phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-08-02 00:34:44 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into feat/litellm_sambanova_usage

Browse source
This commit is contained in:
jhpiedrahitao 2025-05-05 11:49:58 -05:00
commit b7f16ac7a6
535 changed files with 23539 additions and 8112 deletions
Download patch file Download diff file Expand all files Collapse all files

6
.coveragerc Normal file
View file

@ -0,0 +1,6 @@
[run]
omit =
*/tests/*
*/llama_stack/providers/*
*/llama_stack/templates/*
.venv/*

26
.github/workflows/install-script-ci.yml vendored Normal file
View file

@ -0,0 +1,26 @@
name: Installer CI
on:
pull_request:
paths:
- 'install.sh'
push:
paths:
- 'install.sh'
schedule:
- cron: '0 2 * * *' # every day at 02:00 UTC
jobs:
lint:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # 4.2.2
- name: Run ShellCheck on install.sh
run: shellcheck install.sh
smoke-test:
needs: lint
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # 4.2.2
- name: Run installer end-to-end
run: ./install.sh

136
.github/workflows/integration-auth-tests.yml vendored Normal file
View file

@ -0,0 +1,136 @@
name: Integration Auth Tests
on:
push:
branches: [ main ]
pull_request:
branches: [ main ]
paths:
- 'distributions/**'
- 'llama_stack/**'
- 'tests/integration/**'
- 'uv.lock'
- 'pyproject.toml'
- 'requirements.txt'
- '.github/workflows/integration-auth-tests.yml' # This workflow
concurrency:
group: ${{ github.workflow }}-${{ github.ref }}
cancel-in-progress: true
jobs:
test-matrix:
runs-on: ubuntu-latest
strategy:
matrix:
auth-provider: [kubernetes]
fail-fast: false # we want to run all tests regardless of failure
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install uv
uses: astral-sh/setup-uv@v5
with:
python-version: "3.10"
- name: Set Up Environment and Install Dependencies
run: |
uv sync --extra dev --extra test
uv pip install -e .
llama stack build --template ollama --image-type venv
- name: Install minikube
if: ${{ matrix.auth-provider == 'kubernetes' }}
uses: medyagh/setup-minikube@latest
- name: Start minikube
if: ${{ matrix.auth-provider == 'kubernetes' }}
run: |
minikube start
kubectl get pods -A
- name: Configure Kube Auth
if: ${{ matrix.auth-provider == 'kubernetes' }}
run: |
kubectl create namespace llama-stack
kubectl create serviceaccount llama-stack-auth -n llama-stack
kubectl create rolebinding llama-stack-auth-rolebinding --clusterrole=admin --serviceaccount=llama-stack:llama-stack-auth -n llama-stack
kubectl create token llama-stack-auth -n llama-stack > llama-stack-auth-token
- name: Set Kubernetes Config
if: ${{ matrix.auth-provider == 'kubernetes' }}
run: |
echo "KUBERNETES_API_SERVER_URL=$(kubectl config view --minify -o jsonpath='{.clusters[0].cluster.server}')" >> $GITHUB_ENV
echo "KUBERNETES_CA_CERT_PATH=$(kubectl config view --minify -o jsonpath='{.clusters[0].cluster.certificate-authority}')" >> $GITHUB_ENV
- name: Set Kube Auth Config and run server
env:
INFERENCE_MODEL: "meta-llama/Llama-3.2-3B-Instruct"
if: ${{ matrix.auth-provider == 'kubernetes' }}
run: |
run_dir=$(mktemp -d)
cat <<'EOF' > $run_dir/run.yaml
version: '2'
image_name: kube
apis:
- agents
- datasetio
- eval
- inference
- safety
- scoring
- telemetry
- tool_runtime
- vector_io
providers:
agents:
- provider_id: meta-reference
provider_type: inline::meta-reference
config:
persistence_store:
type: sqlite
namespace: null
db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/ollama}/agents_store.db
telemetry:
- provider_id: meta-reference
provider_type: inline::meta-reference
config:
service_name: "${env.OTEL_SERVICE_NAME:\u200B}"
sinks: ${env.TELEMETRY_SINKS:console,sqlite}
sqlite_db_path: ${env.SQLITE_DB_PATH:~/.llama/distributions/ollama/trace_store.db}
server:
port: 8321
EOF
yq eval '.server.auth = {"provider_type": "${{ matrix.auth-provider }}"}' -i $run_dir/run.yaml
yq eval '.server.auth.config = {"api_server_url": "${{ env.KUBERNETES_API_SERVER_URL }}", "ca_cert_path": "${{ env.KUBERNETES_CA_CERT_PATH }}"}' -i $run_dir/run.yaml
cat $run_dir/run.yaml
source .venv/bin/activate
nohup uv run llama stack run $run_dir/run.yaml --image-type venv > server.log 2>&1 &
- name: Wait for Llama Stack server to be ready
run: |
echo "Waiting for Llama Stack server..."
for i in {1..30}; do
if curl -s -L -H "Authorization: Bearer $(cat llama-stack-auth-token)" http://localhost:8321/v1/health | grep -q "OK"; then
echo "Llama Stack server is up!"
if grep -q "Enabling authentication with provider: ${{ matrix.auth-provider }}" server.log; then
echo "Llama Stack server is configured to use ${{ matrix.auth-provider }} auth"
exit 0
else
echo "Llama Stack server is not configured to use ${{ matrix.auth-provider }} auth"
cat server.log
exit 1
fi
fi
sleep 1
done
echo "Llama Stack server failed to start"
cat server.log
exit 1
- name: Test auth
run: |
curl -s -L -H "Authorization: Bearer $(cat llama-stack-auth-token)" http://127.0.0.1:8321/v1/providers|jq

25
.github/workflows/integration-tests.yml vendored
View file

@ -6,7 +6,6 @@ on:
pull_request:
branches: [ main ]
paths:
- 'distributions/**'
- 'llama_stack/**'
- 'tests/integration/**'
- 'uv.lock'
@ -34,19 +33,24 @@ jobs:
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Install uv
uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1
uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
with:
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
- name: Pull Ollama image
# 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: |
# TODO: cache the model. OLLAMA_MODELS defaults to ~ollama/.ollama/models.
ollama pull llama3.2:3b-instruct-fp16
ollama pull all-minilm:latest
- name: Set Up Environment and Install Dependencies
run: |
@ -106,3 +110,16 @@ jobs:
-k "not(builtin_tool or safety_with_image or code_interpreter or test_rag)" \
--text-model="meta-llama/Llama-3.2-3B-Instruct" \
--embedding-model=all-MiniLM-L6-v2
- name: Write ollama logs to file
run: |
sudo journalctl -u ollama.service > ollama.log
- name: Upload all logs to artifacts
if: always()
uses: actions/upload-artifact@v4
with:
name: logs-${{ github.run_id }}-${{ github.run_attempt }}-${{ matrix.client-type }}-${{ matrix.test-type }}
path: |
*.log
retention-days: 1

4
.github/workflows/pre-commit.yml vendored
View file

@ -18,7 +18,7 @@ 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: '3.11'
cache: pip
@ -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: |

121
.github/workflows/providers-build.yml vendored
View file

@ -51,12 +51,12 @@ 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'
- name: Install uv
uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1
uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
with:
python-version: "3.10"
@ -81,3 +81,120 @@ jobs:
run: |
source test/bin/activate
uv pip list
build-single-provider:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Set up Python
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: '3.10'
- name: Install uv
uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
with:
python-version: "3.10"
- name: Install LlamaStack
run: |
uv venv
source .venv/bin/activate
uv pip install -e .
- name: Build a single provider
run: |
USE_COPY_NOT_MOUNT=true LLAMA_STACK_DIR=. uv run llama stack build --image-type venv --image-name test --providers inference=remote::ollama
build-custom-container-distribution:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Set up Python
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: '3.10'
- name: Install uv
uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
with:
python-version: "3.10"
- name: Install LlamaStack
run: |
uv venv
source .venv/bin/activate
uv pip install -e .
- name: Build a single provider
run: |
yq -i '.image_type = "container"' llama_stack/templates/dev/build.yaml
yq -i '.image_name = "test"' llama_stack/templates/dev/build.yaml
USE_COPY_NOT_MOUNT=true LLAMA_STACK_DIR=. uv run llama stack build --config llama_stack/templates/dev/build.yaml
- name: Inspect the container image entrypoint
run: |
IMAGE_ID=$(docker images --format "{{.Repository}}:{{.Tag}}" | head -n 1)
entrypoint=$(docker inspect --format '{{ .Config.Entrypoint }}' $IMAGE_ID)
echo "Entrypoint: $entrypoint"
if [ "$entrypoint" != "[python -m llama_stack.distribution.server.server --config /app/run.yaml]" ]; then
echo "Entrypoint is not correct"
exit 1
fi
build-ubi9-container-distribution:
runs-on: ubuntu-latest
steps:
- name: Checkout repository
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Set up Python
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: '3.10'
- name: Install uv
uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1
with:
python-version: "3.10"
- name: Install LlamaStack
run: |
uv venv
source .venv/bin/activate
uv pip install -e .
- name: Pin template to UBI9 base
run: |
yq -i '
.image_type = "container" |
.image_name = "ubi9-test" |
.distribution_spec.container_image = "registry.access.redhat.com/ubi9:latest"
' llama_stack/templates/dev/build.yaml
- name: Build dev container (UBI9)
env:
USE_COPY_NOT_MOUNT: "true"
LLAMA_STACK_DIR: "."
run: |
uv run llama stack build --config llama_stack/templates/dev/build.yaml
- name: Inspect UBI9 image
run: |
IMAGE_ID=$(docker images --format "{{.Repository}}:{{.Tag}}" | head -n 1)
entrypoint=$(docker inspect --format '{{ .Config.Entrypoint }}' $IMAGE_ID)
echo "Entrypoint: $entrypoint"
if [ "$entrypoint" != "[python -m llama_stack.distribution.server.server --config /app/run.yaml]" ]; then
echo "Entrypoint is not correct"
exit 1
fi
echo "Checking /etc/os-release in $IMAGE_ID"
docker run --rm --entrypoint sh "$IMAGE_ID" -c \
'source /etc/os-release && echo "$ID"' \
| grep -qE '^(rhel|ubi)$' \
|| { echo "Base image is not UBI 9!"; exit 1; }

79
.github/workflows/test-external-providers.yml vendored
View file

@ -5,89 +5,74 @@ on:
branches: [ main ]
pull_request:
branches: [ main ]
paths:
- 'llama_stack/**'
- 'tests/integration/**'
- 'uv.lock'
- 'pyproject.toml'
- 'requirements.txt'
- '.github/workflows/test-external-providers.yml' # This workflow
jobs:
test-external-providers:
runs-on: ubuntu-latest
strategy:
matrix:
image-type: [venv]
# We don't do container yet, it's tricky to install a package from the host into the
# container and point 'uv pip install' to the correct path...
steps:
- name: Checkout repository
uses: actions/checkout@v4
- name: Install uv
uses: astral-sh/setup-uv@v5
uses: astral-sh/setup-uv@v6
with:
python-version: "3.10"
- name: Install Ollama
run: |
curl -fsSL https://ollama.com/install.sh | sh
- name: Pull Ollama image
run: |
ollama pull llama3.2:3b-instruct-fp16
- name: Start Ollama in background
run: |
nohup ollama run llama3.2:3b-instruct-fp16 --keepalive=30m > ollama.log 2>&1 &
- name: Set Up Environment and Install Dependencies
run: |
uv sync --extra dev --extra test
uv pip install -e .
- name: Install Ollama custom provider
- name: Apply image type to config file
run: |
yq -i '.image_type = "${{ matrix.image-type }}"' tests/external-provider/llama-stack-provider-ollama/custom-distro.yaml
cat tests/external-provider/llama-stack-provider-ollama/custom-distro.yaml
- name: Setup directory for Ollama custom provider
run: |
mkdir -p tests/external-provider/llama-stack-provider-ollama/src/
cp -a llama_stack/providers/remote/inference/ollama/ tests/external-provider/llama-stack-provider-ollama/src/llama_stack_provider_ollama
uv pip install tests/external-provider/llama-stack-provider-ollama
- name: Create provider configuration
run: |
mkdir -p /tmp/providers.d/remote/inference
cp tests/external-provider/llama-stack-provider-ollama/custom_ollama.yaml /tmp/providers.d/remote/inference/custom_ollama.yaml
- name: Wait for Ollama to start
- name: Build distro from config file
run: |
echo "Waiting for Ollama..."
for i in {1..30}; do
if curl -s http://localhost:11434 | grep -q "Ollama is running"; then
echo "Ollama is running!"
exit 0
fi
sleep 1
done
echo "Ollama failed to start"
ollama ps
ollama.log
exit 1
USE_COPY_NOT_MOUNT=true LLAMA_STACK_DIR=. uv run llama stack build --config tests/external-provider/llama-stack-provider-ollama/custom-distro.yaml
- name: Start Llama Stack server in background
if: ${{ matrix.image-type }} == 'venv'
env:
INFERENCE_MODEL: "meta-llama/Llama-3.2-3B-Instruct"
run: |
source .venv/bin/activate
nohup uv run llama stack run tests/external-provider/llama-stack-provider-ollama/run.yaml --image-type venv > server.log 2>&1 &
source ci-test/bin/activate
uv run pip list
nohup uv run --active llama stack run tests/external-provider/llama-stack-provider-ollama/run.yaml --image-type ${{ matrix.image-type }} > server.log 2>&1 &
- name: Wait for Llama Stack server to be ready
run: |
echo "Waiting for Llama Stack server..."
for i in {1..30}; do
if curl -s http://localhost:8321/v1/health | grep -q "OK"; then
echo "Llama Stack server is up!"
if grep -q "remote::custom_ollama from /tmp/providers.d/remote/inference/custom_ollama.yaml" server.log; then
echo "Llama Stack server is using custom Ollama provider"
exit 0
else
echo "Llama Stack server is not using custom Ollama provider"
exit 1
fi
if ! grep -q "remote::custom_ollama from /tmp/providers.d/remote/inference/custom_ollama.yaml" server.log; then
echo "Waiting for Llama Stack server to load the provider..."
sleep 1
else
echo "Provider loaded"
exit 0
fi
sleep 1
done
echo "Llama Stack server failed to start"
cat server.log
echo "Provider failed to load"
exit 1
- name: run inference tests
run: |
uv run pytest -v tests/integration/inference/test_text_inference.py --stack-config="http://localhost:8321" --text-model="meta-llama/Llama-3.2-3B-Instruct" --embedding-model=all-MiniLM-L6-v2

5
.github/workflows/unit-tests.yml vendored
View file

@ -6,7 +6,6 @@ on:
pull_request:
branches: [ main ]
paths:
- 'distributions/**'
- 'llama_stack/**'
- 'tests/unit/**'
- 'uv.lock'
@ -34,11 +33,11 @@ jobs:
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
- name: Set up Python ${{ matrix.python }}
uses: actions/setup-python@8d9ed9ac5c53483de85588cdf95a591a75ab9f55 # v5.5.0
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
with:
python-version: ${{ matrix.python }}
- uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1
- uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
with:
python-version: ${{ matrix.python }}
enable-cache: false

4
.github/workflows/update-readthedocs.yml vendored
View file

@ -36,12 +36,12 @@ 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: '3.11'
- name: Install the latest version of uv
uses: astral-sh/setup-uv@0c5e2b8115b80b4c7c5ddf6ffdd634974642d182 # v5.4.1
uses: astral-sh/setup-uv@c7f87aa956e4c323abf06d5dec078e358f6b4d04 # v6.0.0
- name: Sync with uv
run: uv sync --extra docs

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

590
CHANGELOG.md
View file

@ -1,5 +1,33 @@
# Changelog
# 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
---
# 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
---
# v0.2.1
Published on: 2025-04-05T23:13:00Z
@ -10,10 +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
---
@ -21,58 +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
---
@ -80,73 +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
---
@ -154,176 +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
---
@ -337,8 +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
---
@ -374,39 +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

9
README.md
View file

@ -70,6 +70,13 @@ As more providers start supporting Llama 4, you can use them in Llama Stack as w
</details>
### 🚀 One-Line Installer 🚀
To try Llama Stack locally, run:
```bash
curl -LsSf https://github.com/meta-llama/llama-stack/raw/main/install.sh | sh
```
### Overview
@ -119,6 +126,7 @@ Here is a list of the various API providers and available distributions that can
| OpenAI | Hosted | | ✅ | | | |
| Anthropic | Hosted | | ✅ | | | |
| Gemini | Hosted | | ✅ | | | |
| watsonx | Hosted | | ✅ | | | |
### Distributions
@ -128,7 +136,6 @@ A Llama Stack Distribution (or "distro") is a pre-configured bundle of provider
| **Distribution** | **Llama Stack Docker** | Start This Distribution |
|:---------------------------------------------:|:-------------------------------------------------------------------------------------------------------------------------------------------------------------:|:------------------------------------------------------------------------------------------------------------------------:|
| Meta Reference | [llamastack/distribution-meta-reference-gpu](https://hub.docker.com/repository/docker/llamastack/distribution-meta-reference-gpu/general) | [Guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/meta-reference-gpu.html) |
| Meta Reference Quantized | [llamastack/distribution-meta-reference-quantized-gpu](https://hub.docker.com/repository/docker/llamastack/distribution-meta-reference-quantized-gpu/general) | [Guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/meta-reference-quantized-gpu.html) |
| SambaNova | [llamastack/distribution-sambanova](https://hub.docker.com/repository/docker/llamastack/distribution-sambanova/general) | [Guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/sambanova.html) |
| Cerebras | [llamastack/distribution-cerebras](https://hub.docker.com/repository/docker/llamastack/distribution-cerebras/general) | [Guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/cerebras.html) |
| Ollama | [llamastack/distribution-ollama](https://hub.docker.com/repository/docker/llamastack/distribution-ollama/general) | [Guide](https://llama-stack.readthedocs.io/en/latest/distributions/self_hosted_distro/ollama.html) |

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;
}

29
docs/_static/js/detect_theme.js vendored
View file

@ -1,9 +1,32 @@
document.addEventListener("DOMContentLoaded", function () {
const prefersDark = window.matchMedia("(prefers-color-scheme: dark)").matches;
const htmlElement = document.documentElement;
if (prefersDark) {
htmlElement.setAttribute("data-theme", "dark");
// Check if theme is saved in localStorage
const savedTheme = localStorage.getItem("sphinx-rtd-theme");
if (savedTheme) {
// Use the saved theme preference
htmlElement.setAttribute("data-theme", savedTheme);
document.body.classList.toggle("dark", savedTheme === "dark");
} else {
htmlElement.setAttribute("data-theme", "light");
// Fall back to system preference
const theme = prefersDark ? "dark" : "light";
htmlElement.setAttribute("data-theme", theme);
document.body.classList.toggle("dark", theme === "dark");
// Save initial preference
localStorage.setItem("sphinx-rtd-theme", theme);
}
// Listen for theme changes from the existing toggle
const observer = new MutationObserver(function(mutations) {
mutations.forEach(function(mutation) {
if (mutation.attributeName === "data-theme") {
const currentTheme = htmlElement.getAttribute("data-theme");
localStorage.setItem("sphinx-rtd-theme", currentTheme);
}
});
});
observer.observe(htmlElement, { attributes: true });
});

537
docs/_static/llama-stack-spec.html vendored
View file

@ -497,6 +497,54 @@
}
}
},
"/v1/openai/v1/responses": {
"post": {
"responses": {
"200": {
"description": "Runtime representation of an annotated type.",
"content": {
"application/json": {
"schema": {
"$ref": "#/components/schemas/OpenAIResponseObject"
}
},
"text/event-stream": {
"schema": {
"$ref": "#/components/schemas/OpenAIResponseObjectStream"
}
}
}
},
"400": {
"$ref": "#/components/responses/BadRequest400"
},
"429": {
"$ref": "#/components/responses/TooManyRequests429"
},
"500": {
"$ref": "#/components/responses/InternalServerError500"
},
"default": {
"$ref": "#/components/responses/DefaultError"
}
},
"tags": [
"Agents"
],
"description": "Create a new OpenAI response.",
"parameters": [],
"requestBody": {
"content": {
"application/json": {
"schema": {
"$ref": "#/components/schemas/CreateOpenaiResponseRequest"
}
}
},
"required": true
}
}
},
"/v1/files": {
"get": {
"responses": {
@ -1278,6 +1326,49 @@
]
}
},
"/v1/openai/v1/responses/{id}": {
"get": {
"responses": {
"200": {
"description": "An OpenAIResponseObject.",
"content": {
"application/json": {
"schema": {
"$ref": "#/components/schemas/OpenAIResponseObject"
}
}
}
},
"400": {
"$ref": "#/components/responses/BadRequest400"
},
"429": {
"$ref": "#/components/responses/TooManyRequests429"
},
"500": {
"$ref": "#/components/responses/InternalServerError500"
},
"default": {
"$ref": "#/components/responses/DefaultError"
}
},
"tags": [
"Agents"
],
"description": "Retrieve an OpenAI response by its ID.",
"parameters": [
{
"name": "id",
"in": "path",
"description": "The ID of the OpenAI response to retrieve.",
"required": true,
"schema": {
"type": "string"
}
}
]
}
},
"/v1/scoring-functions/{scoring_fn_id}": {
"get": {
"responses": {
@ -5221,17 +5312,25 @@
"default": 10
},
"model": {
"type": "string"
"type": "string",
"description": "The model identifier to use for the agent"
},
"instructions": {
"type": "string"
"type": "string",
"description": "The system instructions for the agent"
},
"name": {
"type": "string",
"description": "Optional name for the agent, used in telemetry and identification"
},
"enable_session_persistence": {
"type": "boolean",
"default": false
"default": false,
"description": "Optional flag indicating whether session data has to be persisted"
},
"response_format": {
"$ref": "#/components/schemas/ResponseFormat"
"$ref": "#/components/schemas/ResponseFormat",
"description": "Optional response format configuration"
}
},
"additionalProperties": false,
@ -5239,7 +5338,8 @@
"model",
"instructions"
],
"title": "AgentConfig"
"title": "AgentConfig",
"description": "Configuration for an agent."
},
"AgentTool": {
"oneOf": [
@ -6183,6 +6283,430 @@
],
"title": "AgentTurnResponseTurnStartPayload"
},
"OpenAIResponseInputMessage": {
"type": "object",
"properties": {
"content": {
"oneOf": [
{
"type": "string"
},
{
"type": "array",
"items": {
"$ref": "#/components/schemas/OpenAIResponseInputMessageContent"
}
}
]
},
"role": {
"oneOf": [
{
"type": "string",
"const": "system"
},
{
"type": "string",
"const": "developer"
},
{
"type": "string",
"const": "user"
},
{
"type": "string",
"const": "assistant"
}
]
},
"type": {
"type": "string",
"const": "message",
"default": "message"
}
},
"additionalProperties": false,
"required": [
"content",
"role"
],
"title": "OpenAIResponseInputMessage"
},
"OpenAIResponseInputMessageContent": {
"oneOf": [
{
"$ref": "#/components/schemas/OpenAIResponseInputMessageContentText"
},
{
"$ref": "#/components/schemas/OpenAIResponseInputMessageContentImage"
}
],
"discriminator": {
"propertyName": "type",
"mapping": {
"input_text": "#/components/schemas/OpenAIResponseInputMessageContentText",
"input_image": "#/components/schemas/OpenAIResponseInputMessageContentImage"
}
}
},
"OpenAIResponseInputMessageContentImage": {
"type": "object",
"properties": {
"detail": {
"oneOf": [
{
"type": "string",
"const": "low"
},
{
"type": "string",
"const": "high"
},
{
"type": "string",
"const": "auto"
}
],
"default": "auto"
},
"type": {
"type": "string",
"const": "input_image",
"default": "input_image"
},
"image_url": {
"type": "string"
}
},
"additionalProperties": false,
"required": [
"detail",
"type"
],
"title": "OpenAIResponseInputMessageContentImage"
},
"OpenAIResponseInputMessageContentText": {
"type": "object",
"properties": {
"text": {
"type": "string"
},
"type": {
"type": "string",
"const": "input_text",
"default": "input_text"
}
},
"additionalProperties": false,
"required": [
"text",
"type"
],
"title": "OpenAIResponseInputMessageContentText"
},
"OpenAIResponseInputTool": {
"type": "object",
"properties": {
"type": {
"oneOf": [
{
"type": "string",
"const": "web_search"
},
{
"type": "string",
"const": "web_search_preview_2025_03_11"
}
],
"default": "web_search"
},
"search_context_size": {
"type": "string",
"default": "medium"
}
},
"additionalProperties": false,
"required": [
"type"
],
"title": "OpenAIResponseInputToolWebSearch"
},
"CreateOpenaiResponseRequest": {
"type": "object",
"properties": {
"input": {
"oneOf": [
{
"type": "string"
},
{
"type": "array",
"items": {
"$ref": "#/components/schemas/OpenAIResponseInputMessage"
}
}
],
"description": "Input message(s) to create the response."
},
"model": {
"type": "string",
"description": "The underlying LLM used for completions."
},
"previous_response_id": {
"type": "string",
"description": "(Optional) if specified, the new response will be a continuation of the previous response. This can be used to easily fork-off new responses from existing responses."
},
"store": {
"type": "boolean"
},
"stream": {
"type": "boolean"
},
"temperature": {
"type": "number"
},
"tools": {
"type": "array",
"items": {
"$ref": "#/components/schemas/OpenAIResponseInputTool"
}
}
},
"additionalProperties": false,
"required": [
"input",
"model"
],
"title": "CreateOpenaiResponseRequest"
},
"OpenAIResponseError": {
"type": "object",
"properties": {
"code": {
"type": "string"
},
"message": {
"type": "string"
}
},
"additionalProperties": false,
"required": [
"code",
"message"
],
"title": "OpenAIResponseError"
},
"OpenAIResponseObject": {
"type": "object",
"properties": {
"created_at": {
"type": "integer"
},
"error": {
"$ref": "#/components/schemas/OpenAIResponseError"
},
"id": {
"type": "string"
},
"model": {
"type": "string"
},
"object": {
"type": "string",
"const": "response",
"default": "response"
},
"output": {
"type": "array",
"items": {
"$ref": "#/components/schemas/OpenAIResponseOutput"
}
},
"parallel_tool_calls": {
"type": "boolean",
"default": false
},
"previous_response_id": {
"type": "string"
},
"status": {
"type": "string"
},
"temperature": {
"type": "number"
},
"top_p": {
"type": "number"
},
"truncation": {
"type": "string"
},
"user": {
"type": "string"
}
},
"additionalProperties": false,
"required": [
"created_at",
"id",
"model",
"object",
"output",
"parallel_tool_calls",
"status"
],
"title": "OpenAIResponseObject"
},
"OpenAIResponseOutput": {
"oneOf": [
{
"$ref": "#/components/schemas/OpenAIResponseOutputMessage"
},
{
"$ref": "#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall"
}
],
"discriminator": {
"propertyName": "type",
"mapping": {
"message": "#/components/schemas/OpenAIResponseOutputMessage",
"web_search_call": "#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall"
}
}
},
"OpenAIResponseOutputMessage": {
"type": "object",
"properties": {
"id": {
"type": "string"
},
"content": {
"type": "array",
"items": {
"$ref": "#/components/schemas/OpenAIResponseOutputMessageContent"
}
},
"role": {
"type": "string",
"const": "assistant",
"default": "assistant"
},
"status": {
"type": "string"
},
"type": {
"type": "string",
"const": "message",
"default": "message"
}
},
"additionalProperties": false,
"required": [
"id",
"content",
"role",
"status",
"type"
],
"title": "OpenAIResponseOutputMessage"
},
"OpenAIResponseOutputMessageContent": {
"type": "object",
"properties": {
"text": {
"type": "string"
},
"type": {
"type": "string",
"const": "output_text",
"default": "output_text"
}
},
"additionalProperties": false,
"required": [
"text",
"type"
],
"title": "OpenAIResponseOutputMessageContentOutputText"
},
"OpenAIResponseOutputMessageWebSearchToolCall": {
"type": "object",
"properties": {
"id": {
"type": "string"
},
"status": {
"type": "string"
},
"type": {
"type": "string",
"const": "web_search_call",
"default": "web_search_call"
}
},
"additionalProperties": false,
"required": [
"id",
"status",
"type"
],
"title": "OpenAIResponseOutputMessageWebSearchToolCall"
},
"OpenAIResponseObjectStream": {
"oneOf": [
{
"$ref": "#/components/schemas/OpenAIResponseObjectStreamResponseCreated"
},
{
"$ref": "#/components/schemas/OpenAIResponseObjectStreamResponseCompleted"
}
],
"discriminator": {
"propertyName": "type",
"mapping": {
"response.created": "#/components/schemas/OpenAIResponseObjectStreamResponseCreated",
"response.completed": "#/components/schemas/OpenAIResponseObjectStreamResponseCompleted"
}
}
},
"OpenAIResponseObjectStreamResponseCompleted": {
"type": "object",
"properties": {
"response": {
"$ref": "#/components/schemas/OpenAIResponseObject"
},
"type": {
"type": "string",
"const": "response.completed",
"default": "response.completed"
}
},
"additionalProperties": false,
"required": [
"response",
"type"
],
"title": "OpenAIResponseObjectStreamResponseCompleted"
},
"OpenAIResponseObjectStreamResponseCreated": {
"type": "object",
"properties": {
"response": {
"$ref": "#/components/schemas/OpenAIResponseObject"
},
"type": {
"type": "string",
"const": "response.created",
"default": "response.created"
}
},
"additionalProperties": false,
"required": [
"response",
"type"
],
"title": "OpenAIResponseObjectStreamResponseCreated"
},
"CreateUploadSessionRequest": {
"type": "object",
"properties": {
@ -8891,8 +9415,7 @@
},
"additionalProperties": false,
"required": [
"role",
"content"
"role"
],
"title": "OpenAIAssistantMessageParam",
"description": "A message containing the model's (assistant) response in an OpenAI-compatible chat completion request."

364
docs/_static/llama-stack-spec.yaml vendored
View file

@ -330,6 +330,39 @@ paths:
schema:
$ref: '#/components/schemas/CreateAgentTurnRequest'
required: true
/v1/openai/v1/responses:
post:
responses:
'200':
description: >-
Runtime representation of an annotated type.
content:
application/json:
schema:
$ref: '#/components/schemas/OpenAIResponseObject'
text/event-stream:
schema:
$ref: '#/components/schemas/OpenAIResponseObjectStream'
'400':
$ref: '#/components/responses/BadRequest400'
'429':
$ref: >-
#/components/responses/TooManyRequests429
'500':
$ref: >-
#/components/responses/InternalServerError500
default:
$ref: '#/components/responses/DefaultError'
tags:
- Agents
description: Create a new OpenAI response.
parameters: []
requestBody:
content:
application/json:
schema:
$ref: '#/components/schemas/CreateOpenaiResponseRequest'
required: true
/v1/files:
get:
responses:
@ -875,6 +908,36 @@ paths:
required: true
schema:
type: string
/v1/openai/v1/responses/{id}:
get:
responses:
'200':
description: An OpenAIResponseObject.
content:
application/json:
schema:
$ref: '#/components/schemas/OpenAIResponseObject'
'400':
$ref: '#/components/responses/BadRequest400'
'429':
$ref: >-
#/components/responses/TooManyRequests429
'500':
$ref: >-
#/components/responses/InternalServerError500
default:
$ref: '#/components/responses/DefaultError'
tags:
- Agents
description: Retrieve an OpenAI response by its ID.
parameters:
- name: id
in: path
description: >-
The ID of the OpenAI response to retrieve.
required: true
schema:
type: string
/v1/scoring-functions/{scoring_fn_id}:
get:
responses:
@ -3686,18 +3749,29 @@ components:
default: 10
model:
type: string
description: >-
The model identifier to use for the agent
instructions:
type: string
description: The system instructions for the agent
name:
type: string
description: >-
Optional name for the agent, used in telemetry and identification
enable_session_persistence:
type: boolean
default: false
description: >-
Optional flag indicating whether session data has to be persisted
response_format:
$ref: '#/components/schemas/ResponseFormat'
description: Optional response format configuration
additionalProperties: false
required:
- model
- instructions
title: AgentConfig
description: Configuration for an agent.
AgentTool:
oneOf:
- type: string
@ -4318,6 +4392,295 @@ components:
- event_type
- turn_id
title: AgentTurnResponseTurnStartPayload
OpenAIResponseInputMessage:
type: object
properties:
content:
oneOf:
- type: string
- type: array
items:
$ref: '#/components/schemas/OpenAIResponseInputMessageContent'
role:
oneOf:
- type: string
const: system
- type: string
const: developer
- type: string
const: user
- type: string
const: assistant
type:
type: string
const: message
default: message
additionalProperties: false
required:
- content
- role
title: OpenAIResponseInputMessage
OpenAIResponseInputMessageContent:
oneOf:
- $ref: '#/components/schemas/OpenAIResponseInputMessageContentText'
- $ref: '#/components/schemas/OpenAIResponseInputMessageContentImage'
discriminator:
propertyName: type
mapping:
input_text: '#/components/schemas/OpenAIResponseInputMessageContentText'
input_image: '#/components/schemas/OpenAIResponseInputMessageContentImage'
OpenAIResponseInputMessageContentImage:
type: object
properties:
detail:
oneOf:
- type: string
const: low
- type: string
const: high
- type: string
const: auto
default: auto
type:
type: string
const: input_image
default: input_image
image_url:
type: string
additionalProperties: false
required:
- detail
- type
title: OpenAIResponseInputMessageContentImage
OpenAIResponseInputMessageContentText:
type: object
properties:
text:
type: string
type:
type: string
const: input_text
default: input_text
additionalProperties: false
required:
- text
- type
title: OpenAIResponseInputMessageContentText
OpenAIResponseInputTool:
type: object
properties:
type:
oneOf:
- type: string
const: web_search
- type: string
const: web_search_preview_2025_03_11
default: web_search
search_context_size:
type: string
default: medium
additionalProperties: false
required:
- type
title: OpenAIResponseInputToolWebSearch
CreateOpenaiResponseRequest:
type: object
properties:
input:
oneOf:
- type: string
- type: array
items:
$ref: '#/components/schemas/OpenAIResponseInputMessage'
description: Input message(s) to create the response.
model:
type: string
description: The underlying LLM used for completions.
previous_response_id:
type: string
description: >-
(Optional) if specified, the new response will be a continuation of the
previous response. This can be used to easily fork-off new responses from
existing responses.
store:
type: boolean
stream:
type: boolean
temperature:
type: number
tools:
type: array
items:
$ref: '#/components/schemas/OpenAIResponseInputTool'
additionalProperties: false
required:
- input
- model
title: CreateOpenaiResponseRequest
OpenAIResponseError:
type: object
properties:
code:
type: string
message:
type: string
additionalProperties: false
required:
- code
- message
title: OpenAIResponseError
OpenAIResponseObject:
type: object
properties:
created_at:
type: integer
error:
$ref: '#/components/schemas/OpenAIResponseError'
id:
type: string
model:
type: string
object:
type: string
const: response
default: response
output:
type: array
items:
$ref: '#/components/schemas/OpenAIResponseOutput'
parallel_tool_calls:
type: boolean
default: false
previous_response_id:
type: string
status:
type: string
temperature:
type: number
top_p:
type: number
truncation:
type: string
user:
type: string
additionalProperties: false
required:
- created_at
- id
- model
- object
- output
- parallel_tool_calls
- status
title: OpenAIResponseObject
OpenAIResponseOutput:
oneOf:
- $ref: '#/components/schemas/OpenAIResponseOutputMessage'
- $ref: '#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall'
discriminator:
propertyName: type
mapping:
message: '#/components/schemas/OpenAIResponseOutputMessage'
web_search_call: '#/components/schemas/OpenAIResponseOutputMessageWebSearchToolCall'
OpenAIResponseOutputMessage:
type: object
properties:
id:
type: string
content:
type: array
items:
$ref: '#/components/schemas/OpenAIResponseOutputMessageContent'
role:
type: string
const: assistant
default: assistant
status:
type: string
type:
type: string
const: message
default: message
additionalProperties: false
required:
- id
- content
- role
- status
- type
title: OpenAIResponseOutputMessage
OpenAIResponseOutputMessageContent:
type: object
properties:
text:
type: string
type:
type: string
const: output_text
default: output_text
additionalProperties: false
required:
- text
- type
title: >-
OpenAIResponseOutputMessageContentOutputText
"OpenAIResponseOutputMessageWebSearchToolCall":
type: object
properties:
id:
type: string
status:
type: string
type:
type: string
const: web_search_call
default: web_search_call
additionalProperties: false
required:
- id
- status
- type
title: >-
OpenAIResponseOutputMessageWebSearchToolCall
OpenAIResponseObjectStream:
oneOf:
- $ref: '#/components/schemas/OpenAIResponseObjectStreamResponseCreated'
- $ref: '#/components/schemas/OpenAIResponseObjectStreamResponseCompleted'
discriminator:
propertyName: type
mapping:
response.created: '#/components/schemas/OpenAIResponseObjectStreamResponseCreated'
response.completed: '#/components/schemas/OpenAIResponseObjectStreamResponseCompleted'
"OpenAIResponseObjectStreamResponseCompleted":
type: object
properties:
response:
$ref: '#/components/schemas/OpenAIResponseObject'
type:
type: string
const: response.completed
default: response.completed
additionalProperties: false
required:
- response
- type
title: >-
OpenAIResponseObjectStreamResponseCompleted
"OpenAIResponseObjectStreamResponseCreated":
type: object
properties:
response:
$ref: '#/components/schemas/OpenAIResponseObject'
type:
type: string
const: response.created
default: response.created
additionalProperties: false
required:
- response
- type
title: >-
OpenAIResponseObjectStreamResponseCreated
CreateUploadSessionRequest:
type: object
properties:
@ -6097,7 +6460,6 @@ components:
additionalProperties: false
required:
- role
- content
title: OpenAIAssistantMessageParam
description: >-
A message containing the model's (assistant) response in an OpenAI-compatible

907
docs/getting_started_llama_api.ipynb Normal file
View file

File diff suppressed because one or more lines are too long

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

5
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
@ -179,7 +180,7 @@ class ContentBuilder:
"Creates the content subtree for a request or response."
def is_iterator_type(t):
return "StreamChunk" in str(t)
return "StreamChunk" in str(t) or "OpenAIResponseObjectStream" in str(t)
def get_media_type(t):
if is_generic_list(t):
@ -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

38
docs/source/building_applications/rag.md
View file

@ -68,7 +68,8 @@ chunks_response = client.vector_io.query(
### Using the RAG Tool
A better way to ingest documents is to use the RAG Tool. This tool allows you to ingest documents from URLs, files, etc.
and automatically chunks them into smaller pieces.
and automatically chunks them into smaller pieces. More examples for how to format a RAGDocument can be found in the
[appendix](#more-ragdocument-examples).
```python
from llama_stack_client import RAGDocument
@ -178,3 +179,38 @@ for vector_db_id in client.vector_dbs.list():
print(f"Unregistering vector database: {vector_db_id.identifier}")
client.vector_dbs.unregister(vector_db_id=vector_db_id.identifier)
```
### Appendix
#### More RAGDocument Examples
```python
from llama_stack_client import RAGDocument
import base64
RAGDocument(document_id="num-0", content={"uri": "file://path/to/file"})
RAGDocument(document_id="num-1", content="plain text")
RAGDocument(
document_id="num-2",
content={
"type": "text",
"text": "plain text input",
}, # for inputs that should be treated as text explicitly
)
RAGDocument(
document_id="num-3",
content={
"type": "image",
"image": {"url": {"uri": "https://mywebsite.com/image.jpg"}},
},
)
B64_ENCODED_IMAGE = base64.b64encode(
requests.get(
"https://raw.githubusercontent.com/meta-llama/llama-stack/refs/heads/main/docs/_static/llama-stack.png"
).content
)
RAGDocuemnt(
document_id="num-4",
content={"type": "image", "image": {"data": B64_ENCODED_IMAGE}},
)
```
for more strongly typed interaction use the typed dicts found [here](https://github.com/meta-llama/llama-stack-client-python/blob/38cd91c9e396f2be0bec1ee96a19771582ba6f17/src/llama_stack_client/types/shared_params/document.py).

91
docs/source/building_applications/tools.md
View file

@ -41,30 +41,9 @@ client.toolgroups.register(
The tool requires an API key which can be provided either in the configuration or through the request header `X-LlamaStack-Provider-Data`. The format of the header is `{"<provider_name>_api_key": <your api key>}`.
> **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
The WolframAlpha tool provides access to computational knowledge through the WolframAlpha API.
@ -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
@ -214,3 +193,69 @@ response = agent.create_turn(
session_id=session_id,
)
```
## Simple Example 2: Using an Agent with the Web Search Tool
1. Start by registering a Tavily API key at [Tavily](https://tavily.com/).
2. [Optional] Provide the API key directly to the Llama Stack server
```bash
export TAVILY_SEARCH_API_KEY="your key"
```
```bash
--env TAVILY_SEARCH_API_KEY=${TAVILY_SEARCH_API_KEY}
```
3. Run the following script.
```python
from llama_stack_client.lib.agents.agent import Agent
from llama_stack_client.types.agent_create_params import AgentConfig
from llama_stack_client.lib.agents.event_logger import EventLogger
from llama_stack_client import LlamaStackClient
client = LlamaStackClient(
base_url=f"http://localhost:8321",
provider_data={
"tavily_search_api_key": "your_TAVILY_SEARCH_API_KEY"
}, # Set this from the client side. No need to provide it if it has already been configured on the Llama Stack server.
)
agent = Agent(
client,
model="meta-llama/Llama-3.2-3B-Instruct",
instructions=(
"You are a web search assistant, must use websearch tool to look up the most current and precise information available. "
),
tools=["builtin::websearch"],
)
session_id = agent.create_session("websearch-session")
response = agent.create_turn(
messages=[
{"role": "user", "content": "How did the USA perform in the last Olympics?"}
],
session_id=session_id,
)
for log in EventLogger().log(response):
log.print()
```
## Simple Example3: Using an Agent with the WolframAlpha Tool
1. Start by registering for a WolframAlpha API key at [WolframAlpha Developer Portal](https://developer.wolframalpha.com/access).
2. Provide the API key either when starting the Llama Stack server:
```bash
--env WOLFRAM_ALPHA_API_KEY=${WOLFRAM_ALPHA_API_KEY}
```
or from the client side:
```python
client = LlamaStackClient(
base_url="http://localhost:8321",
provider_data={"wolfram_alpha_api_key": wolfram_api_key},
)
```
3. Configure the tools in the Agent by setting `tools=["builtin::wolfram_alpha"]`.
4. Example user query:
```python
response = agent.create_turn(
messages=[{"role": "user", "content": "Solve x^2 + 2x + 1 = 0 using WolframAlpha"}],
session_id=session_id,
)
```
```

57
docs/source/distributions/building_distro.md
View file

@ -109,8 +109,6 @@ llama stack build --list-templates
+------------------------------+-----------------------------------------------------------------------------+
| nvidia | Use NVIDIA NIM for running LLM inference |
+------------------------------+-----------------------------------------------------------------------------+
| meta-reference-quantized-gpu | Use Meta Reference with fp8, int4 quantization for running LLM inference |
+------------------------------+-----------------------------------------------------------------------------+
| cerebras | Use Cerebras for running LLM inference |
+------------------------------+-----------------------------------------------------------------------------+
| ollama | Use (an external) Ollama server for running LLM inference |
@ -176,7 +174,11 @@ distribution_spec:
safety: inline::llama-guard
agents: inline::meta-reference
telemetry: inline::meta-reference
image_name: ollama
image_type: conda
# If some providers are external, you can specify the path to the implementation
external_providers_dir: /etc/llama-stack/providers.d
```
```
@ -184,6 +186,57 @@ llama stack build --config llama_stack/templates/ollama/build.yaml
```
:::
:::{tab-item} Building with External Providers
Llama Stack supports external providers that live outside of the main codebase. This allows you to create and maintain your own providers independently or use community-provided providers.
To build a distribution with external providers, you need to:
1. Configure the `external_providers_dir` in your build configuration file:
```yaml
# Example my-external-stack.yaml with external providers
version: '2'
distribution_spec:
description: Custom distro for CI tests
providers:
inference:
- remote::custom_ollama
# Add more providers as needed
image_type: container
image_name: ci-test
# Path to external provider implementations
external_providers_dir: /etc/llama-stack/providers.d
```
Here's an example for a custom Ollama provider:
```yaml
adapter:
adapter_type: custom_ollama
pip_packages:
- ollama
- aiohttp
- llama-stack-provider-ollama # This is the provider package
config_class: llama_stack_ollama_provider.config.OllamaImplConfig
module: llama_stack_ollama_provider
api_dependencies: []
optional_api_dependencies: []
```
The `pip_packages` section lists the Python packages required by the provider, as well as the
provider package itself. The package must be available on PyPI or can be provided from a local
directory or a git repository (git must be installed on the build environment).
2. Build your distribution using the config file:
```
llama stack build --config my-external-stack.yaml
```
For more information on external providers, including directory structure, provider types, and implementation requirements, see the [External Providers documentation](../providers/external.md).
:::
:::{tab-item} Building Container
```{admonition} Podman Alternative

106
docs/source/distributions/configuration.md
View file

@ -53,6 +53,13 @@ models:
provider_id: ollama
provider_model_id: null
shields: []
server:
port: 8321
auth:
provider_type: "kubernetes"
config:
api_server_url: "https://kubernetes.default.svc"
ca_cert_path: "/path/to/ca.crt"
```
Let's break this down into the different sections. The first section specifies the set of APIs that the stack server will serve:
@ -102,6 +109,105 @@ A Model is an instance of a "Resource" (see [Concepts](../concepts/index)) and i
What's with the `provider_model_id` field? This is an identifier for the model inside the provider's model catalog. Contrast it with `model_id` which is the identifier for the same model for Llama Stack's purposes. For example, you may want to name "llama3.2:vision-11b" as "image_captioning_model" when you use it in your Stack interactions. When omitted, the server will set `provider_model_id` to be the same as `model_id`.
## Server Configuration
The `server` section configures the HTTP server that serves the Llama Stack APIs:
```yaml
server:
port: 8321 # Port to listen on (default: 8321)
tls_certfile: "/path/to/cert.pem" # Optional: Path to TLS certificate for HTTPS
tls_keyfile: "/path/to/key.pem" # Optional: Path to TLS key for HTTPS
auth: # Optional: Authentication configuration
provider_type: "kubernetes" # Type of auth provider
config: # Provider-specific configuration
api_server_url: "https://kubernetes.default.svc"
ca_cert_path: "/path/to/ca.crt" # Optional: Path to CA certificate
```
### Authentication Configuration
The `auth` section configures authentication for the server. When configured, all API requests must include a valid Bearer token in the Authorization header:
```
Authorization: Bearer <token>
```
The server supports multiple authentication providers:
#### Kubernetes Provider
The Kubernetes cluster must be configured to use a service account for authentication.
```bash
kubectl create namespace llama-stack
kubectl create serviceaccount llama-stack-auth -n llama-stack
kubectl create rolebinding llama-stack-auth-rolebinding --clusterrole=admin --serviceaccount=llama-stack:llama-stack-auth -n llama-stack
kubectl create token llama-stack-auth -n llama-stack > llama-stack-auth-token
```
Validates tokens against the Kubernetes API server:
```yaml
server:
auth:
provider_type: "kubernetes"
config:
api_server_url: "https://kubernetes.default.svc" # URL of the Kubernetes API server
ca_cert_path: "/path/to/ca.crt" # Optional: Path to CA certificate
```
The provider extracts user information from the JWT token:
- Username from the `sub` claim becomes a role
- Kubernetes groups become teams
You can easily validate a request by running:
```bash
curl -s -L -H "Authorization: Bearer $(cat llama-stack-auth-token)" http://127.0.0.1:8321/v1/providers
```
#### Custom Provider
Validates tokens against a custom authentication endpoint:
```yaml
server:
auth:
provider_type: "custom"
config:
endpoint: "https://auth.example.com/validate" # URL of the auth endpoint
```
The custom endpoint receives a POST request with:
```json
{
"api_key": "<token>",
"request": {
"path": "/api/v1/endpoint",
"headers": {
"content-type": "application/json",
"user-agent": "curl/7.64.1"
},
"params": {
"key": ["value"]
}
}
}
```
And must respond with:
```json
{
"access_attributes": {
"roles": ["admin", "user"],
"teams": ["ml-team", "nlp-team"],
"projects": ["llama-3", "project-x"],
"namespaces": ["research"]
},
"message": "Authentication successful"
}
```
If no access attributes are returned, the token is used as a namespace.
## Extending to handle Safety
Configuring Safety can be a little involved so it is instructive to go through an example.

12
docs/source/distributions/ondevice_distro/android_sdk.md
View file

@ -24,7 +24,7 @@ The key files in the app are `ExampleLlamaStackLocalInference.kt`, `ExampleLlama
Add the following dependency in your `build.gradle.kts` file:
```
dependencies {
implementation("com.llama.llamastack:llama-stack-client-kotlin:0.1.4.2")
implementation("com.llama.llamastack:llama-stack-client-kotlin:0.2.2")
}
```
This will download jar files in your gradle cache in a directory like `~/.gradle/caches/modules-2/files-2.1/com.llama.llamastack/`
@ -37,11 +37,7 @@ For local inferencing, it is required to include the ExecuTorch library into you
Include the ExecuTorch library by:
1. Download the `download-prebuilt-et-lib.sh` script file from the [llama-stack-client-kotlin-client-local](https://github.com/meta-llama/llama-stack-client-kotlin/tree/latest-release/llama-stack-client-kotlin-client-local/download-prebuilt-et-lib.sh) directory to your local machine.
2. Move the script to the top level of your Android app where the app directory resides:
<p align="center">
<img src="https://github.com/meta-llama/llama-stack-client-kotlin/blob/latest-release/doc/img/example_android_app_directory.png" style="width:300px">
</p>
2. Move the script to the top level of your Android app where the `app` directory resides.
3. Run `sh download-prebuilt-et-lib.sh` to create an `app/libs` directory and download the `executorch.aar` in that path. This generates an ExecuTorch library for the XNNPACK delegate.
4. Add the `executorch.aar` dependency in your `build.gradle.kts` file:
```
@ -52,6 +48,8 @@ dependencies {
}
```
See other dependencies for the local RAG in Android app [README](https://github.com/meta-llama/llama-stack-client-kotlin/tree/latest-release/examples/android_app#quick-start).
## Llama Stack APIs in Your Android App
Breaking down the demo app, this section will show the core pieces that are used to initialize and run inference with Llama Stack using the Kotlin library.
@ -60,7 +58,7 @@ Start a Llama Stack server on localhost. Here is an example of how you can do th
```
conda create -n stack-fireworks python=3.10
conda activate stack-fireworks
pip install --no-cache llama-stack==0.1.4
pip install --no-cache llama-stack==0.2.2
llama stack build --template fireworks --image-type conda
export FIREWORKS_API_KEY=<SOME_KEY>
llama stack run fireworks --port 5050

88
docs/source/distributions/remote_hosted_distro/nvidia.md
View file

@ -1,88 +0,0 @@
<!-- This file was auto-generated by distro_codegen.py, please edit source -->
# NVIDIA Distribution
The `llamastack/distribution-nvidia` distribution consists of the following provider configurations.
| API | Provider(s) |
|-----|-------------|
| agents | `inline::meta-reference` |
| datasetio | `inline::localfs` |
| eval | `inline::meta-reference` |
| inference | `remote::nvidia` |
| post_training | `remote::nvidia` |
| safety | `remote::nvidia` |
| scoring | `inline::basic` |
| telemetry | `inline::meta-reference` |
| tool_runtime | `inline::rag-runtime` |
| vector_io | `inline::faiss` |
### Environment Variables
The following environment variables can be configured:
- `NVIDIA_API_KEY`: NVIDIA API Key (default: ``)
- `NVIDIA_USER_ID`: NVIDIA User ID (default: `llama-stack-user`)
- `NVIDIA_DATASET_NAMESPACE`: NVIDIA Dataset Namespace (default: `default`)
- `NVIDIA_ACCESS_POLICIES`: NVIDIA Access Policies (default: `{}`)
- `NVIDIA_PROJECT_ID`: NVIDIA Project ID (default: `test-project`)
- `NVIDIA_CUSTOMIZER_URL`: NVIDIA Customizer URL (default: `https://customizer.api.nvidia.com`)
- `NVIDIA_OUTPUT_MODEL_DIR`: NVIDIA Output Model Directory (default: `test-example-model@v1`)
- `GUARDRAILS_SERVICE_URL`: URL for the NeMo Guardrails Service (default: `http://0.0.0.0:7331`)
- `INFERENCE_MODEL`: Inference model (default: `Llama3.1-8B-Instruct`)
- `SAFETY_MODEL`: Name of the model to use for safety (default: `meta/llama-3.1-8b-instruct`)
### Models
The following models are available by default:
- `meta/llama3-8b-instruct (aliases: meta-llama/Llama-3-8B-Instruct)`
- `meta/llama3-70b-instruct (aliases: meta-llama/Llama-3-70B-Instruct)`
- `meta/llama-3.1-8b-instruct (aliases: meta-llama/Llama-3.1-8B-Instruct)`
- `meta/llama-3.1-70b-instruct (aliases: meta-llama/Llama-3.1-70B-Instruct)`
- `meta/llama-3.1-405b-instruct (aliases: meta-llama/Llama-3.1-405B-Instruct-FP8)`
- `meta/llama-3.2-1b-instruct (aliases: meta-llama/Llama-3.2-1B-Instruct)`
- `meta/llama-3.2-3b-instruct (aliases: meta-llama/Llama-3.2-3B-Instruct)`
- `meta/llama-3.2-11b-vision-instruct (aliases: meta-llama/Llama-3.2-11B-Vision-Instruct)`
- `meta/llama-3.2-90b-vision-instruct (aliases: meta-llama/Llama-3.2-90B-Vision-Instruct)`
- `nvidia/llama-3.2-nv-embedqa-1b-v2 `
- `nvidia/nv-embedqa-e5-v5 `
- `nvidia/nv-embedqa-mistral-7b-v2 `
- `snowflake/arctic-embed-l `
### Prerequisite: API Keys
Make sure you have access to a NVIDIA API Key. You can get one by visiting [https://build.nvidia.com/](https://build.nvidia.com/).
## Running Llama Stack with NVIDIA
You can do this via Conda (build code) or Docker which has a pre-built image.
### Via Docker
This method allows you to get started quickly without having to build the distribution code.
```bash
LLAMA_STACK_PORT=8321
docker run \
-it \
--pull always \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ./run.yaml:/root/my-run.yaml \
llamastack/distribution-nvidia \
--yaml-config /root/my-run.yaml \
--port $LLAMA_STACK_PORT \
--env NVIDIA_API_KEY=$NVIDIA_API_KEY
```
### Via Conda
```bash
llama stack build --template nvidia --image-type conda
llama stack run ./run.yaml \
--port 8321 \
--env NVIDIA_API_KEY=$NVIDIA_API_KEY
--env INFERENCE_MODEL=$INFERENCE_MODEL
```

88
docs/source/distributions/remote_hosted_distro/watsonx.md Normal file
View file

@ -0,0 +1,88 @@
---
orphan: true
---
<!-- This file was auto-generated by distro_codegen.py, please edit source -->
# watsonx Distribution
```{toctree}
:maxdepth: 2
:hidden:
self
```
The `llamastack/distribution-watsonx` distribution consists of the following provider configurations.
| API | Provider(s) |
|-----|-------------|
| agents | `inline::meta-reference` |
| datasetio | `remote::huggingface`, `inline::localfs` |
| eval | `inline::meta-reference` |
| inference | `remote::watsonx` |
| 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::rag-runtime`, `remote::model-context-protocol` |
| vector_io | `inline::faiss` |
### Environment Variables
The following environment variables can be configured:
- `LLAMASTACK_PORT`: Port for the Llama Stack distribution server (default: `5001`)
- `WATSONX_API_KEY`: watsonx API Key (default: ``)
- `WATSONX_PROJECT_ID`: watsonx Project ID (default: ``)
### Models
The following models are available by default:
- `meta-llama/llama-3-3-70b-instruct (aliases: meta-llama/Llama-3.3-70B-Instruct)`
- `meta-llama/llama-2-13b-chat (aliases: meta-llama/Llama-2-13b)`
- `meta-llama/llama-3-1-70b-instruct (aliases: meta-llama/Llama-3.1-70B-Instruct)`
- `meta-llama/llama-3-1-8b-instruct (aliases: meta-llama/Llama-3.1-8B-Instruct)`
- `meta-llama/llama-3-2-11b-vision-instruct (aliases: meta-llama/Llama-3.2-11B-Vision-Instruct)`
- `meta-llama/llama-3-2-1b-instruct (aliases: meta-llama/Llama-3.2-1B-Instruct)`
- `meta-llama/llama-3-2-3b-instruct (aliases: meta-llama/Llama-3.2-3B-Instruct)`
- `meta-llama/llama-3-2-90b-vision-instruct (aliases: meta-llama/Llama-3.2-90B-Vision-Instruct)`
- `meta-llama/llama-guard-3-11b-vision (aliases: meta-llama/Llama-Guard-3-11B-Vision)`
### Prerequisite: API Keys
Make sure you have access to a watsonx API Key. You can get one by referring [watsonx.ai](https://www.ibm.com/docs/en/masv-and-l/maximo-manage/continuous-delivery?topic=setup-create-watsonx-api-key).
## Running Llama Stack with watsonx
You can do this via Conda (build code), venv or Docker which has a pre-built image.
### Via Docker
This method allows you to get started quickly without having to build the distribution code.
```bash
LLAMA_STACK_PORT=5001
docker run \
-it \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ./run.yaml:/root/my-run.yaml \
llamastack/distribution-watsonx \
--yaml-config /root/my-run.yaml \
--port $LLAMA_STACK_PORT \
--env WATSONX_API_KEY=$WATSONX_API_KEY \
--env WATSONX_PROJECT_ID=$WATSONX_PROJECT_ID \
--env WATSONX_BASE_URL=$WATSONX_BASE_URL
```
### Via Conda
```bash
llama stack build --template watsonx --image-type conda
llama stack run ./run.yaml \
--port $LLAMA_STACK_PORT \
--env WATSONX_API_KEY=$WATSONX_API_KEY \
--env WATSONX_PROJECT_ID=$WATSONX_PROJECT_ID
```

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` |

4
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` |
@ -81,6 +81,7 @@ LLAMA_STACK_PORT=8321
docker run \
-it \
--pull always \
--gpu all \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ~/.llama:/root/.llama \
llamastack/distribution-meta-reference-gpu \
@ -94,6 +95,7 @@ If you are using Llama Stack Safety / Shield APIs, use:
docker run \
-it \
--pull always \
--gpu all \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ~/.llama:/root/.llama \
llamastack/distribution-meta-reference-gpu \

123
docs/source/distributions/self_hosted_distro/meta-reference-quantized-gpu.md
View file

@ -1,123 +0,0 @@
---
orphan: true
---
<!-- This file was auto-generated by distro_codegen.py, please edit source -->
# Meta Reference Quantized Distribution
```{toctree}
:maxdepth: 2
:hidden:
self
```
The `llamastack/distribution-meta-reference-quantized-gpu` distribution consists of the following provider configurations:
| API | Provider(s) |
|-----|-------------|
| agents | `inline::meta-reference` |
| datasetio | `remote::huggingface`, `inline::localfs` |
| eval | `inline::meta-reference` |
| inference | `inline::meta-reference-quantized` |
| 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` |
| vector_io | `inline::faiss`, `remote::chromadb`, `remote::pgvector` |
The only difference vs. the `meta-reference-gpu` distribution is that it has support for more efficient inference -- with fp8, int4 quantization, etc.
Note that you need access to nvidia GPUs to run this distribution. This distribution is not compatible with CPU-only machines or machines with AMD GPUs.
### Environment Variables
The following environment variables can be configured:
- `LLAMA_STACK_PORT`: Port for the Llama Stack distribution server (default: `8321`)
- `INFERENCE_MODEL`: Inference model loaded into the Meta Reference server (default: `meta-llama/Llama-3.2-3B-Instruct`)
- `INFERENCE_CHECKPOINT_DIR`: Directory containing the Meta Reference model checkpoint (default: `null`)
## Prerequisite: Downloading Models
Please use `llama model list --downloaded` to check that you have llama model checkpoints downloaded in `~/.llama` before proceeding. See [installation guide](https://llama-stack.readthedocs.io/en/latest/references/llama_cli_reference/download_models.html) here to download the models. Run `llama model list` to see the available models to download, and `llama model download` to download the checkpoints.
```
$ llama model list --downloaded
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┓
┃ Model ┃ Size ┃ Modified Time ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━┩
│ Llama3.2-1B-Instruct:int4-qlora-eo8 │ 1.53 GB │ 2025-02-26 11:22:28 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama3.2-1B │ 2.31 GB │ 2025-02-18 21:48:52 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Prompt-Guard-86M │ 0.02 GB │ 2025-02-26 11:29:28 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama3.2-3B-Instruct:int4-spinquant-eo8 │ 3.69 GB │ 2025-02-26 11:37:41 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama3.2-3B │ 5.99 GB │ 2025-02-18 21:51:26 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama3.1-8B │ 14.97 GB │ 2025-02-16 10:36:37 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama3.2-1B-Instruct:int4-spinquant-eo8 │ 1.51 GB │ 2025-02-26 11:35:02 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama-Guard-3-1B │ 2.80 GB │ 2025-02-26 11:20:46 │
├─────────────────────────────────────────┼──────────┼─────────────────────┤
│ Llama-Guard-3-1B:int4 │ 0.43 GB │ 2025-02-26 11:33:33 │
└─────────────────────────────────────────┴──────────┴─────────────────────┘
```
## Running the Distribution
You can do this via Conda (build code) or Docker which has a pre-built image.
### Via Docker
This method allows you to get started quickly without having to build the distribution code.
```bash
LLAMA_STACK_PORT=8321
docker run \
-it \
--pull always \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ~/.llama:/root/.llama \
llamastack/distribution-meta-reference-quantized-gpu \
--port $LLAMA_STACK_PORT \
--env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
```
If you are using Llama Stack Safety / Shield APIs, use:
```bash
docker run \
-it \
--pull always \
-p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-v ~/.llama:/root/.llama \
llamastack/distribution-meta-reference-quantized-gpu \
--port $LLAMA_STACK_PORT \
--env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct \
--env SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
```
### Via Conda
Make sure you have done `uv pip install llama-stack` and have the Llama Stack CLI available.
```bash
llama stack build --template meta-reference-quantized-gpu --image-type conda
llama stack run distributions/meta-reference-quantized-gpu/run.yaml \
--port $LLAMA_STACK_PORT \
--env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
```
If you are using Llama Stack Safety / Shield APIs, use:
```bash
llama stack run distributions/meta-reference-quantized-gpu/run-with-safety.yaml \
--port $LLAMA_STACK_PORT \
--env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct \
--env SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
```

101
docs/source/distributions/self_hosted_distro/nvidia.md
View file

@ -6,8 +6,8 @@ The `llamastack/distribution-nvidia` distribution consists of the following prov
| API | Provider(s) |
|-----|-------------|
| agents | `inline::meta-reference` |
| datasetio | `inline::localfs` |
| eval | `inline::meta-reference` |
| datasetio | `inline::localfs`, `remote::nvidia` |
| eval | `remote::nvidia` |
| inference | `remote::nvidia` |
| post_training | `remote::nvidia` |
| safety | `remote::nvidia` |
@ -22,13 +22,13 @@ The `llamastack/distribution-nvidia` distribution consists of the following prov
The following environment variables can be configured:
- `NVIDIA_API_KEY`: NVIDIA API Key (default: ``)
- `NVIDIA_USER_ID`: NVIDIA User ID (default: `llama-stack-user`)
- `NVIDIA_APPEND_API_VERSION`: Whether to append the API version to the base_url (default: `True`)
- `NVIDIA_DATASET_NAMESPACE`: NVIDIA Dataset Namespace (default: `default`)
- `NVIDIA_ACCESS_POLICIES`: NVIDIA Access Policies (default: `{}`)
- `NVIDIA_PROJECT_ID`: NVIDIA Project ID (default: `test-project`)
- `NVIDIA_CUSTOMIZER_URL`: NVIDIA Customizer URL (default: `https://customizer.api.nvidia.com`)
- `NVIDIA_OUTPUT_MODEL_DIR`: NVIDIA Output Model Directory (default: `test-example-model@v1`)
- `GUARDRAILS_SERVICE_URL`: URL for the NeMo Guardrails Service (default: `http://0.0.0.0:7331`)
- `NVIDIA_EVALUATOR_URL`: URL for the NeMo Evaluator Service (default: `http://0.0.0.0:7331`)
- `INFERENCE_MODEL`: Inference model (default: `Llama3.1-8B-Instruct`)
- `SAFETY_MODEL`: Name of the model to use for safety (default: `meta/llama-3.1-8b-instruct`)
@ -45,20 +45,91 @@ The following models are available by default:
- `meta/llama-3.2-3b-instruct (aliases: meta-llama/Llama-3.2-3B-Instruct)`
- `meta/llama-3.2-11b-vision-instruct (aliases: meta-llama/Llama-3.2-11B-Vision-Instruct)`
- `meta/llama-3.2-90b-vision-instruct (aliases: meta-llama/Llama-3.2-90B-Vision-Instruct)`
- `meta/llama-3.3-70b-instruct (aliases: meta-llama/Llama-3.3-70B-Instruct)`
- `nvidia/llama-3.2-nv-embedqa-1b-v2 `
- `nvidia/nv-embedqa-e5-v5 `
- `nvidia/nv-embedqa-mistral-7b-v2 `
- `snowflake/arctic-embed-l `
### Prerequisite: API Keys
## Prerequisites
### NVIDIA API Keys
Make sure you have access to a NVIDIA API Key. You can get one by visiting [https://build.nvidia.com/](https://build.nvidia.com/).
Make sure you have access to a NVIDIA API Key. You can get one by visiting [https://build.nvidia.com/](https://build.nvidia.com/). Use this key for the `NVIDIA_API_KEY` environment variable.
### Deploy NeMo Microservices Platform
The NVIDIA NeMo microservices platform supports end-to-end microservice deployment of a complete AI flywheel on your Kubernetes cluster through the NeMo Microservices Helm Chart. Please reference the [NVIDIA NeMo Microservices documentation](https://docs.nvidia.com/nemo/microservices/latest/about/index.html) for platform prerequisites and instructions to install and deploy the platform.
## Supported Services
Each Llama Stack API corresponds to a specific NeMo microservice. The core microservices (Customizer, Evaluator, Guardrails) are exposed by the same endpoint. The platform components (Data Store) are each exposed by separate endpoints.
### Inference: NVIDIA NIM
NVIDIA NIM is used for running inference with registered models. There are two ways to access NVIDIA NIMs:
1. Hosted (default): Preview APIs hosted at https://integrate.api.nvidia.com (Requires an API key)
2. Self-hosted: NVIDIA NIMs that run on your own infrastructure.
The deployed platform includes the NIM Proxy microservice, which is the service that provides to access your NIMs (for example, to run inference on a model). Set the `NVIDIA_BASE_URL` environment variable to use your NVIDIA NIM Proxy deployment.
### Datasetio API: NeMo Data Store
The NeMo Data Store microservice serves as the default file storage solution for the NeMo microservices platform. It exposts APIs compatible with the Hugging Face Hub client (`HfApi`), so you can use the client to interact with Data Store. The `NVIDIA_DATASETS_URL` environment variable should point to your NeMo Data Store endpoint.
See the [NVIDIA Datasetio docs](/llama_stack/providers/remote/datasetio/nvidia/README.md) for supported features and example usage.
### Eval API: NeMo Evaluator
The NeMo Evaluator microservice supports evaluation of LLMs. Launching an Evaluation job with NeMo Evaluator requires an Evaluation Config (an object that contains metadata needed by the job). A Llama Stack Benchmark maps to an Evaluation Config, so registering a Benchmark creates an Evaluation Config in NeMo Evaluator. The `NVIDIA_EVALUATOR_URL` environment variable should point to your NeMo Microservices endpoint.
See the [NVIDIA Eval docs](/llama_stack/providers/remote/eval/nvidia/README.md) for supported features and example usage.
### Post-Training API: NeMo Customizer
The NeMo Customizer microservice supports fine-tuning models. You can reference [this list of supported models](/llama_stack/providers/remote/post_training/nvidia/models.py) that can be fine-tuned using Llama Stack. The `NVIDIA_CUSTOMIZER_URL` environment variable should point to your NeMo Microservices endpoint.
See the [NVIDIA Post-Training docs](/llama_stack/providers/remote/post_training/nvidia/README.md) for supported features and example usage.
### Safety API: NeMo Guardrails
The NeMo Guardrails microservice sits between your application and the LLM, and adds checks and content moderation to a model. The `GUARDRAILS_SERVICE_URL` environment variable should point to your NeMo Microservices endpoint.
See the NVIDIA Safety docs for supported features and example usage.
## Deploying models
In order to use a registered model with the Llama Stack APIs, ensure the corresponding NIM is deployed to your environment. For example, you can use the NIM Proxy microservice to deploy `meta/llama-3.2-1b-instruct`.
Note: For improved inference speeds, we need to use NIM with `fast_outlines` guided decoding system (specified in the request body). This is the default if you deployed the platform with the NeMo Microservices Helm Chart.
```sh
# URL to NeMo NIM Proxy service
export NEMO_URL="http://nemo.test"
curl --location "$NEMO_URL/v1/deployment/model-deployments" \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"name": "llama-3.2-1b-instruct",
"namespace": "meta",
"config": {
"model": "meta/llama-3.2-1b-instruct",
"nim_deployment": {
"image_name": "nvcr.io/nim/meta/llama-3.2-1b-instruct",
"image_tag": "1.8.3",
"pvc_size": "25Gi",
"gpu": 1,
"additional_envs": {
"NIM_GUIDED_DECODING_BACKEND": "fast_outlines"
}
}
}
}'
```
This NIM deployment should take approximately 10 minutes to go live. [See the docs](https://docs.nvidia.com/nemo/microservices/latest/get-started/tutorials/deploy-nims.html) for more information on how to deploy a NIM and verify it's available for inference.
You can also remove a deployed NIM to free up GPU resources, if needed.
```sh
export NEMO_URL="http://nemo.test"
curl -X DELETE "$NEMO_URL/v1/deployment/model-deployments/meta/llama-3.1-8b-instruct"
```
## Running Llama Stack with NVIDIA
You can do this via Conda (build code) or Docker which has a pre-built image.
You can do this via Conda or venv (build code), or Docker which has a pre-built image.
### Via Docker
@ -80,9 +151,23 @@ docker run \
### Via Conda
```bash
INFERENCE_MODEL=meta-llama/Llama-3.1-8b-Instruct
llama stack build --template nvidia --image-type conda
llama stack run ./run.yaml \
--port 8321 \
--env NVIDIA_API_KEY=$NVIDIA_API_KEY
--env NVIDIA_API_KEY=$NVIDIA_API_KEY \
--env INFERENCE_MODEL=$INFERENCE_MODEL
```
### Via venv
If you've set up your local development environment, you can also build the image using your local virtual environment.
```bash
INFERENCE_MODEL=meta-llama/Llama-3.1-8b-Instruct
llama stack build --template nvidia --image-type venv
llama stack run ./run.yaml \
--port 8321 \
--env NVIDIA_API_KEY=$NVIDIA_API_KEY \
--env INFERENCE_MODEL=$INFERENCE_MODEL
```

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` |

55
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` |
@ -41,10 +41,10 @@ The following environment variables can be configured:
## Setting up vLLM server
In the following sections, we'll use either AMD and NVIDIA GPUs to serve as hardware accelerators for the vLLM
In the following sections, we'll use AMD, NVIDIA or Intel GPUs to serve as hardware accelerators for the vLLM
server, which acts as both the LLM inference provider and the safety provider. Note that vLLM also
[supports many other hardware accelerators](https://docs.vllm.ai/en/latest/getting_started/installation.html) and
that we only use GPUs here for demonstration purposes.
that we only use GPUs here for demonstration purposes. Note that if you run into issues, you can include the environment variable `--env VLLM_DEBUG_LOG_API_SERVER_RESPONSE=true` (available in vLLM v0.8.3 and above) in the `docker run` command to enable log response from API server for debugging.
### Setting up vLLM server on AMD GPU
@ -162,6 +162,55 @@ docker run \
--port $SAFETY_PORT
```
### Setting up vLLM server on Intel GPU
Refer to [vLLM Documentation for XPU](https://docs.vllm.ai/en/v0.8.2/getting_started/installation/gpu.html?device=xpu) to get a vLLM endpoint. In addition to vLLM side setup which guides towards installing vLLM from sources orself-building vLLM Docker container, Intel provides prebuilt vLLM container to use on systems with Intel GPUs supported by PyTorch XPU backend:
- [intel/vllm](https://hub.docker.com/r/intel/vllm)
Here is a sample script to start a vLLM server locally via Docker using Intel provided container:
```bash
export INFERENCE_PORT=8000
export INFERENCE_MODEL=meta-llama/Llama-3.2-1B-Instruct
export ZE_AFFINITY_MASK=0
docker run \
--pull always \
--device /dev/dri \
-v /dev/dri/by-path:/dev/dri/by-path \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
--env ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK \
-p $INFERENCE_PORT:$INFERENCE_PORT \
--ipc=host \
intel/vllm:xpu \
--gpu-memory-utilization 0.7 \
--model $INFERENCE_MODEL \
--port $INFERENCE_PORT
```
If you are using Llama Stack Safety / Shield APIs, then you will need to also run another instance of a vLLM with a corresponding safety model like `meta-llama/Llama-Guard-3-1B` using a script like:
```bash
export SAFETY_PORT=8081
export SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
export ZE_AFFINITY_MASK=1
docker run \
--pull always \
--device /dev/dri \
-v /dev/dri/by-path:/dev/dri/by-path \
-v ~/.cache/huggingface:/root/.cache/huggingface \
--env "HUGGING_FACE_HUB_TOKEN=$HF_TOKEN" \
--env ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK \
-p $SAFETY_PORT:$SAFETY_PORT \
--ipc=host \
intel/vllm:xpu \
--gpu-memory-utilization 0.7 \
--model $SAFETY_MODEL \
--port $SAFETY_PORT
```
## Running Llama Stack
Now you are ready to run Llama Stack with vLLM as the inference provider. You can do this via Conda (build code) or Docker which has a pre-built image.

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`, `inline::sentence-transformers` |
| safety | `inline::llama-guard` |
| 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/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` |

8
docs/source/providers/external.md
View file

@ -50,9 +50,11 @@ Llama Stack supports two types of external providers:
Here's a list of known external providers that you can use with Llama Stack:
| Type | Name | Description | Repository |
|------|------|-------------|------------|
| Remote | KubeFlow Training | Train models with KubeFlow | [llama-stack-provider-kft](https://github.com/opendatahub-io/llama-stack-provider-kft) |
| 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) |
### Remote Provider Specification

4
docs/zero_to_hero_guide/00_Inference101.ipynb
View file

@ -389,5 +389,7 @@
"pygments_lexer": "ipython3",
"version": "3.10.15"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/01_Local_Cloud_Inference101.ipynb
View file

@ -256,5 +256,7 @@
"pygments_lexer": "ipython3",
"version": "3.10.15"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/02_Prompt_Engineering101.ipynb
View file

@ -301,5 +301,7 @@
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/03_Image_Chat101.ipynb
View file

@ -200,5 +200,7 @@
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/04_Tool_Calling101.ipynb
View file

@ -355,5 +355,7 @@
"pygments_lexer": "ipython3",
"version": "3.10.15"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/05_Memory101.ipynb
View file

@ -398,5 +398,7 @@
"pygments_lexer": "ipython3",
"version": "3.10.15"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/06_Safety101.ipynb
View file

@ -132,5 +132,7 @@
"pygments_lexer": "ipython3",
"version": "3.11.10"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

4
docs/zero_to_hero_guide/07_Agents101.ipynb
View file

@ -188,5 +188,7 @@
"pygments_lexer": "ipython3",
"version": "3.10.15"
}
}
},
"nbformat": 4,
"nbformat_minor": 5
}

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**:

145
install.sh Executable file
View file

@ -0,0 +1,145 @@
#!/usr/bin/env bash
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
set -Eeuo pipefail
PORT=8321
OLLAMA_PORT=11434
MODEL_ALIAS="llama3.2:3b"
SERVER_IMAGE="llamastack/distribution-ollama:0.2.2"
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"
elif command -v podman &> /dev/null; then
ENGINE="podman"
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" > /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…"
$ENGINE run -d "${PLATFORM_OPTS[@]}" --name ollama-server \
--network llama-net \
-p "${OLLAMA_PORT}:${OLLAMA_PORT}" \
ollama/ollama > /dev/null 2>&1
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 --tail 200 ollama-server
die "Ollama startup failed"
fi
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:"
$ENGINE logs --tail 200 ollama-server
die "Model pull failed"
fi
###############################################################################
# 2. Llama‑Stack
###############################################################################
cmd=( run -d "${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 "🦙 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:"
$ENGINE logs --tail 200 llama-stack
die "Llama-Stack startup failed"
fi
###############################################################################
# Done
###############################################################################
log ""
log "🎉 Llama‑Stack is ready!"
log "👉 API endpoint: http://localhost:${PORT}"
log "📖 Documentation: https://llama-stack.readthedocs.io/en/latest/references/index.html"
log "💻 To access the llama‑stack CLI, exec into the container:"
log " $ENGINE exec -ti llama-stack bash"
log ""

195
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
@ -38,6 +28,13 @@ from llama_stack.apis.safety import SafetyViolation
from llama_stack.apis.tools import ToolDef
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
from .openai_responses import (
OpenAIResponseInputMessage,
OpenAIResponseInputTool,
OpenAIResponseObject,
OpenAIResponseObjectStream,
)
class Attachment(BaseModel):
"""An attachment to an agent turn.
@ -72,8 +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):
@ -113,8 +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
@ -125,7 +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
@ -143,12 +140,7 @@ class MemoryRetrievalStep(StepCommon):
Step = Annotated[
Union[
InferenceStep,
ToolExecutionStep,
ShieldCallStep,
MemoryRetrievalStep,
],
InferenceStep | ToolExecutionStep | ShieldCallStep | MemoryRetrievalStep,
Field(discriminator="step_type"),
]
@ -159,18 +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
@ -179,34 +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:
@ -225,10 +209,20 @@ class AgentConfigCommon(BaseModel):
@json_schema_type
class AgentConfig(AgentConfigCommon):
"""Configuration for an agent.
:param model: The model identifier to use for the agent
:param instructions: The system instructions for the agent
:param name: Optional name for the agent, used in telemetry and identification
:param enable_session_persistence: Optional flag indicating whether session data has to be persisted
:param response_format: Optional response format configuration
"""
model: str
instructions: str
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
@ -240,16 +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):
@ -267,7 +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
@ -310,14 +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")
@ -346,18 +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
@ -365,8 +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
@ -412,17 +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.
@ -446,9 +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.
@ -521,7 +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.
@ -583,3 +565,40 @@ class Agents(Protocol):
:returns: A ListAgentSessionsResponse.
"""
...
# We situate the OpenAI Responses API in the Agents API just like we did things
# for Inference. The Responses API, in its intent, serves the same purpose as
# the Agents API above -- it is essentially a lightweight "agentic loop" with
# integrated tool calling.
#
# Both of these APIs are inherently stateful.
@webmethod(route="/openai/v1/responses/{id}", method="GET")
async def get_openai_response(
self,
id: str,
) -> OpenAIResponseObject:
"""Retrieve an OpenAI response by its ID.
:param id: The ID of the OpenAI response to retrieve.
:returns: An OpenAIResponseObject.
"""
...
@webmethod(route="/openai/v1/responses", method="POST")
async def create_openai_response(
self,
input: str | list[OpenAIResponseInputMessage],
model: str,
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.
:param model: The underlying LLM used for completions.
:param previous_response_id: (Optional) if specified, the new response will be a continuation of the previous response. This can be used to easily fork-off new responses from existing responses.
"""

133
llama_stack/apis/agents/openai_responses.py Normal file
View file

@ -0,0 +1,133 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Annotated, Literal
from pydantic import BaseModel, Field
from llama_stack.schema_utils import json_schema_type, register_schema
@json_schema_type
class OpenAIResponseError(BaseModel):
code: str
message: str
@json_schema_type
class OpenAIResponseOutputMessageContentOutputText(BaseModel):
text: str
type: Literal["output_text"] = "output_text"
OpenAIResponseOutputMessageContent = Annotated[
OpenAIResponseOutputMessageContentOutputText,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutputMessageContent, name="OpenAIResponseOutputMessageContent")
@json_schema_type
class OpenAIResponseOutputMessage(BaseModel):
id: str
content: list[OpenAIResponseOutputMessageContent]
role: Literal["assistant"] = "assistant"
status: str
type: Literal["message"] = "message"
@json_schema_type
class OpenAIResponseOutputMessageWebSearchToolCall(BaseModel):
id: str
status: str
type: Literal["web_search_call"] = "web_search_call"
OpenAIResponseOutput = Annotated[
OpenAIResponseOutputMessage | OpenAIResponseOutputMessageWebSearchToolCall,
Field(discriminator="type"),
]
register_schema(OpenAIResponseOutput, name="OpenAIResponseOutput")
@json_schema_type
class OpenAIResponseObject(BaseModel):
created_at: int
error: OpenAIResponseError | None = None
id: str
model: str
object: Literal["response"] = "response"
output: list[OpenAIResponseOutput]
parallel_tool_calls: bool = False
previous_response_id: str | None = None
status: str
temperature: float | None = None
top_p: float | None = None
truncation: str | None = None
user: str | None = None
@json_schema_type
class OpenAIResponseObjectStreamResponseCreated(BaseModel):
response: OpenAIResponseObject
type: Literal["response.created"] = "response.created"
@json_schema_type
class OpenAIResponseObjectStreamResponseCompleted(BaseModel):
response: OpenAIResponseObject
type: Literal["response.completed"] = "response.completed"
OpenAIResponseObjectStream = Annotated[
OpenAIResponseObjectStreamResponseCreated | OpenAIResponseObjectStreamResponseCompleted,
Field(discriminator="type"),
]
register_schema(OpenAIResponseObjectStream, name="OpenAIResponseObjectStream")
@json_schema_type
class OpenAIResponseInputMessageContentText(BaseModel):
text: str
type: Literal["input_text"] = "input_text"
@json_schema_type
class OpenAIResponseInputMessageContentImage(BaseModel):
detail: Literal["low"] | Literal["high"] | Literal["auto"] = "auto"
type: Literal["input_image"] = "input_image"
# TODO: handle file_id
image_url: str | None = None
# TODO: handle file content types
OpenAIResponseInputMessageContent = Annotated[
OpenAIResponseInputMessageContentText | OpenAIResponseInputMessageContentImage,
Field(discriminator="type"),
]
register_schema(OpenAIResponseInputMessageContent, name="OpenAIResponseInputMessageContent")
@json_schema_type
class OpenAIResponseInputMessage(BaseModel):
content: str | list[OpenAIResponseInputMessageContent]
role: Literal["system"] | Literal["developer"] | Literal["user"] | Literal["assistant"]
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: str | None = Field(default="medium", pattern="^low|medium|high$")
# TODO: add user_location
OpenAIResponseInputTool = Annotated[
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: OpenAIChatCompletionMessageContent
name: Optional[str] = None
tool_calls: Optional[List[OpenAIChatCompletionToolCall]] = Field(default_factory=list)
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: ...

27
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)
@ -460,15 +459,17 @@ def run_download_cmd(args: argparse.Namespace, parser: argparse.ArgumentParser):
from llama_stack.models.llama.sku_list import llama_meta_net_info, resolve_model
from .model.safety_models import (
prompt_guard_download_info,
prompt_guard_model_sku,
prompt_guard_download_info_map,
prompt_guard_model_sku_map,
)
prompt_guard = prompt_guard_model_sku()
prompt_guard_model_sku_map = prompt_guard_model_sku_map()
prompt_guard_download_info_map = prompt_guard_download_info_map()
for model_id in model_ids:
if model_id == prompt_guard.model_id:
model = prompt_guard
info = prompt_guard_download_info()
if model_id in prompt_guard_model_sku_map.keys():
model = prompt_guard_model_sku_map[model_id]
info = prompt_guard_download_info_map[model_id]
else:
model = resolve_model(model_id)
if model is None:

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

@ -36,11 +36,11 @@ class ModelDescribe(Subcommand):
)
def _run_model_describe_cmd(self, args: argparse.Namespace) -> None:
from .safety_models import prompt_guard_model_sku
from .safety_models import prompt_guard_model_sku_map
prompt_guard = prompt_guard_model_sku()
if args.model_id == prompt_guard.model_id:
model = prompt_guard
prompt_guard_model_map = prompt_guard_model_sku_map()
if args.model_id in prompt_guard_model_map.keys():
model = prompt_guard_model_map[args.model_id]
else:
model = resolve_model(args.model_id)

4
llama_stack/cli/model/list.py
View file

@ -84,7 +84,7 @@ class ModelList(Subcommand):
)
def _run_model_list_cmd(self, args: argparse.Namespace) -> None:
from .safety_models import prompt_guard_model_sku
from .safety_models import prompt_guard_model_skus
if args.downloaded:
return _run_model_list_downloaded_cmd()
@ -96,7 +96,7 @@ class ModelList(Subcommand):
]
rows = []
for model in all_registered_models() + [prompt_guard_model_sku()]:
for model in all_registered_models() + prompt_guard_model_skus():
if not args.show_all and not model.is_featured:
continue

9
llama_stack/cli/model/remove.py
View file

@ -42,11 +42,12 @@ class ModelRemove(Subcommand):
)
def _run_model_remove_cmd(self, args: argparse.Namespace) -> None:
from .safety_models import prompt_guard_model_sku
from .safety_models import prompt_guard_model_sku_map
prompt_guard = prompt_guard_model_sku()
if args.model == prompt_guard.model_id:
model = prompt_guard
prompt_guard_model_map = prompt_guard_model_sku_map()
if args.model in prompt_guard_model_map.keys():
model = prompt_guard_model_map[args.model]
else:
model = resolve_model(args.model)

53
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
@ -15,14 +15,14 @@ from llama_stack.models.llama.sku_types import CheckpointQuantizationFormat
class PromptGuardModel(BaseModel):
"""Make a 'fake' Model-like object for Prompt Guard. Eventually this will be removed."""
model_id: str = "Prompt-Guard-86M"
model_id: str
huggingface_repo: str
description: str = "Prompt Guard. NOTE: this model will not be provided via `llama` CLI soon."
is_featured: bool = False
huggingface_repo: str = "meta-llama/Prompt-Guard-86M"
max_seq_length: int = 2048
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
@ -30,18 +30,35 @@ class PromptGuardModel(BaseModel):
model_config = ConfigDict(protected_namespaces=())
def prompt_guard_model_sku():
return PromptGuardModel()
def prompt_guard_model_skus():
return [
PromptGuardModel(model_id="Prompt-Guard-86M", huggingface_repo="meta-llama/Prompt-Guard-86M"),
PromptGuardModel(
model_id="Llama-Prompt-Guard-2-86M",
huggingface_repo="meta-llama/Llama-Prompt-Guard-2-86M",
),
PromptGuardModel(
model_id="Llama-Prompt-Guard-2-22M",
huggingface_repo="meta-llama/Llama-Prompt-Guard-2-22M",
),
]
def prompt_guard_download_info():
return LlamaDownloadInfo(
folder="Prompt-Guard",
files=[
"model.safetensors",
"special_tokens_map.json",
"tokenizer.json",
"tokenizer_config.json",
],
pth_size=1,
)
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]:
return {
model.model_id: LlamaDownloadInfo(
folder="Prompt-Guard" if model.model_id == "Prompt-Guard-86M" else model.model_id,
files=[
"model.safetensors",
"special_tokens_map.json",
"tokenizer.json",
"tokenizer_config.json",
],
pth_size=1,
)
for model in prompt_guard_model_skus()
}

127
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
@ -89,6 +88,43 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
color="red",
)
sys.exit(1)
elif args.providers:
providers = dict()
for api_provider in args.providers.split(","):
if "=" not in api_provider:
cprint(
"Could not parse `--providers`. Please ensure the list is in the format api1=provider1,api2=provider2",
color="red",
)
sys.exit(1)
api, provider = api_provider.split("=")
providers_for_api = get_provider_registry().get(Api(api), None)
if providers_for_api is None:
cprint(
f"{api} is not a valid API.",
color="red",
)
sys.exit(1)
if provider in providers_for_api:
providers.setdefault(api, []).append(provider)
else:
cprint(
f"{provider} is not a valid provider for the {api} API.",
color="red",
)
sys.exit(1)
distribution_spec = DistributionSpec(
providers=providers,
description=",".join(args.providers),
)
if not args.image_type:
cprint(
f"Please specify a image-type (container | conda | venv) for {args.template}",
color="red",
)
sys.exit(1)
build_config = BuildConfig(image_type=args.image_type, distribution_spec=distribution_spec)
elif not args.config and not args.template:
name = prompt(
"> Enter a name for your Llama Stack (e.g. my-local-stack): ",
@ -99,12 +135,13 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
)
image_type = prompt(
f"> Enter the image type you want your Llama Stack to be built as ({' or '.join(e.value for e in ImageType)}): ",
"> Enter the image type you want your Llama Stack to be built as (use <TAB> to see options): ",
completer=WordCompleter([e.value for e in ImageType]),
complete_while_typing=True,
validator=Validator.from_callable(
lambda x: x in [e.value for e in ImageType],
error_message=f"Invalid image type, please enter {' or '.join(e.value for e in ImageType)}",
error_message="Invalid image type. Use <TAB> to see options",
),
default=ImageType.CONDA.value,
)
if image_type == ImageType.CONDA.value:
@ -140,7 +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(
@ -163,7 +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:
@ -173,16 +210,9 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
)
sys.exit(1)
if build_config.image_type == LlamaStackImageType.CONTAINER.value and not args.image_name:
cprint(
"Please specify --image-name when building a container from a config file",
color="red",
)
sys.exit(1)
if args.print_deps_only:
print(f"# Dependencies for {args.template or args.config or image_name}")
normal_deps, special_deps = get_provider_dependencies(build_config.distribution_spec.providers)
normal_deps, special_deps = get_provider_dependencies(build_config)
normal_deps += SERVER_DEPENDENCIES
print(f"uv pip install {' '.join(normal_deps)}")
for special_dep in special_deps:
@ -198,10 +228,14 @@ def run_stack_build_command(args: argparse.Namespace) -> None:
)
except (Exception, RuntimeError) as exc:
import traceback
cprint(
f"Error building stack: {exc}",
color="red",
)
cprint("Stack trace:", color="red")
traceback.print_exc()
sys.exit(1)
if run_config is None:
cprint(
@ -233,9 +267,10 @@ def _generate_run_config(
image_name=image_name,
apis=apis,
providers={},
external_providers_dir=build_config.external_providers_dir if build_config.external_providers_dir else None,
)
# build providers dict
provider_registry = get_provider_registry()
provider_registry = get_provider_registry(build_config)
for api in apis:
run_config.providers[api] = []
provider_types = build_config.distribution_spec.providers[api]
@ -249,8 +284,22 @@ def _generate_run_config(
if p.deprecation_error:
raise InvalidProviderError(p.deprecation_error)
config_type = instantiate_class_type(provider_registry[Api(api)][provider_type].config_class)
if hasattr(config_type, "sample_run_config"):
try:
config_type = instantiate_class_type(provider_registry[Api(api)][provider_type].config_class)
except ModuleNotFoundError:
# HACK ALERT:
# This code executes after building is done, the import cannot work since the
# package is either available in the venv or container - not available on the host.
# TODO: use a "is_external" flag in ProviderSpec to check if the provider is
# external
cprint(
f"Failed to import provider {provider_type} for API {api} - assuming it's external, skipping",
color="yellow",
)
# Set config_type to None to avoid UnboundLocalError
config_type = None
if config_type is not None and hasattr(config_type, "sample_run_config"):
config = config_type.sample_run_config(__distro_dir__=f"~/.llama/distributions/{image_name}")
else:
config = {}
@ -268,20 +317,25 @@ def _generate_run_config(
to_write = json.loads(run_config.model_dump_json())
f.write(yaml.dump(to_write, sort_keys=False))
# this path is only invoked when no template is provided
cprint(
f"You can now run your stack with `llama stack run {run_config_file}`",
color="green",
)
# Only print this message for non-container builds since it will be displayed before the
# container is built
# For non-container builds, the run.yaml is generated at the very end of the build process so it
# makes sense to display this message
if build_config.image_type != LlamaStackImageType.CONTAINER.value:
cprint(
f"You can now run your stack with `llama stack run {run_config_file}`",
color="green",
)
return run_config_file
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:
if template_name:
image_name = f"distribution-{template_name}"
@ -305,6 +359,13 @@ def _run_stack_build_command_from_build_config(
build_file_path = build_dir / f"{image_name}-build.yaml"
os.makedirs(build_dir, exist_ok=True)
run_config_file = None
# Generate the run.yaml so it can be included in the container image with the proper entrypoint
# Only do this if we're building a container image and we're not using a template
if build_config.image_type == LlamaStackImageType.CONTAINER.value and not template_name and config_path:
cprint("Generating run.yaml file", color="green")
run_config_file = _generate_run_config(build_config, build_dir, image_name)
with open(build_file_path, "w") as f:
to_write = json.loads(build_config.model_dump_json())
f.write(yaml.dump(to_write, sort_keys=False))
@ -313,7 +374,8 @@ def _run_stack_build_command_from_build_config(
build_config,
build_file_path,
image_name,
template_or_config=template_name or config_path,
template_or_config=template_name or config_path or str(build_file_path),
run_config=run_config_file,
)
if return_code != 0:
raise RuntimeError(f"Failed to build image {image_name}")
@ -326,6 +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)

6
llama_stack/cli/stack/build.py
View file

@ -75,6 +75,12 @@ the build. If not specified, currently active environment will be used if found.
default=False,
help="Run the stack after building using the same image type, name, and other applicable arguments",
)
self.parser.add_argument(
"--providers",
type=str,
default=None,
help="Build a config for a list of providers and only those providers. This list is formatted like: api1=provider1,api2=provider2. Where there can be multiple providers per API.",
)
def _run_stack_build_command(self, args: argparse.Namespace) -> None:
# always keep implementation completely silo-ed away from CLI so CLI

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.

32
llama_stack/distribution/build.py
View file

@ -7,16 +7,16 @@
import importlib.resources
import logging
from pathlib import Path
from typing import Dict, List
from pydantic import BaseModel
from termcolor import cprint
from llama_stack.distribution.datatypes import BuildConfig, Provider
from llama_stack.distribution.datatypes import BuildConfig
from llama_stack.distribution.distribution import get_provider_registry
from llama_stack.distribution.utils.exec import run_command
from llama_stack.distribution.utils.image_types import LlamaStackImageType
from llama_stack.providers.datatypes import Api
from llama_stack.templates.template import DistributionTemplate
log = logging.getLogger(__name__)
@ -37,19 +37,23 @@ class ApiInput(BaseModel):
def get_provider_dependencies(
config_providers: Dict[str, List[Provider]],
config: BuildConfig | DistributionTemplate,
) -> tuple[list[str], list[str]]:
"""Get normal and special dependencies from provider configuration."""
all_providers = get_provider_registry()
# Extract providers based on config type
if isinstance(config, DistributionTemplate):
providers = config.providers
elif isinstance(config, BuildConfig):
providers = config.distribution_spec.providers
deps = []
for api_str, provider_or_providers in config_providers.items():
providers_for_api = all_providers[Api(api_str)]
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:
@ -71,8 +75,8 @@ def get_provider_dependencies(
return list(set(normal_deps)), list(set(special_deps))
def print_pip_install_help(providers: Dict[str, List[Provider]]):
normal_deps, special_deps = get_provider_dependencies(providers)
def print_pip_install_help(config: BuildConfig):
normal_deps, special_deps = get_provider_dependencies(config)
cprint(
f"Please install needed dependencies using the following commands:\n\nuv pip install {' '.join(normal_deps)}",
@ -88,10 +92,11 @@ def build_image(
build_file_path: Path,
image_name: str,
template_or_config: str,
run_config: str | None = None,
):
container_base = build_config.distribution_spec.container_image or "python:3.10-slim"
normal_deps, special_deps = get_provider_dependencies(build_config.distribution_spec.providers)
normal_deps, special_deps = get_provider_dependencies(build_config)
normal_deps += SERVER_DEPENDENCIES
if build_config.image_type == LlamaStackImageType.CONTAINER.value:
@ -103,6 +108,11 @@ def build_image(
container_base,
" ".join(normal_deps),
]
# When building from a config file (not a template), include the run config path in the
# build arguments
if run_config is not None:
args.append(run_config)
elif build_config.image_type == LlamaStackImageType.CONDA.value:
script = str(importlib.resources.files("llama_stack") / "distribution/build_conda_env.sh")
args = [

94
llama_stack/distribution/build_container.sh
View file

@ -19,12 +19,16 @@ UV_HTTP_TIMEOUT=${UV_HTTP_TIMEOUT:-500}
# mounting is not supported by docker buildx, so we use COPY instead
USE_COPY_NOT_MOUNT=${USE_COPY_NOT_MOUNT:-}
# Path to the run.yaml file in the container
RUN_CONFIG_PATH=/app/run.yaml
BUILD_CONTEXT_DIR=$(pwd)
if [ "$#" -lt 4 ]; then
# This only works for templates
echo "Usage: $0 <template_or_config> <image_name> <container_base> <pip_dependencies> [<special_pip_deps>]" >&2
echo "Usage: $0 <template_or_config> <image_name> <container_base> <pip_dependencies> [<run_config>] [<special_pip_deps>]" >&2
exit 1
fi
set -euo pipefail
template_or_config="$1"
@ -35,8 +39,27 @@ container_base="$1"
shift
pip_dependencies="$1"
shift
special_pip_deps="${1:-}"
# Handle optional arguments
run_config=""
special_pip_deps=""
# Check if there are more arguments
# The logics is becoming cumbersom, we should refactor it if we can do better
if [ $# -gt 0 ]; then
# Check if the argument ends with .yaml
if [[ "$1" == *.yaml ]]; then
run_config="$1"
shift
# If there's another argument after .yaml, it must be special_pip_deps
if [ $# -gt 0 ]; then
special_pip_deps="$1"
fi
else
# If it's not .yaml, it must be special_pip_deps
special_pip_deps="$1"
fi
fi
# Define color codes
RED='\033[0;31m'
@ -72,9 +95,13 @@ if [[ $container_base == *"registry.access.redhat.com/ubi9"* ]]; then
FROM $container_base
WORKDIR /app
RUN dnf -y update && dnf install -y iputils net-tools wget \
# We install the Python 3.11 dev headers and build tools so that any
# C‑extension wheels (e.g. polyleven, faiss‑cpu) can compile successfully.
RUN dnf -y update && dnf install -y iputils git net-tools wget \
vim-minimal python3.11 python3.11-pip python3.11-wheel \
python3.11-setuptools && ln -s /bin/pip3.11 /bin/pip && ln -s /bin/python3.11 /bin/python && dnf clean all
python3.11-setuptools python3.11-devel gcc make && \
ln -s /bin/pip3.11 /bin/pip && ln -s /bin/python3.11 /bin/python && dnf clean all
ENV UV_SYSTEM_PYTHON=1
RUN pip install uv
@ -86,7 +113,7 @@ WORKDIR /app
RUN apt-get update && apt-get install -y \
iputils-ping net-tools iproute2 dnsutils telnet \
curl wget telnet \
curl wget telnet git\
procps psmisc lsof \
traceroute \
bubblewrap \
@ -115,6 +142,45 @@ EOF
done
fi
# Function to get Python command
get_python_cmd() {
if is_command_available python; then
echo "python"
elif is_command_available python3; then
echo "python3"
else
echo "Error: Neither python nor python3 is installed. Please install Python to continue." >&2
exit 1
fi
}
if [ -n "$run_config" ]; then
# Copy the run config to the build context since it's an absolute path
cp "$run_config" "$BUILD_CONTEXT_DIR/run.yaml"
add_to_container << EOF
COPY run.yaml $RUN_CONFIG_PATH
EOF
# Parse the run.yaml configuration to identify external provider directories
# If external providers are specified, copy their directory to the container
# and update the configuration to reference the new container path
python_cmd=$(get_python_cmd)
external_providers_dir=$($python_cmd -c "import yaml; config = yaml.safe_load(open('$run_config')); print(config.get('external_providers_dir') or '')")
if [ -n "$external_providers_dir" ]; then
echo "Copying external providers directory: $external_providers_dir"
add_to_container << EOF
COPY $external_providers_dir /app/providers.d
EOF
# Edit the run.yaml file to change the external_providers_dir to /app/providers.d
if [ "$(uname)" = "Darwin" ]; then
sed -i.bak -e 's|external_providers_dir:.*|external_providers_dir: /app/providers.d|' "$BUILD_CONTEXT_DIR/run.yaml"
rm -f "$BUILD_CONTEXT_DIR/run.yaml.bak"
else
sed -i 's|external_providers_dir:.*|external_providers_dir: /app/providers.d|' "$BUILD_CONTEXT_DIR/run.yaml"
fi
fi
fi
stack_mount="/app/llama-stack-source"
client_mount="/app/llama-stack-client-source"
@ -174,15 +240,16 @@ fi
RUN pip uninstall -y uv
EOF
# if template_or_config ends with .yaml, it is not a template and we should not use the --template flag
if [[ "$template_or_config" != *.yaml ]]; then
# If a run config is provided, we use the --config flag
if [[ -n "$run_config" ]]; then
add_to_container << EOF
ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server", "--config", "$RUN_CONFIG_PATH"]
EOF
# If a template is provided (not a yaml file), we use the --template flag
elif [[ "$template_or_config" != *.yaml ]]; then
add_to_container << EOF
ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server", "--template", "$template_or_config"]
EOF
else
add_to_container << EOF
ENTRYPOINT ["python", "-m", "llama_stack.distribution.server.server"]
EOF
fi
# Add other require item commands genearic to all containers
@ -254,9 +321,10 @@ $CONTAINER_BINARY build \
"${CLI_ARGS[@]}" \
-t "$image_tag" \
-f "$TEMP_DIR/Containerfile" \
"."
"$BUILD_CONTEXT_DIR"
# clean up tmp/configs
rm -f "$BUILD_CONTEXT_DIR/run.yaml"
set +x
echo "Success!"

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)

133
llama_stack/distribution/datatypes.py
View file

@ -4,7 +4,8 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Annotated, Any, Dict, List, Optional, Union
from enum import Enum
from typing import Annotated, Any
from pydantic import BaseModel, Field
@ -29,7 +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):
@ -46,17 +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"
)
@ -105,7 +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
@ -141,41 +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
@ -183,15 +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")
@ -199,20 +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
@ -224,21 +205,32 @@ 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""",
)
class AuthProviderType(str, Enum):
"""Supported authentication provider types."""
KUBERNETES = "kubernetes"
CUSTOM = "custom"
class AuthenticationConfig(BaseModel):
endpoint: str = Field(
provider_type: AuthProviderType = Field(
...,
description="Endpoint URL to validate authentication tokens",
description="Type of authentication provider (e.g., 'kubernetes', 'custom')",
)
config: dict[str, str] = Field(
...,
description="Provider-specific configuration",
)
@ -249,15 +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",
)
@ -273,23 +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,
@ -297,22 +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.",
)
@ -326,3 +318,12 @@ class BuildConfig(BaseModel):
default="conda",
description="Type of package to build (conda | container | venv)",
)
image_name: str | None = Field(
default=None,
description="Name of the distribution to build",
)
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.",
)

24
llama_stack/distribution/distribution.py
View file

@ -7,12 +7,11 @@
import glob
import importlib
import os
from typing import Any, Dict, List
from typing import Any
import yaml
from pydantic import BaseModel
from llama_stack.distribution.datatypes import StackRunConfig
from llama_stack.log import get_logger
from llama_stack.providers.datatypes import (
AdapterSpec,
@ -25,7 +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)
@ -34,7 +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,
@ -67,12 +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,
@ -82,7 +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}",
@ -97,7 +96,9 @@ def _load_inline_provider_spec(spec_data: Dict[str, Any], api: Api, provider_nam
return spec
def get_provider_registry(config: StackRunConfig | None = None) -> Dict[Api, Dict[str, ProviderSpec]]:
def get_provider_registry(
config=None,
) -> 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.
@ -122,7 +123,7 @@ def get_provider_registry(config: StackRunConfig | None = None) -> Dict[Api, Dic
llama-guard.yaml
Args:
config: Optional StackRunConfig containing the external providers directory path
config: Optional object containing the external providers directory path
Returns:
A dictionary mapping APIs to their available providers
@ -132,7 +133,7 @@ def get_provider_registry(config: StackRunConfig | None = None) -> Dict[Api, Dic
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}")
@ -142,7 +143,8 @@ def get_provider_registry(config: StackRunConfig | None = None) -> Dict[Api, Dic
except ImportError as e:
logger.warning(f"Failed to import module {name}: {e}")
if config and config.external_providers_dir:
# Check if config has the external_providers_dir attribute
if config and hasattr(config, "external_providers_dir") and config.external_providers_dir:
external_providers_dir = os.path.abspath(config.external_providers_dir)
if not os.path.exists(external_providers_dir):
raise FileNotFoundError(f"External providers directory not found: {external_providers_dir}")

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 {}

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