaml
You can run the new Llama Stack Distro via: llama stack run ~/.llama/distributions/ollama/ollama-run.yaml --image-type conda
```
3. **Set the ENV variables by exporting them to the terminal**:

107
install.sh
View file

@ -16,61 +16,120 @@ WAIT_TIMEOUT=300
log(){ printf "\e[1;32m%s\e[0m\n" "$*"; }
die(){ printf "\e[1;31m❌ %s\e[0m\n" "$*" >&2; exit 1; }
wait_for_service() {
local url="$1"
local pattern="$2"
local timeout="$3"
local name="$4"
local start ts
log "⏳ Waiting for ${name}…"
start=$(date +%s)
while true; do
if curl --retry 5 --retry-delay 1 --retry-max-time "$timeout" --retry-all-errors --silent --fail "$url" 2>/dev/null | grep -q "$pattern"; then
break
fi
ts=$(date +%s)
if (( ts - start >= timeout )); then
return 1
fi
printf '.'
sleep 1
done
return 0
}
if command -v docker &> /dev/null; then
ENGINE="docker"
HOST_DNS="host.docker.internal"
elif command -v podman &> /dev/null; then
ENGINE="podman"
HOST_DNS="host.containers.internal"
else
die "Docker or Podman is required. Install Docker: https://docs.docker.com/get-docker/ or Podman: https://podman.io/getting-started/installation"
fi
# Explicitly set the platform for the host architecture
HOST_ARCH="$(uname -m)"
if [ "$HOST_ARCH" = "arm64" ]; then
if [ "$ENGINE" = "docker" ]; then
PLATFORM_OPTS=( --platform linux/amd64 )
else
PLATFORM_OPTS=( --os linux --arch amd64 )
fi
else
PLATFORM_OPTS=()
fi
# macOS + Podman: ensure VM is running before we try to launch containers
# If you need GPU passthrough under Podman on macOS, init the VM with libkrun:
# CONTAINERS_MACHINE_PROVIDER=libkrun podman machine init
if [ "$ENGINE" = "podman" ] && [ "$(uname -s)" = "Darwin" ]; then
if ! podman info &>/dev/null; then
log "⌛️ Initializing Podman VM…"
podman machine init &>/dev/null || true
podman machine start &>/dev/null || true
log "⌛️ Waiting for Podman API…"
until podman info &>/dev/null; do
sleep 1
done
log "✅ Podman VM is up"
fi
fi
# Clean up any leftovers from earlier runs
for name in ollama-server llama-stack; do
ids=$($ENGINE ps -aq --filter "name=^${name}$")
if [ -n "$ids" ]; then
log "⚠️ Found existing container(s) for '${name}', removing..."
$ENGINE rm -f "$ids"
log "⚠️ Found existing container(s) for '${name}', removing…"
$ENGINE rm -f "$ids" > /dev/null 2>&1
fi
done
###############################################################################
# 0. Create a shared network
###############################################################################
if ! $ENGINE network inspect llama-net >/dev/null 2>&1; then
log "🌐 Creating network…"
$ENGINE network create llama-net >/dev/null 2>&1
fi
###############################################################################
# 1. Ollama
###############################################################################
log "🦙 Starting Ollama…"
$ENGINE run -d --name ollama-server \
-p "${OLLAMA_PORT}:11434" \
$ENGINE run -d "${PLATFORM_OPTS[@]}" --name ollama-server \
--network llama-net \
-p "${OLLAMA_PORT}:${OLLAMA_PORT}" \
ollama/ollama > /dev/null 2>&1
log "⏳ Waiting for Ollama daemon…"
if ! timeout "$WAIT_TIMEOUT" bash -c \
"until curl -fsS http://localhost:${OLLAMA_PORT}/ 2>/dev/null | grep -q 'Ollama'; do sleep 1; done"; then
if ! wait_for_service "http://localhost:${OLLAMA_PORT}/" "Ollama" "$WAIT_TIMEOUT" "Ollama daemon"; then
log "❌ Ollama daemon did not become ready in ${WAIT_TIMEOUT}s; dumping container logs:"
$ENGINE logs ollama-server --tail=200
$ENGINE logs --tail 200 ollama-server
die "Ollama startup failed"
fi
log "📦 Ensuring model is pulled: ${MODEL_ALIAS}..."
$ENGINE exec ollama-server ollama pull "${MODEL_ALIAS}" > /dev/null 2>&1
log "📦 Ensuring model is pulled: ${MODEL_ALIAS}…"
if ! $ENGINE exec ollama-server ollama pull "${MODEL_ALIAS}" > /dev/null 2>&1; then
log "❌ Failed to pull model ${MODEL_ALIAS}; dumping container logs:"
$ENGINE logs --tail 200 ollama-server
die "Model pull failed"
fi
###############################################################################
# 2. Llama‑Stack
###############################################################################
log "🦙📦 Starting Llama‑Stack…"
$ENGINE run -d --name llama-stack \
-p "${PORT}:${PORT}" \
--add-host="${HOST_DNS}:host-gateway" \
"${SERVER_IMAGE}" \
--port "${PORT}" \
--env INFERENCE_MODEL="${MODEL_ALIAS}" \
--env OLLAMA_URL="http://${HOST_DNS}:${OLLAMA_PORT}" > /dev/null 2>&1
cmd=( run -d "${PLATFORM_OPTS[@]}" --name llama-stack \
--network llama-net \
-p "${PORT}:${PORT}" \
"${SERVER_IMAGE}" --port "${PORT}" \
--env INFERENCE_MODEL="${MODEL_ALIAS}" \
--env OLLAMA_URL="http://ollama-server:${OLLAMA_PORT}" )
log "⏳ Waiting for Llama-Stack API…"
if ! timeout "$WAIT_TIMEOUT" bash -c \
"until curl -fsS http://localhost:${PORT}/v1/health 2>/dev/null | grep -q 'OK'; do sleep 1; done"; then
log "🦙 Starting Llama‑Stack…"
$ENGINE "${cmd[@]}" > /dev/null 2>&1
if ! wait_for_service "http://127.0.0.1:${PORT}/v1/health" "OK" "$WAIT_TIMEOUT" "Llama-Stack API"; then
log "❌ Llama-Stack did not become ready in ${WAIT_TIMEOUT}s; dumping container logs:"
$ENGINE logs llama-stack --tail=200
$ENGINE logs --tail 200 llama-stack
die "Llama-Stack startup failed"
fi

156
llama_stack/apis/agents/agents.py
View file

@ -4,20 +4,10 @@
# 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 AsyncIterator
from datetime import datetime
from enum import Enum
from typing import (
Annotated,
Any,
AsyncIterator,
Dict,
List,
Literal,
Optional,
Protocol,
Union,
runtime_checkable,
)
from typing import Annotated, Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, ConfigDict, Field
@ -79,8 +69,8 @@ class StepCommon(BaseModel):
turn_id: str
step_id: str
started_at: Optional[datetime] = None
completed_at: Optional[datetime] = None
started_at: datetime | None = None
completed_at: datetime | None = None
class StepType(Enum):
@ -120,8 +110,8 @@ class ToolExecutionStep(StepCommon):
"""
step_type: Literal[StepType.tool_execution.value] = StepType.tool_execution.value
tool_calls: List[ToolCall]
tool_responses: List[ToolResponse]
tool_calls: list[ToolCall]
tool_responses: list[ToolResponse]
@json_schema_type
@ -132,7 +122,7 @@ class ShieldCallStep(StepCommon):
"""
step_type: Literal[StepType.shield_call.value] = StepType.shield_call.value
violation: Optional[SafetyViolation]
violation: SafetyViolation | None
@json_schema_type
@ -150,12 +140,7 @@ class MemoryRetrievalStep(StepCommon):
Step = Annotated[
Union[
InferenceStep,
ToolExecutionStep,
ShieldCallStep,
MemoryRetrievalStep,
],
InferenceStep | ToolExecutionStep | ShieldCallStep | MemoryRetrievalStep,
Field(discriminator="step_type"),
]
@ -166,18 +151,13 @@ class Turn(BaseModel):
turn_id: str
session_id: str
input_messages: List[
Union[
UserMessage,
ToolResponseMessage,
]
]
steps: List[Step]
input_messages: list[UserMessage | ToolResponseMessage]
steps: list[Step]
output_message: CompletionMessage
output_attachments: Optional[List[Attachment]] = Field(default_factory=list)
output_attachments: list[Attachment] | None = Field(default_factory=list)
started_at: datetime
completed_at: Optional[datetime] = None
completed_at: datetime | None = None
@json_schema_type
@ -186,34 +166,31 @@ class Session(BaseModel):
session_id: str
session_name: str
turns: List[Turn]
turns: list[Turn]
started_at: datetime
class AgentToolGroupWithArgs(BaseModel):
name: str
args: Dict[str, Any]
args: dict[str, Any]
AgentToolGroup = Union[
str,
AgentToolGroupWithArgs,
]
AgentToolGroup = str | AgentToolGroupWithArgs
register_schema(AgentToolGroup, name="AgentTool")
class AgentConfigCommon(BaseModel):
sampling_params: Optional[SamplingParams] = Field(default_factory=SamplingParams)
sampling_params: SamplingParams | None = Field(default_factory=SamplingParams)
input_shields: Optional[List[str]] = Field(default_factory=list)
output_shields: Optional[List[str]] = Field(default_factory=list)
toolgroups: Optional[List[AgentToolGroup]] = Field(default_factory=list)
client_tools: Optional[List[ToolDef]] = Field(default_factory=list)
tool_choice: Optional[ToolChoice] = Field(default=None, deprecated="use tool_config instead")
tool_prompt_format: Optional[ToolPromptFormat] = Field(default=None, deprecated="use tool_config instead")
tool_config: Optional[ToolConfig] = Field(default=None)
input_shields: list[str] | None = Field(default_factory=list)
output_shields: list[str] | None = Field(default_factory=list)
toolgroups: list[AgentToolGroup] | None = Field(default_factory=list)
client_tools: list[ToolDef] | None = Field(default_factory=list)
tool_choice: ToolChoice | None = Field(default=None, deprecated="use tool_config instead")
tool_prompt_format: ToolPromptFormat | None = Field(default=None, deprecated="use tool_config instead")
tool_config: ToolConfig | None = Field(default=None)
max_infer_iters: Optional[int] = 10
max_infer_iters: int | None = 10
def model_post_init(self, __context):
if self.tool_config:
@ -243,9 +220,9 @@ class AgentConfig(AgentConfigCommon):
model: str
instructions: str
name: Optional[str] = None
enable_session_persistence: Optional[bool] = False
response_format: Optional[ResponseFormat] = None
name: str | None = None
enable_session_persistence: bool | None = False
response_format: ResponseFormat | None = None
@json_schema_type
@ -257,16 +234,16 @@ class Agent(BaseModel):
@json_schema_type
class ListAgentsResponse(BaseModel):
data: List[Agent]
data: list[Agent]
@json_schema_type
class ListAgentSessionsResponse(BaseModel):
data: List[Session]
data: list[Session]
class AgentConfigOverridablePerTurn(AgentConfigCommon):
instructions: Optional[str] = None
instructions: str | None = None
class AgentTurnResponseEventType(Enum):
@ -284,7 +261,7 @@ class AgentTurnResponseStepStartPayload(BaseModel):
event_type: Literal[AgentTurnResponseEventType.step_start.value] = AgentTurnResponseEventType.step_start.value
step_type: StepType
step_id: str
metadata: Optional[Dict[str, Any]] = Field(default_factory=dict)
metadata: dict[str, Any] | None = Field(default_factory=dict)
@json_schema_type
@ -327,14 +304,12 @@ class AgentTurnResponseTurnAwaitingInputPayload(BaseModel):
AgentTurnResponseEventPayload = Annotated[
Union[
AgentTurnResponseStepStartPayload,
AgentTurnResponseStepProgressPayload,
AgentTurnResponseStepCompletePayload,
AgentTurnResponseTurnStartPayload,
AgentTurnResponseTurnCompletePayload,
AgentTurnResponseTurnAwaitingInputPayload,
],
AgentTurnResponseStepStartPayload
| AgentTurnResponseStepProgressPayload
| AgentTurnResponseStepCompletePayload
| AgentTurnResponseTurnStartPayload
| AgentTurnResponseTurnCompletePayload
| AgentTurnResponseTurnAwaitingInputPayload,
Field(discriminator="event_type"),
]
register_schema(AgentTurnResponseEventPayload, name="AgentTurnResponseEventPayload")
@ -363,18 +338,13 @@ class AgentTurnCreateRequest(AgentConfigOverridablePerTurn):
# TODO: figure out how we can simplify this and make why
# ToolResponseMessage needs to be here (it is function call
# execution from outside the system)
messages: List[
Union[
UserMessage,
ToolResponseMessage,
]
]
messages: list[UserMessage | ToolResponseMessage]
documents: Optional[List[Document]] = None
toolgroups: Optional[List[AgentToolGroup]] = None
documents: list[Document] | None = None
toolgroups: list[AgentToolGroup] | None = None
stream: Optional[bool] = False
tool_config: Optional[ToolConfig] = None
stream: bool | None = False
tool_config: ToolConfig | None = None
@json_schema_type
@ -382,8 +352,8 @@ class AgentTurnResumeRequest(BaseModel):
agent_id: str
session_id: str
turn_id: str
tool_responses: List[ToolResponse]
stream: Optional[bool] = False
tool_responses: list[ToolResponse]
stream: bool | None = False
@json_schema_type
@ -429,17 +399,12 @@ class Agents(Protocol):
self,
agent_id: str,
session_id: str,
messages: List[
Union[
UserMessage,
ToolResponseMessage,
]
],
stream: Optional[bool] = False,
documents: Optional[List[Document]] = None,
toolgroups: Optional[List[AgentToolGroup]] = None,
tool_config: Optional[ToolConfig] = None,
) -> Union[Turn, AsyncIterator[AgentTurnResponseStreamChunk]]:
messages: list[UserMessage | ToolResponseMessage],
stream: bool | None = False,
documents: list[Document] | None = None,
toolgroups: list[AgentToolGroup] | None = None,
tool_config: ToolConfig | None = None,
) -> Turn | AsyncIterator[AgentTurnResponseStreamChunk]:
"""Create a new turn for an agent.
:param agent_id: The ID of the agent to create the turn for.
@ -463,9 +428,9 @@ class Agents(Protocol):
agent_id: str,
session_id: str,
turn_id: str,
tool_responses: List[ToolResponse],
stream: Optional[bool] = False,
) -> Union[Turn, AsyncIterator[AgentTurnResponseStreamChunk]]:
tool_responses: list[ToolResponse],
stream: bool | None = False,
) -> Turn | AsyncIterator[AgentTurnResponseStreamChunk]:
"""Resume an agent turn with executed tool call responses.
When a Turn has the status `awaiting_input` due to pending input from client side tool calls, this endpoint can be used to submit the outputs from the tool calls once they are ready.
@ -538,7 +503,7 @@ class Agents(Protocol):
self,
session_id: str,
agent_id: str,
turn_ids: Optional[List[str]] = None,
turn_ids: list[str] | None = None,
) -> Session:
"""Retrieve an agent session by its ID.
@ -623,13 +588,14 @@ class Agents(Protocol):
@webmethod(route="/openai/v1/responses", method="POST")
async def create_openai_response(
self,
input: Union[str, List[OpenAIResponseInputMessage]],
input: str | list[OpenAIResponseInputMessage],
model: str,
previous_response_id: Optional[str] = None,
store: Optional[bool] = True,
stream: Optional[bool] = False,
tools: Optional[List[OpenAIResponseInputTool]] = None,
) -> Union[OpenAIResponseObject, AsyncIterator[OpenAIResponseObjectStream]]:
previous_response_id: str | None = None,
store: bool | None = True,
stream: bool | None = False,
temperature: float | None = None,
tools: list[OpenAIResponseInputTool] | None = None,
) -> OpenAIResponseObject | AsyncIterator[OpenAIResponseObjectStream]:
"""Create a new OpenAI response.
:param input: Input message(s) to create the response.

43
llama_stack/apis/agents/openai_responses.py
View file

@ -4,10 +4,9 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List, Literal, Optional, Union
from typing import Annotated, Literal
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.schema_utils import json_schema_type, register_schema
@ -25,7 +24,7 @@ class OpenAIResponseOutputMessageContentOutputText(BaseModel):
OpenAIResponseOutputMessageContent = Annotated[
Union[OpenAIResponseOutputMessageContentOutputText,],
OpenAIResponseOutputMessageContentOutputText,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutputMessageContent, name="OpenAIResponseOutputMessageContent")
@ -34,7 +33,7 @@ register_schema(OpenAIResponseOutputMessageContent, name="OpenAIResponseOutputMe
@json_schema_type
class OpenAIResponseOutputMessage(BaseModel):
id: str
content: List[OpenAIResponseOutputMessageContent]
content: list[OpenAIResponseOutputMessageContent]
role: Literal["assistant"] = "assistant"
status: str
type: Literal["message"] = "message"
@ -48,10 +47,7 @@ class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
OpenAIResponseOutput = Annotated[
Union[
OpenAIResponseOutputMessage,
OpenAIResponseOutputMessageWebSearchToolCall,
],
OpenAIResponseOutputMessage | OpenAIResponseOutputMessageWebSearchToolCall,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
@ -60,18 +56,18 @@ register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
@json_schema_type
class OpenAIResponseObject(BaseModel):
created_at: int
error: Optional[OpenAIResponseError] = None
error: OpenAIResponseError | None = None
id: str
model: str
object: Literal["response"] = "response"
output: List[OpenAIResponseOutput]
output: list[OpenAIResponseOutput]
parallel_tool_calls: bool = False
previous_response_id: Optional[str] = None
previous_response_id: str | None = None
status: str
temperature: Optional[float] = None
top_p: Optional[float] = None
truncation: Optional[str] = None
user: Optional[str] = None
temperature: float | None = None
top_p: float | None = None
truncation: str | None = None
user: str | None = None
@json_schema_type
@ -87,10 +83,7 @@ class OpenAIResponseObjectStreamResponseCompleted(BaseModel):
OpenAIResponseObjectStream = Annotated[
Union[
OpenAIResponseObjectStreamResponseCreated,
OpenAIResponseObjectStreamResponseCompleted,
],
OpenAIResponseObjectStreamResponseCreated | OpenAIResponseObjectStreamResponseCompleted,
Field(discriminator="type"),
]
register_schema(OpenAIResponseObjectStream, name="OpenAIResponseObjectStream")
@ -107,12 +100,12 @@ class OpenAIResponseInputMessageContentImage(BaseModel):
detail: Literal["low"] | Literal["high"] | Literal["auto"] = "auto"
type: Literal["input_image"] = "input_image"
# TODO: handle file_id
image_url: Optional[str] = None
image_url: str | None = None
# TODO: handle file content types
OpenAIResponseInputMessageContent = Annotated[
Union[OpenAIResponseInputMessageContentText, OpenAIResponseInputMessageContentImage],
OpenAIResponseInputMessageContentText | OpenAIResponseInputMessageContentImage,
Field(discriminator="type"),
]
register_schema(OpenAIResponseInputMessageContent, name="OpenAIResponseInputMessageContent")
@ -120,21 +113,21 @@ register_schema(OpenAIResponseInputMessageContent, name="OpenAIResponseInputMess
@json_schema_type
class OpenAIResponseInputMessage(BaseModel):
content: Union[str, List[OpenAIResponseInputMessageContent]]
content: str | list[OpenAIResponseInputMessageContent]
role: Literal["system"] | Literal["developer"] | Literal["user"] | Literal["assistant"]
type: Optional[Literal["message"]] = "message"
type: Literal["message"] | None = "message"
@json_schema_type
class OpenAIResponseInputToolWebSearch(BaseModel):
type: Literal["web_search"] | Literal["web_search_preview_2025_03_11"] = "web_search"
# TODO: actually use search_context_size somewhere...
search_context_size: Optional[str] = Field(default="medium", pattern="^low|medium|high$")
search_context_size: str | None = Field(default="medium", pattern="^low|medium|high$")
# TODO: add user_location
OpenAIResponseInputTool = Annotated[
Union[OpenAIResponseInputToolWebSearch,],
OpenAIResponseInputToolWebSearch,
Field(discriminator="type"),
]
register_schema(OpenAIResponseInputTool, name="OpenAIResponseInputTool")

24
llama_stack/apis/batch_inference/batch_inference.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List, Optional, Protocol, runtime_checkable
from typing import Protocol, runtime_checkable
from llama_stack.apis.common.job_types import Job
from llama_stack.apis.inference import (
@ -34,22 +34,22 @@ class BatchInference(Protocol):
async def completion(
self,
model: str,
content_batch: List[InterleavedContent],
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
content_batch: list[InterleavedContent],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> Job: ...
@webmethod(route="/batch-inference/chat-completion", method="POST")
async def chat_completion(
self,
model: str,
messages_batch: List[List[Message]],
sampling_params: Optional[SamplingParams] = None,
messages_batch: list[list[Message]],
sampling_params: SamplingParams | None = None,
# zero-shot tool definitions as input to the model
tools: Optional[List[ToolDefinition]] = None,
tool_choice: Optional[ToolChoice] = ToolChoice.auto,
tool_prompt_format: Optional[ToolPromptFormat] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
tools: list[ToolDefinition] | None = None,
tool_choice: ToolChoice | None = ToolChoice.auto,
tool_prompt_format: ToolPromptFormat | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> Job: ...

20
llama_stack/apis/benchmarks/benchmarks.py
View file

@ -3,7 +3,7 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Literal, Optional, Protocol, runtime_checkable
from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, Field
@ -13,8 +13,8 @@ from llama_stack.schema_utils import json_schema_type, webmethod
class CommonBenchmarkFields(BaseModel):
dataset_id: str
scoring_functions: List[str]
metadata: Dict[str, Any] = Field(
scoring_functions: list[str]
metadata: dict[str, Any] = Field(
default_factory=dict,
description="Metadata for this evaluation task",
)
@ -35,12 +35,12 @@ class Benchmark(CommonBenchmarkFields, Resource):
class BenchmarkInput(CommonBenchmarkFields, BaseModel):
benchmark_id: str
provider_id: Optional[str] = None
provider_benchmark_id: Optional[str] = None
provider_id: str | None = None
provider_benchmark_id: str | None = None
class ListBenchmarksResponse(BaseModel):
data: List[Benchmark]
data: list[Benchmark]
@runtime_checkable
@ -59,8 +59,8 @@ class Benchmarks(Protocol):
self,
benchmark_id: str,
dataset_id: str,
scoring_functions: List[str],
provider_benchmark_id: Optional[str] = None,
provider_id: Optional[str] = None,
metadata: Optional[Dict[str, Any]] = None,
scoring_functions: list[str],
provider_benchmark_id: str | None = None,
provider_id: str | None = None,
metadata: dict[str, Any] | None = None,
) -> None: ...

14
llama_stack/apis/common/content_types.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Annotated, List, Literal, Optional, Union
from typing import Annotated, Literal
from pydantic import BaseModel, Field, model_validator
@ -26,9 +26,9 @@ class _URLOrData(BaseModel):
:param data: base64 encoded image data as string
"""
url: Optional[URL] = None
url: URL | None = None
# data is a base64 encoded string, hint with contentEncoding=base64
data: Optional[str] = Field(contentEncoding="base64", default=None)
data: str | None = Field(contentEncoding="base64", default=None)
@model_validator(mode="before")
@classmethod
@ -64,13 +64,13 @@ class TextContentItem(BaseModel):
# other modalities can be added here
InterleavedContentItem = Annotated[
Union[ImageContentItem, TextContentItem],
ImageContentItem | TextContentItem,
Field(discriminator="type"),
]
register_schema(InterleavedContentItem, name="InterleavedContentItem")
# accept a single "str" as a special case since it is common
InterleavedContent = Union[str, InterleavedContentItem, List[InterleavedContentItem]]
InterleavedContent = str | InterleavedContentItem | list[InterleavedContentItem]
register_schema(InterleavedContent, name="InterleavedContent")
@ -100,13 +100,13 @@ class ToolCallDelta(BaseModel):
# you either send an in-progress tool call so the client can stream a long
# code generation or you send the final parsed tool call at the end of the
# stream
tool_call: Union[str, ToolCall]
tool_call: str | ToolCall
parse_status: ToolCallParseStatus
# streaming completions send a stream of ContentDeltas
ContentDelta = Annotated[
Union[TextDelta, ImageDelta, ToolCallDelta],
TextDelta | ImageDelta | ToolCallDelta,
Field(discriminator="type"),
]
register_schema(ContentDelta, name="ContentDelta")

8
llama_stack/apis/common/deployment_types.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, Optional
from typing import Any
from pydantic import BaseModel
@ -25,6 +25,6 @@ class RestAPIMethod(Enum):
class RestAPIExecutionConfig(BaseModel):
url: URL
method: RestAPIMethod
params: Optional[Dict[str, Any]] = None
headers: Optional[Dict[str, Any]] = None
body: Optional[Dict[str, Any]] = None
params: dict[str, Any] | None = None
headers: dict[str, Any] | None = None
body: dict[str, Any] | None = None

4
llama_stack/apis/common/responses.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List
from typing import Any
from pydantic import BaseModel
@ -19,5 +19,5 @@ class PaginatedResponse(BaseModel):
:param has_more: Whether there are more items available after this set
"""
data: List[Dict[str, Any]]
data: list[dict[str, Any]]
has_more: bool

3
llama_stack/apis/common/training_types.py
View file

@ -5,7 +5,6 @@
# the root directory of this source tree.
from datetime import datetime
from typing import Optional
from pydantic import BaseModel
@ -27,4 +26,4 @@ class Checkpoint(BaseModel):
epoch: int
post_training_job_id: str
path: str
training_metrics: Optional[PostTrainingMetric] = None
training_metrics: PostTrainingMetric | None = None

25
llama_stack/apis/common/type_system.py
View file

@ -4,10 +4,9 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Literal, Union
from typing import Annotated, Literal
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.schema_utils import json_schema_type, register_schema
@ -73,18 +72,16 @@ class DialogType(BaseModel):
ParamType = Annotated[
Union[
StringType,
NumberType,
BooleanType,
ArrayType,
ObjectType,
JsonType,
UnionType,
ChatCompletionInputType,
CompletionInputType,
AgentTurnInputType,
],
StringType
| NumberType
| BooleanType
| ArrayType
| ObjectType
| JsonType
| UnionType
| ChatCompletionInputType
| CompletionInputType
| AgentTurnInputType,
Field(discriminator="type"),
]
register_schema(ParamType, name="ParamType")

8
llama_stack/apis/datasetio/datasetio.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
from typing import Any, Protocol, runtime_checkable
from llama_stack.apis.common.responses import PaginatedResponse
from llama_stack.apis.datasets import Dataset
@ -24,8 +24,8 @@ class DatasetIO(Protocol):
async def iterrows(
self,
dataset_id: str,
start_index: Optional[int] = None,
limit: Optional[int] = None,
start_index: int | None = None,
limit: int | None = None,
) -> PaginatedResponse:
"""Get a paginated list of rows from a dataset.
@ -44,4 +44,4 @@ class DatasetIO(Protocol):
...
@webmethod(route="/datasetio/append-rows/{dataset_id:path}", method="POST")
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None: ...
async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None: ...

14
llama_stack/apis/datasets/datasets.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Annotated, Any, Dict, List, Literal, Optional, Protocol, Union
from typing import Annotated, Any, Literal, Protocol
from pydantic import BaseModel, Field
@ -81,11 +81,11 @@ class RowsDataSource(BaseModel):
"""
type: Literal["rows"] = "rows"
rows: List[Dict[str, Any]]
rows: list[dict[str, Any]]
DataSource = Annotated[
Union[URIDataSource, RowsDataSource],
URIDataSource | RowsDataSource,
Field(discriminator="type"),
]
register_schema(DataSource, name="DataSource")
@ -98,7 +98,7 @@ class CommonDatasetFields(BaseModel):
purpose: DatasetPurpose
source: DataSource
metadata: Dict[str, Any] = Field(
metadata: dict[str, Any] = Field(
default_factory=dict,
description="Any additional metadata for this dataset",
)
@ -122,7 +122,7 @@ class DatasetInput(CommonDatasetFields, BaseModel):
class ListDatasetsResponse(BaseModel):
data: List[Dataset]
data: list[Dataset]
class Datasets(Protocol):
@ -131,8 +131,8 @@ class Datasets(Protocol):
self,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
dataset_id: Optional[str] = None,
metadata: dict[str, Any] | None = None,
dataset_id: str | None = None,
) -> Dataset:
"""
Register a new dataset.

3
llama_stack/apis/datatypes.py
View file

@ -5,7 +5,6 @@
# the root directory of this source tree.
from enum import Enum
from typing import Optional
from pydantic import BaseModel
@ -54,4 +53,4 @@ class Error(BaseModel):
status: int
title: str
detail: str
instance: Optional[str] = None
instance: str | None = None

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

@ -4,10 +4,9 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Literal, Optional, Protocol, Union
from typing import Annotated, Any, Literal, Protocol
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.apis.agents import AgentConfig
from llama_stack.apis.common.job_types import Job
@ -29,7 +28,7 @@ class ModelCandidate(BaseModel):
type: Literal["model"] = "model"
model: str
sampling_params: SamplingParams
system_message: Optional[SystemMessage] = None
system_message: SystemMessage | None = None
@json_schema_type
@ -43,7 +42,7 @@ class AgentCandidate(BaseModel):
config: AgentConfig
EvalCandidate = Annotated[Union[ModelCandidate, AgentCandidate], Field(discriminator="type")]
EvalCandidate = Annotated[ModelCandidate | AgentCandidate, Field(discriminator="type")]
register_schema(EvalCandidate, name="EvalCandidate")
@ -57,11 +56,11 @@ class BenchmarkConfig(BaseModel):
"""
eval_candidate: EvalCandidate
scoring_params: Dict[str, ScoringFnParams] = Field(
scoring_params: dict[str, ScoringFnParams] = Field(
description="Map between scoring function id and parameters for each scoring function you want to run",
default_factory=dict,
)
num_examples: Optional[int] = Field(
num_examples: int | None = Field(
description="Number of examples to evaluate (useful for testing), if not provided, all examples in the dataset will be evaluated",
default=None,
)
@ -76,9 +75,9 @@ class EvaluateResponse(BaseModel):
:param scores: The scores from the evaluation.
"""
generations: List[Dict[str, Any]]
generations: list[dict[str, Any]]
# each key in the dict is a scoring function name
scores: Dict[str, ScoringResult]
scores: dict[str, ScoringResult]
class Eval(Protocol):
@ -101,8 +100,8 @@ class Eval(Protocol):
async def evaluate_rows(
self,
benchmark_id: str,
input_rows: List[Dict[str, Any]],
scoring_functions: List[str],
input_rows: list[dict[str, Any]],
scoring_functions: list[str],
benchmark_config: BenchmarkConfig,
) -> EvaluateResponse:
"""Evaluate a list of rows on a benchmark.

8
llama_stack/apis/files/files.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List, Optional, Protocol, runtime_checkable
from typing import Protocol, runtime_checkable
from pydantic import BaseModel
@ -42,7 +42,7 @@ class ListBucketResponse(BaseModel):
:param data: List of FileResponse entries
"""
data: List[BucketResponse]
data: list[BucketResponse]
@json_schema_type
@ -74,7 +74,7 @@ class ListFileResponse(BaseModel):
:param data: List of FileResponse entries
"""
data: List[FileResponse]
data: list[FileResponse]
@runtime_checkable
@ -102,7 +102,7 @@ class Files(Protocol):
async def upload_content_to_session(
self,
upload_id: str,
) -> Optional[FileResponse]:
) -> FileResponse | None:
"""
Upload file content to an existing upload session.
On the server, request body will have the raw bytes that are uploaded.

329
llama_stack/apis/inference/inference.py
View file

@ -4,21 +4,18 @@
# 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 AsyncIterator
from enum import Enum
from typing import (
Annotated,
Any,
AsyncIterator,
Dict,
List,
Literal,
Optional,
Protocol,
Union,
runtime_checkable,
)
from pydantic import BaseModel, Field, field_validator
from typing_extensions import Annotated, TypedDict
from typing_extensions import TypedDict
from llama_stack.apis.common.content_types import ContentDelta, InterleavedContent, InterleavedContentItem
from llama_stack.apis.models import Model
@ -47,8 +44,8 @@ class GreedySamplingStrategy(BaseModel):
@json_schema_type
class TopPSamplingStrategy(BaseModel):
type: Literal["top_p"] = "top_p"
temperature: Optional[float] = Field(..., gt=0.0)
top_p: Optional[float] = 0.95
temperature: float | None = Field(..., gt=0.0)
top_p: float | None = 0.95
@json_schema_type
@ -58,7 +55,7 @@ class TopKSamplingStrategy(BaseModel):
SamplingStrategy = Annotated[
Union[GreedySamplingStrategy, TopPSamplingStrategy, TopKSamplingStrategy],
GreedySamplingStrategy | TopPSamplingStrategy | TopKSamplingStrategy,
Field(discriminator="type"),
]
register_schema(SamplingStrategy, name="SamplingStrategy")
@ -79,9 +76,9 @@ class SamplingParams(BaseModel):
strategy: SamplingStrategy = Field(default_factory=GreedySamplingStrategy)
max_tokens: Optional[int] = 0
repetition_penalty: Optional[float] = 1.0
stop: Optional[List[str]] = None
max_tokens: int | None = 0
repetition_penalty: float | None = 1.0
stop: list[str] | None = None
class LogProbConfig(BaseModel):
@ -90,7 +87,7 @@ class LogProbConfig(BaseModel):
:param top_k: How many tokens (for each position) to return log probabilities for.
"""
top_k: Optional[int] = 0
top_k: int | None = 0
class QuantizationType(Enum):
@ -125,11 +122,11 @@ class Int4QuantizationConfig(BaseModel):
"""
type: Literal["int4_mixed"] = "int4_mixed"
scheme: Optional[str] = "int4_weight_int8_dynamic_activation"
scheme: str | None = "int4_weight_int8_dynamic_activation"
QuantizationConfig = Annotated[
Union[Bf16QuantizationConfig, Fp8QuantizationConfig, Int4QuantizationConfig],
Bf16QuantizationConfig | Fp8QuantizationConfig | Int4QuantizationConfig,
Field(discriminator="type"),
]
@ -145,7 +142,7 @@ class UserMessage(BaseModel):
role: Literal["user"] = "user"
content: InterleavedContent
context: Optional[InterleavedContent] = None
context: InterleavedContent | None = None
@json_schema_type
@ -190,16 +187,11 @@ class CompletionMessage(BaseModel):
role: Literal["assistant"] = "assistant"
content: InterleavedContent
stop_reason: StopReason
tool_calls: Optional[List[ToolCall]] = Field(default_factory=list)
tool_calls: list[ToolCall] | None = Field(default_factory=list)
Message = Annotated[
Union[
UserMessage,
SystemMessage,
ToolResponseMessage,
CompletionMessage,
],
UserMessage | SystemMessage | ToolResponseMessage | CompletionMessage,
Field(discriminator="role"),
]
register_schema(Message, name="Message")
@ -208,9 +200,9 @@ register_schema(Message, name="Message")
@json_schema_type
class ToolResponse(BaseModel):
call_id: str
tool_name: Union[BuiltinTool, str]
tool_name: BuiltinTool | str
content: InterleavedContent
metadata: Optional[Dict[str, Any]] = None
metadata: dict[str, Any] | None = None
@field_validator("tool_name", mode="before")
@classmethod
@ -243,7 +235,7 @@ class TokenLogProbs(BaseModel):
:param logprobs_by_token: Dictionary mapping tokens to their log probabilities
"""
logprobs_by_token: Dict[str, float]
logprobs_by_token: dict[str, float]
class ChatCompletionResponseEventType(Enum):
@ -271,8 +263,8 @@ class ChatCompletionResponseEvent(BaseModel):
event_type: ChatCompletionResponseEventType
delta: ContentDelta
logprobs: Optional[List[TokenLogProbs]] = None
stop_reason: Optional[StopReason] = None
logprobs: list[TokenLogProbs] | None = None
stop_reason: StopReason | None = None
class ResponseFormatType(Enum):
@ -295,7 +287,7 @@ class JsonSchemaResponseFormat(BaseModel):
"""
type: Literal[ResponseFormatType.json_schema.value] = ResponseFormatType.json_schema.value
json_schema: Dict[str, Any]
json_schema: dict[str, Any]
@json_schema_type
@ -307,11 +299,11 @@ class GrammarResponseFormat(BaseModel):
"""
type: Literal[ResponseFormatType.grammar.value] = ResponseFormatType.grammar.value
bnf: Dict[str, Any]
bnf: dict[str, Any]
ResponseFormat = Annotated[
Union[JsonSchemaResponseFormat, GrammarResponseFormat],
JsonSchemaResponseFormat | GrammarResponseFormat,
Field(discriminator="type"),
]
register_schema(ResponseFormat, name="ResponseFormat")
@ -321,10 +313,10 @@ register_schema(ResponseFormat, name="ResponseFormat")
class CompletionRequest(BaseModel):
model: str
content: InterleavedContent
sampling_params: Optional[SamplingParams] = Field(default_factory=SamplingParams)
response_format: Optional[ResponseFormat] = None
stream: Optional[bool] = False
logprobs: Optional[LogProbConfig] = None
sampling_params: SamplingParams | None = Field(default_factory=SamplingParams)
response_format: ResponseFormat | None = None
stream: bool | None = False
logprobs: LogProbConfig | None = None
@json_schema_type
@ -338,7 +330,7 @@ class CompletionResponse(MetricResponseMixin):
content: str
stop_reason: StopReason
logprobs: Optional[List[TokenLogProbs]] = None
logprobs: list[TokenLogProbs] | None = None
@json_schema_type
@ -351,8 +343,8 @@ class CompletionResponseStreamChunk(MetricResponseMixin):
"""
delta: str
stop_reason: Optional[StopReason] = None
logprobs: Optional[List[TokenLogProbs]] = None
stop_reason: StopReason | None = None
logprobs: list[TokenLogProbs] | None = None
class SystemMessageBehavior(Enum):
@ -383,9 +375,9 @@ class ToolConfig(BaseModel):
'{{function_definitions}}' to indicate where the function definitions should be inserted.
"""
tool_choice: Optional[ToolChoice | str] = Field(default=ToolChoice.auto)
tool_prompt_format: Optional[ToolPromptFormat] = Field(default=None)
system_message_behavior: Optional[SystemMessageBehavior] = Field(default=SystemMessageBehavior.append)
tool_choice: ToolChoice | str | None = Field(default=ToolChoice.auto)
tool_prompt_format: ToolPromptFormat | None = Field(default=None)
system_message_behavior: SystemMessageBehavior | None = Field(default=SystemMessageBehavior.append)
def model_post_init(self, __context: Any) -> None:
if isinstance(self.tool_choice, str):
@ -399,15 +391,15 @@ class ToolConfig(BaseModel):
@json_schema_type
class ChatCompletionRequest(BaseModel):
model: str
messages: List[Message]
sampling_params: Optional[SamplingParams] = Field(default_factory=SamplingParams)
messages: list[Message]
sampling_params: SamplingParams | None = Field(default_factory=SamplingParams)
tools: Optional[List[ToolDefinition]] = Field(default_factory=list)
tool_config: Optional[ToolConfig] = Field(default_factory=ToolConfig)
tools: list[ToolDefinition] | None = Field(default_factory=list)
tool_config: ToolConfig | None = Field(default_factory=ToolConfig)
response_format: Optional[ResponseFormat] = None
stream: Optional[bool] = False
logprobs: Optional[LogProbConfig] = None
response_format: ResponseFormat | None = None
stream: bool | None = False
logprobs: LogProbConfig | None = None
@json_schema_type
@ -429,7 +421,7 @@ class ChatCompletionResponse(MetricResponseMixin):
"""
completion_message: CompletionMessage
logprobs: Optional[List[TokenLogProbs]] = None
logprobs: list[TokenLogProbs] | None = None
@json_schema_type
@ -439,7 +431,7 @@ class EmbeddingsResponse(BaseModel):
:param embeddings: List of embedding vectors, one per input content. Each embedding is a list of floats. The dimensionality of the embedding is model-specific; you can check model metadata using /models/{model_id}
"""
embeddings: List[List[float]]
embeddings: list[list[float]]
@json_schema_type
@ -451,7 +443,7 @@ class OpenAIChatCompletionContentPartTextParam(BaseModel):
@json_schema_type
class OpenAIImageURL(BaseModel):
url: str
detail: Optional[str] = None
detail: str | None = None
@json_schema_type
@ -461,16 +453,13 @@ class OpenAIChatCompletionContentPartImageParam(BaseModel):
OpenAIChatCompletionContentPartParam = Annotated[
Union[
OpenAIChatCompletionContentPartTextParam,
OpenAIChatCompletionContentPartImageParam,
],
OpenAIChatCompletionContentPartTextParam | OpenAIChatCompletionContentPartImageParam,
Field(discriminator="type"),
]
register_schema(OpenAIChatCompletionContentPartParam, name="OpenAIChatCompletionContentPartParam")
OpenAIChatCompletionMessageContent = Union[str, List[OpenAIChatCompletionContentPartParam]]
OpenAIChatCompletionMessageContent = str | list[OpenAIChatCompletionContentPartParam]
@json_schema_type
@ -484,7 +473,7 @@ class OpenAIUserMessageParam(BaseModel):
role: Literal["user"] = "user"
content: OpenAIChatCompletionMessageContent
name: Optional[str] = None
name: str | None = None
@json_schema_type
@ -498,21 +487,21 @@ class OpenAISystemMessageParam(BaseModel):
role: Literal["system"] = "system"
content: OpenAIChatCompletionMessageContent
name: Optional[str] = None
name: str | None = None
@json_schema_type
class OpenAIChatCompletionToolCallFunction(BaseModel):
name: Optional[str] = None
arguments: Optional[str] = None
name: str | None = None
arguments: str | None = None
@json_schema_type
class OpenAIChatCompletionToolCall(BaseModel):
index: Optional[int] = None
id: Optional[str] = None
index: int | None = None
id: str | None = None
type: Literal["function"] = "function"
function: Optional[OpenAIChatCompletionToolCallFunction] = None
function: OpenAIChatCompletionToolCallFunction | None = None
@json_schema_type
@ -526,9 +515,9 @@ class OpenAIAssistantMessageParam(BaseModel):
"""
role: Literal["assistant"] = "assistant"
content: Optional[OpenAIChatCompletionMessageContent] = None
name: Optional[str] = None
tool_calls: Optional[List[OpenAIChatCompletionToolCall]] = None
content: OpenAIChatCompletionMessageContent | None = None
name: str | None = None
tool_calls: list[OpenAIChatCompletionToolCall] | None = None
@json_schema_type
@ -556,17 +545,15 @@ class OpenAIDeveloperMessageParam(BaseModel):
role: Literal["developer"] = "developer"
content: OpenAIChatCompletionMessageContent
name: Optional[str] = None
name: str | None = None
OpenAIMessageParam = Annotated[
Union[
OpenAIUserMessageParam,
OpenAISystemMessageParam,
OpenAIAssistantMessageParam,
OpenAIToolMessageParam,
OpenAIDeveloperMessageParam,
],
OpenAIUserMessageParam
| OpenAISystemMessageParam
| OpenAIAssistantMessageParam
| OpenAIToolMessageParam
| OpenAIDeveloperMessageParam,
Field(discriminator="role"),
]
register_schema(OpenAIMessageParam, name="OpenAIMessageParam")
@ -580,14 +567,14 @@ class OpenAIResponseFormatText(BaseModel):
@json_schema_type
class OpenAIJSONSchema(TypedDict, total=False):
name: str
description: Optional[str] = None
strict: Optional[bool] = None
description: str | None = None
strict: bool | None = None
# Pydantic BaseModel cannot be used with a schema param, since it already
# has one. And, we don't want to alias here because then have to handle
# that alias when converting to OpenAI params. So, to support schema,
# we use a TypedDict.
schema: Optional[Dict[str, Any]] = None
schema: dict[str, Any] | None = None
@json_schema_type
@ -602,11 +589,7 @@ class OpenAIResponseFormatJSONObject(BaseModel):
OpenAIResponseFormatParam = Annotated[
Union[
OpenAIResponseFormatText,
OpenAIResponseFormatJSONSchema,
OpenAIResponseFormatJSONObject,
],
OpenAIResponseFormatText | OpenAIResponseFormatJSONSchema | OpenAIResponseFormatJSONObject,
Field(discriminator="type"),
]
register_schema(OpenAIResponseFormatParam, name="OpenAIResponseFormatParam")
@ -622,7 +605,7 @@ class OpenAITopLogProb(BaseModel):
"""
token: str
bytes: Optional[List[int]] = None
bytes: list[int] | None = None
logprob: float
@ -637,9 +620,9 @@ class OpenAITokenLogProb(BaseModel):
"""
token: str
bytes: Optional[List[int]] = None
bytes: list[int] | None = None
logprob: float
top_logprobs: List[OpenAITopLogProb]
top_logprobs: list[OpenAITopLogProb]
@json_schema_type
@ -650,8 +633,8 @@ class OpenAIChoiceLogprobs(BaseModel):
:param refusal: (Optional) The log probabilities for the tokens in the message
"""
content: Optional[List[OpenAITokenLogProb]] = None
refusal: Optional[List[OpenAITokenLogProb]] = None
content: list[OpenAITokenLogProb] | None = None
refusal: list[OpenAITokenLogProb] | None = None
@json_schema_type
@ -664,10 +647,10 @@ class OpenAIChoiceDelta(BaseModel):
:param tool_calls: (Optional) The tool calls of the delta
"""
content: Optional[str] = None
refusal: Optional[str] = None
role: Optional[str] = None
tool_calls: Optional[List[OpenAIChatCompletionToolCall]] = None
content: str | None = None
refusal: str | None = None
role: str | None = None
tool_calls: list[OpenAIChatCompletionToolCall] | None = None
@json_schema_type
@ -683,7 +666,7 @@ class OpenAIChunkChoice(BaseModel):
delta: OpenAIChoiceDelta
finish_reason: str
index: int
logprobs: Optional[OpenAIChoiceLogprobs] = None
logprobs: OpenAIChoiceLogprobs | None = None
@json_schema_type
@ -699,7 +682,7 @@ class OpenAIChoice(BaseModel):
message: OpenAIMessageParam
finish_reason: str
index: int
logprobs: Optional[OpenAIChoiceLogprobs] = None
logprobs: OpenAIChoiceLogprobs | None = None
@json_schema_type
@ -714,7 +697,7 @@ class OpenAIChatCompletion(BaseModel):
"""
id: str
choices: List[OpenAIChoice]
choices: list[OpenAIChoice]
object: Literal["chat.completion"] = "chat.completion"
created: int
model: str
@ -732,7 +715,7 @@ class OpenAIChatCompletionChunk(BaseModel):
"""
id: str
choices: List[OpenAIChunkChoice]
choices: list[OpenAIChunkChoice]
object: Literal["chat.completion.chunk"] = "chat.completion.chunk"
created: int
model: str
@ -748,10 +731,10 @@ class OpenAICompletionLogprobs(BaseModel):
:top_logprobs: (Optional) The top log probabilities for the tokens
"""
text_offset: Optional[List[int]] = None
token_logprobs: Optional[List[float]] = None
tokens: Optional[List[str]] = None
top_logprobs: Optional[List[Dict[str, float]]] = None
text_offset: list[int] | None = None
token_logprobs: list[float] | None = None
tokens: list[str] | None = None
top_logprobs: list[dict[str, float]] | None = None
@json_schema_type
@ -767,7 +750,7 @@ class OpenAICompletionChoice(BaseModel):
finish_reason: str
text: str
index: int
logprobs: Optional[OpenAIChoiceLogprobs] = None
logprobs: OpenAIChoiceLogprobs | None = None
@json_schema_type
@ -782,7 +765,7 @@ class OpenAICompletion(BaseModel):
"""
id: str
choices: List[OpenAICompletionChoice]
choices: list[OpenAICompletionChoice]
created: int
model: str
object: Literal["text_completion"] = "text_completion"
@ -818,12 +801,12 @@ class EmbeddingTaskType(Enum):
@json_schema_type
class BatchCompletionResponse(BaseModel):
batch: List[CompletionResponse]
batch: list[CompletionResponse]
@json_schema_type
class BatchChatCompletionResponse(BaseModel):
batch: List[ChatCompletionResponse]
batch: list[ChatCompletionResponse]
@runtime_checkable
@ -843,11 +826,11 @@ class Inference(Protocol):
self,
model_id: str,
content: InterleavedContent,
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
) -> Union[CompletionResponse, AsyncIterator[CompletionResponseStreamChunk]]:
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
) -> CompletionResponse | AsyncIterator[CompletionResponseStreamChunk]:
"""Generate a completion for the given content using the specified model.
:param model_id: The identifier of the model to use. The model must be registered with Llama Stack and available via the /models endpoint.
@ -865,10 +848,10 @@ class Inference(Protocol):
async def batch_completion(
self,
model_id: str,
content_batch: List[InterleavedContent],
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
content_batch: list[InterleavedContent],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> BatchCompletionResponse:
raise NotImplementedError("Batch completion is not implemented")
@ -876,16 +859,16 @@ class Inference(Protocol):
async def chat_completion(
self,
model_id: str,
messages: List[Message],
sampling_params: Optional[SamplingParams] = None,
tools: Optional[List[ToolDefinition]] = None,
tool_choice: Optional[ToolChoice] = ToolChoice.auto,
tool_prompt_format: Optional[ToolPromptFormat] = None,
response_format: Optional[ResponseFormat] = None,
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
tool_config: Optional[ToolConfig] = None,
) -> Union[ChatCompletionResponse, AsyncIterator[ChatCompletionResponseStreamChunk]]:
messages: list[Message],
sampling_params: SamplingParams | None = None,
tools: list[ToolDefinition] | None = None,
tool_choice: ToolChoice | None = ToolChoice.auto,
tool_prompt_format: ToolPromptFormat | None = None,
response_format: ResponseFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
tool_config: ToolConfig | None = None,
) -> ChatCompletionResponse | AsyncIterator[ChatCompletionResponseStreamChunk]:
"""Generate a chat completion for the given messages using the specified model.
:param model_id: The identifier of the model to use. The model must be registered with Llama Stack and available via the /models endpoint.
@ -916,12 +899,12 @@ class Inference(Protocol):
async def batch_chat_completion(
self,
model_id: str,
messages_batch: List[List[Message]],
sampling_params: Optional[SamplingParams] = None,
tools: Optional[List[ToolDefinition]] = None,
tool_config: Optional[ToolConfig] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
messages_batch: list[list[Message]],
sampling_params: SamplingParams | None = None,
tools: list[ToolDefinition] | None = None,
tool_config: ToolConfig | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> BatchChatCompletionResponse:
raise NotImplementedError("Batch chat completion is not implemented")
@ -929,10 +912,10 @@ class Inference(Protocol):
async def embeddings(
self,
model_id: str,
contents: List[str] | List[InterleavedContentItem],
text_truncation: Optional[TextTruncation] = TextTruncation.none,
output_dimension: Optional[int] = None,
task_type: Optional[EmbeddingTaskType] = None,
contents: list[str] | list[InterleavedContentItem],
text_truncation: TextTruncation | None = TextTruncation.none,
output_dimension: int | None = None,
task_type: EmbeddingTaskType | None = None,
) -> EmbeddingsResponse:
"""Generate embeddings for content pieces using the specified model.
@ -950,25 +933,25 @@ class Inference(Protocol):
self,
# Standard OpenAI completion parameters
model: str,
prompt: Union[str, List[str], List[int], List[List[int]]],
best_of: Optional[int] = None,
echo: Optional[bool] = None,
frequency_penalty: Optional[float] = None,
logit_bias: Optional[Dict[str, float]] = None,
logprobs: Optional[bool] = None,
max_tokens: Optional[int] = None,
n: Optional[int] = None,
presence_penalty: Optional[float] = None,
seed: Optional[int] = None,
stop: Optional[Union[str, List[str]]] = None,
stream: Optional[bool] = None,
stream_options: Optional[Dict[str, Any]] = None,
temperature: Optional[float] = None,
top_p: Optional[float] = None,
user: Optional[str] = None,
prompt: str | list[str] | list[int] | list[list[int]],
best_of: int | None = None,
echo: bool | None = None,
frequency_penalty: float | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_tokens: int | None = None,
n: int | None = None,
presence_penalty: float | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
top_p: float | None = None,
user: str | None = None,
# vLLM-specific parameters
guided_choice: Optional[List[str]] = None,
prompt_logprobs: Optional[int] = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
) -> OpenAICompletion:
"""Generate an OpenAI-compatible completion for the given prompt using the specified model.
@ -996,29 +979,29 @@ class Inference(Protocol):
async def openai_chat_completion(
self,
model: str,
messages: List[OpenAIMessageParam],
frequency_penalty: Optional[float] = None,
function_call: Optional[Union[str, Dict[str, Any]]] = None,
functions: Optional[List[Dict[str, Any]]] = None,
logit_bias: Optional[Dict[str, float]] = None,
logprobs: Optional[bool] = None,
max_completion_tokens: Optional[int] = None,
max_tokens: Optional[int] = None,
n: Optional[int] = None,
parallel_tool_calls: Optional[bool] = None,
presence_penalty: Optional[float] = None,
response_format: Optional[OpenAIResponseFormatParam] = None,
seed: Optional[int] = None,
stop: Optional[Union[str, List[str]]] = None,
stream: Optional[bool] = None,
stream_options: Optional[Dict[str, Any]] = None,
temperature: Optional[float] = None,
tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
tools: Optional[List[Dict[str, Any]]] = None,
top_logprobs: Optional[int] = None,
top_p: Optional[float] = None,
user: Optional[str] = None,
) -> Union[OpenAIChatCompletion, AsyncIterator[OpenAIChatCompletionChunk]]:
messages: list[OpenAIMessageParam],
frequency_penalty: float | None = None,
function_call: str | dict[str, Any] | None = None,
functions: list[dict[str, Any]] | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_completion_tokens: int | None = None,
max_tokens: int | None = None,
n: int | None = None,
parallel_tool_calls: bool | None = None,
presence_penalty: float | None = None,
response_format: OpenAIResponseFormatParam | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
tool_choice: str | dict[str, Any] | None = None,
tools: list[dict[str, Any]] | None = None,
top_logprobs: int | None = None,
top_p: float | None = None,
user: str | None = None,
) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
"""Generate an OpenAI-compatible chat completion for the given messages using the specified model.
:param model: The identifier of the model to use. The model must be registered with Llama Stack and available via the /models endpoint.

6
llama_stack/apis/inspect/inspect.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List, Protocol, runtime_checkable
from typing import Protocol, runtime_checkable
from pydantic import BaseModel
@ -16,7 +16,7 @@ from llama_stack.schema_utils import json_schema_type, webmethod
class RouteInfo(BaseModel):
route: str
method: str
provider_types: List[str]
provider_types: list[str]
@json_schema_type
@ -30,7 +30,7 @@ class VersionInfo(BaseModel):
class ListRoutesResponse(BaseModel):
data: List[RouteInfo]
data: list[RouteInfo]
@runtime_checkable

22
llama_stack/apis/models/models.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, List, Literal, Optional, Protocol, runtime_checkable
from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel, ConfigDict, Field
@ -15,7 +15,7 @@ from llama_stack.schema_utils import json_schema_type, webmethod
class CommonModelFields(BaseModel):
metadata: Dict[str, Any] = Field(
metadata: dict[str, Any] = Field(
default_factory=dict,
description="Any additional metadata for this model",
)
@ -46,14 +46,14 @@ class Model(CommonModelFields, Resource):
class ModelInput(CommonModelFields):
model_id: str
provider_id: Optional[str] = None
provider_model_id: Optional[str] = None
model_type: Optional[ModelType] = ModelType.llm
provider_id: str | None = None
provider_model_id: str | None = None
model_type: ModelType | None = ModelType.llm
model_config = ConfigDict(protected_namespaces=())
class ListModelsResponse(BaseModel):
data: List[Model]
data: list[Model]
@json_schema_type
@ -73,7 +73,7 @@ class OpenAIModel(BaseModel):
class OpenAIListModelsResponse(BaseModel):
data: List[OpenAIModel]
data: list[OpenAIModel]
@runtime_checkable
@ -95,10 +95,10 @@ class Models(Protocol):
async def register_model(
self,
model_id: str,
provider_model_id: Optional[str] = None,
provider_id: Optional[str] = None,
metadata: Optional[Dict[str, Any]] = None,
model_type: Optional[ModelType] = None,
provider_model_id: str | None = None,
provider_id: str | None = None,
metadata: dict[str, Any] | None = None,
model_type: ModelType | None = None,
) -> Model: ...
@webmethod(route="/models/{model_id:path}", method="DELETE")

69
llama_stack/apis/post_training/post_training.py
View file

@ -6,10 +6,9 @@
from datetime import datetime
from enum import Enum
from typing import Any, Dict, List, Literal, Optional, Protocol, Union
from typing import Annotated, Any, Literal, Protocol
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.job_types import JobStatus
@ -36,9 +35,9 @@ class DataConfig(BaseModel):
batch_size: int
shuffle: bool
data_format: DatasetFormat
validation_dataset_id: Optional[str] = None
packed: Optional[bool] = False
train_on_input: Optional[bool] = False
validation_dataset_id: str | None = None
packed: bool | None = False
train_on_input: bool | None = False
@json_schema_type
@ -51,10 +50,10 @@ class OptimizerConfig(BaseModel):
@json_schema_type
class EfficiencyConfig(BaseModel):
enable_activation_checkpointing: Optional[bool] = False
enable_activation_offloading: Optional[bool] = False
memory_efficient_fsdp_wrap: Optional[bool] = False
fsdp_cpu_offload: Optional[bool] = False
enable_activation_checkpointing: bool | None = False
enable_activation_offloading: bool | None = False
memory_efficient_fsdp_wrap: bool | None = False
fsdp_cpu_offload: bool | None = False
@json_schema_type
@ -62,23 +61,23 @@ class TrainingConfig(BaseModel):
n_epochs: int
max_steps_per_epoch: int = 1
gradient_accumulation_steps: int = 1
max_validation_steps: Optional[int] = 1
data_config: Optional[DataConfig] = None
optimizer_config: Optional[OptimizerConfig] = None
efficiency_config: Optional[EfficiencyConfig] = None
dtype: Optional[str] = "bf16"
max_validation_steps: int | None = 1
data_config: DataConfig | None = None
optimizer_config: OptimizerConfig | None = None
efficiency_config: EfficiencyConfig | None = None
dtype: str | None = "bf16"
@json_schema_type
class LoraFinetuningConfig(BaseModel):
type: Literal["LoRA"] = "LoRA"
lora_attn_modules: List[str]
lora_attn_modules: list[str]
apply_lora_to_mlp: bool
apply_lora_to_output: bool
rank: int
alpha: int
use_dora: Optional[bool] = False
quantize_base: Optional[bool] = False
use_dora: bool | None = False
quantize_base: bool | None = False
@json_schema_type
@ -88,7 +87,7 @@ class QATFinetuningConfig(BaseModel):
group_size: int
AlgorithmConfig = Annotated[Union[LoraFinetuningConfig, QATFinetuningConfig], Field(discriminator="type")]
AlgorithmConfig = Annotated[LoraFinetuningConfig | QATFinetuningConfig, Field(discriminator="type")]
register_schema(AlgorithmConfig, name="AlgorithmConfig")
@ -97,7 +96,7 @@ class PostTrainingJobLogStream(BaseModel):
"""Stream of logs from a finetuning job."""
job_uuid: str
log_lines: List[str]
log_lines: list[str]
@json_schema_type
@ -131,8 +130,8 @@ class PostTrainingRLHFRequest(BaseModel):
training_config: TrainingConfig
# TODO: define these
hyperparam_search_config: Dict[str, Any]
logger_config: Dict[str, Any]
hyperparam_search_config: dict[str, Any]
logger_config: dict[str, Any]
class PostTrainingJob(BaseModel):
@ -146,17 +145,17 @@ class PostTrainingJobStatusResponse(BaseModel):
job_uuid: str
status: JobStatus
scheduled_at: Optional[datetime] = None
started_at: Optional[datetime] = None
completed_at: Optional[datetime] = None
scheduled_at: datetime | None = None
started_at: datetime | None = None
completed_at: datetime | None = None
resources_allocated: Optional[Dict[str, Any]] = None
resources_allocated: dict[str, Any] | None = None
checkpoints: List[Checkpoint] = Field(default_factory=list)
checkpoints: list[Checkpoint] = Field(default_factory=list)
class ListPostTrainingJobsResponse(BaseModel):
data: List[PostTrainingJob]
data: list[PostTrainingJob]
@json_schema_type
@ -164,7 +163,7 @@ class PostTrainingJobArtifactsResponse(BaseModel):
"""Artifacts of a finetuning job."""
job_uuid: str
checkpoints: List[Checkpoint] = Field(default_factory=list)
checkpoints: list[Checkpoint] = Field(default_factory=list)
# TODO(ashwin): metrics, evals
@ -175,14 +174,14 @@ class PostTraining(Protocol):
self,
job_uuid: str,
training_config: TrainingConfig,
hyperparam_search_config: Dict[str, Any],
logger_config: Dict[str, Any],
model: Optional[str] = Field(
hyperparam_search_config: dict[str, Any],
logger_config: dict[str, Any],
model: str | None = Field(
default=None,
description="Model descriptor for training if not in provider config`",
),
checkpoint_dir: Optional[str] = None,
algorithm_config: Optional[AlgorithmConfig] = None,
checkpoint_dir: str | None = None,
algorithm_config: AlgorithmConfig | None = None,
) -> PostTrainingJob: ...
@webmethod(route="/post-training/preference-optimize", method="POST")
@ -192,8 +191,8 @@ class PostTraining(Protocol):
finetuned_model: str,
algorithm_config: DPOAlignmentConfig,
training_config: TrainingConfig,
hyperparam_search_config: Dict[str, Any],
logger_config: Dict[str, Any],
hyperparam_search_config: dict[str, Any],
logger_config: dict[str, Any],
) -> PostTrainingJob: ...
@webmethod(route="/post-training/jobs", method="GET")

6
llama_stack/apis/providers/providers.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Protocol, runtime_checkable
from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel
@ -17,12 +17,12 @@ class ProviderInfo(BaseModel):
api: str
provider_id: str
provider_type: str
config: Dict[str, Any]
config: dict[str, Any]
health: HealthResponse
class ListProvidersResponse(BaseModel):
data: List[ProviderInfo]
data: list[ProviderInfo]
@runtime_checkable

12
llama_stack/apis/safety/safety.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel, Field
@ -27,16 +27,16 @@ class SafetyViolation(BaseModel):
violation_level: ViolationLevel
# what message should you convey to the user
user_message: Optional[str] = None
user_message: str | None = None
# additional metadata (including specific violation codes) more for
# debugging, telemetry
metadata: Dict[str, Any] = Field(default_factory=dict)
metadata: dict[str, Any] = Field(default_factory=dict)
@json_schema_type
class RunShieldResponse(BaseModel):
violation: Optional[SafetyViolation] = None
violation: SafetyViolation | None = None
class ShieldStore(Protocol):
@ -52,6 +52,6 @@ class Safety(Protocol):
async def run_shield(
self,
shield_id: str,
messages: List[Message],
params: Dict[str, Any] = None,
messages: list[Message],
params: dict[str, Any] = None,
) -> RunShieldResponse: ...

20
llama_stack/apis/scoring/scoring.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel
@ -12,7 +12,7 @@ from llama_stack.apis.scoring_functions import ScoringFn, ScoringFnParams
from llama_stack.schema_utils import json_schema_type, webmethod
# mapping of metric to value
ScoringResultRow = Dict[str, Any]
ScoringResultRow = dict[str, Any]
@json_schema_type
@ -24,15 +24,15 @@ class ScoringResult(BaseModel):
:param aggregated_results: Map of metric name to aggregated value
"""
score_rows: List[ScoringResultRow]
score_rows: list[ScoringResultRow]
# aggregated metrics to value
aggregated_results: Dict[str, Any]
aggregated_results: dict[str, Any]
@json_schema_type
class ScoreBatchResponse(BaseModel):
dataset_id: Optional[str] = None
results: Dict[str, ScoringResult]
dataset_id: str | None = None
results: dict[str, ScoringResult]
@json_schema_type
@ -44,7 +44,7 @@ class ScoreResponse(BaseModel):
"""
# each key in the dict is a scoring function name
results: Dict[str, ScoringResult]
results: dict[str, ScoringResult]
class ScoringFunctionStore(Protocol):
@ -59,15 +59,15 @@ class Scoring(Protocol):
async def score_batch(
self,
dataset_id: str,
scoring_functions: Dict[str, Optional[ScoringFnParams]],
scoring_functions: dict[str, ScoringFnParams | None],
save_results_dataset: bool = False,
) -> ScoreBatchResponse: ...
@webmethod(route="/scoring/score", method="POST")
async def score(
self,
input_rows: List[Dict[str, Any]],
scoring_functions: Dict[str, Optional[ScoringFnParams]],
input_rows: list[dict[str, Any]],
scoring_functions: dict[str, ScoringFnParams | None],
) -> ScoreResponse:
"""Score a list of rows.

42
llama_stack/apis/scoring_functions/scoring_functions.py
View file

@ -6,18 +6,14 @@
from enum import Enum
from typing import (
Annotated,
Any,
Dict,
List,
Literal,
Optional,
Protocol,
Union,
runtime_checkable,
)
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.apis.common.type_system import ParamType
from llama_stack.apis.resource import Resource, ResourceType
@ -46,12 +42,12 @@ class AggregationFunctionType(Enum):
class LLMAsJudgeScoringFnParams(BaseModel):
type: Literal[ScoringFnParamsType.llm_as_judge.value] = ScoringFnParamsType.llm_as_judge.value
judge_model: str
prompt_template: Optional[str] = None
judge_score_regexes: Optional[List[str]] = Field(
prompt_template: str | None = None
judge_score_regexes: list[str] | None = Field(
description="Regexes to extract the answer from generated response",
default_factory=list,
)
aggregation_functions: Optional[List[AggregationFunctionType]] = Field(
aggregation_functions: list[AggregationFunctionType] | None = Field(
description="Aggregation functions to apply to the scores of each row",
default_factory=list,
)
@ -60,11 +56,11 @@ class LLMAsJudgeScoringFnParams(BaseModel):
@json_schema_type
class RegexParserScoringFnParams(BaseModel):
type: Literal[ScoringFnParamsType.regex_parser.value] = ScoringFnParamsType.regex_parser.value
parsing_regexes: Optional[List[str]] = Field(
parsing_regexes: list[str] | None = Field(
description="Regex to extract the answer from generated response",
default_factory=list,
)
aggregation_functions: Optional[List[AggregationFunctionType]] = Field(
aggregation_functions: list[AggregationFunctionType] | None = Field(
description="Aggregation functions to apply to the scores of each row",
default_factory=list,
)
@ -73,33 +69,29 @@ class RegexParserScoringFnParams(BaseModel):
@json_schema_type
class BasicScoringFnParams(BaseModel):
type: Literal[ScoringFnParamsType.basic.value] = ScoringFnParamsType.basic.value
aggregation_functions: Optional[List[AggregationFunctionType]] = Field(
aggregation_functions: list[AggregationFunctionType] | None = Field(
description="Aggregation functions to apply to the scores of each row",
default_factory=list,
)
ScoringFnParams = Annotated[
Union[
LLMAsJudgeScoringFnParams,
RegexParserScoringFnParams,
BasicScoringFnParams,
],
LLMAsJudgeScoringFnParams | RegexParserScoringFnParams | BasicScoringFnParams,
Field(discriminator="type"),
]
register_schema(ScoringFnParams, name="ScoringFnParams")
class CommonScoringFnFields(BaseModel):
description: Optional[str] = None
metadata: Dict[str, Any] = Field(
description: str | None = None
metadata: dict[str, Any] = Field(
default_factory=dict,
description="Any additional metadata for this definition",
)
return_type: ParamType = Field(
description="The return type of the deterministic function",
)
params: Optional[ScoringFnParams] = Field(
params: ScoringFnParams | None = Field(
description="The parameters for the scoring function for benchmark eval, these can be overridden for app eval",
default=None,
)
@ -120,12 +112,12 @@ class ScoringFn(CommonScoringFnFields, Resource):
class ScoringFnInput(CommonScoringFnFields, BaseModel):
scoring_fn_id: str
provider_id: Optional[str] = None
provider_scoring_fn_id: Optional[str] = None
provider_id: str | None = None
provider_scoring_fn_id: str | None = None
class ListScoringFunctionsResponse(BaseModel):
data: List[ScoringFn]
data: list[ScoringFn]
@runtime_checkable
@ -142,7 +134,7 @@ class ScoringFunctions(Protocol):
scoring_fn_id: str,
description: str,
return_type: ParamType,
provider_scoring_fn_id: Optional[str] = None,
provider_id: Optional[str] = None,
params: Optional[ScoringFnParams] = None,
provider_scoring_fn_id: str | None = None,
provider_id: str | None = None,
params: ScoringFnParams | None = None,
) -> None: ...

16
llama_stack/apis/shields/shields.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Literal, Optional, Protocol, runtime_checkable
from typing import Any, Literal, Protocol, runtime_checkable
from pydantic import BaseModel
@ -14,7 +14,7 @@ from llama_stack.schema_utils import json_schema_type, webmethod
class CommonShieldFields(BaseModel):
params: Optional[Dict[str, Any]] = None
params: dict[str, Any] | None = None
@json_schema_type
@ -34,12 +34,12 @@ class Shield(CommonShieldFields, Resource):
class ShieldInput(CommonShieldFields):
shield_id: str
provider_id: Optional[str] = None
provider_shield_id: Optional[str] = None
provider_id: str | None = None
provider_shield_id: str | None = None
class ListShieldsResponse(BaseModel):
data: List[Shield]
data: list[Shield]
@runtime_checkable
@ -55,7 +55,7 @@ class Shields(Protocol):
async def register_shield(
self,
shield_id: str,
provider_shield_id: Optional[str] = None,
provider_id: Optional[str] = None,
params: Optional[Dict[str, Any]] = None,
provider_shield_id: str | None = None,
provider_id: str | None = None,
params: dict[str, Any] | None = None,
) -> Shield: ...

16
llama_stack/apis/synthetic_data_generation/synthetic_data_generation.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, List, Optional, Protocol, Union
from typing import Any, Protocol
from pydantic import BaseModel
@ -28,24 +28,24 @@ class FilteringFunction(Enum):
class SyntheticDataGenerationRequest(BaseModel):
"""Request to generate synthetic data. A small batch of prompts and a filtering function"""
dialogs: List[Message]
dialogs: list[Message]
filtering_function: FilteringFunction = FilteringFunction.none
model: Optional[str] = None
model: str | None = None
@json_schema_type
class SyntheticDataGenerationResponse(BaseModel):
"""Response from the synthetic data generation. Batch of (prompt, response, score) tuples that pass the threshold."""
synthetic_data: List[Dict[str, Any]]
statistics: Optional[Dict[str, Any]] = None
synthetic_data: list[dict[str, Any]]
statistics: dict[str, Any] | None = None
class SyntheticDataGeneration(Protocol):
@webmethod(route="/synthetic-data-generation/generate")
def synthetic_data_generate(
self,
dialogs: List[Message],
dialogs: list[Message],
filtering_function: FilteringFunction = FilteringFunction.none,
model: Optional[str] = None,
) -> Union[SyntheticDataGenerationResponse]: ...
model: str | None = None,
) -> SyntheticDataGenerationResponse: ...

69
llama_stack/apis/telemetry/telemetry.py
View file

@ -7,18 +7,14 @@
from datetime import datetime
from enum import Enum
from typing import (
Annotated,
Any,
Dict,
List,
Literal,
Optional,
Protocol,
Union,
runtime_checkable,
)
from pydantic import BaseModel, Field
from typing_extensions import Annotated
from llama_stack.models.llama.datatypes import Primitive
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@ -37,11 +33,11 @@ class SpanStatus(Enum):
class Span(BaseModel):
span_id: str
trace_id: str
parent_span_id: Optional[str] = None
parent_span_id: str | None = None
name: str
start_time: datetime
end_time: Optional[datetime] = None
attributes: Optional[Dict[str, Any]] = Field(default_factory=dict)
end_time: datetime | None = None
attributes: dict[str, Any] | None = Field(default_factory=dict)
def set_attribute(self, key: str, value: Any):
if self.attributes is None:
@ -54,7 +50,7 @@ class Trace(BaseModel):
trace_id: str
root_span_id: str
start_time: datetime
end_time: Optional[datetime] = None
end_time: datetime | None = None
@json_schema_type
@ -78,7 +74,7 @@ class EventCommon(BaseModel):
trace_id: str
span_id: str
timestamp: datetime
attributes: Optional[Dict[str, Primitive]] = Field(default_factory=dict)
attributes: dict[str, Primitive] | None = Field(default_factory=dict)
@json_schema_type
@ -92,15 +88,15 @@ class UnstructuredLogEvent(EventCommon):
class MetricEvent(EventCommon):
type: Literal[EventType.METRIC.value] = EventType.METRIC.value
metric: str # this would be an enum
value: Union[int, float]
value: int | float
unit: str
@json_schema_type
class MetricInResponse(BaseModel):
metric: str
value: Union[int, float]
unit: Optional[str] = None
value: int | float
unit: str | None = None
# This is a short term solution to allow inference API to return metrics
@ -124,7 +120,7 @@ class MetricInResponse(BaseModel):
class MetricResponseMixin(BaseModel):
metrics: Optional[List[MetricInResponse]] = None
metrics: list[MetricInResponse] | None = None
@json_schema_type
@ -137,7 +133,7 @@ class StructuredLogType(Enum):
class SpanStartPayload(BaseModel):
type: Literal[StructuredLogType.SPAN_START.value] = StructuredLogType.SPAN_START.value
name: str
parent_span_id: Optional[str] = None
parent_span_id: str | None = None
@json_schema_type
@ -147,10 +143,7 @@ class SpanEndPayload(BaseModel):
StructuredLogPayload = Annotated[
Union[
SpanStartPayload,
SpanEndPayload,
],
SpanStartPayload | SpanEndPayload,
Field(discriminator="type"),
]
register_schema(StructuredLogPayload, name="StructuredLogPayload")
@ -163,11 +156,7 @@ class StructuredLogEvent(EventCommon):
Event = Annotated[
Union[
UnstructuredLogEvent,
MetricEvent,
StructuredLogEvent,
],
UnstructuredLogEvent | MetricEvent | StructuredLogEvent,
Field(discriminator="type"),
]
register_schema(Event, name="Event")
@ -184,7 +173,7 @@ class EvalTrace(BaseModel):
@json_schema_type
class SpanWithStatus(Span):
status: Optional[SpanStatus] = None
status: SpanStatus | None = None
@json_schema_type
@ -203,15 +192,15 @@ class QueryCondition(BaseModel):
class QueryTracesResponse(BaseModel):
data: List[Trace]
data: list[Trace]
class QuerySpansResponse(BaseModel):
data: List[Span]
data: list[Span]
class QuerySpanTreeResponse(BaseModel):
data: Dict[str, SpanWithStatus]
data: dict[str, SpanWithStatus]
@runtime_checkable
@ -222,10 +211,10 @@ class Telemetry(Protocol):
@webmethod(route="/telemetry/traces", method="POST")
async def query_traces(
self,
attribute_filters: Optional[List[QueryCondition]] = None,
limit: Optional[int] = 100,
offset: Optional[int] = 0,
order_by: Optional[List[str]] = None,
attribute_filters: list[QueryCondition] | None = None,
limit: int | None = 100,
offset: int | None = 0,
order_by: list[str] | None = None,
) -> QueryTracesResponse: ...
@webmethod(route="/telemetry/traces/{trace_id:path}", method="GET")
@ -238,23 +227,23 @@ class Telemetry(Protocol):
async def get_span_tree(
self,
span_id: str,
attributes_to_return: Optional[List[str]] = None,
max_depth: Optional[int] = None,
attributes_to_return: list[str] | None = None,
max_depth: int | None = None,
) -> QuerySpanTreeResponse: ...
@webmethod(route="/telemetry/spans", method="POST")
async def query_spans(
self,
attribute_filters: List[QueryCondition],
attributes_to_return: List[str],
max_depth: Optional[int] = None,
attribute_filters: list[QueryCondition],
attributes_to_return: list[str],
max_depth: int | None = None,
) -> QuerySpansResponse: ...
@webmethod(route="/telemetry/spans/export", method="POST")
async def save_spans_to_dataset(
self,
attribute_filters: List[QueryCondition],
attributes_to_save: List[str],
attribute_filters: list[QueryCondition],
attributes_to_save: list[str],
dataset_id: str,
max_depth: Optional[int] = None,
max_depth: int | None = None,
) -> None: ...

21
llama_stack/apis/tools/rag_tool.py
View file

@ -5,10 +5,10 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, List, Literal, Optional, Union
from typing import Annotated, Any, Literal
from pydantic import BaseModel, Field
from typing_extensions import Annotated, Protocol, runtime_checkable
from typing_extensions import Protocol, runtime_checkable
from llama_stack.apis.common.content_types import URL, InterleavedContent
from llama_stack.providers.utils.telemetry.trace_protocol import trace_protocol
@ -29,13 +29,13 @@ class RAGDocument(BaseModel):
document_id: str
content: InterleavedContent | URL
mime_type: str | None = None
metadata: Dict[str, Any] = Field(default_factory=dict)
metadata: dict[str, Any] = Field(default_factory=dict)
@json_schema_type
class RAGQueryResult(BaseModel):
content: Optional[InterleavedContent] = None
metadata: Dict[str, Any] = Field(default_factory=dict)
content: InterleavedContent | None = None
metadata: dict[str, Any] = Field(default_factory=dict)
@json_schema_type
@ -59,10 +59,7 @@ class LLMRAGQueryGeneratorConfig(BaseModel):
RAGQueryGeneratorConfig = Annotated[
Union[
DefaultRAGQueryGeneratorConfig,
LLMRAGQueryGeneratorConfig,
],
DefaultRAGQueryGeneratorConfig | LLMRAGQueryGeneratorConfig,
Field(discriminator="type"),
]
register_schema(RAGQueryGeneratorConfig, name="RAGQueryGeneratorConfig")
@ -83,7 +80,7 @@ class RAGToolRuntime(Protocol):
@webmethod(route="/tool-runtime/rag-tool/insert", method="POST")
async def insert(
self,
documents: List[RAGDocument],
documents: list[RAGDocument],
vector_db_id: str,
chunk_size_in_tokens: int = 512,
) -> None:
@ -94,8 +91,8 @@ class RAGToolRuntime(Protocol):
async def query(
self,
content: InterleavedContent,
vector_db_ids: List[str],
query_config: Optional[RAGQueryConfig] = None,
vector_db_ids: list[str],
query_config: RAGQueryConfig | None = None,
) -> RAGQueryResult:
"""Query the RAG system for context; typically invoked by the agent"""
...

44
llama_stack/apis/tools/tools.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Any, Dict, List, Literal, Optional
from typing import Any, Literal
from pydantic import BaseModel, Field
from typing_extensions import Protocol, runtime_checkable
@ -24,7 +24,7 @@ class ToolParameter(BaseModel):
parameter_type: str
description: str
required: bool = Field(default=True)
default: Optional[Any] = None
default: Any | None = None
@json_schema_type
@ -40,39 +40,39 @@ class Tool(Resource):
toolgroup_id: str
tool_host: ToolHost
description: str
parameters: List[ToolParameter]
metadata: Optional[Dict[str, Any]] = None
parameters: list[ToolParameter]
metadata: dict[str, Any] | None = None
@json_schema_type
class ToolDef(BaseModel):
name: str
description: Optional[str] = None
parameters: Optional[List[ToolParameter]] = None
metadata: Optional[Dict[str, Any]] = None
description: str | None = None
parameters: list[ToolParameter] | None = None
metadata: dict[str, Any] | None = None
@json_schema_type
class ToolGroupInput(BaseModel):
toolgroup_id: str
provider_id: str
args: Optional[Dict[str, Any]] = None
mcp_endpoint: Optional[URL] = None
args: dict[str, Any] | None = None
mcp_endpoint: URL | None = None
@json_schema_type
class ToolGroup(Resource):
type: Literal[ResourceType.tool_group.value] = ResourceType.tool_group.value
mcp_endpoint: Optional[URL] = None
args: Optional[Dict[str, Any]] = None
mcp_endpoint: URL | None = None
args: dict[str, Any] | None = None
@json_schema_type
class ToolInvocationResult(BaseModel):
content: Optional[InterleavedContent] = None
error_message: Optional[str] = None
error_code: Optional[int] = None
metadata: Optional[Dict[str, Any]] = None
content: InterleavedContent | None = None
error_message: str | None = None
error_code: int | None = None
metadata: dict[str, Any] | None = None
class ToolStore(Protocol):
@ -81,11 +81,11 @@ class ToolStore(Protocol):
class ListToolGroupsResponse(BaseModel):
data: List[ToolGroup]
data: list[ToolGroup]
class ListToolsResponse(BaseModel):
data: List[Tool]
data: list[Tool]
class ListToolDefsResponse(BaseModel):
@ -100,8 +100,8 @@ class ToolGroups(Protocol):
self,
toolgroup_id: str,
provider_id: str,
mcp_endpoint: Optional[URL] = None,
args: Optional[Dict[str, Any]] = None,
mcp_endpoint: URL | None = None,
args: dict[str, Any] | None = None,
) -> None:
"""Register a tool group"""
...
@ -118,7 +118,7 @@ class ToolGroups(Protocol):
...
@webmethod(route="/tools", method="GET")
async def list_tools(self, toolgroup_id: Optional[str] = None) -> ListToolsResponse:
async def list_tools(self, toolgroup_id: str | None = None) -> ListToolsResponse:
"""List tools with optional tool group"""
...
@ -151,10 +151,10 @@ class ToolRuntime(Protocol):
# TODO: This needs to be renamed once OPEN API generator name conflict issue is fixed.
@webmethod(route="/tool-runtime/list-tools", method="GET")
async def list_runtime_tools(
self, tool_group_id: Optional[str] = None, mcp_endpoint: Optional[URL] = None
self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
) -> ListToolDefsResponse: ...
@webmethod(route="/tool-runtime/invoke", method="POST")
async def invoke_tool(self, tool_name: str, kwargs: Dict[str, Any]) -> ToolInvocationResult:
async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> ToolInvocationResult:
"""Run a tool with the given arguments"""
...

12
llama_stack/apis/vector_dbs/vector_dbs.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import List, Literal, Optional, Protocol, runtime_checkable
from typing import Literal, Protocol, runtime_checkable
from pydantic import BaseModel
@ -33,11 +33,11 @@ class VectorDBInput(BaseModel):
vector_db_id: str
embedding_model: str
embedding_dimension: int
provider_vector_db_id: Optional[str] = None
provider_vector_db_id: str | None = None
class ListVectorDBsResponse(BaseModel):
data: List[VectorDB]
data: list[VectorDB]
@runtime_checkable
@ -57,9 +57,9 @@ class VectorDBs(Protocol):
self,
vector_db_id: str,
embedding_model: str,
embedding_dimension: Optional[int] = 384,
provider_id: Optional[str] = None,
provider_vector_db_id: Optional[str] = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorDB: ...
@webmethod(route="/vector-dbs/{vector_db_id:path}", method="DELETE")

16
llama_stack/apis/vector_io/vector_io.py
View file

@ -8,7 +8,7 @@
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, List, Optional, Protocol, runtime_checkable
from typing import Any, Protocol, runtime_checkable
from pydantic import BaseModel, Field
@ -20,17 +20,17 @@ from llama_stack.schema_utils import json_schema_type, webmethod
class Chunk(BaseModel):
content: InterleavedContent
metadata: Dict[str, Any] = Field(default_factory=dict)
metadata: dict[str, Any] = Field(default_factory=dict)
@json_schema_type
class QueryChunksResponse(BaseModel):
chunks: List[Chunk]
scores: List[float]
chunks: list[Chunk]
scores: list[float]
class VectorDBStore(Protocol):
def get_vector_db(self, vector_db_id: str) -> Optional[VectorDB]: ...
def get_vector_db(self, vector_db_id: str) -> VectorDB | None: ...
@runtime_checkable
@ -44,8 +44,8 @@ class VectorIO(Protocol):
async def insert_chunks(
self,
vector_db_id: str,
chunks: List[Chunk],
ttl_seconds: Optional[int] = None,
chunks: list[Chunk],
ttl_seconds: int | None = None,
) -> None: ...
@webmethod(route="/vector-io/query", method="POST")
@ -53,5 +53,5 @@ class VectorIO(Protocol):
self,
vector_db_id: str,
query: InterleavedContent,
params: Optional[Dict[str, Any]] = None,
params: dict[str, Any] | None = None,
) -> QueryChunksResponse: ...

13
llama_stack/cli/download.py
View file

@ -13,7 +13,6 @@ from dataclasses import dataclass
from datetime import datetime, timezone
from functools import partial
from pathlib import Path
from typing import Dict, List, Optional
import httpx
from pydantic import BaseModel, ConfigDict
@ -102,7 +101,7 @@ class DownloadTask:
output_file: str
total_size: int = 0
downloaded_size: int = 0
task_id: Optional[int] = None
task_id: int | None = None
retries: int = 0
max_retries: int = 3
@ -262,7 +261,7 @@ class ParallelDownloader:
self.progress.update(task.task_id, description=f"[red]Failed: {task.output_file}[/red]")
raise DownloadError(f"Download failed for {task.output_file}: {str(e)}") from e
def has_disk_space(self, tasks: List[DownloadTask]) -> bool:
def has_disk_space(self, tasks: list[DownloadTask]) -> bool:
try:
total_remaining_size = sum(task.total_size - task.downloaded_size for task in tasks)
dir_path = os.path.dirname(os.path.abspath(tasks[0].output_file))
@ -282,7 +281,7 @@ class ParallelDownloader:
except Exception as e:
raise DownloadError(f"Failed to check disk space: {str(e)}") from e
async def download_all(self, tasks: List[DownloadTask]) -> None:
async def download_all(self, tasks: list[DownloadTask]) -> None:
if not tasks:
raise ValueError("No download tasks provided")
@ -391,20 +390,20 @@ def _meta_download(
class ModelEntry(BaseModel):
model_id: str
files: Dict[str, str]
files: dict[str, str]
model_config = ConfigDict(protected_namespaces=())
class Manifest(BaseModel):
models: List[ModelEntry]
models: list[ModelEntry]
expires_on: datetime
def _download_from_manifest(manifest_file: str, max_concurrent_downloads: int):
from llama_stack.distribution.utils.model_utils import model_local_dir
with open(manifest_file, "r") as f:
with open(manifest_file) as f:
d = json.load(f)
manifest = Manifest(**d)

8
llama_stack/cli/model/safety_models.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from typing import Any
from pydantic import BaseModel, ConfigDict, Field
@ -22,7 +22,7 @@ class PromptGuardModel(BaseModel):
max_seq_length: int = 512
is_instruct_model: bool = False
quantization_format: CheckpointQuantizationFormat = CheckpointQuantizationFormat.bf16
arch_args: Dict[str, Any] = Field(default_factory=dict)
arch_args: dict[str, Any] = Field(default_factory=dict)
def descriptor(self) -> str:
return self.model_id
@ -44,11 +44,11 @@ def prompt_guard_model_skus():
]
def prompt_guard_model_sku_map() -> Dict[str, Any]:
def prompt_guard_model_sku_map() -> dict[str, Any]:
return {model.model_id: model for model in prompt_guard_model_skus()}
def prompt_guard_download_info_map() -> Dict[str, LlamaDownloadInfo]:
def prompt_guard_download_info_map() -> dict[str, LlamaDownloadInfo]:
return {
model.model_id: LlamaDownloadInfo(
folder="Prompt-Guard" if model.model_id == "Prompt-Guard-86M" else model.model_id,

24
llama_stack/cli/stack/_build.py
View file

@ -13,13 +13,12 @@ import sys
import textwrap
from functools import lru_cache
from pathlib import Path
from typing import Dict, Optional
import yaml
from prompt_toolkit import prompt
from prompt_toolkit.completion import WordCompleter
from prompt_toolkit.validation import Validator
from termcolor import cprint
from termcolor import colored, cprint
from llama_stack.cli.stack.utils import ImageType
from llama_stack.cli.table import print_table
@ -46,14 +45,14 @@ from llama_stack.providers.datatypes import Api
TEMPLATES_PATH = Path(__file__).parent.parent.parent / "templates"
@lru_cache()
def available_templates_specs() -> Dict[str, BuildConfig]:
@lru_cache
def available_templates_specs() -> dict[str, BuildConfig]:
import yaml
template_specs = {}
for p in TEMPLATES_PATH.rglob("*build.yaml"):
template_name = p.parent.name
with open(p, "r") as f:
with open(p) as f:
build_config = BuildConfig(**yaml.safe_load(f))
template_specs[template_name] = build_config
return template_specs
@ -178,7 +177,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
if not available_providers:
continue
api_provider = prompt(
"> Enter provider for API {}: ".format(api.value),
f"> Enter provider for API {api.value}: ",
completer=WordCompleter(available_providers),
complete_while_typing=True,
validator=Validator.from_callable(
@ -201,7 +200,7 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
build_config = BuildConfig(image_type=image_type, distribution_spec=distribution_spec)
else:
with open(args.config, "r") as f:
with open(args.config) as f:
try:
build_config = BuildConfig(**yaml.safe_load(f))
except Exception as e:
@ -332,9 +331,9 @@ def _generate_run_config(
def _run_stack_build_command_from_build_config(
build_config: BuildConfig,
image_name: Optional[str] = None,
template_name: Optional[str] = None,
config_path: Optional[str] = None,
image_name: str | None = None,
template_name: str | None = None,
config_path: str | None = None,
) -> str:
image_name = image_name or build_config.image_name
if build_config.image_type == LlamaStackImageType.CONTAINER.value:
@ -389,6 +388,11 @@ def _run_stack_build_command_from_build_config(
shutil.copy(path, run_config_file)
cprint("Build Successful!", color="green")
cprint("You can find the newly-built template here: " + colored(template_path, "light_blue"))
cprint(
"You can run the new Llama Stack distro via: "
+ colored(f"llama stack run {template_path} --image-type {build_config.image_type}", "light_blue")
)
return template_path
else:
return _generate_run_config(build_config, build_dir, image_name)

2
llama_stack/cli/stack/run.py
View file

@ -119,7 +119,7 @@ class StackRun(Subcommand):
if not config_file.is_file():
self.parser.error(
f"Config file must be a valid file path, '{config_file}’ is not a file: type={type(config_file)}"
f"Config file must be a valid file path, '{config_file}' is not a file: type={type(config_file)}"
)
logger.info(f"Using run configuration: {config_file}")

2
llama_stack/cli/table.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Iterable
from collections.abc import Iterable
from rich.console import Console
from rich.table import Table

9
llama_stack/cli/verify_download.py
View file

@ -9,7 +9,6 @@ import hashlib
from dataclasses import dataclass
from functools import partial
from pathlib import Path
from typing import Dict, List, Optional
from rich.console import Console
from rich.progress import Progress, SpinnerColumn, TextColumn
@ -21,7 +20,7 @@ from llama_stack.cli.subcommand import Subcommand
class VerificationResult:
filename: str
expected_hash: str
actual_hash: Optional[str]
actual_hash: str | None
exists: bool
matches: bool
@ -60,9 +59,9 @@ def calculate_md5(filepath: Path, chunk_size: int = 8192) -> str:
return md5_hash.hexdigest()
def load_checksums(checklist_path: Path) -> Dict[str, str]:
def load_checksums(checklist_path: Path) -> dict[str, str]:
checksums = {}
with open(checklist_path, "r") as f:
with open(checklist_path) as f:
for line in f:
if line.strip():
md5sum, filepath = line.strip().split(" ", 1)
@ -72,7 +71,7 @@ def load_checksums(checklist_path: Path) -> Dict[str, str]:
return checksums
def verify_files(model_dir: Path, checksums: Dict[str, str], console: Console) -> List[VerificationResult]:
def verify_files(model_dir: Path, checksums: dict[str, str], console: Console) -> list[VerificationResult]:
results = []
with Progress(

6
llama_stack/distribution/access_control.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict, Optional
from typing import Any
from llama_stack.distribution.datatypes import AccessAttributes
from llama_stack.log import get_logger
@ -14,8 +14,8 @@ logger = get_logger(__name__, category="core")
def check_access(
obj_identifier: str,
obj_attributes: Optional[AccessAttributes],
user_attributes: Optional[Dict[str, Any]] = None,
obj_attributes: AccessAttributes | None,
user_attributes: dict[str, Any] | None = None,
) -> bool:
"""Check if the current user has access to the given object, based on access attributes.

3
llama_stack/distribution/build.py
View file

@ -47,14 +47,13 @@ def get_provider_dependencies(
providers = config.distribution_spec.providers
deps = []
registry = get_provider_registry(config)
for api_str, provider_or_providers in providers.items():
providers_for_api = registry[Api(api_str)]
providers = provider_or_providers if isinstance(provider_or_providers, list) else [provider_or_providers]
for provider in providers:
# Providers from BuildConfig and RunConfig are subtly different – not great
# Providers from BuildConfig and RunConfig are subtly different - not great
provider_type = provider if isinstance(provider, str) else provider.provider_type
if provider_type not in providers_for_api:

4
llama_stack/distribution/client.py
View file

@ -8,7 +8,7 @@ import inspect
import json
from collections.abc import AsyncIterator
from enum import Enum
from typing import Any, Type, Union, get_args, get_origin
from typing import Any, Union, get_args, get_origin
import httpx
from pydantic import BaseModel, parse_obj_as
@ -27,7 +27,7 @@ async def get_client_impl(protocol, config: RemoteProviderConfig, _deps: Any):
return impl
def create_api_client_class(protocol) -> Type:
def create_api_client_class(protocol) -> type:
if protocol in _CLIENT_CLASSES:
return _CLIENT_CLASSES[protocol]

2
llama_stack/distribution/common.sh
View file

@ -1,3 +1,5 @@
#!/usr/bin/env bash
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#

10
llama_stack/distribution/configure.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
import logging
import textwrap
from typing import Any, Dict
from typing import Any
from llama_stack.distribution.datatypes import (
LLAMA_STACK_RUN_CONFIG_VERSION,
@ -24,7 +24,7 @@ from llama_stack.providers.datatypes import Api, ProviderSpec
logger = logging.getLogger(__name__)
def configure_single_provider(registry: Dict[str, ProviderSpec], provider: Provider) -> Provider:
def configure_single_provider(registry: dict[str, ProviderSpec], provider: Provider) -> Provider:
provider_spec = registry[provider.provider_type]
config_type = instantiate_class_type(provider_spec.config_class)
try:
@ -120,8 +120,8 @@ def configure_api_providers(config: StackRunConfig, build_spec: DistributionSpec
def upgrade_from_routing_table(
config_dict: Dict[str, Any],
) -> Dict[str, Any]:
config_dict: dict[str, Any],
) -> dict[str, Any]:
def get_providers(entries):
return [
Provider(
@ -163,7 +163,7 @@ def upgrade_from_routing_table(
return config_dict
def parse_and_maybe_upgrade_config(config_dict: Dict[str, Any]) -> StackRunConfig:
def parse_and_maybe_upgrade_config(config_dict: dict[str, Any]) -> StackRunConfig:
version = config_dict.get("version", None)
if version == LLAMA_STACK_RUN_CONFIG_VERSION:
return StackRunConfig(**config_dict)

114
llama_stack/distribution/datatypes.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
from enum import Enum
from typing import Annotated, Any, Dict, List, Optional, Union
from typing import Annotated, Any
from pydantic import BaseModel, Field
@ -30,7 +30,7 @@ LLAMA_STACK_BUILD_CONFIG_VERSION = "2"
LLAMA_STACK_RUN_CONFIG_VERSION = "2"
RoutingKey = Union[str, List[str]]
RoutingKey = str | list[str]
class AccessAttributes(BaseModel):
@ -47,17 +47,17 @@ class AccessAttributes(BaseModel):
"""
# Standard attribute categories - the minimal set we need now
roles: Optional[List[str]] = Field(
roles: list[str] | None = Field(
default=None, description="Role-based attributes (e.g., 'admin', 'data-scientist', 'user')"
)
teams: Optional[List[str]] = Field(default=None, description="Team-based attributes (e.g., 'ml-team', 'nlp-team')")
teams: list[str] | None = Field(default=None, description="Team-based attributes (e.g., 'ml-team', 'nlp-team')")
projects: Optional[List[str]] = Field(
projects: list[str] | None = Field(
default=None, description="Project-based access attributes (e.g., 'llama-3', 'customer-insights')"
)
namespaces: Optional[List[str]] = Field(
namespaces: list[str] | None = Field(
default=None, description="Namespace-based access control for resource isolation"
)
@ -106,7 +106,7 @@ class ResourceWithACL(Resource):
# ^ User must have access to the customer-insights project AND have confidential namespace
"""
access_attributes: Optional[AccessAttributes] = None
access_attributes: AccessAttributes | None = None
# Use the extended Resource for all routable objects
@ -142,41 +142,21 @@ class ToolGroupWithACL(ToolGroup, ResourceWithACL):
pass
RoutableObject = Union[
Model,
Shield,
VectorDB,
Dataset,
ScoringFn,
Benchmark,
Tool,
ToolGroup,
]
RoutableObject = Model | Shield | VectorDB | Dataset | ScoringFn | Benchmark | Tool | ToolGroup
RoutableObjectWithProvider = Annotated[
Union[
ModelWithACL,
ShieldWithACL,
VectorDBWithACL,
DatasetWithACL,
ScoringFnWithACL,
BenchmarkWithACL,
ToolWithACL,
ToolGroupWithACL,
],
ModelWithACL
| ShieldWithACL
| VectorDBWithACL
| DatasetWithACL
| ScoringFnWithACL
| BenchmarkWithACL
| ToolWithACL
| ToolGroupWithACL,
Field(discriminator="type"),
]
RoutedProtocol = Union[
Inference,
Safety,
VectorIO,
DatasetIO,
Scoring,
Eval,
ToolRuntime,
]
RoutedProtocol = Inference | Safety | VectorIO | DatasetIO | Scoring | Eval | ToolRuntime
# Example: /inference, /safety
@ -184,15 +164,15 @@ class AutoRoutedProviderSpec(ProviderSpec):
provider_type: str = "router"
config_class: str = ""
container_image: Optional[str] = None
container_image: str | None = None
routing_table_api: Api
module: str
provider_data_validator: Optional[str] = Field(
provider_data_validator: str | None = Field(
default=None,
)
@property
def pip_packages(self) -> List[str]:
def pip_packages(self) -> list[str]:
raise AssertionError("Should not be called on AutoRoutedProviderSpec")
@ -200,20 +180,20 @@ class AutoRoutedProviderSpec(ProviderSpec):
class RoutingTableProviderSpec(ProviderSpec):
provider_type: str = "routing_table"
config_class: str = ""
container_image: Optional[str] = None
container_image: str | None = None
router_api: Api
module: str
pip_packages: List[str] = Field(default_factory=list)
pip_packages: list[str] = Field(default_factory=list)
class DistributionSpec(BaseModel):
description: Optional[str] = Field(
description: str | None = Field(
default="",
description="Description of the distribution",
)
container_image: Optional[str] = None
providers: Dict[str, Union[str, List[str]]] = Field(
container_image: str | None = None
providers: dict[str, str | list[str]] = Field(
default_factory=dict,
description="""
Provider Types for each of the APIs provided by this distribution. If you
@ -225,12 +205,12 @@ in the runtime configuration to help route to the correct provider.""",
class Provider(BaseModel):
provider_id: str
provider_type: str
config: Dict[str, Any]
config: dict[str, Any]
class LoggingConfig(BaseModel):
category_levels: Dict[str, str] = Field(
default_factory=Dict,
category_levels: dict[str, str] = Field(
default_factory=dict,
description="""
Dictionary of different logging configurations for different portions (ex: core, server) of llama stack""",
)
@ -248,7 +228,7 @@ class AuthenticationConfig(BaseModel):
...,
description="Type of authentication provider (e.g., 'kubernetes', 'custom')",
)
config: Dict[str, str] = Field(
config: dict[str, str] = Field(
...,
description="Provider-specific configuration",
)
@ -261,15 +241,15 @@ class ServerConfig(BaseModel):
ge=1024,
le=65535,
)
tls_certfile: Optional[str] = Field(
tls_certfile: str | None = Field(
default=None,
description="Path to TLS certificate file for HTTPS",
)
tls_keyfile: Optional[str] = Field(
tls_keyfile: str | None = Field(
default=None,
description="Path to TLS key file for HTTPS",
)
auth: Optional[AuthenticationConfig] = Field(
auth: AuthenticationConfig | None = Field(
default=None,
description="Authentication configuration for the server",
)
@ -285,23 +265,23 @@ Reference to the distribution this package refers to. For unregistered (adhoc) p
this could be just a hash
""",
)
container_image: Optional[str] = Field(
container_image: str | None = Field(
default=None,
description="Reference to the container image if this package refers to a container",
)
apis: List[str] = Field(
apis: list[str] = Field(
default_factory=list,
description="""
The list of APIs to serve. If not specified, all APIs specified in the provider_map will be served""",
)
providers: Dict[str, List[Provider]] = Field(
providers: dict[str, list[Provider]] = Field(
description="""
One or more providers to use for each API. The same provider_type (e.g., meta-reference)
can be instantiated multiple times (with different configs) if necessary.
""",
)
metadata_store: Optional[KVStoreConfig] = Field(
metadata_store: KVStoreConfig | None = Field(
default=None,
description="""
Configuration for the persistence store used by the distribution registry. If not specified,
@ -309,22 +289,22 @@ a default SQLite store will be used.""",
)
# registry of "resources" in the distribution
models: List[ModelInput] = Field(default_factory=list)
shields: List[ShieldInput] = Field(default_factory=list)
vector_dbs: List[VectorDBInput] = Field(default_factory=list)
datasets: List[DatasetInput] = Field(default_factory=list)
scoring_fns: List[ScoringFnInput] = Field(default_factory=list)
benchmarks: List[BenchmarkInput] = Field(default_factory=list)
tool_groups: List[ToolGroupInput] = Field(default_factory=list)
models: list[ModelInput] = Field(default_factory=list)
shields: list[ShieldInput] = Field(default_factory=list)
vector_dbs: list[VectorDBInput] = Field(default_factory=list)
datasets: list[DatasetInput] = Field(default_factory=list)
scoring_fns: list[ScoringFnInput] = Field(default_factory=list)
benchmarks: list[BenchmarkInput] = Field(default_factory=list)
tool_groups: list[ToolGroupInput] = Field(default_factory=list)
logging: Optional[LoggingConfig] = Field(default=None, description="Configuration for Llama Stack Logging")
logging: LoggingConfig | None = Field(default=None, description="Configuration for Llama Stack Logging")
server: ServerConfig = Field(
default_factory=ServerConfig,
description="Configuration for the HTTP(S) server",
)
external_providers_dir: Optional[str] = Field(
external_providers_dir: str | None = Field(
default=None,
description="Path to directory containing external provider implementations. The providers code and dependencies must be installed on the system.",
)
@ -338,11 +318,11 @@ class BuildConfig(BaseModel):
default="conda",
description="Type of package to build (conda | container | venv)",
)
image_name: Optional[str] = Field(
image_name: str | None = Field(
default=None,
description="Name of the distribution to build",
)
external_providers_dir: Optional[str] = Field(
external_providers_dir: str | None = Field(
default=None,
description="Path to directory containing external provider implementations. The providers packages will be resolved from this directory. "
"pip_packages MUST contain the provider package name.",

16
llama_stack/distribution/distribution.py
View file

@ -7,7 +7,7 @@
import glob
import importlib
import os
from typing import Any, Dict, List
from typing import Any
import yaml
from pydantic import BaseModel
@ -24,7 +24,7 @@ from llama_stack.providers.datatypes import (
logger = get_logger(name=__name__, category="core")
def stack_apis() -> List[Api]:
def stack_apis() -> list[Api]:
return list(Api)
@ -33,7 +33,7 @@ class AutoRoutedApiInfo(BaseModel):
router_api: Api
def builtin_automatically_routed_apis() -> List[AutoRoutedApiInfo]:
def builtin_automatically_routed_apis() -> list[AutoRoutedApiInfo]:
return [
AutoRoutedApiInfo(
routing_table_api=Api.models,
@ -66,12 +66,12 @@ def builtin_automatically_routed_apis() -> List[AutoRoutedApiInfo]:
]
def providable_apis() -> List[Api]:
def providable_apis() -> list[Api]:
routing_table_apis = {x.routing_table_api for x in builtin_automatically_routed_apis()}
return [api for api in Api if api not in routing_table_apis and api != Api.inspect and api != Api.providers]
def _load_remote_provider_spec(spec_data: Dict[str, Any], api: Api) -> ProviderSpec:
def _load_remote_provider_spec(spec_data: dict[str, Any], api: Api) -> ProviderSpec:
adapter = AdapterSpec(**spec_data["adapter"])
spec = remote_provider_spec(
api=api,
@ -81,7 +81,7 @@ def _load_remote_provider_spec(spec_data: Dict[str, Any], api: Api) -> ProviderS
return spec
def _load_inline_provider_spec(spec_data: Dict[str, Any], api: Api, provider_name: str) -> ProviderSpec:
def _load_inline_provider_spec(spec_data: dict[str, Any], api: Api, provider_name: str) -> ProviderSpec:
spec = InlineProviderSpec(
api=api,
provider_type=f"inline::{provider_name}",
@ -98,7 +98,7 @@ def _load_inline_provider_spec(spec_data: Dict[str, Any], api: Api, provider_nam
def get_provider_registry(
config=None,
) -> Dict[Api, Dict[str, ProviderSpec]]:
) -> dict[Api, dict[str, ProviderSpec]]:
"""Get the provider registry, optionally including external providers.
This function loads both built-in providers and external providers from YAML files.
@ -133,7 +133,7 @@ def get_provider_registry(
ValueError: If any provider spec is invalid
"""
ret: Dict[Api, Dict[str, ProviderSpec]] = {}
ret: dict[Api, dict[str, ProviderSpec]] = {}
for api in providable_apis():
name = api.name.lower()
logger.debug(f"Importing module {name}")

12
llama_stack/distribution/library_client.py
View file

@ -12,7 +12,7 @@ import os
from concurrent.futures import ThreadPoolExecutor
from enum import Enum
from pathlib import Path
from typing import Any, Optional, TypeVar, Union, get_args, get_origin
from typing import Any, TypeVar, Union, get_args, get_origin
import httpx
import yaml
@ -119,8 +119,8 @@ class LlamaStackAsLibraryClient(LlamaStackClient):
self,
config_path_or_template_name: str,
skip_logger_removal: bool = False,
custom_provider_registry: Optional[ProviderRegistry] = None,
provider_data: Optional[dict[str, Any]] = None,
custom_provider_registry: ProviderRegistry | None = None,
provider_data: dict[str, Any] | None = None,
):
super().__init__()
self.async_client = AsyncLlamaStackAsLibraryClient(
@ -181,8 +181,8 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
def __init__(
self,
config_path_or_template_name: str,
custom_provider_registry: Optional[ProviderRegistry] = None,
provider_data: Optional[dict[str, Any]] = None,
custom_provider_registry: ProviderRegistry | None = None,
provider_data: dict[str, Any] | None = None,
):
super().__init__()
# when using the library client, we should not log to console since many
@ -371,7 +371,7 @@ class AsyncLlamaStackAsLibraryClient(AsyncLlamaStackClient):
)
return await response.parse()
def _convert_body(self, path: str, method: str, body: Optional[dict] = None) -> dict:
def _convert_body(self, path: str, method: str, body: dict | None = None) -> dict:
if not body:
return {}

6
llama_stack/distribution/providers.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
import asyncio
from typing import Any, Dict
from typing import Any
from pydantic import BaseModel
@ -73,14 +73,14 @@ class ProviderImpl(Providers):
raise ValueError(f"Provider {provider_id} not found")
async def get_providers_health(self) -> Dict[str, Dict[str, HealthResponse]]:
async def get_providers_health(self) -> dict[str, dict[str, HealthResponse]]:
"""Get health status for all providers.
Returns:
Dict[str, Dict[str, HealthResponse]]: A dictionary mapping API names to provider health statuses.
Each API maps to a dictionary of provider IDs to their health responses.
"""
providers_health: Dict[str, Dict[str, HealthResponse]] = {}
providers_health: dict[str, dict[str, HealthResponse]] = {}
timeout = 1.0
async def check_provider_health(impl: Any) -> tuple[str, HealthResponse] | None:

15
llama_stack/distribution/request_headers.py
View file

@ -7,7 +7,8 @@
import contextvars
import json
import logging
from typing import Any, ContextManager, Dict, List, Optional
from contextlib import AbstractContextManager
from typing import Any
from .utils.dynamic import instantiate_class_type
@ -17,11 +18,11 @@ log = logging.getLogger(__name__)
PROVIDER_DATA_VAR = contextvars.ContextVar("provider_data", default=None)
class RequestProviderDataContext(ContextManager):
class RequestProviderDataContext(AbstractContextManager):
"""Context manager for request provider data"""
def __init__(
self, provider_data: Optional[Dict[str, Any]] = None, auth_attributes: Optional[Dict[str, List[str]]] = None
self, provider_data: dict[str, Any] | None = None, auth_attributes: dict[str, list[str]] | None = None
):
self.provider_data = provider_data or {}
if auth_attributes:
@ -63,7 +64,7 @@ class NeedsRequestProviderData:
return None
def parse_request_provider_data(headers: Dict[str, str]) -> Optional[Dict[str, Any]]:
def parse_request_provider_data(headers: dict[str, str]) -> dict[str, Any] | None:
"""Parse provider data from request headers"""
keys = [
"X-LlamaStack-Provider-Data",
@ -86,14 +87,14 @@ def parse_request_provider_data(headers: Dict[str, str]) -> Optional[Dict[str, A
def request_provider_data_context(
headers: Dict[str, str], auth_attributes: Optional[Dict[str, List[str]]] = None
) -> ContextManager:
headers: dict[str, str], auth_attributes: dict[str, list[str]] | None = None
) -> AbstractContextManager:
"""Context manager that sets request provider data from headers and auth attributes for the duration of the context"""
provider_data = parse_request_provider_data(headers)
return RequestProviderDataContext(provider_data, auth_attributes)
def get_auth_attributes() -> Optional[Dict[str, List[str]]]:
def get_auth_attributes() -> dict[str, list[str]] | None:
"""Helper to retrieve auth attributes from the provider data context"""
provider_data = PROVIDER_DATA_VAR.get()
if not provider_data:

50
llama_stack/distribution/resolver.py
View file

@ -5,7 +5,7 @@
# the root directory of this source tree.
import importlib
import inspect
from typing import Any, Dict, List, Set, Tuple
from typing import Any
from llama_stack.apis.agents import Agents
from llama_stack.apis.benchmarks import Benchmarks
@ -58,7 +58,7 @@ class InvalidProviderError(Exception):
pass
def api_protocol_map() -> Dict[Api, Any]:
def api_protocol_map() -> dict[Api, Any]:
return {
Api.providers: ProvidersAPI,
Api.agents: Agents,
@ -83,7 +83,7 @@ def api_protocol_map() -> Dict[Api, Any]:
}
def additional_protocols_map() -> Dict[Api, Any]:
def additional_protocols_map() -> dict[Api, Any]:
return {
Api.inference: (ModelsProtocolPrivate, Models, Api.models),
Api.tool_groups: (ToolsProtocolPrivate, ToolGroups, Api.tool_groups),
@ -104,14 +104,14 @@ class ProviderWithSpec(Provider):
spec: ProviderSpec
ProviderRegistry = Dict[Api, Dict[str, ProviderSpec]]
ProviderRegistry = dict[Api, dict[str, ProviderSpec]]
async def resolve_impls(
run_config: StackRunConfig,
provider_registry: ProviderRegistry,
dist_registry: DistributionRegistry,
) -> Dict[Api, Any]:
) -> dict[Api, Any]:
"""
Resolves provider implementations by:
1. Validating and organizing providers.
@ -136,7 +136,7 @@ async def resolve_impls(
return await instantiate_providers(sorted_providers, router_apis, dist_registry)
def specs_for_autorouted_apis(apis_to_serve: List[str] | Set[str]) -> Dict[str, Dict[str, ProviderWithSpec]]:
def specs_for_autorouted_apis(apis_to_serve: list[str] | set[str]) -> dict[str, dict[str, ProviderWithSpec]]:
"""Generates specifications for automatically routed APIs."""
specs = {}
for info in builtin_automatically_routed_apis():
@ -178,10 +178,10 @@ def specs_for_autorouted_apis(apis_to_serve: List[str] | Set[str]) -> Dict[str,
def validate_and_prepare_providers(
run_config: StackRunConfig, provider_registry: ProviderRegistry, routing_table_apis: Set[Api], router_apis: Set[Api]
) -> Dict[str, Dict[str, ProviderWithSpec]]:
run_config: StackRunConfig, provider_registry: ProviderRegistry, routing_table_apis: set[Api], router_apis: set[Api]
) -> dict[str, dict[str, ProviderWithSpec]]:
"""Validates providers, handles deprecations, and organizes them into a spec dictionary."""
providers_with_specs: Dict[str, Dict[str, ProviderWithSpec]] = {}
providers_with_specs: dict[str, dict[str, ProviderWithSpec]] = {}
for api_str, providers in run_config.providers.items():
api = Api(api_str)
@ -222,10 +222,10 @@ def validate_provider(provider: Provider, api: Api, provider_registry: ProviderR
def sort_providers_by_deps(
providers_with_specs: Dict[str, Dict[str, ProviderWithSpec]], run_config: StackRunConfig
) -> List[Tuple[str, ProviderWithSpec]]:
providers_with_specs: dict[str, dict[str, ProviderWithSpec]], run_config: StackRunConfig
) -> list[tuple[str, ProviderWithSpec]]:
"""Sorts providers based on their dependencies."""
sorted_providers: List[Tuple[str, ProviderWithSpec]] = topological_sort(
sorted_providers: list[tuple[str, ProviderWithSpec]] = topological_sort(
{k: list(v.values()) for k, v in providers_with_specs.items()}
)
@ -236,11 +236,11 @@ def sort_providers_by_deps(
async def instantiate_providers(
sorted_providers: List[Tuple[str, ProviderWithSpec]], router_apis: Set[Api], dist_registry: DistributionRegistry
) -> Dict:
sorted_providers: list[tuple[str, ProviderWithSpec]], router_apis: set[Api], dist_registry: DistributionRegistry
) -> dict:
"""Instantiates providers asynchronously while managing dependencies."""
impls: Dict[Api, Any] = {}
inner_impls_by_provider_id: Dict[str, Dict[str, Any]] = {f"inner-{x.value}": {} for x in router_apis}
impls: dict[Api, Any] = {}
inner_impls_by_provider_id: dict[str, dict[str, Any]] = {f"inner-{x.value}": {} for x in router_apis}
for api_str, provider in sorted_providers:
deps = {a: impls[a] for a in provider.spec.api_dependencies}
for a in provider.spec.optional_api_dependencies:
@ -263,9 +263,9 @@ async def instantiate_providers(
def topological_sort(
providers_with_specs: Dict[str, List[ProviderWithSpec]],
) -> List[Tuple[str, ProviderWithSpec]]:
def dfs(kv, visited: Set[str], stack: List[str]):
providers_with_specs: dict[str, list[ProviderWithSpec]],
) -> list[tuple[str, ProviderWithSpec]]:
def dfs(kv, visited: set[str], stack: list[str]):
api_str, providers = kv
visited.add(api_str)
@ -280,8 +280,8 @@ def topological_sort(
stack.append(api_str)
visited: Set[str] = set()
stack: List[str] = []
visited: set[str] = set()
stack: list[str] = []
for api_str, providers in providers_with_specs.items():
if api_str not in visited:
@ -298,8 +298,8 @@ def topological_sort(
# returns a class implementing the protocol corresponding to the Api
async def instantiate_provider(
provider: ProviderWithSpec,
deps: Dict[Api, Any],
inner_impls: Dict[str, Any],
deps: dict[Api, Any],
inner_impls: dict[str, Any],
dist_registry: DistributionRegistry,
):
protocols = api_protocol_map()
@ -391,8 +391,8 @@ def check_protocol_compliance(obj: Any, protocol: Any) -> None:
async def resolve_remote_stack_impls(
config: RemoteProviderConfig,
apis: List[str],
) -> Dict[Api, Any]:
apis: list[str],
) -> dict[Api, Any]:
protocols = api_protocol_map()
additional_protocols = additional_protocols_map()

6
llama_stack/distribution/routers/__init__.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from typing import Any
from llama_stack.distribution.datatypes import RoutedProtocol
from llama_stack.distribution.store import DistributionRegistry
@ -23,7 +23,7 @@ from .routing_tables import (
async def get_routing_table_impl(
api: Api,
impls_by_provider_id: Dict[str, RoutedProtocol],
impls_by_provider_id: dict[str, RoutedProtocol],
_deps,
dist_registry: DistributionRegistry,
) -> Any:
@ -45,7 +45,7 @@ async def get_routing_table_impl(
return impl
async def get_auto_router_impl(api: Api, routing_table: RoutingTable, deps: Dict[str, Any]) -> Any:
async def get_auto_router_impl(api: Api, routing_table: RoutingTable, deps: dict[str, Any]) -> Any:
from .routers import (
DatasetIORouter,
EvalRouter,

216
llama_stack/distribution/routers/routers.py
View file

@ -6,12 +6,12 @@
import asyncio
import time
from typing import Any, AsyncGenerator, AsyncIterator, Dict, List, Optional, Union
from collections.abc import AsyncGenerator, AsyncIterator
from typing import Annotated, Any
from openai.types.chat import ChatCompletionToolChoiceOptionParam as OpenAIChatCompletionToolChoiceOptionParam
from openai.types.chat import ChatCompletionToolParam as OpenAIChatCompletionToolParam
from pydantic import Field, TypeAdapter
from typing_extensions import Annotated
from llama_stack.apis.common.content_types import (
URL,
@ -100,9 +100,9 @@ class VectorIORouter(VectorIO):
self,
vector_db_id: str,
embedding_model: str,
embedding_dimension: Optional[int] = 384,
provider_id: Optional[str] = None,
provider_vector_db_id: Optional[str] = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> None:
logger.debug(f"VectorIORouter.register_vector_db: {vector_db_id}, {embedding_model}")
await self.routing_table.register_vector_db(
@ -116,8 +116,8 @@ class VectorIORouter(VectorIO):
async def insert_chunks(
self,
vector_db_id: str,
chunks: List[Chunk],
ttl_seconds: Optional[int] = None,
chunks: list[Chunk],
ttl_seconds: int | None = None,
) -> None:
logger.debug(
f"VectorIORouter.insert_chunks: {vector_db_id}, {len(chunks)} chunks, ttl_seconds={ttl_seconds}, chunk_ids={[chunk.metadata['document_id'] for chunk in chunks[:3]]}{' and more...' if len(chunks) > 3 else ''}",
@ -128,7 +128,7 @@ class VectorIORouter(VectorIO):
self,
vector_db_id: str,
query: InterleavedContent,
params: Optional[Dict[str, Any]] = None,
params: dict[str, Any] | None = None,
) -> QueryChunksResponse:
logger.debug(f"VectorIORouter.query_chunks: {vector_db_id}")
return await self.routing_table.get_provider_impl(vector_db_id).query_chunks(vector_db_id, query, params)
@ -140,7 +140,7 @@ class InferenceRouter(Inference):
def __init__(
self,
routing_table: RoutingTable,
telemetry: Optional[Telemetry] = None,
telemetry: Telemetry | None = None,
) -> None:
logger.debug("Initializing InferenceRouter")
self.routing_table = routing_table
@ -160,10 +160,10 @@ class InferenceRouter(Inference):
async def register_model(
self,
model_id: str,
provider_model_id: Optional[str] = None,
provider_id: Optional[str] = None,
metadata: Optional[Dict[str, Any]] = None,
model_type: Optional[ModelType] = None,
provider_model_id: str | None = None,
provider_id: str | None = None,
metadata: dict[str, Any] | None = None,
model_type: ModelType | None = None,
) -> None:
logger.debug(
f"InferenceRouter.register_model: {model_id=} {provider_model_id=} {provider_id=} {metadata=} {model_type=}",
@ -176,7 +176,7 @@ class InferenceRouter(Inference):
completion_tokens: int,
total_tokens: int,
model: Model,
) -> List[MetricEvent]:
) -> list[MetricEvent]:
"""Constructs a list of MetricEvent objects containing token usage metrics.
Args:
@ -221,7 +221,7 @@ class InferenceRouter(Inference):
completion_tokens: int,
total_tokens: int,
model: Model,
) -> List[MetricInResponse]:
) -> list[MetricInResponse]:
metrics = self._construct_metrics(prompt_tokens, completion_tokens, total_tokens, model)
if self.telemetry:
for metric in metrics:
@ -230,9 +230,9 @@ class InferenceRouter(Inference):
async def _count_tokens(
self,
messages: List[Message] | InterleavedContent,
tool_prompt_format: Optional[ToolPromptFormat] = None,
) -> Optional[int]:
messages: list[Message] | InterleavedContent,
tool_prompt_format: ToolPromptFormat | None = None,
) -> int | None:
if isinstance(messages, list):
encoded = self.formatter.encode_dialog_prompt(messages, tool_prompt_format)
else:
@ -242,16 +242,16 @@ class InferenceRouter(Inference):
async def chat_completion(
self,
model_id: str,
messages: List[Message],
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
tools: Optional[List[ToolDefinition]] = None,
tool_choice: Optional[ToolChoice] = None,
tool_prompt_format: Optional[ToolPromptFormat] = None,
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
tool_config: Optional[ToolConfig] = None,
) -> Union[ChatCompletionResponse, AsyncIterator[ChatCompletionResponseStreamChunk]]:
messages: list[Message],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
tools: list[ToolDefinition] | None = None,
tool_choice: ToolChoice | None = None,
tool_prompt_format: ToolPromptFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
tool_config: ToolConfig | None = None,
) -> ChatCompletionResponse | AsyncIterator[ChatCompletionResponseStreamChunk]:
logger.debug(
f"InferenceRouter.chat_completion: {model_id=}, {stream=}, {messages=}, {tools=}, {tool_config=}, {response_format=}",
)
@ -351,12 +351,12 @@ class InferenceRouter(Inference):
async def batch_chat_completion(
self,
model_id: str,
messages_batch: List[List[Message]],
tools: Optional[List[ToolDefinition]] = None,
tool_config: Optional[ToolConfig] = None,
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
messages_batch: list[list[Message]],
tools: list[ToolDefinition] | None = None,
tool_config: ToolConfig | None = None,
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> BatchChatCompletionResponse:
logger.debug(
f"InferenceRouter.batch_chat_completion: {model_id=}, {len(messages_batch)=}, {sampling_params=}, {response_format=}, {logprobs=}",
@ -376,10 +376,10 @@ class InferenceRouter(Inference):
self,
model_id: str,
content: InterleavedContent,
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
stream: Optional[bool] = False,
logprobs: Optional[LogProbConfig] = None,
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
stream: bool | None = False,
logprobs: LogProbConfig | None = None,
) -> AsyncGenerator:
if sampling_params is None:
sampling_params = SamplingParams()
@ -439,10 +439,10 @@ class InferenceRouter(Inference):
async def batch_completion(
self,
model_id: str,
content_batch: List[InterleavedContent],
sampling_params: Optional[SamplingParams] = None,
response_format: Optional[ResponseFormat] = None,
logprobs: Optional[LogProbConfig] = None,
content_batch: list[InterleavedContent],
sampling_params: SamplingParams | None = None,
response_format: ResponseFormat | None = None,
logprobs: LogProbConfig | None = None,
) -> BatchCompletionResponse:
logger.debug(
f"InferenceRouter.batch_completion: {model_id=}, {len(content_batch)=}, {sampling_params=}, {response_format=}, {logprobs=}",
@ -453,10 +453,10 @@ class InferenceRouter(Inference):
async def embeddings(
self,
model_id: str,
contents: List[str] | List[InterleavedContentItem],
text_truncation: Optional[TextTruncation] = TextTruncation.none,
output_dimension: Optional[int] = None,
task_type: Optional[EmbeddingTaskType] = None,
contents: list[str] | list[InterleavedContentItem],
text_truncation: TextTruncation | None = TextTruncation.none,
output_dimension: int | None = None,
task_type: EmbeddingTaskType | None = None,
) -> EmbeddingsResponse:
logger.debug(f"InferenceRouter.embeddings: {model_id}")
model = await self.routing_table.get_model(model_id)
@ -475,24 +475,24 @@ class InferenceRouter(Inference):
async def openai_completion(
self,
model: str,
prompt: Union[str, List[str], List[int], List[List[int]]],
best_of: Optional[int] = None,
echo: Optional[bool] = None,
frequency_penalty: Optional[float] = None,
logit_bias: Optional[Dict[str, float]] = None,
logprobs: Optional[bool] = None,
max_tokens: Optional[int] = None,
n: Optional[int] = None,
presence_penalty: Optional[float] = None,
seed: Optional[int] = None,
stop: Optional[Union[str, List[str]]] = None,
stream: Optional[bool] = None,
stream_options: Optional[Dict[str, Any]] = None,
temperature: Optional[float] = None,
top_p: Optional[float] = None,
user: Optional[str] = None,
guided_choice: Optional[List[str]] = None,
prompt_logprobs: Optional[int] = None,
prompt: str | list[str] | list[int] | list[list[int]],
best_of: int | None = None,
echo: bool | None = None,
frequency_penalty: float | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_tokens: int | None = None,
n: int | None = None,
presence_penalty: float | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
top_p: float | None = None,
user: str | None = None,
guided_choice: list[str] | None = None,
prompt_logprobs: int | None = None,
) -> OpenAICompletion:
logger.debug(
f"InferenceRouter.openai_completion: {model=}, {stream=}, {prompt=}",
@ -531,29 +531,29 @@ class InferenceRouter(Inference):
async def openai_chat_completion(
self,
model: str,
messages: Annotated[List[OpenAIMessageParam], Field(..., min_length=1)],
frequency_penalty: Optional[float] = None,
function_call: Optional[Union[str, Dict[str, Any]]] = None,
functions: Optional[List[Dict[str, Any]]] = None,
logit_bias: Optional[Dict[str, float]] = None,
logprobs: Optional[bool] = None,
max_completion_tokens: Optional[int] = None,
max_tokens: Optional[int] = None,
n: Optional[int] = None,
parallel_tool_calls: Optional[bool] = None,
presence_penalty: Optional[float] = None,
response_format: Optional[OpenAIResponseFormatParam] = None,
seed: Optional[int] = None,
stop: Optional[Union[str, List[str]]] = None,
stream: Optional[bool] = None,
stream_options: Optional[Dict[str, Any]] = None,
temperature: Optional[float] = None,
tool_choice: Optional[Union[str, Dict[str, Any]]] = None,
tools: Optional[List[Dict[str, Any]]] = None,
top_logprobs: Optional[int] = None,
top_p: Optional[float] = None,
user: Optional[str] = None,
) -> Union[OpenAIChatCompletion, AsyncIterator[OpenAIChatCompletionChunk]]:
messages: Annotated[list[OpenAIMessageParam], Field(..., min_length=1)],
frequency_penalty: float | None = None,
function_call: str | dict[str, Any] | None = None,
functions: list[dict[str, Any]] | None = None,
logit_bias: dict[str, float] | None = None,
logprobs: bool | None = None,
max_completion_tokens: int | None = None,
max_tokens: int | None = None,
n: int | None = None,
parallel_tool_calls: bool | None = None,
presence_penalty: float | None = None,
response_format: OpenAIResponseFormatParam | None = None,
seed: int | None = None,
stop: str | list[str] | None = None,
stream: bool | None = None,
stream_options: dict[str, Any] | None = None,
temperature: float | None = None,
tool_choice: str | dict[str, Any] | None = None,
tools: list[dict[str, Any]] | None = None,
top_logprobs: int | None = None,
top_p: float | None = None,
user: str | None = None,
) -> OpenAIChatCompletion | AsyncIterator[OpenAIChatCompletionChunk]:
logger.debug(
f"InferenceRouter.openai_chat_completion: {model=}, {stream=}, {messages=}",
)
@ -602,7 +602,7 @@ class InferenceRouter(Inference):
provider = self.routing_table.get_provider_impl(model_obj.identifier)
return await provider.openai_chat_completion(**params)
async def health(self) -> Dict[str, HealthResponse]:
async def health(self) -> dict[str, HealthResponse]:
health_statuses = {}
timeout = 0.5
for provider_id, impl in self.routing_table.impls_by_provider_id.items():
@ -645,9 +645,9 @@ class SafetyRouter(Safety):
async def register_shield(
self,
shield_id: str,
provider_shield_id: Optional[str] = None,
provider_id: Optional[str] = None,
params: Optional[Dict[str, Any]] = None,
provider_shield_id: str | None = None,
provider_id: str | None = None,
params: dict[str, Any] | None = None,
) -> Shield:
logger.debug(f"SafetyRouter.register_shield: {shield_id}")
return await self.routing_table.register_shield(shield_id, provider_shield_id, provider_id, params)
@ -655,8 +655,8 @@ class SafetyRouter(Safety):
async def run_shield(
self,
shield_id: str,
messages: List[Message],
params: Dict[str, Any] = None,
messages: list[Message],
params: dict[str, Any] = None,
) -> RunShieldResponse:
logger.debug(f"SafetyRouter.run_shield: {shield_id}")
return await self.routing_table.get_provider_impl(shield_id).run_shield(
@ -686,8 +686,8 @@ class DatasetIORouter(DatasetIO):
self,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
dataset_id: Optional[str] = None,
metadata: dict[str, Any] | None = None,
dataset_id: str | None = None,
) -> None:
logger.debug(
f"DatasetIORouter.register_dataset: {purpose=} {source=} {metadata=} {dataset_id=}",
@ -702,8 +702,8 @@ class DatasetIORouter(DatasetIO):
async def iterrows(
self,
dataset_id: str,
start_index: Optional[int] = None,
limit: Optional[int] = None,
start_index: int | None = None,
limit: int | None = None,
) -> PaginatedResponse:
logger.debug(
f"DatasetIORouter.iterrows: {dataset_id}, {start_index=} {limit=}",
@ -714,7 +714,7 @@ class DatasetIORouter(DatasetIO):
limit=limit,
)
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
async def append_rows(self, dataset_id: str, rows: list[dict[str, Any]]) -> None:
logger.debug(f"DatasetIORouter.append_rows: {dataset_id}, {len(rows)} rows")
return await self.routing_table.get_provider_impl(dataset_id).append_rows(
dataset_id=dataset_id,
@ -741,7 +741,7 @@ class ScoringRouter(Scoring):
async def score_batch(
self,
dataset_id: str,
scoring_functions: Dict[str, Optional[ScoringFnParams]] = None,
scoring_functions: dict[str, ScoringFnParams | None] = None,
save_results_dataset: bool = False,
) -> ScoreBatchResponse:
logger.debug(f"ScoringRouter.score_batch: {dataset_id}")
@ -762,8 +762,8 @@ class ScoringRouter(Scoring):
async def score(
self,
input_rows: List[Dict[str, Any]],
scoring_functions: Dict[str, Optional[ScoringFnParams]] = None,
input_rows: list[dict[str, Any]],
scoring_functions: dict[str, ScoringFnParams | None] = None,
) -> ScoreResponse:
logger.debug(f"ScoringRouter.score: {len(input_rows)} rows, {len(scoring_functions)} functions")
res = {}
@ -808,8 +808,8 @@ class EvalRouter(Eval):
async def evaluate_rows(
self,
benchmark_id: str,
input_rows: List[Dict[str, Any]],
scoring_functions: List[str],
input_rows: list[dict[str, Any]],
scoring_functions: list[str],
benchmark_config: BenchmarkConfig,
) -> EvaluateResponse:
logger.debug(f"EvalRouter.evaluate_rows: {benchmark_id}, {len(input_rows)} rows")
@ -863,8 +863,8 @@ class ToolRuntimeRouter(ToolRuntime):
async def query(
self,
content: InterleavedContent,
vector_db_ids: List[str],
query_config: Optional[RAGQueryConfig] = None,
vector_db_ids: list[str],
query_config: RAGQueryConfig | None = None,
) -> RAGQueryResult:
logger.debug(f"ToolRuntimeRouter.RagToolImpl.query: {vector_db_ids}")
return await self.routing_table.get_provider_impl("knowledge_search").query(
@ -873,7 +873,7 @@ class ToolRuntimeRouter(ToolRuntime):
async def insert(
self,
documents: List[RAGDocument],
documents: list[RAGDocument],
vector_db_id: str,
chunk_size_in_tokens: int = 512,
) -> None:
@ -904,7 +904,7 @@ class ToolRuntimeRouter(ToolRuntime):
logger.debug("ToolRuntimeRouter.shutdown")
pass
async def invoke_tool(self, tool_name: str, kwargs: Dict[str, Any]) -> Any:
async def invoke_tool(self, tool_name: str, kwargs: dict[str, Any]) -> Any:
logger.debug(f"ToolRuntimeRouter.invoke_tool: {tool_name}")
return await self.routing_table.get_provider_impl(tool_name).invoke_tool(
tool_name=tool_name,
@ -912,7 +912,7 @@ class ToolRuntimeRouter(ToolRuntime):
)
async def list_runtime_tools(
self, tool_group_id: Optional[str] = None, mcp_endpoint: Optional[URL] = None
self, tool_group_id: str | None = None, mcp_endpoint: URL | None = None
) -> ListToolDefsResponse:
logger.debug(f"ToolRuntimeRouter.list_runtime_tools: {tool_group_id}")
return await self.routing_table.get_provider_impl(tool_group_id).list_tools(tool_group_id, mcp_endpoint)

58
llama_stack/distribution/routers/routing_tables.py
View file

@ -7,7 +7,7 @@
import logging
import time
import uuid
from typing import Any, Dict, List, Optional
from typing import Any
from pydantic import TypeAdapter
@ -106,20 +106,20 @@ async def unregister_object_from_provider(obj: RoutableObject, p: Any) -> None:
raise ValueError(f"Unregister not supported for {api}")
Registry = Dict[str, List[RoutableObjectWithProvider]]
Registry = dict[str, list[RoutableObjectWithProvider]]
class CommonRoutingTableImpl(RoutingTable):
def __init__(
self,
impls_by_provider_id: Dict[str, RoutedProtocol],
impls_by_provider_id: dict[str, RoutedProtocol],
dist_registry: DistributionRegistry,
) -> None:
self.impls_by_provider_id = impls_by_provider_id
self.dist_registry = dist_registry
async def initialize(self) -> None:
async def add_objects(objs: List[RoutableObjectWithProvider], provider_id: str, cls) -> None:
async def add_objects(objs: list[RoutableObjectWithProvider], provider_id: str, cls) -> None:
for obj in objs:
if cls is None:
obj.provider_id = provider_id
@ -154,7 +154,7 @@ class CommonRoutingTableImpl(RoutingTable):
for p in self.impls_by_provider_id.values():
await p.shutdown()
def get_provider_impl(self, routing_key: str, provider_id: Optional[str] = None) -> Any:
def get_provider_impl(self, routing_key: str, provider_id: str | None = None) -> Any:
def apiname_object():
if isinstance(self, ModelsRoutingTable):
return ("Inference", "model")
@ -192,7 +192,7 @@ class CommonRoutingTableImpl(RoutingTable):
raise ValueError(f"Provider not found for `{routing_key}`")
async def get_object_by_identifier(self, type: str, identifier: str) -> Optional[RoutableObjectWithProvider]:
async def get_object_by_identifier(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
# Get from disk registry
obj = await self.dist_registry.get(type, identifier)
if not obj:
@ -236,7 +236,7 @@ class CommonRoutingTableImpl(RoutingTable):
await self.dist_registry.register(obj)
return obj
async def get_all_with_type(self, type: str) -> List[RoutableObjectWithProvider]:
async def get_all_with_type(self, type: str) -> list[RoutableObjectWithProvider]:
objs = await self.dist_registry.get_all()
filtered_objs = [obj for obj in objs if obj.type == type]
@ -277,10 +277,10 @@ class ModelsRoutingTable(CommonRoutingTableImpl, Models):
async def register_model(
self,
model_id: str,
provider_model_id: Optional[str] = None,
provider_id: Optional[str] = None,
metadata: Optional[Dict[str, Any]] = None,
model_type: Optional[ModelType] = None,
provider_model_id: str | None = None,
provider_id: str | None = None,
metadata: dict[str, Any] | None = None,
model_type: ModelType | None = None,
) -> Model:
if provider_model_id is None:
provider_model_id = model_id
@ -328,9 +328,9 @@ class ShieldsRoutingTable(CommonRoutingTableImpl, Shields):
async def register_shield(
self,
shield_id: str,
provider_shield_id: Optional[str] = None,
provider_id: Optional[str] = None,
params: Optional[Dict[str, Any]] = None,
provider_shield_id: str | None = None,
provider_id: str | None = None,
params: dict[str, Any] | None = None,
) -> Shield:
if provider_shield_id is None:
provider_shield_id = shield_id
@ -368,9 +368,9 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
self,
vector_db_id: str,
embedding_model: str,
embedding_dimension: Optional[int] = 384,
provider_id: Optional[str] = None,
provider_vector_db_id: Optional[str] = None,
embedding_dimension: int | None = 384,
provider_id: str | None = None,
provider_vector_db_id: str | None = None,
) -> VectorDB:
if provider_vector_db_id is None:
provider_vector_db_id = vector_db_id
@ -423,8 +423,8 @@ class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
self,
purpose: DatasetPurpose,
source: DataSource,
metadata: Optional[Dict[str, Any]] = None,
dataset_id: Optional[str] = None,
metadata: dict[str, Any] | None = None,
dataset_id: str | None = None,
) -> Dataset:
if isinstance(source, dict):
if source["type"] == "uri":
@ -489,9 +489,9 @@ class ScoringFunctionsRoutingTable(CommonRoutingTableImpl, ScoringFunctions):
scoring_fn_id: str,
description: str,
return_type: ParamType,
provider_scoring_fn_id: Optional[str] = None,
provider_id: Optional[str] = None,
params: Optional[ScoringFnParams] = None,
provider_scoring_fn_id: str | None = None,
provider_id: str | None = None,
params: ScoringFnParams | None = None,
) -> None:
if provider_scoring_fn_id is None:
provider_scoring_fn_id = scoring_fn_id
@ -528,10 +528,10 @@ class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
self,
benchmark_id: str,
dataset_id: str,
scoring_functions: List[str],
metadata: Optional[Dict[str, Any]] = None,
provider_benchmark_id: Optional[str] = None,
provider_id: Optional[str] = None,
scoring_functions: list[str],
metadata: dict[str, Any] | None = None,
provider_benchmark_id: str | None = None,
provider_id: str | None = None,
) -> None:
if metadata is None:
metadata = {}
@ -556,7 +556,7 @@ class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
class ToolGroupsRoutingTable(CommonRoutingTableImpl, ToolGroups):
async def list_tools(self, toolgroup_id: Optional[str] = None) -> ListToolsResponse:
async def list_tools(self, toolgroup_id: str | None = None) -> ListToolsResponse:
tools = await self.get_all_with_type("tool")
if toolgroup_id:
tools = [tool for tool in tools if tool.toolgroup_id == toolgroup_id]
@ -578,8 +578,8 @@ class ToolGroupsRoutingTable(CommonRoutingTableImpl, ToolGroups):
self,
toolgroup_id: str,
provider_id: str,
mcp_endpoint: Optional[URL] = None,
args: Optional[Dict[str, Any]] = None,
mcp_endpoint: URL | None = None,
args: dict[str, Any] | None = None,
) -> None:
tools = []
tool_defs = await self.impls_by_provider_id[provider_id].list_runtime_tools(toolgroup_id, mcp_endpoint)

21
llama_stack/distribution/server/auth_providers.py
View file

@ -7,7 +7,6 @@
import json
from abc import ABC, abstractmethod
from enum import Enum
from typing import Dict, List, Optional
from urllib.parse import parse_qs
import httpx
@ -22,7 +21,7 @@ logger = get_logger(name=__name__, category="auth")
class AuthResponse(BaseModel):
"""The format of the authentication response from the auth endpoint."""
access_attributes: Optional[AccessAttributes] = Field(
access_attributes: AccessAttributes | None = Field(
default=None,
description="""
Structured user attributes for attribute-based access control.
@ -44,7 +43,7 @@ class AuthResponse(BaseModel):
""",
)
message: Optional[str] = Field(
message: str | None = Field(
default=None, description="Optional message providing additional context about the authentication result."
)
@ -52,9 +51,9 @@ class AuthResponse(BaseModel):
class AuthRequestContext(BaseModel):
path: str = Field(description="The path of the request being authenticated")
headers: Dict[str, str] = Field(description="HTTP headers from the original request (excluding Authorization)")
headers: dict[str, str] = Field(description="HTTP headers from the original request (excluding Authorization)")
params: Dict[str, List[str]] = Field(
params: dict[str, list[str]] = Field(
description="Query parameters from the original request, parsed as dictionary of lists"
)
@ -76,14 +75,14 @@ class AuthProviderConfig(BaseModel):
"""Base configuration for authentication providers."""
provider_type: AuthProviderType = Field(..., description="Type of authentication provider")
config: Dict[str, str] = Field(..., description="Provider-specific configuration")
config: dict[str, str] = Field(..., description="Provider-specific configuration")
class AuthProvider(ABC):
"""Abstract base class for authentication providers."""
@abstractmethod
async def validate_token(self, token: str, scope: Optional[Dict] = None) -> Optional[AccessAttributes]:
async def validate_token(self, token: str, scope: dict | None = None) -> AccessAttributes | None:
"""Validate a token and return access attributes."""
pass
@ -96,7 +95,7 @@ class AuthProvider(ABC):
class KubernetesAuthProvider(AuthProvider):
"""Kubernetes authentication provider that validates tokens against the Kubernetes API server."""
def __init__(self, config: Dict[str, str]):
def __init__(self, config: dict[str, str]):
self.api_server_url = config["api_server_url"]
self.ca_cert_path = config.get("ca_cert_path")
self._client = None
@ -120,7 +119,7 @@ class KubernetesAuthProvider(AuthProvider):
self._client = ApiClient(configuration)
return self._client
async def validate_token(self, token: str, scope: Optional[Dict] = None) -> Optional[AccessAttributes]:
async def validate_token(self, token: str, scope: dict | None = None) -> AccessAttributes | None:
"""Validate a Kubernetes token and return access attributes."""
try:
client = await self._get_client()
@ -166,11 +165,11 @@ class KubernetesAuthProvider(AuthProvider):
class CustomAuthProvider(AuthProvider):
"""Custom authentication provider that uses an external endpoint."""
def __init__(self, config: Dict[str, str]):
def __init__(self, config: dict[str, str]):
self.endpoint = config["endpoint"]
self._client = None
async def validate_token(self, token: str, scope: Optional[Dict] = None) -> Optional[AccessAttributes]:
async def validate_token(self, token: str, scope: dict | None = None) -> AccessAttributes | None:
"""Validate a token using the custom authentication endpoint."""
if not self.endpoint:
raise ValueError("Authentication endpoint not configured")

3
llama_stack/distribution/server/endpoints.py
View file

@ -6,7 +6,6 @@
import inspect
import re
from typing import Dict, List
from pydantic import BaseModel
@ -29,7 +28,7 @@ def toolgroup_protocol_map():
}
def get_all_api_endpoints() -> Dict[Api, List[ApiEndpoint]]:
def get_all_api_endpoints() -> dict[Api, list[ApiEndpoint]]:
apis = {}
protocols = api_protocol_map()

11
llama_stack/distribution/server/server.py
View file

@ -15,7 +15,7 @@ import warnings
from contextlib import asynccontextmanager
from importlib.metadata import version as parse_version
from pathlib import Path
from typing import Any, List, Optional, Union
from typing import Annotated, Any
import yaml
from fastapi import Body, FastAPI, HTTPException, Request
@ -24,7 +24,6 @@ from fastapi.exceptions import RequestValidationError
from fastapi.responses import JSONResponse, StreamingResponse
from openai import BadRequestError
from pydantic import BaseModel, ValidationError
from typing_extensions import Annotated
from llama_stack.distribution.datatypes import LoggingConfig, StackRunConfig
from llama_stack.distribution.distribution import builtin_automatically_routed_apis
@ -91,7 +90,7 @@ async def global_exception_handler(request: Request, exc: Exception):
return JSONResponse(status_code=http_exc.status_code, content={"error": {"detail": http_exc.detail}})
def translate_exception(exc: Exception) -> Union[HTTPException, RequestValidationError]:
def translate_exception(exc: Exception) -> HTTPException | RequestValidationError:
if isinstance(exc, ValidationError):
exc = RequestValidationError(exc.errors())
@ -315,7 +314,7 @@ class ClientVersionMiddleware:
return await self.app(scope, receive, send)
def main(args: Optional[argparse.Namespace] = None):
def main(args: argparse.Namespace | None = None):
"""Start the LlamaStack server."""
parser = argparse.ArgumentParser(description="Start the LlamaStack server.")
parser.add_argument(
@ -385,7 +384,7 @@ def main(args: Optional[argparse.Namespace] = None):
raise ValueError("Either --yaml-config or --template must be provided")
logger_config = None
with open(config_file, "r") as fp:
with open(config_file) as fp:
config_contents = yaml.safe_load(fp)
if isinstance(config_contents, dict) and (cfg := config_contents.get("logging_config")):
logger_config = LoggingConfig(**cfg)
@ -517,7 +516,7 @@ def main(args: Optional[argparse.Namespace] = None):
uvicorn.run(**uvicorn_config)
def extract_path_params(route: str) -> List[str]:
def extract_path_params(route: str) -> list[str]:
segments = route.split("/")
params = [seg[1:-1] for seg in segments if seg.startswith("{") and seg.endswith("}")]
# to handle path params like {param:path}

12
llama_stack/distribution/stack.py
View file

@ -8,7 +8,7 @@ import importlib.resources
import os
import re
import tempfile
from typing import Any, Dict, Optional
from typing import Any
import yaml
@ -90,7 +90,7 @@ RESOURCES = [
]
async def register_resources(run_config: StackRunConfig, impls: Dict[Api, Any]):
async def register_resources(run_config: StackRunConfig, impls: dict[Api, Any]):
for rsrc, api, register_method, list_method in RESOURCES:
objects = getattr(run_config, rsrc)
if api not in impls:
@ -197,7 +197,7 @@ def validate_env_pair(env_pair: str) -> tuple[str, str]:
) from e
def add_internal_implementations(impls: Dict[Api, Any], run_config: StackRunConfig) -> None:
def add_internal_implementations(impls: dict[Api, Any], run_config: StackRunConfig) -> None:
"""Add internal implementations (inspect and providers) to the implementations dictionary.
Args:
@ -220,8 +220,8 @@ def add_internal_implementations(impls: Dict[Api, Any], run_config: StackRunConf
# Produces a stack of providers for the given run config. Not all APIs may be
# asked for in the run config.
async def construct_stack(
run_config: StackRunConfig, provider_registry: Optional[ProviderRegistry] = None
) -> Dict[Api, Any]:
run_config: StackRunConfig, provider_registry: ProviderRegistry | None = None
) -> dict[Api, Any]:
dist_registry, _ = await create_dist_registry(run_config.metadata_store, run_config.image_name)
impls = await resolve_impls(run_config, provider_registry or get_provider_registry(run_config), dist_registry)
@ -244,7 +244,7 @@ def get_stack_run_config_from_template(template: str) -> StackRunConfig:
def run_config_from_adhoc_config_spec(
adhoc_config_spec: str, provider_registry: Optional[ProviderRegistry] = None
adhoc_config_spec: str, provider_registry: ProviderRegistry | None = None
) -> StackRunConfig:
"""
Create an adhoc distribution from a list of API providers.

28
llama_stack/distribution/store/registry.py
View file

@ -6,7 +6,7 @@
import asyncio
from contextlib import asynccontextmanager
from typing import Dict, List, Optional, Protocol, Tuple
from typing import Protocol
import pydantic
@ -20,13 +20,13 @@ logger = get_logger(__name__, category="core")
class DistributionRegistry(Protocol):
async def get_all(self) -> List[RoutableObjectWithProvider]: ...
async def get_all(self) -> list[RoutableObjectWithProvider]: ...
async def initialize(self) -> None: ...
async def get(self, identifier: str) -> Optional[RoutableObjectWithProvider]: ...
async def get(self, identifier: str) -> RoutableObjectWithProvider | None: ...
def get_cached(self, identifier: str) -> Optional[RoutableObjectWithProvider]: ...
def get_cached(self, identifier: str) -> RoutableObjectWithProvider | None: ...
async def update(self, obj: RoutableObjectWithProvider) -> RoutableObjectWithProvider: ...
@ -40,13 +40,13 @@ KEY_VERSION = "v8"
KEY_FORMAT = f"{REGISTER_PREFIX}:{KEY_VERSION}::" + "{type}:{identifier}"
def _get_registry_key_range() -> Tuple[str, str]:
def _get_registry_key_range() -> tuple[str, str]:
"""Returns the start and end keys for the registry range query."""
start_key = f"{REGISTER_PREFIX}:{KEY_VERSION}"
return start_key, f"{start_key}\xff"
def _parse_registry_values(values: List[str]) -> List[RoutableObjectWithProvider]:
def _parse_registry_values(values: list[str]) -> list[RoutableObjectWithProvider]:
"""Utility function to parse registry values into RoutableObjectWithProvider objects."""
all_objects = []
for value in values:
@ -67,16 +67,16 @@ class DiskDistributionRegistry(DistributionRegistry):
async def initialize(self) -> None:
pass
def get_cached(self, type: str, identifier: str) -> Optional[RoutableObjectWithProvider]:
def get_cached(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
# Disk registry does not have a cache
raise NotImplementedError("Disk registry does not have a cache")
async def get_all(self) -> List[RoutableObjectWithProvider]:
async def get_all(self) -> list[RoutableObjectWithProvider]:
start_key, end_key = _get_registry_key_range()
values = await self.kvstore.range(start_key, end_key)
return _parse_registry_values(values)
async def get(self, type: str, identifier: str) -> Optional[RoutableObjectWithProvider]:
async def get(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
json_str = await self.kvstore.get(KEY_FORMAT.format(type=type, identifier=identifier))
if not json_str:
return None
@ -113,7 +113,7 @@ class DiskDistributionRegistry(DistributionRegistry):
class CachedDiskDistributionRegistry(DiskDistributionRegistry):
def __init__(self, kvstore: KVStore):
super().__init__(kvstore)
self.cache: Dict[Tuple[str, str], RoutableObjectWithProvider] = {}
self.cache: dict[tuple[str, str], RoutableObjectWithProvider] = {}
self._initialized = False
self._initialize_lock = asyncio.Lock()
self._cache_lock = asyncio.Lock()
@ -147,15 +147,15 @@ class CachedDiskDistributionRegistry(DiskDistributionRegistry):
async def initialize(self) -> None:
await self._ensure_initialized()
def get_cached(self, type: str, identifier: str) -> Optional[RoutableObjectWithProvider]:
def get_cached(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
return self.cache.get((type, identifier), None)
async def get_all(self) -> List[RoutableObjectWithProvider]:
async def get_all(self) -> list[RoutableObjectWithProvider]:
await self._ensure_initialized()
async with self._locked_cache() as cache:
return list(cache.values())
async def get(self, type: str, identifier: str) -> Optional[RoutableObjectWithProvider]:
async def get(self, type: str, identifier: str) -> RoutableObjectWithProvider | None:
await self._ensure_initialized()
cache_key = (type, identifier)
@ -189,7 +189,7 @@ class CachedDiskDistributionRegistry(DiskDistributionRegistry):
async def create_dist_registry(
metadata_store: Optional[KVStoreConfig],
metadata_store: KVStoreConfig | None,
image_name: str,
) -> tuple[CachedDiskDistributionRegistry, KVStore]:
# instantiate kvstore for storing and retrieving distribution metadata

3
llama_stack/distribution/ui/modules/api.py
View file

@ -5,7 +5,6 @@
# the root directory of this source tree.
import os
from typing import Optional
from llama_stack_client import LlamaStackClient
@ -23,7 +22,7 @@ class LlamaStackApi:
},
)
def run_scoring(self, row, scoring_function_ids: list[str], scoring_params: Optional[dict]):
def run_scoring(self, row, scoring_function_ids: list[str], scoring_params: dict | None):
"""Run scoring on a single row"""
if not scoring_params:
scoring_params = {fn_id: None for fn_id in scoring_function_ids}

8
llama_stack/distribution/ui/requirements.txt
View file

@ -1,5 +1,5 @@
streamlit
pandas
llama-stack-client>=0.2.1
streamlit-option-menu
llama-stack>=0.2.1
llama-stack-client>=0.2.1
pandas
streamlit
streamlit-option-menu

6
llama_stack/distribution/utils/config.py
View file

@ -4,10 +4,10 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from typing import Any
def redact_sensitive_fields(data: Dict[str, Any]) -> Dict[str, Any]:
def redact_sensitive_fields(data: dict[str, Any]) -> dict[str, Any]:
"""Redact sensitive information from config before printing."""
sensitive_patterns = ["api_key", "api_token", "password", "secret"]
@ -18,7 +18,7 @@ def redact_sensitive_fields(data: Dict[str, Any]) -> Dict[str, Any]:
return [_redact_value(i) for i in v]
return v
def _redact_dict(d: Dict[str, Any]) -> Dict[str, Any]:
def _redact_dict(d: dict[str, Any]) -> dict[str, Any]:
result = {}
for k, v in d.items():
if any(pattern in k.lower() for pattern in sensitive_patterns):

5
llama_stack/distribution/utils/context.py
View file

@ -4,14 +4,15 @@
# 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 AsyncGenerator
from contextvars import ContextVar
from typing import AsyncGenerator, List, TypeVar
from typing import TypeVar
T = TypeVar("T")
def preserve_contexts_async_generator(
gen: AsyncGenerator[T, None], context_vars: List[ContextVar]
gen: AsyncGenerator[T, None], context_vars: list[ContextVar]
) -> AsyncGenerator[T, None]:
"""
Wraps an async generator to preserve context variables across iterations.

9
llama_stack/distribution/utils/prompt_for_config.py
View file

@ -8,12 +8,11 @@ import inspect
import json
import logging
from enum import Enum
from typing import Any, List, Literal, Optional, Type, Union, get_args, get_origin
from typing import Annotated, Any, Literal, Union, get_args, get_origin
from pydantic import BaseModel
from pydantic.fields import FieldInfo
from pydantic_core import PydanticUndefinedType
from typing_extensions import Annotated
log = logging.getLogger(__name__)
@ -21,7 +20,7 @@ log = logging.getLogger(__name__)
def is_list_of_primitives(field_type):
"""Check if a field type is a List of primitive types."""
origin = get_origin(field_type)
if origin is List or origin is list:
if origin is list or origin is list:
args = get_args(field_type)
if len(args) == 1 and args[0] in (int, float, str, bool):
return True
@ -53,7 +52,7 @@ def get_non_none_type(field_type):
return next(arg for arg in get_args(field_type) if arg is not type(None))
def manually_validate_field(model: Type[BaseModel], field_name: str, value: Any):
def manually_validate_field(model: type[BaseModel], field_name: str, value: Any):
validators = model.__pydantic_decorators__.field_validators
for _name, validator in validators.items():
if field_name in validator.info.fields:
@ -126,7 +125,7 @@ def prompt_for_discriminated_union(
#
# doesn't support List[nested_class] yet or Dicts of any kind. needs a bunch of
# unit tests for coverage.
def prompt_for_config(config_type: type[BaseModel], existing_config: Optional[BaseModel] = None) -> BaseModel:
def prompt_for_config(config_type: type[BaseModel], existing_config: BaseModel | None = None) -> BaseModel:
"""
Recursively prompt the user for configuration values based on a Pydantic BaseModel.

13
llama_stack/log.py
View file

@ -7,7 +7,6 @@
import logging
import os
from logging.config import dictConfig
from typing import Dict, Optional
from rich.console import Console
from rich.errors import MarkupError
@ -33,7 +32,7 @@ CATEGORIES = [
]
# Initialize category levels with default level
_category_levels: Dict[str, int] = {category: DEFAULT_LOG_LEVEL for category in CATEGORIES}
_category_levels: dict[str, int] = {category: DEFAULT_LOG_LEVEL for category in CATEGORIES}
def config_to_category_levels(category: str, level: str):
@ -49,7 +48,7 @@ def config_to_category_levels(category: str, level: str):
Dict[str, int]: A dictionary mapping categories to their log levels.
"""
category_levels: Dict[str, int] = {}
category_levels: dict[str, int] = {}
level_value = logging._nameToLevel.get(str(level).upper())
if level_value is None:
logging.warning(f"Unknown log level '{level}' for category '{category}'. Falling back to default 'INFO'.")
@ -69,7 +68,7 @@ def config_to_category_levels(category: str, level: str):
return category_levels
def parse_yaml_config(yaml_config: LoggingConfig) -> Dict[str, int]:
def parse_yaml_config(yaml_config: LoggingConfig) -> dict[str, int]:
"""
Helper function to parse a yaml logging configuration found in the run.yaml
@ -86,7 +85,7 @@ def parse_yaml_config(yaml_config: LoggingConfig) -> Dict[str, int]:
return category_levels
def parse_environment_config(env_config: str) -> Dict[str, int]:
def parse_environment_config(env_config: str) -> dict[str, int]:
"""
Parse the LLAMA_STACK_LOGGING environment variable and return a dictionary of category log levels.
@ -131,7 +130,7 @@ class CustomRichHandler(RichHandler):
self.markup = original_markup
def setup_logging(category_levels: Dict[str, int], log_file: str | None) -> None:
def setup_logging(category_levels: dict[str, int], log_file: str | None) -> None:
"""
Configure logging based on the provided category log levels and an optional log file.
@ -211,7 +210,7 @@ def setup_logging(category_levels: Dict[str, int], log_file: str | None) -> None
def get_logger(
name: str, category: str = "uncategorized", config: Optional[LoggingConfig] | None = None
name: str, category: str = "uncategorized", config: LoggingConfig | None | None = None
) -> logging.LoggerAdapter:
"""
Returns a logger with the specified name and category.

22
llama_stack/models/llama/checkpoint.py
View file

@ -7,14 +7,14 @@
import concurrent.futures
import re
from pathlib import Path
from typing import Any, Dict, List, Optional, Union
from typing import Any
import numpy as np
import torch
from fairscale.nn.model_parallel.initialize import get_model_parallel_rank, get_model_parallel_world_size
def map_mp_rank(old_mp_size: int, new_mp_size: int, new_mp_rank: int) -> List[int]:
def map_mp_rank(old_mp_size: int, new_mp_size: int, new_mp_rank: int) -> list[int]:
"""Map a new MP rank to a list of old MP ranks given a change in MP size."""
if new_mp_size % old_mp_size == 0:
# Read old MP shard and split it into smaller ones
@ -31,12 +31,12 @@ def map_mp_rank(old_mp_size: int, new_mp_size: int, new_mp_rank: int) -> List[in
def maybe_reshard_state_dict(
ckpt_paths: List[Path],
ckpt_paths: list[Path],
n_kv_heads: int,
moe_num_experts: Optional[int] = None,
map_location: Union[str, torch.device] = "cpu",
moe_num_experts: int | None = None,
map_location: str | torch.device = "cpu",
mmap: bool = True,
) -> Dict[str, torch.Tensor]:
) -> dict[str, torch.Tensor]:
if str(map_location) == "cpu":
torch.set_default_tensor_type(torch.BFloat16Tensor)
else:
@ -97,18 +97,18 @@ _MOE_WEIGHT_COLUMN_KEY = {"feed_forward.experts.moe_w_out_eF_D"}
def reshard_mp(
state_dicts: List[Dict[str, torch.Tensor]],
state_dicts: list[dict[str, torch.Tensor]],
size: int,
rank: int,
repeat_qk_qv: int = 1,
) -> Dict[str, torch.Tensor]:
) -> dict[str, torch.Tensor]:
"""
Reshard a list of state dicts into a single state dict given a change in MP size.
If the list has more than one state dict, we concatenate the values of the same
key across all state dicts. Otherwise, we just slice it for the current MP rank.
"""
def concat_or_chunk(tensors: List[torch.Tensor], dim: int) -> torch.Tensor:
def concat_or_chunk(tensors: list[torch.Tensor], dim: int) -> torch.Tensor:
if len(tensors) > 1:
return torch.cat(tensors, dim=dim)
return tensors[0].chunk(size, dim=dim)[rank].clone()
@ -144,7 +144,7 @@ def reshard_mp(
column_regex = re.compile("|".join(column_keys))
row_regex = re.compile("|".join(row_keys))
output: Dict[str, torch.Tensor] = {}
output: dict[str, torch.Tensor] = {}
with concurrent.futures.ThreadPoolExecutor() as executor:
# Note: only processes keys in the first state dict.
# Assumes keys are the same across all state dicts.
@ -154,7 +154,7 @@ def reshard_mp(
return output
def convert_moe_weights(state_dict: Dict[str, Any], num_experts: int) -> Dict[str, Any]:
def convert_moe_weights(state_dict: dict[str, Any], num_experts: int) -> dict[str, Any]:
routed_keys = _MOE_WEIGHT_ROW_KEY | _MOE_WEIGHT_COLUMN_KEY
routed_regex = re.compile("|".join(routed_keys))
keys = list(state_dict.keys())

39
llama_stack/models/llama/datatypes.py
View file

@ -7,10 +7,9 @@
import base64
from enum import Enum
from io import BytesIO
from typing import Any, Dict, List, Literal, Optional, Union
from typing import Annotated, Any, Literal
from pydantic import BaseModel, ConfigDict, Field, field_serializer, field_validator
from typing_extensions import Annotated
# The goal is that these set of types are relevant for all Llama models.
# That isn't the current state yet -- e.g., BuiltinTool is somewhat specific to
@ -31,21 +30,21 @@ class BuiltinTool(Enum):
code_interpreter = "code_interpreter"
Primitive = Union[str, int, float, bool, None]
RecursiveType = Union[Primitive, List[Primitive], Dict[str, Primitive]]
Primitive = str | int | float | bool | None
RecursiveType = Primitive | list[Primitive] | dict[str, Primitive]
class ToolCall(BaseModel):
call_id: str
tool_name: Union[BuiltinTool, str]
tool_name: BuiltinTool | str
# Plan is to deprecate the Dict in favor of a JSON string
# that is parsed on the client side instead of trying to manage
# the recursive type here.
# Making this a union so that client side can start prepping for this change.
# Eventually, we will remove both the Dict and arguments_json field,
# and arguments will just be a str
arguments: Union[str, Dict[str, RecursiveType]]
arguments_json: Optional[str] = None
arguments: str | dict[str, RecursiveType]
arguments_json: str | None = None
@field_validator("tool_name", mode="before")
@classmethod
@ -91,15 +90,15 @@ class StopReason(Enum):
class ToolParamDefinition(BaseModel):
param_type: str
description: Optional[str] = None
required: Optional[bool] = True
default: Optional[Any] = None
description: str | None = None
required: bool | None = True
default: Any | None = None
class ToolDefinition(BaseModel):
tool_name: Union[BuiltinTool, str]
description: Optional[str] = None
parameters: Optional[Dict[str, ToolParamDefinition]] = None
tool_name: BuiltinTool | str
description: str | None = None
parameters: dict[str, ToolParamDefinition] | None = None
@field_validator("tool_name", mode="before")
@classmethod
@ -119,7 +118,7 @@ class RawMediaItem(BaseModel):
model_config = ConfigDict(arbitrary_types_allowed=True)
@field_serializer("data")
def serialize_data(self, data: Optional[bytes], _info):
def serialize_data(self, data: bytes | None, _info):
if data is None:
return None
return base64.b64encode(data).decode("utf-8")
@ -137,9 +136,9 @@ class RawTextItem(BaseModel):
text: str
RawContentItem = Annotated[Union[RawTextItem, RawMediaItem], Field(discriminator="type")]
RawContentItem = Annotated[RawTextItem | RawMediaItem, Field(discriminator="type")]
RawContent = str | RawContentItem | List[RawContentItem]
RawContent = str | RawContentItem | list[RawContentItem]
class RawMessage(BaseModel):
@ -147,17 +146,17 @@ class RawMessage(BaseModel):
content: RawContent
# This is for RAG but likely should be absorbed into content
context: Optional[RawContent] = None
context: RawContent | None = None
# These are for the output message coming from the assistant
stop_reason: Optional[StopReason] = None
tool_calls: List[ToolCall] = Field(default_factory=list)
stop_reason: StopReason | None = None
tool_calls: list[ToolCall] = Field(default_factory=list)
class GenerationResult(BaseModel):
token: int
text: str
logprobs: Optional[List[float]] = None
logprobs: list[float] | None = None
source: Literal["input"] | Literal["output"]

13
llama_stack/models/llama/llama3/args.py
View file

@ -6,7 +6,6 @@
from dataclasses import dataclass
from enum import Enum
from typing import Optional
class QuantizationScheme(Enum):
@ -15,8 +14,8 @@ class QuantizationScheme(Enum):
@dataclass
class QuantizationArgs:
scheme: Optional[QuantizationScheme] = None
group_size: Optional[int] = None
scheme: QuantizationScheme | None = None
group_size: int | None = None
spinquant: bool = False
def __init__(self, **kwargs):
@ -39,10 +38,10 @@ class ModelArgs:
dim: int = 4096
n_layers: int = 32
n_heads: int = 32
n_kv_heads: Optional[int] = None
n_kv_heads: int | None = None
vocab_size: int = -1
multiple_of: int = 256 # make SwiGLU hidden layer size multiple of large power of 2
ffn_dim_multiplier: Optional[float] = None
ffn_dim_multiplier: float | None = None
norm_eps: float = 1e-5
rope_theta: float = 500000
use_scaled_rope: bool = False
@ -55,8 +54,8 @@ class ModelArgs:
vision_max_num_chunks: int = 4
vision_num_cross_attention_layers: int = -1
quantization_args: Optional[QuantizationArgs] = None
lora_args: Optional[LoRAArgs] = None
quantization_args: QuantizationArgs | None = None
lora_args: LoRAArgs | None = None
def __init__(self, **kwargs):
for k, v in kwargs.items():

29
llama_stack/models/llama/llama3/chat_format.py
View file

@ -8,7 +8,6 @@ import io
import json
import uuid
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
from PIL import Image as PIL_Image
@ -29,14 +28,14 @@ from .tool_utils import ToolUtils
@dataclass
class VisionInput:
mask: List[List[int]]
images: List[PIL_Image.Image]
mask: list[list[int]]
images: list[PIL_Image.Image]
@dataclass
class LLMInput:
tokens: List[int]
vision: Optional[VisionInput] = None
tokens: list[int]
vision: VisionInput | None = None
def role_str(role: Role) -> str:
@ -50,7 +49,7 @@ def role_str(role: Role) -> str:
class ChatFormat:
possible_headers: Dict[Role, str]
possible_headers: dict[Role, str]
def __init__(self, tokenizer: Tokenizer):
self.tokenizer = tokenizer
@ -58,7 +57,7 @@ class ChatFormat:
self.possible_headers = {role: f"<|start_header_id|>{role_str(role)}<|end_header_id|>\n\n" for role in Role}
self.vision_token = self.tokenizer.special_tokens["<|image|>"]
def _encode_header(self, role: str) -> List[int]:
def _encode_header(self, role: str) -> list[int]:
tokens = []
tokens.append(self.tokenizer.special_tokens["<|start_header_id|>"])
tokens.extend(self.tokenizer.encode("ipython" if role == "tool" else role, bos=False, eos=False))
@ -70,7 +69,7 @@ class ChatFormat:
tokens, images = self._encode_content(content, bos=True)
return self._model_input_from_tokens_images(tokens, images)
def _encode_content(self, content: RawContent, bos: bool = False) -> Tuple[List[int], List[PIL_Image.Image]]:
def _encode_content(self, content: RawContent, bos: bool = False) -> tuple[list[int], list[PIL_Image.Image]]:
tokens = []
images = []
@ -107,7 +106,7 @@ class ChatFormat:
def encode_message(
self, message: RawMessage, tool_prompt_format: ToolPromptFormat
) -> Tuple[List[int], List[PIL_Image.Image]]:
) -> tuple[list[int], list[PIL_Image.Image]]:
tokens = self._encode_header(message.role)
images = []
@ -145,8 +144,8 @@ class ChatFormat:
def encode_dialog_prompt(
self,
messages: List[RawMessage],
tool_prompt_format: Optional[ToolPromptFormat] = None,
messages: list[RawMessage],
tool_prompt_format: ToolPromptFormat | None = None,
) -> LLMInput:
tool_prompt_format = tool_prompt_format or ToolPromptFormat.json
tokens = []
@ -163,7 +162,7 @@ class ChatFormat:
return self._model_input_from_tokens_images(tokens, images)
# TODO(this should be generic, not only for assistant messages)
def decode_assistant_message(self, tokens: List[int], stop_reason: StopReason) -> RawMessage:
def decode_assistant_message(self, tokens: list[int], stop_reason: StopReason) -> RawMessage:
content = self.tokenizer.decode(tokens)
return self.decode_assistant_message_from_content(content, stop_reason)
@ -234,7 +233,7 @@ class ChatFormat:
tool_calls=tool_calls,
)
def _model_input_from_tokens_images(self, tokens: List[int], images: List[PIL_Image.Image]) -> LLMInput:
def _model_input_from_tokens_images(self, tokens: list[int], images: list[PIL_Image.Image]) -> LLMInput:
vision_input = None
if len(images) > 0:
vision_input = VisionInput(
@ -249,9 +248,9 @@ class ChatFormat:
def create_vision_mask(
tokens: List[int],
tokens: list[int],
vision_token: int,
) -> List[List[int]]:
) -> list[list[int]]:
vision_token_locations = [i for i, token in enumerate(tokens) if token == vision_token]
if len(vision_token_locations) == 0:
return []

28
llama_stack/models/llama/llama3/generation.py
View file

@ -15,8 +15,8 @@ import json
import os
import sys
import time
from collections.abc import Callable, Generator
from pathlib import Path
from typing import Callable, Generator, List, Optional
import torch
import torch.nn.functional as F
@ -41,8 +41,8 @@ class Llama3:
ckpt_dir: str,
max_seq_len: int,
max_batch_size: int,
world_size: Optional[int] = None,
quantization_mode: Optional[QuantizationMode] = None,
world_size: int | None = None,
quantization_mode: QuantizationMode | None = None,
seed: int = 1,
device: str = "cuda",
):
@ -82,7 +82,7 @@ class Llama3:
ckpt_paths = sorted(Path(ckpt_dir).glob("*.pth"))
assert len(ckpt_paths) > 0, f"no checkpoint files found in {ckpt_dir}"
print(f"Loading a checkpoint (shards={len(ckpt_paths)}, current-mp-size={world_size})")
with open(Path(ckpt_dir) / "params.json", "r") as f:
with open(Path(ckpt_dir) / "params.json") as f:
params = json.loads(f.read())
model_args: ModelArgs = ModelArgs(
@ -154,15 +154,15 @@ class Llama3:
@torch.inference_mode()
def generate(
self,
llm_inputs: List[LLMInput],
llm_inputs: list[LLMInput],
temperature: float = 0.6,
top_p: float = 0.9,
max_gen_len: Optional[int] = None,
max_gen_len: int | None = None,
logprobs: bool = False,
echo: bool = False,
print_model_input: bool = False,
logits_processor: Optional[Callable[[torch.Tensor, torch.Tensor], torch.Tensor]] = None,
) -> Generator[List[GenerationResult], None, None]:
logits_processor: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] | None = None,
) -> Generator[list[GenerationResult], None, None]:
if max_gen_len is None or max_gen_len == 0 or max_gen_len >= self.args.max_seq_len:
max_gen_len = self.args.max_seq_len - 1
params = self.model.params
@ -302,13 +302,13 @@ class Llama3:
def completion(
self,
contents: List[RawContent],
contents: list[RawContent],
temperature: float = 0.6,
top_p: float = 0.9,
max_gen_len: Optional[int] = None,
max_gen_len: int | None = None,
logprobs: bool = False,
echo: bool = False,
) -> Generator[List[GenerationResult], None, None]:
) -> Generator[list[GenerationResult], None, None]:
model_inputs = [self.formatter.encode_content(c) for c in contents]
for result in self.generate(
model_inputs=model_inputs,
@ -324,14 +324,14 @@ class Llama3:
def chat_completion(
self,
messages_batch: List[List[RawMessage]],
messages_batch: list[list[RawMessage]],
temperature: float = 0.6,
top_p: float = 0.9,
max_gen_len: Optional[int] = None,
max_gen_len: int | None = None,
logprobs: bool = False,
tool_prompt_format: ToolPromptFormat = ToolPromptFormat.json,
echo: bool = False,
) -> Generator[List[GenerationResult], None, None]:
) -> Generator[list[GenerationResult], None, None]:
model_inputs = [self.formatter.encode_dialog_prompt(messages) for messages in messages_batch]
for result in self.generate(
model_inputs=model_inputs,

19
llama_stack/models/llama/llama3/interface.py
View file

@ -12,7 +12,6 @@
# the top-level of this source tree.
from pathlib import Path
from typing import List, Optional
from termcolor import colored
@ -131,7 +130,7 @@ class LLama31Interface:
self.formatter = ChatFormat(self.tokenizer)
self.tool_prompt_format = tool_prompt_format
def get_tokens(self, messages: List[RawMessage]) -> List[int]:
def get_tokens(self, messages: list[RawMessage]) -> list[int]:
model_input = self.formatter.encode_dialog_prompt(
messages,
self.tool_prompt_format,
@ -149,10 +148,10 @@ class LLama31Interface:
def system_messages(
self,
builtin_tools: List[BuiltinTool],
custom_tools: List[ToolDefinition],
instruction: Optional[str] = None,
) -> List[RawMessage]:
builtin_tools: list[BuiltinTool],
custom_tools: list[ToolDefinition],
instruction: str | None = None,
) -> list[RawMessage]:
messages = []
default_gen = SystemDefaultGenerator()
@ -194,8 +193,8 @@ class LLama31Interface:
self,
content: str,
stop_reason: StopReason,
tool_call: Optional[ToolCall] = None,
) -> List[RawMessage]:
tool_call: ToolCall | None = None,
) -> list[RawMessage]:
tool_calls = []
if tool_call:
tool_calls.append(tool_call)
@ -208,7 +207,7 @@ class LLama31Interface:
)
]
def user_message(self, content: str) -> List[RawMessage]:
def user_message(self, content: str) -> list[RawMessage]:
return [RawMessage(role="user", content=content)]
def display_message_as_tokens(self, message: RawMessage) -> None:
@ -228,7 +227,7 @@ class LLama31Interface:
print("\n", end="")
def list_jinja_templates() -> List[Template]:
def list_jinja_templates() -> list[Template]:
return TEMPLATES

9
llama_stack/models/llama/llama3/model.py
View file

@ -5,7 +5,6 @@
# the root directory of this source tree.
import math
from typing import Optional, Tuple
import fairscale.nn.model_parallel.initialize as fs_init
import torch
@ -80,7 +79,7 @@ def apply_rotary_emb(
xq: torch.Tensor,
xk: torch.Tensor,
freqs_cis: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
) -> tuple[torch.Tensor, torch.Tensor]:
xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
@ -162,7 +161,7 @@ class Attention(nn.Module):
x: torch.Tensor,
start_pos: int,
freqs_cis: torch.Tensor,
mask: Optional[torch.Tensor],
mask: torch.Tensor | None,
):
bsz, seqlen, _ = x.shape
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
@ -204,7 +203,7 @@ class FeedForward(nn.Module):
dim: int,
hidden_dim: int,
multiple_of: int,
ffn_dim_multiplier: Optional[float],
ffn_dim_multiplier: float | None,
):
super().__init__()
hidden_dim = int(2 * hidden_dim / 3)
@ -243,7 +242,7 @@ class TransformerBlock(nn.Module):
x: torch.Tensor,
start_pos: int,
freqs_cis: torch.Tensor,
mask: Optional[torch.Tensor],
mask: torch.Tensor | None,
):
h = x + self.attention(self.attention_norm(x), start_pos, freqs_cis, mask)
out = h + self.feed_forward(self.ffn_norm(h))

22
llama_stack/models/llama/llama3/multimodal/image_transform.py
View file

@ -14,7 +14,7 @@
import math
from collections import defaultdict
from logging import getLogger
from typing import Any, Optional, Set, Tuple
from typing import Any
import torch
import torchvision.transforms as tv
@ -26,7 +26,7 @@ IMAGE_RES = 224
logger = getLogger()
class VariableSizeImageTransform(object):
class VariableSizeImageTransform:
"""
This class accepts images of any size and dynamically resize, pads and chunks it
based on the image aspect ratio and the number of image chunks we allow.
@ -75,7 +75,7 @@ class VariableSizeImageTransform(object):
self.resample = tv.InterpolationMode.BILINEAR
@staticmethod
def get_factors(n: int) -> Set[int]:
def get_factors(n: int) -> set[int]:
"""
Calculate all factors of a given number, i.e. a dividor that leaves
no remainder. For example, if n=12, it will return {1, 2, 3, 4, 6, 12}.
@ -145,9 +145,9 @@ class VariableSizeImageTransform(object):
@staticmethod
def get_max_res_without_distortion(
image_size: Tuple[int, int],
target_size: Tuple[int, int],
) -> Tuple[int, int]:
image_size: tuple[int, int],
target_size: tuple[int, int],
) -> tuple[int, int]:
"""
Determines the maximum resolution to which an image can be resized to without distorting its
aspect ratio, based on the target resolution.
@ -198,8 +198,8 @@ class VariableSizeImageTransform(object):
def resize_without_distortion(
self,
image: torch.Tensor,
target_size: Tuple[int, int],
max_upscaling_size: Optional[int],
target_size: tuple[int, int],
max_upscaling_size: int | None,
) -> torch.Tensor:
"""
Used to resize an image to target_resolution, without distortion.
@ -261,10 +261,10 @@ class VariableSizeImageTransform(object):
def get_best_fit(
self,
image_size: Tuple[int, int],
image_size: tuple[int, int],
possible_resolutions: torch.Tensor,
resize_to_max_canvas: bool = False,
) -> Tuple[int, int]:
) -> tuple[int, int]:
"""
Determines the best canvas possible from a list of possible resolutions to, without distortion,
resize an image to.
@ -364,7 +364,7 @@ class VariableSizeImageTransform(object):
max_num_chunks: int,
normalize_img: bool = True,
resize_to_max_canvas: bool = False,
) -> Tuple[Any, Any]:
) -> tuple[Any, Any]:
"""
Args:
image (PIL.Image): Image to be resized.

41
llama_stack/models/llama/llama3/multimodal/model.py
View file

@ -6,8 +6,9 @@
import logging
import math
from collections.abc import Callable
from functools import partial
from typing import Any, Callable, Dict, List, Optional, Tuple, Union
from typing import Any
import fairscale.nn.model_parallel.initialize as fs_init
import torch
@ -104,9 +105,9 @@ class ColumnParallelConv2dPatch(torch.nn.Module):
self,
in_channels: int,
out_channels: int,
kernel_size: Union[int, Tuple[int, int]],
stride: Union[int, Tuple[int, int]],
bias: Optional[bool] = False,
kernel_size: int | tuple[int, int],
stride: int | tuple[int, int],
bias: bool | None = False,
) -> None:
super().__init__()
if isinstance(kernel_size, int):
@ -390,13 +391,13 @@ class VisionEncoder(nn.Module):
def load_hook(
self,
state_dict: Dict[str, Any],
state_dict: dict[str, Any],
prefix: str,
local_metadata: Dict[str, Any],
local_metadata: dict[str, Any],
strict: bool = True,
missing_keys: List[str] = None,
unexpected_keys: List[str] = None,
error_msgs: List[str] = None,
missing_keys: list[str] = None,
unexpected_keys: list[str] = None,
error_msgs: list[str] = None,
return_state_dict: bool = False,
) -> None:
orig_pos_embed = state_dict.get(prefix + "positional_embedding")
@ -641,7 +642,7 @@ class FeedForward(nn.Module):
dim: int,
hidden_dim: int,
multiple_of: int,
ffn_dim_multiplier: Optional[float],
ffn_dim_multiplier: float | None,
):
"""
Initialize the FeedForward module.
@ -983,7 +984,7 @@ class CrossAttentionTransformerBlock(torch.nn.Module):
self,
x: torch.Tensor,
xattn_mask: torch.Tensor,
full_text_row_masked_out_mask: Tuple[torch.Tensor, torch.Tensor],
full_text_row_masked_out_mask: tuple[torch.Tensor, torch.Tensor],
xattn_cache: torch.Tensor,
) -> torch.Tensor:
_attn_out = self.attention(
@ -1144,7 +1145,7 @@ class CrossAttentionTransformerText(torch.nn.Module):
def _init_fusion_schedule(
self,
num_layers: int,
) -> List[int]:
) -> list[int]:
llama_layers = list(range(self.n_llama_layers))
# uniformly spread the layers
@ -1231,7 +1232,7 @@ class CrossAttentionTransformerText(torch.nn.Module):
text_dtype,
vision_tokens,
cross_attention_masks,
) -> Tuple[Tensor, Tensor]:
) -> tuple[Tensor, Tensor]:
assert vision_tokens is not None, "Vision tokens must be provided"
vision_seqlen = vision_tokens.shape[3]
assert vision_tokens.shape[1] == cross_attention_masks.shape[2], (
@ -1280,11 +1281,11 @@ class CrossAttentionTransformer(torch.nn.Module):
def compute_vision_tokens_masks(
self,
batch_images: List[List[PIL_Image.Image]],
batch_masks: List[List[List[int]]],
batch_images: list[list[PIL_Image.Image]],
batch_masks: list[list[list[int]]],
total_len: int,
device: torch.device,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
skip_vision_encoder = False
assert len(batch_images) == len(batch_masks), "Images and masks must have the same length"
@ -1371,11 +1372,11 @@ class CrossAttentionTransformer(torch.nn.Module):
def _stack_images(
images: List[List[PIL_Image.Image]],
images: list[list[PIL_Image.Image]],
max_num_chunks: int,
image_res: int,
max_num_images: int,
) -> Tuple[torch.Tensor, List[int]]:
) -> tuple[torch.Tensor, list[int]]:
"""
Takes a list of list of images and stacks them into a tensor.
This function is needed since images can be of completely
@ -1400,8 +1401,8 @@ def _stack_images(
def _pad_masks(
all_masks: List[List[List[int]]],
all_num_chunks: List[List[int]],
all_masks: list[list[list[int]]],
all_num_chunks: list[list[int]],
total_len: int,
max_num_chunks: int,
) -> torch.Tensor:

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