# What does this PR do?
- remove auto-download of ollama embedding models
- add embedding model metadata to dynamic listing w/ unit test
- add support and tests for allowed_models
- removed inference provider models.py files where dynamic listing is
enabled
- store embedding metadata in embedding_model_metadata field on
inference providers
- make model_entries optional on ModelRegistryHelper and
LiteLLMOpenAIMixin
- make OpenAIMixin a ModelRegistryHelper
- skip base64 embedding test for remote::ollama, always returns floats
- only use OpenAI client for ollama model listing
- remove unused build_model_entry function
- remove unused get_huggingface_repo function
## Test Plan
ci w/ new tests
# What does this PR do?
*Add dynamic authentication token forwarding support for vLLM provider*
This enables per-request authentication tokens for vLLM providers,
supporting use cases like RAG operations where different requests may
need different authentication tokens. The implementation follows the
same pattern as other providers like Together AI, Fireworks, and
Passthrough.
- Add LiteLLMOpenAIMixin that manages the vllm_api_token properly
Usage:
- Static: VLLM_API_TOKEN env var or config.api_token
- Dynamic: X-LlamaStack-Provider-Data header with vllm_api_token
All existing functionality is preserved while adding new dynamic
capabilities.
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
<!-- 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.* -->
```
curl -X POST "http://localhost:8000/v1/chat/completions" -H "Authorization: Bearer my-dynamic-token" \
-H "X-LlamaStack-Provider-Data: {\"vllm_api_token\": \"Bearer my-dynamic-token\", \"vllm_url\": \"http://dynamic-server:8000\"}" \
-H "Content-Type: application/json" \
-d '{"model": "llama-3.1-8b", "messages": [{"role": "user", "content": "Hello!"}]}'
```
---------
Signed-off-by: Akram Ben Aissi <akram.benaissi@gmail.com>
# 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
```
# 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>
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.
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.
# What does this PR do?
This commit significantly improves the environment variable substitution
functionality in Llama Stack configuration files:
* The version field in configuration files has been changed from string
to integer type for better type consistency across build and run
configurations.
* The environment variable substitution system for ${env.FOO:} was fixed
and properly returns an error
* The environment variable substitution system for ${env.FOO+} returns
None instead of an empty strings, it better matches type annotations in
config fields
* The system includes automatic type conversion for boolean, integer,
and float values.
* The error messages have been enhanced to provide clearer guidance when
environment variables are missing, including suggestions for using
default values or conditional syntax.
* Comprehensive documentation has been added to the configuration guide
explaining all supported syntax patterns, best practices, and runtime
override capabilities.
* Multiple provider configurations have been updated to use the new
conditional syntax for optional API keys, making the system more
flexible for different deployment scenarios. The telemetry configuration
has been improved to properly handle optional endpoints with appropriate
validation, ensuring that required endpoints are specified when their
corresponding sinks are enabled.
* There were many instances of ${env.NVIDIA_API_KEY:} that should have
caused the code to fail. However, due to a bug, the distro server was
still being started, and early validation wasn’t triggered. As a result,
failures were likely being handled downstream by the providers. I’ve
maintained similar behavior by using ${env.NVIDIA_API_KEY:+}, though I
believe this is incorrect for many configurations. I’ll leave it to each
provider to correct it as needed.
* Environment variable substitution now uses the same syntax as Bash
parameter expansion.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
* Given that our API packages use "import *" in `__init.py__` we don't
need to do `from llama_stack.apis.models.models` but simply from
llama_stack.apis.models. The decision to use `import *` is debatable and
should probably be revisited at one point.
* Remove unneeded Ruff F401 rule
* Consolidate Ruff F403 rule in the pyprojectfrom
llama_stack.apis.models.models
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
Our starter distro required Ollama to be running (and a large list of
models available in that Ollama) to successfully start. This adjusts
things so that Ollama does not have to be running to use the starter
template / distro.
To accomplish this, a few changes were needed:
* The Ollama provider is now configurable whether it raises an Exception
or just logs a warning when it cannot reach the Ollama server on
startup. The default is to raise an exception (same as previous
behavior), but in the starter template we adjust this to just log a
warning so that we can bring the stack up without needing a running
Ollama server.
* The starter template no longer specifies a default list of models for
Ollama, as any models specified there need to actually be pulled and
available in Ollama. Instead, it adds a new
`OLLAMA_INFERENCE_MODEL` environment variable where users can provide an
optional model to register with the Ollama provider on startup.
Additional models can also be registered via the typical
`models.register(...)` at runtime.
* The vLLM template was adjusted to also allow an optional
`VLLM_INFERENCE_MODEL` specified on startup, so that the behavior
between vLLM and Ollama was consistent here to make it easy to get up
and running quickly.
* The default vector store was changed from sqlite-vec to faiss.
sqlite-vec can enabled via setting the `ENABLE_SQLITE_VEC` environment
variable, like we do for chromadb and pgvector. This is due to
sqlite-vec not shipping proper arm64 binaries, like we previously fixed
in #1530 for the ollama distribution.
## Test Plan
With this change, the following scenarios now work with the starter
template that did not before:
* no Ollama running
* Ollama running but not all of the Llama models pulled locally
* Ollama running with a custom model registered on startup
* vLLM running with a custom model registered on startup
* running the starter template on linux/arm64, like when running
containers on Mac without rosetta emulation
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
- Implement OpenAI-compatible embeddings endpoint in vLLM provider
- Support both float and base64 encoding formats
- Add proper error handling and response formatting
<!-- If resolving an issue, uncomment and update the line below -->
Closes#2447
## 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: Varsha Prasad Narsing <varshaprasad96@gmail.com>
For code completion apps need "fill in the middle" capabilities.
Added option of `suffix` to `openai_completion` to enable this.
Updated ollama provider to showcase the same.
### Test Plan
```
pytest -sv --stack-config="inference=ollama" tests/integration/inference/test_openai_completion.py --text-model qwen2.5-coder:1.5b -k test_openai_completion_non_streaming_suffix
```
### OpenAI Sample script
```
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8321/v1/openai/v1")
response = client.completions.create(
model="qwen2.5-coder:1.5b",
prompt="The capital of ",
suffix="is Paris.",
max_tokens=10,
)
print(response.choices[0].text)
```
### Output
```
France is ____.
To answer this question, we
```
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
To add health status check for remote VLLM
<!-- 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.* -->
PR includes the unit test to test the added health check implementation
feature.
# What does this PR do?
Adds a new endpoint that is compatible with OpenAI for embeddings api.
`/openai/v1/embeddings`
Added providers for OpenAI, LiteLLM and SentenceTransformer.
## Test Plan
```
LLAMA_STACK_CONFIG=http://localhost:8321 pytest -sv tests/integration/inference/test_openai_embeddings.py --embedding-model all-MiniLM-L6-v2,text-embedding-3-small,gemini/text-embedding-004
```
# What does this PR do?
Handles the case where the vllm config `tls_verify` is set to `false` or
`true`.
Closes: https://github.com/meta-llama/llama-stack/issues/2283
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
The `tls_verify` can now receive a path to a certificate file if the
endpoint requires it.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
This fixes an issue in how we used the tool_call_buf from streaming tool
calls in the remote-vllm provider where it would end up concatenating
parameters from multiple different tool call results instead of
aggregating the results from each tool call separately.
It also fixes an issue found while digging into that where we were
accidentally mixing the json string form of tool call parameters with
the string representation of the python form, which mean we'd end up
with single quotes in what should be double-quoted json strings.
Closes#1120
## Test Plan
The following tests are now passing 100% for the remote-vllm provider,
where some of the test_text_inference were failing before this change:
```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_text_inference.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_vision_inference.py --vision-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"
```
All but one of the agent tests are passing (including the multi-tool
one). See the PR at https://github.com/vllm-project/vllm/pull/17917 and
a gist at
https://gist.github.com/bbrowning/4734240ce96b4264340caa9584e47c9e for
changes needed there, which will have to get made upstream in vLLM.
Agent tests:
```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/agents/test_agents.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic"
````
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
Closes#2113.
Closes#1783.
Fixes a bug in handling the end of tool execution request stream where
no `finish_reason` is provided by the model.
## Test Plan
1. Ran existing unit tests
2. Added a dedicated test verifying correct behavior in this edge case
3. Ran the code snapshot from #2113
[//]: # (## Documentation)
# What does this PR do?
Closes#2111.
Fixes an error causing Llama Stack to just return `<tool_call>` and
complete the turn without actually executing the tool. See the issue
description for more detail.
## Test Plan
1) Ran existing unit tests
2) Added a dedicated test verifying correct behavior in this edge case
3) Ran the code snapshot from #2111
# What does this PR do?
Mainly tried to cover the entire llama_stack/apis directory, we only
have one left. Some excludes were just noop.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
The goal of this PR is code base modernization.
Schema reflection code needed a minor adjustment to handle UnionTypes
and collections.abc.AsyncIterator. (Both are preferred for latest Python
releases.)
Note to reviewers: almost all changes here are automatically generated
by pyupgrade. Some additional unused imports were cleaned up. The only
change worth of note can be found under `docs/openapi_generator` and
`llama_stack/strong_typing/schema.py` where reflection code was updated
to deal with "newer" types.
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
Closes#1968.
The asynchronous client in `VLLMInferenceAdapter` is now initialized
directly before first use and not in `VLLMInferenceAdapter.initialize`.
This prevents issues arising due to accessing an expired event loop from
a completed `asyncio.run`.
## Test Plan
Ran unit tests, including `test_remote_vllm.py`.
Ran the code snippet mentioned in #1968.
---------
Co-authored-by: Sébastien Han <seb@redhat.com>
Fixes: #1955
Since 0.2.0, the vLLM gets an empty list (vs ``None``in 0.1.9 and
before) when there are no tools configured which causes the issue
described in #1955 p. This patch avoids sending the 'tools' param to the
vLLM altogether instead of an empty list.
It also adds a small unit test to avoid regressions.
The OpenAI
[specification](https://platform.openai.com/docs/api-reference/chat/create)
does not explicitly state that the list cannot be empty but I found this
out through experimentation and it might depend on the actual remote
vllm. In any case, as this parameter is Optional, is best to skip it
altogether if there's no tools configured.
Signed-off-by: Daniel Alvarez <dalvarez@redhat.com>
# What does this PR do?
TLDR: Changes needed to get 100% passing tests for OpenAI API
verification tests when run against Llama Stack with the `together`,
`fireworks`, and `openai` providers. And `groq` is better than before,
at 88% passing.
This cleans up the OpenAI API support for image message types
(specifically `image_url` types) and handling of the `response_format`
chat completion parameter. Both of these required a few more Pydantic
model definitions in our Inference API, just to move from the
not-quite-right stubs I had in place to something fleshed out to match
the actual OpenAI API specs.
As part of testing this, I also found and fixed a bug in the litellm
implementation of openai_completion and openai_chat_completion, so the
providers based on those should actually be working now.
The method `prepare_openai_completion_params` in
`llama_stack/providers/utils/inference/openai_compat.py` was improved to
actually recursively clean up input parameters, including handling of
lists, dicts, and dumping of Pydantic models to dicts. These changes
were required to get to 100% passing tests on the OpenAI API
verification against the `openai` provider.
With the above, the together.ai provider was passing as well as it is
without Llama Stack. But, since we have Llama Stack in the middle, I
took the opportunity to clean up the together.ai provider so that it now
also passes the OpenAI API spec tests we have at 100%. That means
together.ai is now passing our verification test better when using an
OpenAI client talking to Llama Stack than it is when hitting together.ai
directly, without Llama Stack in the middle.
And, another round of work for Fireworks to improve translation of
incoming OpenAI chat completion requests to Llama Stack chat completion
requests gets the fireworks provider passing at 100%. The server-side
fireworks.ai tool calling support with OpenAI chat completions and Llama
4 models isn't great yet, but by pointing the OpenAI clients at Llama
Stack's API we can clean things up and get everything working as
expected for Llama 4 models.
## Test Plan
### OpenAI API Verification Tests
I ran the OpenAI API verification tests as below and 100% of the tests
passed.
First, start a Llama Stack server that runs the `openai` provider with
the `gpt-4o` and `gpt-4o-mini` models deployed. There's not a template
setup to do this out of the box, so I added a
`tests/verifications/openai-api-verification-run.yaml` to do this.
First, ensure you have the necessary API key environment variables set:
```
export TOGETHER_API_KEY="..."
export FIREWORKS_API_KEY="..."
export OPENAI_API_KEY="..."
```
Then, run a Llama Stack server that serves up all these providers:
```
llama stack run \
--image-type venv \
tests/verifications/openai-api-verification-run.yaml
```
Finally, generate a new verification report against all these providers,
both with and without the Llama Stack server in the middle.
```
python tests/verifications/generate_report.py \
--run-tests \
--provider \
together \
fireworks \
groq \
openai \
together-llama-stack \
fireworks-llama-stack \
groq-llama-stack \
openai-llama-stack
```
You'll see that most of the configurations with Llama Stack in the
middle now pass at 100%, even though some of them do not pass at 100%
when hitting the backend provider's API directly with an OpenAI client.
### OpenAI Completion Integration Tests with vLLM:
I also ran the smaller `test_openai_completion.py` test suite (that's
not yet merged with the verification tests) on multiple of the
providers, since I had to adjust the method signature of
openai_chat_completion a bit and thus had to touch lots of these
providers to match. Here's the tests I ran there, all passing:
```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" llama stack build --template remote-vllm --image-type venv --run
```
in another terminal
```
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "meta-llama/Llama-3.2-3B-Instruct"
```
### OpenAI Completion Integration Tests with ollama
```
INFERENCE_MODEL="llama3.2:3b-instruct-q8_0" llama stack build --template ollama --image-type venv --run
```
in another terminal
```
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="llama3.2:3b-instruct-q8_0" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "llama3.2:3b-instruct-q8_0"
```
### OpenAI Completion Integration Tests with together.ai
```
INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct-Turbo" llama stack build --template together --image-type venv --run
```
in another terminal
```
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct-Turbo" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "meta-llama/Llama-3.2-3B-Instruct-Turbo"
```
### OpenAI Completion Integration Tests with fireworks.ai
```
INFERENCE_MODEL="meta-llama/Llama-3.1-8B-Instruct" llama stack build --template fireworks --image-type venv --run
```
in another terminal
```
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="meta-llama/Llama-3.1-8B-Instruct" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "meta-llama/Llama-3.1-8B-Instruct"
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
This PR adds two methods to the Inference API:
- `batch_completion`
- `batch_chat_completion`
The motivation is for evaluations targeting a local inference engine
(like meta-reference or vllm) where batch APIs provide for a substantial
amount of acceleration.
Why did I not add this to `Api.batch_inference` though? That just
resulted in a _lot_ more book-keeping given the structure of Llama
Stack. Had I done that, I would have needed to create a notion of a
"batch model" resource, setup routing based on that, etc. This does not
sound ideal.
So what's the future of the batch inference API? I am not sure. Maybe we
can keep it for true _asynchronous_ execution. So you can submit
requests, and it can return a Job instance, etc.
## Test Plan
Run meta-reference-gpu using:
```bash
export INFERENCE_MODEL=meta-llama/Llama-4-Scout-17B-16E-Instruct
export INFERENCE_CHECKPOINT_DIR=../checkpoints/Llama-4-Scout-17B-16E-Instruct-20250331210000
export MODEL_PARALLEL_SIZE=4
export MAX_BATCH_SIZE=32
export MAX_SEQ_LEN=6144
LLAMA_MODELS_DEBUG=1 llama stack run meta-reference-gpu
```
Then run the batch inference test case.
# What does this PR do?
This stubs in some OpenAI server-side compatibility with three new
endpoints:
/v1/openai/v1/models
/v1/openai/v1/completions
/v1/openai/v1/chat/completions
This gives common inference apps using OpenAI clients the ability to
talk to Llama Stack using an endpoint like
http://localhost:8321/v1/openai/v1 .
The two "v1" instances in there isn't awesome, but the thinking is that
Llama Stack's API is v1 and then our OpenAI compatibility layer is
compatible with OpenAI V1. And, some OpenAI clients implicitly assume
the URL ends with "v1", so this gives maximum compatibility.
The openai models endpoint is implemented in the routing layer, and just
returns all the models Llama Stack knows about.
The following providers should be working with the new OpenAI
completions and chat/completions API:
* remote::anthropic (untested)
* remote::cerebras-openai-compat (untested)
* remote::fireworks (tested)
* remote::fireworks-openai-compat (untested)
* remote::gemini (untested)
* remote::groq-openai-compat (untested)
* remote::nvidia (tested)
* remote::ollama (tested)
* remote::openai (untested)
* remote::passthrough (untested)
* remote::sambanova-openai-compat (untested)
* remote::together (tested)
* remote::together-openai-compat (untested)
* remote::vllm (tested)
The goal to support this for every inference provider - proxying
directly to the provider's OpenAI endpoint for OpenAI-compatible
providers. For providers that don't have an OpenAI-compatible API, we'll
add a mixin to translate incoming OpenAI requests to Llama Stack
inference requests and translate the Llama Stack inference responses to
OpenAI responses.
This is related to #1817 but is a bit larger in scope than just chat
completions, as I have real use-cases that need the older completions
API as well.
## Test Plan
### vLLM
```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" llama stack build --template remote-vllm --image-type venv --run
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "meta-llama/Llama-3.2-3B-Instruct"
```
### ollama
```
INFERENCE_MODEL="llama3.2:3b-instruct-q8_0" llama stack build --template ollama --image-type venv --run
LLAMA_STACK_CONFIG=http://localhost:8321 INFERENCE_MODEL="llama3.2:3b-instruct-q8_0" python -m pytest -v tests/integration/inference/test_openai_completion.py --text-model "llama3.2:3b-instruct-q8_0"
```
## Documentation
Run a Llama Stack distribution that uses one of the providers mentioned
in the list above. Then, use your favorite OpenAI client to send
completion or chat completion requests with the base_url set to
http://localhost:8321/v1/openai/v1 . Replace "localhost:8321" with the
host and port of your Llama Stack server, if different.
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
- **chore: mypy for strong_typing**
- **chore: mypy for remote::vllm**
- **chore: mypy for remote::ollama**
- **chore: mypy for providers.datatype**
---------
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
### What does this PR do?
Currently, `ToolCall.arguments` is a `Dict[str, RecursiveType]`.
However, on the client SDK side -- the `RecursiveType` gets deserialized
into a number ( both int and float get collapsed ) and hence when params
are `int` they get converted to float which might break client side
tools that might be doing type checking.
Closes: https://github.com/meta-llama/llama-stack/issues/1683
### Test Plan
Stainless changes --
https://github.com/meta-llama/llama-stack-client-python/pull/204
```
pytest -s -v --stack-config=fireworks tests/integration/agents/test_agents.py --text-model meta-llama/Llama-3.1-8B-Instruct
```
# What does this PR do?
Add the option to not verify SSL certificates for the remote-vllm
provider. This allows llama stack server to talk to remote LLMs which
have self-signed certificates
Partially addresses #1545
# What does this PR do?
This switches from an OpenAI client to the AsyncOpenAI client in the
remote vllm provider. The main benefit of this is that instead of each
client call being a blocking operation that was blocking our server
event loop, the client calls are now async operations that do not block
the event loop.
The actual fix is quite simple and straightforward. Creating a reliable
reproducer of this with a unit test that verifies we were blocking the
event loop before and are not blocking it any longer was a bit harder.
Some other inference providers have this same issue, so we may want to
make that simple delayed http server a bit more generic and pull it into
a common place as other inference providers get fixed.
(Closes#1457)
## Test Plan
I verified the unit tests and test_text_inference tests pass with this
change like below:
```
python -m pytest -v tests/unit
```
```
VLLM_URL="http://localhost:8000/v1" \
INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" \
LLAMA_STACK_CONFIG=remote-vllm \
python -m pytest -v -s \
tests/integration/inference/test_text_inference.py \
--text-model "meta-llama/Llama-3.2-3B-Instruct"
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
The commit addresses the Ruff warning B008 by refactoring the code to
avoid calling SamplingParams() directly in function argument defaults.
Instead, it either uses Field(default_factory=SamplingParams) for
Pydantic models or sets the default to None and instantiates
SamplingParams inside the function body when the argument is None.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
This gracefully handles the case where the vLLM server responded to a
completion request with no choices, which can happen in certain vLLM
error situations. Previously, we'd error out with a stack trace about a
list index out of range. Now, we just log a warning to the user and move
past any chunks with an empty choices list.
A specific example of the type of stack trace this fixes:
```
File "/app/llama-stack-source/llama_stack/providers/remote/inference/vllm/vllm.py", line 170, in _process_vllm_chat_completion_stream_response
choice = chunk.choices[0]
~~~~~~~~~~~~~^^^
IndexError: list index out of range
```
Now, instead of erroring out with that stack trace, we log a warning
that vLLM failed to generate any completions and alert the user to check
the vLLM server logs for details.
This is related to #1277 and addresses the stack trace shown in that
issue, although does not in and of itself change the functional behavior
of vLLM tool calling.
## Test Plan
As part of this fix, I added new unit tests to trigger this same error
and verify it no longer happens. That is
`test_process_vllm_chat_completion_stream_response_no_choices` in the
new `tests/unit/providers/inference/test_remote_vllm.py`. I also added a
couple of more tests to trigger and verify the last couple of remote
vllm provider bug fixes - specifically a test for #1236 (builtin tool
calling) and #1325 (vLLM <= v0.6.3).
This required fixing the signature of
`_process_vllm_chat_completion_stream_response` to accept the actual
type of chunks it was getting passed - specifically changing from our
openai_compat `OpenAICompatCompletionResponse` to
`openai.types.chat.chat_completion_chunk.ChatCompletionChunk`. It was
not actually getting passed `OpenAICompatCompletionResponse` objects
before, and was using attributes that didn't exist on those objects. So,
the signature now matches the type of object it's actually passed.
Run these new unit tests like this:
```
pytest tests/unit/providers/inference/test_remote_vllm.py
```
Additionally, I ensured the existing `test_text_inference.py` tests
passed via:
```
VLLM_URL="http://localhost:8000/v1" \
INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" \
LLAMA_STACK_CONFIG=remote-vllm \
python -m pytest -v tests/integration/inference/test_text_inference.py \
--inference-model "meta-llama/Llama-3.2-3B-Instruct" \
--vision-inference-model ""
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
This is to be consistent with OpenAI API and support vLLM <= v0.6.3
References:
*
https://platform.openai.com/docs/api-reference/chat/create#chat-create-tool_choice
* https://github.com/vllm-project/vllm/pull/10000
This fixes the error when running older versions of vLLM:
```
00:50:19.834 [START] /v1/inference/chat-completion
INFO 2025-02-28 00:50:20,203 httpx:1025: HTTP Request: POST https://api-xeai-granite-3-1-8b-instruct.apps.int.stc.ai.preprod.us-east-1.aws.paas.redhat.com/v1/chat/completions "HTTP/1.1 400 Bad Request"
Traceback (most recent call last):
File "/usr/local/lib/python3.10/site-packages/llama_stack/distribution/server/server.py", line 235, in endpoint
return await maybe_await(value)
File "/usr/local/lib/python3.10/site-packages/llama_stack/distribution/server/server.py", line 201, in maybe_await
return await value
File "/usr/local/lib/python3.10/site-packages/llama_stack/providers/utils/telemetry/trace_protocol.py", line 89, in async_wrapper
result = await method(self, *args, **kwargs)
File "/usr/local/lib/python3.10/site-packages/llama_stack/distribution/routers/routers.py", line 193, in chat_completion
return await provider.chat_completion(**params)
File "/usr/local/lib/python3.10/site-packages/llama_stack/providers/utils/telemetry/trace_protocol.py", line 89, in async_wrapper
result = await method(self, *args, **kwargs)
File "/usr/local/lib/python3.10/site-packages/llama_stack/providers/remote/inference/vllm/vllm.py", line 286, in chat_completion
return await self._nonstream_chat_completion(request, self.client)
File "/usr/local/lib/python3.10/site-packages/llama_stack/providers/remote/inference/vllm/vllm.py", line 292, in _nonstream_chat_completion
r = client.chat.completions.create(**params)
File "/usr/local/lib/python3.10/site-packages/openai/_utils/_utils.py", line 279, in wrapper
return func(*args, **kwargs)
File "/usr/local/lib/python3.10/site-packages/openai/resources/chat/completions/completions.py", line 879, in create
return self._post(
File "/usr/local/lib/python3.10/site-packages/openai/_base_client.py", line 1290, in post
return cast(ResponseT, self.request(cast_to, opts, stream=stream, stream_cls=stream_cls))
File "/usr/local/lib/python3.10/site-packages/openai/_base_client.py", line 967, in request
return self._request(
File "/usr/local/lib/python3.10/site-packages/openai/_base_client.py", line 1071, in _request
raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - {'object': 'error', 'message': "[{'type': 'value_error', 'loc': ('body',), 'msg': 'Value error, When using `tool_choice`, `tools` must be set.', 'input': {'messages': [{'role': 'user', 'content': [{'type': 'text', 'text': 'What model are you?'}]}], 'model': 'granite-3-1-8b-instruct', 'max_tokens': 4096, 'stream': False, 'temperature': 0.0, 'tools': None, 'tool_choice': 'auto'}, 'ctx': {'error': ValueError('When using `tool_choice`, `tools` must be set.')}}]", 'type': 'BadRequestError', 'param': None, 'code': 400}
INFO: 2600:1700:9d20:ac0::49:59736 - "POST /v1/inference/chat-completion HTTP/1.1" 500 Internal Server Error
00:50:20.266 [END] /v1/inference/chat-completion [StatusCode.OK] (431.99ms)
```
## Test Plan
All existing tests pass.
---------
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
# What does this PR do?
This PR makes a couple of changes required to get the test
`tests/client-sdk/agents/test_agents.py::test_builtin_tool_web_search`
passing on the remote-vllm provider.
First, we adjust agent_instance to also pass in the description and
parameters of builtin tools. We need these parameters so we can pass the
tool's expected parameters into vLLM. The meta-reference implementations
may not have needed these for builtin tools, as they are able to take
advantage of the Llama-model specific support for certain builtin tools.
However, with vLLM, our server-side chat templates for tool calling
treat all tools the same and don't separate out Llama builtin vs custom
tools. So, we need to pass the full set of parameter definitions and
list of required parameters for builtin tools as well.
Next, we adjust the vllm streaming chat completion code to fix up some
edge cases where it was returning an extra ChatCompletionResponseEvent
with an empty ToolCall with empty string call_id, tool_name, and
arguments properties. This is a bug discovered after the above fix,
where after a successful tool invocation we were sending extra chunks
back to the client with these empty ToolCalls.
## Test Plan
With these changes, the following test that previously failed now
passes:
```
VLLM_URL="http://localhost:8000/v1" \
INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" \
LLAMA_STACK_CONFIG=remote-vllm \
python -m pytest -v \
tests/client-sdk/agents/test_agents.py::test_builtin_tool_web_search \
--inference-model "meta-llama/Llama-3.2-3B-Instruct"
```
Additionally, I ran the remote-vllm client-sdk and provider inference
tests as below to ensure they all still passed with this change:
```
VLLM_URL="http://localhost:8000/v1" \
INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" \
LLAMA_STACK_CONFIG=remote-vllm \
python -m pytest -v \
tests/client-sdk/inference/test_text_inference.py \
--inference-model "meta-llama/Llama-3.2-3B-Instruct"
```
```
VLLM_URL="http://localhost:8000/v1" \
python -m pytest -s -v \
llama_stack/providers/tests/inference/test_text_inference.py \
--providers "inference=vllm_remote"
```
[//]: # (## Documentation)
Signed-off-by: Ben Browning <bbrownin@redhat.com>
See Issue #922
The change is slightly backwards incompatible but no callsite (in our
client codebases or stack-apps) every passes a depth-2
`List[List[InterleavedContentItem]]` (which is now disallowed.)
## Test Plan
```bash
$ cd llama_stack/providers/tests/inference
$ pytest -s -v -k fireworks test_embeddings.py \
--inference-model nomic-ai/nomic-embed-text-v1.5 --env EMBEDDING_DIMENSION=784
$ pytest -s -v -k together test_embeddings.py \
--inference-model togethercomputer/m2-bert-80M-8k-retrieval --env EMBEDDING_DIMENSION=784
$ pytest -s -v -k ollama test_embeddings.py \
--inference-model all-minilm:latest --env EMBEDDING_DIMENSION=784
```
Also ran `tests/client-sdk/inference/test_embeddings.py`
# What does this PR do?
The `tool_name` attribute of `ToolDefinition` instances can either be a
str or a BuiltinTool enum type. This fixes the remote vLLM provider to
use the value of those BuiltinTool enums when serializing to JSON
instead of attempting to serialize the actual enum to JSON.
Reference of how this is handled in some other areas, since I followed
that same pattern for the remote vLLM provider here:
- [remote nvidia
provider](https://github.com/meta-llama/llama-stack/blob/v0.1.3/llama_stack/providers/remote/inference/nvidia/openai_utils.py#L137-L140)
- [meta reference
provider](https://github.com/meta-llama/llama-stack/blob/v0.1.3/llama_stack/providers/inline/agents/meta_reference/agent_instance.py#L635-L636)
There is opportunity to potentially reconcile the remove nvidia and
remote vllm bits where they are both translating Llama Stack Inference
APIs to OpenAI client requests, but that's a can of worms I didn't want
to open for this bug fix.
This explicitly fixes this error when using the remote vLLM provider and
the agent tests:
```
TypeError: Object of type BuiltinTool is not JSON serializable
```
So, this is related to #1144 and addresses the immediate issue raised
there. With this fix,
`tests/client-sdk/agents/test_agents.py::test_builtin_tool_web_search`
now gets past the JSON serialization error when using the remote vLLM
provider and actually attempts to call the web search tool. I don't have
any API keys setup for the actual web search providers yet, so I cannot
verify everything works after that point.
## Test Plan
I ran the `test_builtin_tool_web_search` locally with the remote vLLM
provider like:
```
VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="meta-llama/Llama-3.2-3B-Instruct" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/client-sdk/agents/test_agents.py::test_builtin_tool_web_search --inference-model "meta-llama/Llama-3.2-3B-Instruct"
```
Before my change, that reproduced the `TypeError: Object of type
BuiltinTool is not JSON serializable` error. After my change, that error
is gone and the test actually attempts the web search. That failed for
me locally, due to lack of API key, but it gets past the JSON
serialization error.
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
This fixes the following issue on the server side when the tool call
response contains empty args. This happens when running
`examples.agents.e2e_loop_with_client_tools` but `get_ticker_data`
returns `[]`:
```
Traceback (most recent call last):
File "/home/yutang/repos/llama-stack/llama_stack/distribution/server/server.py", line 208, in sse_generator
async for item in event_gen:
File "/home/yutang/repos/llama-stack/llama_stack/providers/inline/agents/meta_reference/agents.py", line 169, in _create_agent_turn_streaming
async for event in agent.create_and_execute_turn(request):
File "/home/yutang/repos/llama-stack/llama_stack/providers/inline/agents/meta_reference/agent_instance.py", line 189, in create_and_execute_turn
async for chunk in self.run(
File "/home/yutang/repos/llama-stack/llama_stack/providers/inline/agents/meta_reference/agent_instance.py", line 258, in run
async for res in self._run(
File "/home/yutang/repos/llama-stack/llama_stack/providers/inline/agents/meta_reference/agent_instance.py", line 499, in _run
async for chunk in await self.inference_api.chat_completion(
File "/home/yutang/repos/llama-stack/llama_stack/distribution/routers/routers.py", line 182, in <genexpr>
return (chunk async for chunk in await provider.chat_completion(**params))
File "/home/yutang/repos/llama-stack/llama_stack/providers/remote/inference/vllm/vllm.py", line 296, in _stream_chat_completion
async for chunk in res:
File "/home/yutang/repos/llama-stack/llama_stack/providers/remote/inference/vllm/vllm.py", line 162, in _process_vllm_chat_completion_stream_response
arguments=json.loads(tool_call_buf.arguments),
File "/home/yutang/.conda/envs/distribution-myenv/lib/python3.10/json/__init__.py", line 346, in loads
return _default_decoder.decode(s)
File "/home/yutang/.conda/envs/distribution-myenv/lib/python3.10/json/decoder.py", line 337, in decode
obj, end = self.raw_decode(s, idx=_w(s, 0).end())
File "/home/yutang/.conda/envs/distribution-myenv/lib/python3.10/json/decoder.py", line 355, in raw_decode
raise JSONDecodeError("Expecting value", s, err.value) from None
json.decoder.JSONDecodeError: Expecting value: line 1 column 1 (char 0)
```
## Test Plan
All existing tests in
`tests/client-sdk/inference/test_text_inference.py` passed.
[//]: # (## Documentation)
---------
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
