# What does this PR do?
update vLLM inference provider to use OpenAIMixin for openai-compat
functions
inference recordings from Qwen3-0.6B and vLLM 0.8.3 -
```
docker run --gpus all -v ~/.cache/huggingface:/root/.cache/huggingface -p 8000:8000 --ipc=host \
vllm/vllm-openai:latest \
--model Qwen/Qwen3-0.6B --enable-auto-tool-choice --tool-call-parser hermes
```
## Test Plan
```
./scripts/integration-tests.sh --stack-config server:ci-tests --setup vllm --subdirs inference
```
Fixes#3370
AWS switched to requiring region-prefixed inference profile IDs instead
of foundation model IDs for on-demand throughput. This was causing
ValidationException errors.
Added auto-detection based on boto3 client region to convert model IDs
like meta.llama3-1-70b-instruct-v1:0 to
us.meta.llama3-1-70b-instruct-v1:0 depending on the detected region.
Also handles edge cases like ARNs, case insensitive regions, and None
regions.
Tested with this request.
```json
{
"model_id": "meta.llama3-1-8b-instruct-v1:0",
"messages": [
{
"role": "system",
"content": "You are a helpful assistant."
},
{
"role": "user",
"content": "tell me a riddle"
}
],
"sampling_params": {
"strategy": {
"type": "top_p",
"temperature": 0.7,
"top_p": 0.9
},
"max_tokens": 512
}
}
```
<img width="1488" height="878" alt="image"
src="https://github.com/user-attachments/assets/0d61beec-3869-4a31-8f37-9f554c280b88"
/>
# What does this PR do?
update VertexAI inference provider to use openai-python for
openai-compat functions
## Test Plan
```
$ VERTEX_AI_PROJECT=... uv run llama stack build --image-type venv --providers inference=remote::vertexai --run
...
$ LLAMA_STACK_CONFIG=http://localhost:8321 uv run --group test pytest -v -ra --text-model vertexai/vertex_ai/gemini-2.5-flash tests/integration/inference/test_openai_completion.py
...
```
i don't have an account to test this. `get_api_key` may also need to be
updated per
https://cloud.google.com/vertex-ai/generative-ai/docs/start/openai
---------
Signed-off-by: Sébastien Han <seb@redhat.com>
Co-authored-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
update the Anthropic inference provider to use openai-python for the
openai-compat endpoints
## Test Plan
ci
Co-authored-by: raghotham <rsm@meta.com>
# What does this PR do?
update Groq inference provider to use OpenAIMixin for openai-compat
endpoints
changes on api.groq.com -
- json_schema is now supported for specific models, see
https://console.groq.com/docs/structured-outputs#supported-models
- response_format with streaming is now supported for models that
support response_format
- groq no longer returns a 400 error if tools are provided and
tool_choice is not "required"
## Test Plan
```
$ GROQ_API_KEY=... uv run llama stack build --image-type venv --providers inference=remote::groq --run
...
$ LLAMA_STACK_CONFIG=http://localhost:8321 uv run --group test pytest -v -ra --text-model groq/llama-3.3-70b-versatile tests/integration/inference/test_openai_completion.py -k 'not store'
...
SKIPPED [3] tests/integration/inference/test_openai_completion.py:44: Model groq/llama-3.3-70b-versatile hosted by remote::groq doesn't support OpenAI completions.
SKIPPED [3] tests/integration/inference/test_openai_completion.py:94: Model groq/llama-3.3-70b-versatile hosted by remote::groq doesn't support vllm extra_body parameters.
SKIPPED [4] tests/integration/inference/test_openai_completion.py:73: Model groq/llama-3.3-70b-versatile hosted by remote::groq doesn't support n param.
SKIPPED [1] tests/integration/inference/test_openai_completion.py💯 Model groq/llama-3.3-70b-versatile hosted by remote::groq doesn't support chat completion calls with base64 encoded files.
======================= 8 passed, 11 skipped, 8 deselected, 2 warnings in 5.13s ========================
```
---------
Co-authored-by: raghotham <rsm@meta.com>
# What does this PR do?
update SambaNova inference provider to use OpenAIMixin for openai-compat
endpoints
## Test Plan
```
$ SAMBANOVA_API_KEY=... uv run llama stack build --image-type venv --providers inference=remote::sambanova --run
...
$ LLAMA_STACK_CONFIG=http://localhost:8321 uv run --group test pytest -v -ra --text-model sambanova/Meta-Llama-3.3-70B-Instruct tests/integration/inference -k 'not store'
...
FAILED tests/integration/inference/test_text_inference.py::test_text_chat_completion_tool_calling_tools_not_in_request[txt=sambanova/Meta-Llama-3.3-70B-Instruct-inference:chat_completion:tool_calling_tools_absent-True] - AttributeError: 'NoneType' object has no attribute 'delta'
FAILED tests/integration/inference/test_text_inference.py::test_text_chat_completion_tool_calling_tools_not_in_request[txt=sambanova/Meta-Llama-3.3-70B-Instruct-inference:chat_completion:tool_calling_tools_absent-False] - llama_stack_client.InternalServerError: Error code: 500 - {'detail': 'Internal server error: An une...
=========== 2 failed, 16 passed, 68 skipped, 8 deselected, 3 xfailed, 13 warnings in 15.85s ============
```
the two failures also exist before this change. they are part of the
deprecated inference.chat_completion tests that flow through litellm.
they can be resolved later.
# What does this PR do?
update the Gemini inference provider to use openai-python for the
openai-compat endpoints
partially addresses #3349, does not address /inference/completion or
/inference/chat-completion
## Test Plan
ci
One needed to specify record-replay related environment variables for
running integration tests. We could not use defaults because integration
tests could be run against Ollama instances which could be running
different models. For example, text vs vision tests needed separate
instances of Ollama because a single instance typically cannot serve
both of these models if you assume the standard CI worker configuration
on Github. As a result, `client.list()` as returned by the Ollama client
would be different between these runs and we'd end up overwriting
responses.
This PR "solves" it by adding a small amount of complexity -- we store
model list responses specially, keyed by the hashes of the models they
return. At replay time, we merge all of them and pretend that we have
the union of all models available.
## Test Plan
Re-recorded all the tests using `scripts/integration-tests.sh
--inference-mode record`, including the vision tests.
# What does this PR do?
This PR updates the Watsonx provider dependencies from
`ibm_watson_machine_learning` to `ibm_watsonx_ai`.
The old package `ibm_watson_machine_learning` is in **deprecation mode**
([[PyPI
link](https://pypi.org/project/ibm-watson-machine-learning/)](https://pypi.org/project/ibm-watson-machine-learning/))
and relies on older versions of dependencies such as `pandas`. Updating
to `ibm_watsonx_ai` ensures compatibility with current dependency
versions and ongoing support.
## Test Plan
I verified the update by running an inference using a model provided by
Watsonx. The model ran successfully, confirming that the new dependency
works as expected.
Co-authored-by: are-ces <cpompeia@redhat.com>
# What does this PR do?
closes https://github.com/llamastack/llama-stack/issues/3236
mypy considered our default implementations (raise NotImplementedError)
to be trivial. the result was we implemented the same stubs in
providers.
this change puts enough into the default impls so mypy considers them
non-trivial. this allows us to remove the duplicate implementations.
# What does this PR do?
Context: https://github.com/meta-llama/llama-stack/issues/2937
The API design is inspired by existing offerings, but not exactly the
same:
* `top_n` as the parameter to control number of results, instead of
`top_k`, since `n` is conventional to control number
* `truncation` bool instead of `max_token_per_doc`, since we should just
handle the truncation automatically depending on model capability,
instead of user setting the context length manually.
* `data` field in the response, to be consistent with other OpenAI APIs
(though they don't have a rerank API). Also, it is one less name to
learn in the API.
## Test Plan
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
This PR renames categories of llama_stack loggers.
This PR aligns logging categories as per the package name, as well as
reviews from initial
https://github.com/meta-llama/llama-stack/pull/2868. This is a follow up
to #3061.
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
Replaces https://github.com/meta-llama/llama-stack/pull/2868
Part of https://github.com/meta-llama/llama-stack/issues/2865
cc @leseb @rhuss
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
# What does this PR do?
Currently the embedding integration test cases fail due to a
misalignment in the error type. This PR fixes the embedding integration
test by fixing the error type.
## Test Plan
```
pytest -s -v tests/integration/inference/test_embedding.py --stack-config="inference=nvidia" --embedding-model="nvidia/llama-3.2-nv-embedqa-1b-v2" --env NVIDIA_API_KEY={nvidia_api_key} --env NVIDIA_BASE_URL="https://integrate.api.nvidia.com"
```
# What does this PR do?
- Documentation update and fix for the NVIDIA Inference provider.
- Update the `run_moderation` for safety API with a
`NotImplementedError` placeholder. Otherwise initialization NVIDIA
inference client will raise an error.
## Test Plan
N/A
# What does this PR do?
NVIDIA asymmetric embedding models (e.g.,
`nvidia/llama-3.2-nv-embedqa-1b-v2`) require an `input_type` parameter
not present in the standard OpenAI embeddings API. This PR adds the
`input_type="query"` as default and updates the documentation to suggest
using the `embedding` API for passage embeddings.
<!-- If resolving an issue, uncomment and update the line below -->
Resolves#2892
## Test Plan
```
pytest -s -v tests/integration/inference/test_openai_embeddings.py --stack-config="inference=nvidia" --embedding-model="nvidia/llama-3.2-nv-embedqa-1b-v2" --env NVIDIA_API_KEY={nvidia_api_key} --env NVIDIA_BASE_URL="https://integrate.api.nvidia.com"
```
# What does this PR do?
This PR adds a step in pre-commit to enforce using `llama_stack` logger.
Currently, various parts of the code base uses different loggers. As a
custom `llama_stack` logger exist and used in the codebase, it is better
to standardize its utilization.
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
Co-authored-by: Matthew Farrellee <matt@cs.wisc.edu>
Some fixes to MCP tests. And a bunch of fixes for Vector providers.
I also enabled a bunch of Vector IO tests to be used with
`LlamaStackLibraryClient`
## Test Plan
Run Responses tests with llama stack library client:
```
pytest -s -v tests/integration/non_ci/responses/ --stack-config=server:starter \
--text-model openai/gpt-4o \
--embedding-model=sentence-transformers/all-MiniLM-L6-v2 \
-k "client_with_models"
```
Do the same with `-k openai_client`
The rest should be taken care of by CI.
# What does this PR do?
- Add new Vertex AI remote inference provider with litellm integration
- Support for Gemini models through Google Cloud Vertex AI platform
- Uses Google Cloud Application Default Credentials (ADC) for
authentication
- Added VertexAI models: gemini-2.5-flash, gemini-2.5-pro,
gemini-2.0-flash.
- Updated provider registry to include vertexai provider
- Updated starter template to support Vertex AI configuration
- Added comprehensive documentation and sample configuration
<!-- If resolving an issue, uncomment and update the line below -->
relates to https://github.com/meta-llama/llama-stack/issues/2747
## Test Plan
<!-- Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.* -->
Signed-off-by: Eran Cohen <eranco@redhat.com>
Co-authored-by: Francisco Arceo <arceofrancisco@gmail.com>
PR adds Flash-Lite 2.0 and 2.5 models to the Gemini inference provider
Closes#3046
## Test Plan
I was not able to locate any existing test for this provider, so I
performed manual testing. But the change is really trivial and
straightforward.
A bunch of miscellaneous cleanup focusing on tests, but ended up
speeding up starter distro substantially.
- Pulled llama stack client init for tests into `pytest_sessionstart` so
it does not clobber output
- Profiling of that told me where we were doing lots of heavy imports
for starter, so lazied them
- starter now starts 20seconds+ faster on my Mac
- A few other smallish refactors for `compat_client`
As the title says. Distributions is in, Templates is out.
`llama stack build --template` --> `llama stack build --distro`. For
backward compatibility, the previous option is kept but results in a
warning.
Updated `server.py` to remove the "config_or_template" backward
compatibility since it has been a couple releases since that change.
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
This PR is responsible for removal of Conda support in Llama Stack
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
Closes#2539
## Test Plan
<!-- Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.* -->
# What does this PR do?
closes#2995
update SambaNovaInferenceAdapter to efficiently use LiteLLMOpenAIMixin
## Test Plan
```
$ uv run pytest -s -v tests/integration/inference --stack-config inference=sambanova --text-model sambanova/Meta-Llama-3.1-8B-Instruct
...
======================== 10 passed, 84 skipped, 3 xfailed, 51 warnings in 8.14s ========================
```
# What does this PR do?
- Add base_url field to OpenAIConfig with default
"https://api.openai.com/v1"
- Update sample_run_config to support OPENAI_BASE_URL environment
variable
- Modify get_base_url() to return configured base_url instead of
hardcoded value
- Add comprehensive test suite covering:
- Default base URL behavior
- Custom base URL from config
- Environment variable override
- Config precedence over environment variables
- Client initialization with configured URL
- Model availability checks using configured URL
This enables users to configure custom OpenAI-compatible API endpoints
via environment variables or configuration files.
Closes#2910
## Test Plan
run unit tests
This flips #2823 and #2805 by making the Stack periodically query the
providers for models rather than the providers going behind the back and
calling "register" on to the registry themselves. This also adds support
for model listing for all other providers via `ModelRegistryHelper`.
Once this is done, we do not need to manually list or register models
via `run.yaml` and it will remove both noise and annoyance (setting
`INFERENCE_MODEL` environment variables, for example) from the new user
experience.
In addition, it adds a configuration variable `allowed_models` which can
be used to optionally restrict the set of models exposed from a
provider.
# What does this PR do?
openai/models.py has backward compat entries for litellm model names.
the starter template includes these in the list of registered models.
the inclusion results in duplicate model registrations.
the backward compat is no longer necessary.
## Test Plan
ci
# What does this PR do?
add an `OpenAIMixin` for use by inference providers who remote endpoints
support an OpenAI compatible API.
use is demonstrated by refactoring
- OpenAIInferenceAdapter
- NVIDIAInferenceAdapter (adds embedding support)
- LlamaCompatInferenceAdapter
## Test Plan
existing unit and integration tests
Just like #2805 but for vLLM.
We also make VLLM_URL env variable optional (not required) -- if not
specified, the provider silently sits idle and yells eventually if
someone tries to call a completion on it. This is done so as to allow
this provider to be present in the `starter` distribution.
## Test Plan
Set up vLLM, copy the starter template and set `{ refresh_models: true,
refresh_models_interval: 10 }` for the vllm provider and then run:
```
ENABLE_VLLM=vllm VLLM_URL=http://localhost:8000/v1 \
uv run llama stack run --image-type venv /tmp/starter.yaml
```
Verify that `llama-stack-client models list` brings up the model
correctly from vLLM.
For self-hosted providers like Ollama (or vLLM), the backing server is
running a set of models. That server should be treated as the source of
truth and the Stack registry should just be a cache for those models. Of
course, in production environments, you may not want this (because you
know what model you are running statically) hence there's a config
boolean to control this behavior.
_This is part of a series of PRs aimed at removing the requirement of
needing to set `INFERENCE_MODEL` env variables for running Llama Stack
server._
## Test Plan
Copy and modify the starter.yaml template / config and enable
`refresh_models: true, refresh_models_interval: 10` for the ollama
provider. Then, run:
```
LLAMA_STACK_LOGGING=all=debug \
ENABLE_OLLAMA=ollama uv run llama stack run --image-type venv /tmp/starter.yaml
```
See a gargantuan amount of logs, but verify that the provider is
periodically refreshing models. Stop and prune a model from ollama
server, restart the server. Verify that the model goes away when I call
`uv run llama-stack-client models list`
# What does this PR do?
let's users register models available at
https://integrate.api.nvidia.com/v1/models that isn't already in
llama_stack/providers/remote/inference/nvidia/models.py
## Test Plan
1. run the nvidia distro
2. register a model from https://integrate.api.nvidia.com/v1/models that
isn't already know, as of this writing
nvidia/llama-3.1-nemotron-ultra-253b-v1 is a good example
3. perform inference w/ the model
The vision models are now available at the standard URL, so the
workaround code has been removed. This also simplifies the codebase by
eliminating the need for per-model client caching.
- Remove special URL handling for meta/llama-3.2-11b/90b-vision-instruct
models
- Convert _get_client method to _client property for cleaner API
- Remove unnecessary lru_cache decorator and functools import
- Simplify client creation logic to use single base URL for all models
- fireworks, together do not support Llama-guard 3 8b model anymore
- Need to default to ollama
- current safety shields logic was not correct since the shield_id was
the provider ( which had duplicates )
- Followed similar logic to models
Note: Seems a bit over-engineered but this can now be extended to other
providers and fits in the overall mechanism of how env_vars are used to
manage starter.
### How to test
```
ENABLE_OLLAMA=ollama ENABLE_FIREWORKS=fireworks SAFETY_MODEL=llama-guard3:1b pytest -s -v tests/integration/ --stack-config starter -k 'not(supervised_fine_tune or builtin_tool_code or safety_with_image or code_interpreter_for or rag_and_code or truncation or register_and_unregister)' --text-model fireworks/meta-llama/Llama-3.3-70B-Instruct --vision-model fireworks/meta-llama/Llama-4-Scout-17B-16E-Instruct --safety-shield llama-guard3:1b --embedding-model all-MiniLM-L6-v2
```
### Related but not obvious in this PR
In the llama-stack-ops repo, we run tests before publishing packages and
docker containers.
The actions in that repo were using the fireworks / together distros (
which are non-existent )
So need to update that to run with `starter` and use `ollama`
specifically for safety.
# What does this PR do?
Some of our inference providers support passthrough authentication via
`x-llamastack-provider-data` header values. This fixes the providers
that support passthrough auth to not cache their clients to the backend
providers (mostly OpenAI client instances) so that the client connecting
to Llama Stack has to provide those auth values on each and every
request.
## Test Plan
I added some unit tests to ensure we're not caching clients across
requests for all the fixed providers in this PR.
```
uv run pytest -sv tests/unit/providers/inference/test_inference_client_caching.py
```
I also ran some of our OpenAI compatible API integration tests for each
of the changed providers, just to ensure they still work. Note that
these providers don't actually pass all these tests (for unrelated
reasons due to quirks of the Groq and Together SaaS services), but
enough of the tests passed to confirm the clients are still working as
intended.
### Together
```
ENABLE_TOGETHER="together" \
uv run llama stack run llama_stack/templates/starter/run.yaml
LLAMA_STACK_CONFIG=http://localhost:8321 \
uv run pytest -sv \
tests/integration/inference/test_openai_completion.py \
--text-model "together/meta-llama/Llama-3.1-8B-Instruct"
```
### OpenAI
```
ENABLE_OPENAI="openai" \
uv run llama stack run llama_stack/templates/starter/run.yaml
LLAMA_STACK_CONFIG=http://localhost:8321 \
uv run pytest -sv \
tests/integration/inference/test_openai_completion.py \
--text-model "openai/gpt-4o-mini"
```
### Groq
```
ENABLE_GROQ="groq" \
uv run llama stack run llama_stack/templates/starter/run.yaml
LLAMA_STACK_CONFIG=http://localhost:8321 \
uv run pytest -sv \
tests/integration/inference/test_openai_completion.py \
--text-model "groq/meta-llama/Llama-3.1-8B-Instruct"
```
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
Update the shield register validation of Sambanova not to raise, but
only warn when a model is not available in the base url endpoint used,
also added warnings when model is not available in the base url endpoint
used
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
## Test Plan
<!-- Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.* -->
run starter distro with Sambanova enabled
# What does this PR do?
We are now testing the safety capability with the starter image. This
includes a few changes:
* Enable the safety integration test
* Relax the shield model requirements from llama-guard to make it work
with llama-guard3:8b coming from Ollama
* Expose a shield for each inference provider in the starter distro. The
shield will only be registered if the provider is enabled.
Closes: https://github.com/meta-llama/llama-stack/issues/2528
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
- we are using `all-minilm:l6-v2` but the model we download from ollama
is `all-minilm:latest`
latest: https://ollama.com/library/all-minilm:latest 1b226e2802db
l6-v2: https://ollama.com/library/all-minilm:l6-v2 pin 1b226e2802db
- even currently they are exactly the same model but if
[all-minilm:l12-v2](https://ollama.com/library/all-minilm:l12-v2) is
updated, "latest" might not be the same for l6-v2.
- the only change in this PR is pin the model id in ollama
- also update detailed_tutorial with "starter" to replace deprecated
"ollama".
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
## Test Plan
<!-- Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.* -->
```
>INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct"
>llama stack build --run --template ollama --image-type venv
...
Build Successful!
You can find the newly-built template here: /home/wenzhou/zdtsw-forking/lls/llama-stack/llama_stack/templates/ollama/run.yaml
....
- metadata:
embedding_dimension: 384
model_id: all-MiniLM-L6-v2
model_type: !!python/object/apply:llama_stack.apis.models.models.ModelType
- embedding
provider_id: ollama
provider_model_id: all-minilm:l6-v2
...
```
test
```
>llama-stack-client inference chat-completion --message "Write me a 2-sentence poem about the moon"
INFO:httpx:HTTP Request: GET http://localhost:8321/v1/models "HTTP/1.1 200 OK"
INFO:httpx:HTTP Request: POST http://localhost:8321/v1/openai/v1/chat/completions "HTTP/1.1 200 OK"
OpenAIChatCompletion(
id='chatcmpl-04f99071-3da2-44ba-a19f-03b5b7fc70b7',
choices=[
OpenAIChatCompletionChoice(
finish_reason='stop',
index=0,
message=OpenAIChatCompletionChoiceMessageOpenAIAssistantMessageParam(
role='assistant',
content="Here is a 2-sentence poem about the moon:\n\nSilver crescent in the midnight sky,\nLuna's gentle face, a beauty to the eye.",
name=None,
tool_calls=None,
refusal=None,
annotations=None,
audio=None,
function_call=None
),
logprobs=None
)
],
created=1751644429,
model='llama3.2:3b-instruct-fp16',
object='chat.completion',
service_tier=None,
system_fingerprint='fp_ollama',
usage={'completion_tokens': 33, 'prompt_tokens': 36, 'total_tokens': 69, 'completion_tokens_details': None, 'prompt_tokens_details': None}
)
```
---------
Signed-off-by: Wen Zhou <wenzhou@redhat.com>
# What does this PR do?
* Removes a bunch of distros
* Removed distros were added into the "starter" distribution
* Doc for "starter" has been added
* Partially reverts https://github.com/meta-llama/llama-stack/pull/2482
since inference providers are disabled by default and can be turned on
manually via env variable.
* Disables safety in starter distro
Closes: https://github.com/meta-llama/llama-stack/issues/2502.
~Needs: https://github.com/meta-llama/llama-stack/pull/2482 for Ollama
to work properly in the CI.~
TODO:
- [ ] We can only update `install.sh` when we get a new release.
- [x] Update providers documentation
- [ ] Update notebooks to reference starter instead of ollama
Signed-off-by: Sébastien Han <seb@redhat.com>
