# What does this PR do?
In our OpenAI API verification tests, some providers were still calling
tools even when `tool_choice="none"` was passed in the chat completion
requests. Because they aren't all respecting `tool_choice` properly,
this adjusts our routing implementation to remove the `tools` and
`tool_choice` from the request if `tool_choice="none"` is passed in so
that it does not attempt to call any of those tools. Adjusting this in
the router fixes this across all providers.
This also cleans up the non-streaming together.ai responses for tools,
ensuring it returns `None` instead of an empty list when there are no
tool calls, to exactly match the OpenAI API responses in that case.
## Test Plan
I observed existing failures in our OpenAI API verification suite - see
https://github.com/bbrowning/llama-stack-tests/blob/main/openai-api-verification/2025-04-27.md#together-llama-stack
for the failing `test_chat_*_tool_choice_none` tests. All streaming and
non-streaming variants were failing across all 3 tested models.
After this change, all of those 6 failing tests are now passing with no
regression in the other tests.
I verified this via:
```
llama stack run --image-type venv \
tests/verifications/openai-api-verification-run.yaml
```
```
python -m pytest -s -v \
'tests/verifications/openai_api/test_chat_completion.py' \
--provider=together-llama-stack
```
The entire verification suite is not 100% on together.ai yet, but it's
getting closer.
This also increased the pass rate for fireworks.ai, and did not regress
the groq or openai tests at all.
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
switch sambanova inference adaptor to LiteLLM usage to simplify
integration and solve issues with current adaptor when streaming and
tool calling, models and templates updated
## Test Plan
pytest -s -v tests/integration/inference/test_text_inference.py
--stack-config=sambanova
--text-model=sambanova/Meta-Llama-3.3-70B-Instruct
pytest -s -v tests/integration/inference/test_vision_inference.py
--stack-config=sambanova
--vision-model=sambanova/Llama-3.2-11B-Vision-Instruct
# What does this PR do?
Checks for RAGDocument of type InterleavedContent
I noticed when stepping through the code that the supported types for
`RAGDocument` included `InterleavedContent` as a content type. This type
is not checked against before putting the `doc.content` is regex matched
against. This would cause a runtime error. This change adds an explicit
check for type.
The only other part that I'm unclear on is how to handle the
`ImageContent` type since this would always just return `<image>` which
seems like an undesired behavior. Should the `InterleavedContent` type
be removed from `RAGDocument` and replaced with `URI | str`?
## Test Plan
[//]: # (## Documentation)
---------
Signed-off-by: Kevin <kpostlet@redhat.com>
# 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?
Revert a change that by mistake forced efficiency_config on torchtune
provider
users.
```
fix: Don't require efficiency_config for torchtune
It was enforced by mistake when
0751a960a5 merged.
Other asserts made sense in that the code was written, potentially, to
always expect a non-None value. But not efficiency_config.
```
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>
# What does this PR do?
For the Issue :-
#[2010](https://github.com/meta-llama/llama-stack/issues/2010)
Currently, if we try to connect the Llama stack server to a remote
Milvus instance that has TLS enabled, the connection fails because TLS
support is not implemented in the Llama stack codebase. As a result,
users are unable to use secured Milvus deployments out of the box.
After adding this , the user will be able to connect to remote::Milvus
which is TLS enabled .
if TLS enabled :-
```
vector_io:
- provider_id: milvus
provider_type: remote::milvus
config:
uri: "http://<host>:<port>"
token: "<user>:<password>"
secure: True
server_pem_path: "path/to/server.pem"
```
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
I have already tested it by connecting to a Milvus instance which is TLS
enabled and i was able to start llama stack server .
# What does this PR do?
When converting OpenAI message content for the "system" and "assistant"
roles to Llama Stack inference APIs (used for some providers when
dealing with Llama models via OpenAI API requests to get proper prompt /
tool handling), we were not properly converting any non-string content.
I discovered this while running the new Responses AI verification suite
against the Fireworks provider, but instead of fixing it as part of some
ongoing work there split this out into a separate PR.
This fixes that, by using the `openai_content_to_content` helper we used
elsewhere to ensure content parts were mapped properly.
## Test Plan
I added a couple of new tests to `test_openai_compat` to reproduce this
issue and validate its fix. I ran those as below:
```
python -m pytest -s -v tests/unit/providers/utils/inference/test_openai_compat.py
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
The builtin implementation of code interpreter is not robust and has a
really weak sandboxing shell (the `bubblewrap` container). Given the
availability of better MCP code interpreter servers coming up, we should
use them instead of baking an implementation into the Stack and
expanding the vulnerability surface to the rest of the Stack.
This PR only does the removal. We will add examples with how to
integrate with MCPs in subsequent ones.
## Test Plan
Existing tests.
# 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>
Nullable param type is not supported, e.g. ['string', 'null'], since it
fails type validation.
Tests:
Run inference with
messages:
- content: You are a helpful assistant that can use tools to get
information.
role: system
- content: What's the temperature in San Francisco in celsius?
role: user
tools:
- function:
description: Get current temperature for a given location.
name: get_weather
parameters:
additionalProperties: false
properties:
location:
description: "City and country e.g. Bogot\xE1, Colombia"
type: string
unit:
description: "Unit of temperature, default to celsius"
type: [string, "null"] # <= nullable type
required:
- location
type: object
type: function
Co-authored-by: Eric Huang <erichuang@fb.com>
# What does this PR do?
Add support for the temperature to the responses API
## Test Plan
Manually tested simple case
unit tests added for simple case and tool calls
Signed-off-by: Derek Higgins <derekh@redhat.com>
# What does this PR do?
When the result of a ToolCall gets passed back into vLLM for the model
to handle the tool call result (as is often the case in agentic
tool-calling workflows), we forgot to handle the case where BuiltinTool
calls are not string values but instead instances of the BuiltinTool
enum. This fixes that, properly converting those enums to string values
before trying to serialize them into an OpenAI chat completion request
to vLLM.
PR #1931 fixed a bug where we weren't passing these tool calling results
back into vLLM, but as a side-effect it created this serialization bug
when using BuiltinTools.
Closes#2070
## Test Plan
I added a new unit test to the openai_compat unit tests to cover this
scenario, ensured the new test failed before this fix, and all the
existing tests there plus the new one passed with this fix.
```
python -m pytest -s -v tests/unit/providers/utils/inference/test_openai_compat.py
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
Add several new pre-commit hooks to improve code quality and security:
- no-commit-to-branch: prevent direct commits to protected branches like
`main`
- check-yaml: validate YAML files
- detect-private-key: prevent accidental commit of private keys
- requirements-txt-fixer: maintain consistent requirements.txt format
and sorting
- mixed-line-ending: enforce LF line endings to avoid mixed line endings
- check-executables-have-shebangs: ensure executable scripts have
shebangs
- check-json: validate JSON files
- check-shebang-scripts-are-executable: verify shebang scripts are
executable
- check-symlinks: validate symlinks and report broken ones
- check-toml: validate TOML files mainly for pyproject.toml
The respective fixes have been included.
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
When running a Llama Stack server and invoking the
`/v1/safety/run-shield` endpoint, the NVIDIA Guardrails endpoint in some
cases errors with a `422: Unprocessable Entity` due to malformed input.
For example, given an request body like:
```
{
"model": "test",
"messages": [
{ "role": "user", "content": "You are stupid." }
]
}
```
`convert_pydantic_to_json_value` converts the message to:
```
{ "role": "user", "content": "You are stupid.", "context": null }
```
Which causes NVIDIA Guardrails to return an error `HTTPError: 422 Client
Error: Unprocessable Entity for url:
http://nemo.test/v1/guardrail/checks`, because `context` shouldn't be
included in the body.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
I ran the Llama Stack server locally and manually verified that the
endpoint now succeeds.
```
message = {"role": "user", "content": "You are stupid."}
response = client.safety.run_shield(messages=[message], shield_id=shield_id, params={})
```
Server logs:
```
14:29:09.656 [START] /v1/safety/run-shield
INFO: 127.0.0.1:54616 - "POST /v1/safety/run-shield HTTP/1.1" 200 OK
14:29:09.918 [END] /v1/safety/run-shield [StatusCode.OK] (262.26ms
```
[//]: # (## Documentation)
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.com>
# What does this PR do?
Replaced `${env.OTEL_SERVICE_NAME:\u200B}` and similar variants with
properly formatted `${env.OTEL_SERVICE_NAME:}` across all YAML templates
and TelemetryConfig. This prevents silent parsing issues and ensures
consistent environment variable resolution.
Slipped in https://github.com/meta-llama/llama-stack/pull/2058
Signed-off-by: Sébastien Han <seb@redhat.com>
# What does this PR do?
The telemetry provider configs is the only one who leverages the env var
`SQLITE_DB_PATH` for pointing to persistent data in the respective
templates, whereas usually `SQLITE_STORE_DIR` is used.
This PR modifies the `sqlite_db_path` in various telemetry configuration
files to use the environment variable `SQLITE_STORE_DIR` instead of
`SQLITE_DB_PATH`. This change ensures that _only_ the SQLITE_STORE_DIR
needs to be set to point to a different persistence location for
providers.
All references to `SQLITE_DB_PATH` have been removed.
Another improvement could be to move `sqlite_db_path` to `db_path` in
the telemetry provider config, to align with the other provider
configurations. That could be done by another PR (if wanted).
# What does this PR do?
In our OpenAI API verification tests, ollama was still calling tools
even when `tool_choice="none"` was passed in its chat completion
requests. Because ollama isn't respecting `tool_choice` properly, this
adjusts our provider implementation to remove the `tools` from the
request if `tool_choice="none"` is passed in so that it does not attempt
to call any of those tools.
## Test Plan
I tested this with a couple of Llama models, using both our OpenAI
completions integration tests and our verification test suites.
### OpenAI Completions / Chat Completions integration tests
These all passed before, and still do.
```
INFERENCE_MODEL="llama3.2:3b-instruct-fp16" \
llama stack build --template ollama --image-type venv --run
```
```
LLAMA_STACK_CONFIG=http://localhost:8321 \
python -m pytest -v \
tests/integration/inference/test_openai_completion.py \
--text-model "llama3.2:3b-instruct-fp16"
```
### OpenAI API Verification test suite
test_chat_*_tool_choice_none OpenAI API verification tests pass now,
when they failed before.
See
https://github.com/bbrowning/llama-stack-tests/blob/main/openai-api-verification/2025-04-27.md#ollama-llama-stack
for an example of these failures from a recent nightly CI run.
```
INFERENCE_MODEL="llama3.3:70b-instruct-q3_K_M" \
llama stack build --template ollama --image-type venv --run
```
```
cat <<-EOF > tests/verifications/conf/ollama-llama-stack.yaml
base_url: http://localhost:8321/v1/openai/v1
api_key_var: OPENAI_API_KEY
models:
- llama3.3:70b-instruct-q3_K_M
model_display_names:
llama3.3:70b-instruct-q3_K_M: Llama-3.3-70B-Instruct
test_exclusions:
llama3.3:70b-instruct-q3_K_M:
- test_chat_non_streaming_image
- test_chat_streaming_image
- test_chat_multi_turn_multiple_images
EOF
```
```
python -m pytest -s -v \
'tests/verifications/openai_api/test_chat_completion.py' \
--provider=ollama-llama-stack
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
This provides an initial [OpenAI Responses
API](https://platform.openai.com/docs/api-reference/responses)
implementation. The API is not yet complete, and this is more a
proof-of-concept to show how we can store responses in our key-value
stores and use them to support the Responses API concepts like
`previous_response_id`.
## Test Plan
I've added a new
`tests/integration/openai_responses/test_openai_responses.py` as part of
a test-driven development for this new API. I'm only testing this
locally with the remote-vllm provider for now, but it should work with
any of our inference providers since the only API it requires out of the
inference provider is the `openai_chat_completion` endpoint.
```
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" \
python -m pytest -v \
tests/integration/openai_responses/test_openai_responses.py \
--text-model "meta-llama/Llama-3.2-3B-Instruct"
```
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
Implemetation of NeMO Datastore register, unregister API.
Open Issues:
- provider_id gets set to `localfs` in client.datasets.register() as it
is specified in routing_tables.py: DatasetsRoutingTable
see: #1860
Currently I have passed `"provider_id":"nvidia"` in metadata and have
parsed that in `DatasetsRoutingTable`
(Not the best approach, but just a quick workaround to make it work for
now.)
## Test Plan
- Unit test cases: `pytest
tests/unit/providers/nvidia/test_datastore.py`
```bash
========================================================== test session starts ===========================================================
platform linux -- Python 3.10.0, pytest-8.3.5, pluggy-1.5.0
rootdir: /home/ubuntu/llama-stack
configfile: pyproject.toml
plugins: anyio-4.9.0, asyncio-0.26.0, nbval-0.11.0, metadata-3.1.1, html-4.1.1, cov-6.1.0
asyncio: mode=strict, asyncio_default_fixture_loop_scope=None, asyncio_default_test_loop_scope=function
collected 2 items
tests/unit/providers/nvidia/test_datastore.py .. [100%]
============================================================ warnings summary ============================================================
====================================================== 2 passed, 1 warning in 0.84s ======================================================
```
cc: @dglogo, @mattf, @yanxi0830
# What does this PR do?
There are new changes in repo which needs to add some additional
functions to the inference which is fixed. Also need one additional
params to pass some extra arguments to watsonx.ai
[//]: # (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.*]
[//]: # (## Documentation)
---------
Co-authored-by: Sajikumar JS <sajikumar.js@ibm.com>
# What does this PR do?
This addresses 2 bugs I ran into when launching a fine-tuning job with
the NVIDIA Adapter:
1. Session handling in `_make_request` helper function returns an error.
```
INFO: 127.0.0.1:55831 - "POST /v1/post-training/supervised-fine-tune HTTP/1.1" 500 Internal Server Error
16:11:45.643 [END] /v1/post-training/supervised-fine-tune [StatusCode.OK] (270.44ms)
16:11:45.643 [ERROR] Error executing endpoint route='/v1/post-training/supervised-fine-tune' method='post'
Traceback (most recent call last):
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/distribution/server/server.py", line 201, in endpoint
return await maybe_await(value)
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/distribution/server/server.py", line 161, in maybe_await
return await value
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/providers/remote/post_training/nvidia/post_training.py", line 408, in supervised_fine_tune
response = await self._make_request(
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/providers/remote/post_training/nvidia/post_training.py", line 98, in _make_request
async with self.session.request(method, url, params=params, json=json, **kwargs) as response:
File "/Users/jgulabrai/Projects/forks/llama-stack/.venv/lib/python3.10/site-packages/aiohttp/client.py", line 1425, in __aenter__
self._resp: _RetType = await self._coro
File "/Users/jgulabrai/Projects/forks/llama-stack/.venv/lib/python3.10/site-packages/aiohttp/client.py", line 579, in _request
handle = tm.start()
File "/Users/jgulabrai/Projects/forks/llama-stack/.venv/lib/python3.10/site-packages/aiohttp/helpers.py", line 587, in start
return self._loop.call_at(when, self.__call__)
File "/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/asyncio/base_events.py", line 724, in call_at
self._check_closed()
File "/Library/Frameworks/Python.framework/Versions/3.10/lib/python3.10/asyncio/base_events.py", line 510, in _check_closed
raise RuntimeError('Event loop is closed')
RuntimeError: Event loop is closed
```
Note: This only occurred when initializing the client like so:
```
client = LlamaStackClient(
base_url="http://0.0.0.0:8321"
)
response = client.post_training.supervised_fine_tune(...) # Returns error
```
I didn't run into this issue when using the library client:
```
client = LlamaStackAsLibraryClient("nvidia")
client.initialize()
response = client.post_training.supervised_fine_tune(...) # Works fine
```
2. The `algorithm_config` param in `supervised_fine_tune` is parsed as a
`dict` when run from unit tests, but a Pydantic model when invoked using
the Llama Stack client. So, the call fails outside of unit tests:
```
INFO: 127.0.0.1:54024 - "POST /v1/post-training/supervised-fine-tune HTTP/1.1" 500 Internal Server Error
21:14:02.315 [END] /v1/post-training/supervised-fine-tune [StatusCode.OK] (71.18ms)
21:14:02.314 [ERROR] Error executing endpoint route='/v1/post-training/supervised-fine-tune' method='post'
Traceback (most recent call last):
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/distribution/server/server.py", line 205, in endpoint
return await maybe_await(value)
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/distribution/server/server.py", line 164, in maybe_await
return await value
File "/Users/jgulabrai/Projects/forks/llama-stack/llama_stack/providers/remote/post_training/nvidia/post_training.py", line 407, in supervised_fine_tune
"adapter_dim": algorithm_config.get("adapter_dim"),
File "/Users/jgulabrai/Projects/forks/llama-stack/.venv/lib/python3.10/site-packages/pydantic/main.py", line 891, in __getattr__
raise AttributeError(f'{type(self).__name__!r} object has no attribute {item!r}')
AttributeError: 'LoraFinetuningConfig' object has no attribute 'get'
```
The code assumes `algorithm_config` should be `dict`, so I just handle
both cases.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
1. I ran a local Llama Stack server with the necessary env vars:
```
lama stack run llama_stack/templates/nvidia/run.yaml --port 8321 --env ...
```
And invoked `supervised_fine_tune` to confirm neither of the errors
above occur.
```
client = LlamaStackClient(
base_url="http://0.0.0.0:8321"
)
response = client.post_training.supervised_fine_tune(...)
```
2. I confirmed the unit tests still pass: `./scripts/unit-tests.sh
tests/unit/providers/nvidia/test_supervised_fine_tuning.py`
[//]: # (## Documentation)
---------
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.com>
# What does this PR do?
IBM watsonx ai added as the inference [#1741
](https://github.com/meta-llama/llama-stack/issues/1741)
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
---------
Co-authored-by: Sajikumar JS <sajikumar.js@ibm.com>
# What does this PR do?
Adds custom model registration functionality to NVIDIAInferenceAdapter
which let's the inference happen on:
- post-training model
- non-llama models in API Catalogue(behind
https://integrate.api.nvidia.com and endpoints compatible with
AyncOpenAI)
## Example Usage:
```python
from llama_stack.apis.models import Model, ModelType
from llama_stack.distribution.library_client import LlamaStackAsLibraryClient
client = LlamaStackAsLibraryClient("nvidia")
_ = client.initialize()
client.models.register(
model_id=model_name,
model_type=ModelType.llm,
provider_id="nvidia"
)
response = client.inference.chat_completion(
model_id=model_name,
messages=[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"Write a limerick about the wonders of GPU computing."}],
)
```
## Test Plan
```bash
pytest tests/unit/providers/nvidia/test_supervised_fine_tuning.py
========================================================== test session starts ===========================================================
platform linux -- Python 3.10.0, pytest-8.3.5, pluggy-1.5.0
rootdir: /home/ubuntu/llama-stack
configfile: pyproject.toml
plugins: anyio-4.9.0
collected 6 items
tests/unit/providers/nvidia/test_supervised_fine_tuning.py ...... [100%]
============================================================ warnings summary ============================================================
../miniconda/envs/nvidia-1/lib/python3.10/site-packages/pydantic/fields.py:1076
/home/ubuntu/miniconda/envs/nvidia-1/lib/python3.10/site-packages/pydantic/fields.py:1076: PydanticDeprecatedSince20: Using extra keyword arguments on `Field` is deprecated and will be removed. Use `json_schema_extra` instead. (Extra keys: 'contentEncoding'). Deprecated in Pydantic V2.0 to be removed in V3.0. See Pydantic V2 Migration Guide at https://errors.pydantic.dev/2.11/migration/
warn(
-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
====================================================== 6 passed, 1 warning in 1.51s ======================================================
```
[//]: # (## Documentation)
Updated Readme.md
cc: @dglogo, @sumitb, @mattf
# What does this PR do?
This PR adds support for NVIDIA's NeMo Evaluator API to the Llama Stack
eval module. The integration enables users to evaluate models via the
Llama Stack interface.
## 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.*]
1. Added unit tests and successfully ran from root of project:
`./scripts/unit-tests.sh tests/unit/providers/nvidia/test_eval.py`
```
tests/unit/providers/nvidia/test_eval.py::TestNVIDIAEvalImpl::test_job_cancel PASSED
tests/unit/providers/nvidia/test_eval.py::TestNVIDIAEvalImpl::test_job_result PASSED
tests/unit/providers/nvidia/test_eval.py::TestNVIDIAEvalImpl::test_job_status PASSED
tests/unit/providers/nvidia/test_eval.py::TestNVIDIAEvalImpl::test_register_benchmark PASSED
tests/unit/providers/nvidia/test_eval.py::TestNVIDIAEvalImpl::test_run_eval PASSED
```
2. Verified I could build the Llama Stack image: `LLAMA_STACK_DIR=$(pwd)
llama stack build --template nvidia --image-type venv`
Documentation added to
`llama_stack/providers/remote/eval/nvidia/README.md`
---------
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.com>
Include the tool call details with the chat when doing Rag with Remote
vllm
Fixes: #1929
With this PR the tool call is included in the chat returned to vllm, the
model (meta-llama/Llama-3.1-8B-Instruct) the returns the answer as
expected.
Signed-off-by: Derek Higgins <derekh@redhat.com>
# What does this PR do?
This PR addresses the content dominance problem that frequently arises
with multiple models when executing queries with the RAG tool. When the
retrieved content is too large, it disproportionately influences the
generation process, causing the model to ignore the original question
and to provide meaningless comments on the retrieved information
instead.
This situation is especially common with agentic RAG, which is the
standard way of doing RAG in Llama Stack, since directly manipulating
the prompt combining the query with the retrieved content is not
possible.
This PR appends a grounding message to the results returned by the
knowledge search tool, reminding the model about the original query and
the purpose of the inference call. This makes the problem significantly
less likely to occur.
## Test Plan
Running the following script before the fix demonstrates the content
dominance problem where the model insists to comment on the retrieved
content and refuses to address the question.
Running the script after the fix results in getting the correct answer.
```
import os
import uuid
from llama_stack_client import Agent, AgentEventLogger, RAGDocument, LlamaStackClient
# the server endpoint
LLAMA_STACK_SERVER_URL = "http://localhost:8321"
# inference settings
MODEL_ID = ""meta-llama/Llama-3.1-8B-Instruct"
SYSTEM_PROMPT = "You are a helpful assistant. "
# RAG settings
VECTOR_DB_EMBEDDING_MODEL = "all-MiniLM-L6-v2"
VECTOR_DB_EMBEDDING_DIMENSION = 384
VECTOR_DB_CHUNK_SIZE = 512
# initialize the server connection
client = LlamaStackClient(base_url=os.environ.get("LLAMA_STACK_ENDPOINT", LLAMA_STACK_SERVER_URL))
# init the RAG retrieval parameters
vector_db_id = f"test_vector_db_{uuid.uuid4()}"
vector_providers = [
provider for provider in client.providers.list() if provider.api == "vector_io"
]
vector_provider_to_use = vector_providers[0]
# define and register the document collection to be used
client.vector_dbs.register(
vector_db_id=vector_db_id,
embedding_model=VECTOR_DB_EMBEDDING_MODEL,
embedding_dimension=VECTOR_DB_EMBEDDING_DIMENSION,
provider_id=vector_provider_to_use.provider_id,
)
# ingest the documents into the newly created document collection
urls = [
("https://www.openshift.guide/openshift-guide-screen.pdf", "application/pdf"),
]
documents = [
RAGDocument(
document_id=f"num-{i}",
content=url,
mime_type=url_type,
metadata={},
)
for i, (url, url_type) in enumerate(urls)
]
client.tool_runtime.rag_tool.insert(
documents=documents,
vector_db_id=vector_db_id,
chunk_size_in_tokens=VECTOR_DB_CHUNK_SIZE,
)
queries = [
"How to install OpenShift?",
]
# initializing the agent
agent = Agent(
client,
model=MODEL_ID,
instructions=SYSTEM_PROMPT,
# we make our agent aware of the RAG tool by including builtin::rag/knowledge_search in the list of tools
tools=[
dict(
name="builtin::rag/knowledge_search",
args={
"vector_db_ids": [vector_db_id], # list of IDs of document collections to consider during retrieval
},
)
],
)
for prompt in queries:
print(f"User> {prompt}")
# create a new turn with a new session ID for each prompt
response = agent.create_turn(
messages=[
{
"role": "user",
"content": prompt,
}
],
session_id=agent.create_session(f"rag-session_{uuid.uuid4()}")
)
# print the response, including tool calls output
for log in AgentEventLogger().log(response):
print(log.content, end='')
```
# 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>
# What does this PR do?
The together inference provider was throwing a stack trace every time it
shut down, as it was trying to call a non-existent `close` method on the
AsyncTogether client. While fixing that, I also adjusted its shutdown
logic to close the OpenAI client if we've created one of those, as that
client does have a `close` method.
In testing that, I also realized we were defaulting to treating all
requests as streaming requests instead of defaulting to non-streaming.
So, this flips that default to non-streaming to match how the other
providers work.
## Test Plan
I tested this by ensuring the together inference provider no longer
spits out a long stack trace when shutting it down and by running the
OpenAI API chat completion verification suite to ensure the change in
default streaming logic didn't mess anything else up.
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
We were passing a dict into the compat mixin for OpenAI Completions when
using Llama models with Fireworks, and that was breaking some strong
typing code that was added in openai_compat.py. We shouldn't have been
converting these params to a dict in that case anyway, so this adjusts
things to pass the params in as their actual original types when calling
the OpenAIChatCompletionToLlamaStackMixin.
## Test Plan
All of the fireworks provider verification tests were failing due to
some OpenAI compatibility cleanup in #1962. The changes in that PR were
good to make, and this just cleans up the fireworks provider code to
stop passing in untyped dicts to some of those `openai_compat.py`
methods since we have the original strongly-typed parameters we can pass
in.
```
llama stack run --image-type venv tests/verifications/openai-api-verification-run.yaml
```
```
python -m pytest -s -v tests/verifications/openai_api/test_chat_completion.py --provider=fireworks-llama-stack
```
Before this PR, all of the fireworks OpenAI verification tests were
failing. Now, most of them are passing.
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
NVIDIA Inference provider was using the ModelRegistryHelper to map input
model ids to provider model ids. this updates it to use the model_store.
## Test Plan
`LLAMA_STACK_CONFIG=http://localhost:8321 uv run pytest -v
tests/integration/inference/{test_embedding.py,test_text_inference.py,test_openai_completion.py}
--embedding-model nvidia/llama-3.2-nv-embedqa-1b-v2
--text-model=meta-llama/Llama-3.1-70B-Instruct`
Test plan:
python tests/verifications/generate_report.py --providers
fireworks,together,llama_meta_ref,openai
Co-authored-by: Eric Huang <erichuang@fb.com>
# What does this PR do?
Allow users to name an agent and use the name in telemetry instead of
relying on randomly generated agent_ids. This improves the developer
experience by making it easier to find specific agents in telemetry
logs.
Closes#1832
## Test Plan
- Added tests to verify the agent name is properly stored and retrieved
- Ran `uv run -- pytest -v
tests/integration/telemetry/test_telemetry.py::test_agent_name_filtering`
from the root of the project and made sure the tests pass
- Ran `uv run -- pytest -v
tests/integration/telemetry/test_telemetry.py::test_agent_query_spans`
to verify existing code without agent names still works correctly
## Use Example
```
agent = Agent(
llama_stack_client,
model=text_model_id,
name="CustomerSupportAgent", # New parameter
instructions="You are a helpful customer support assistant"
)
session_id = agent.create_session(f"test-session-{uuid4()}")
```
## Implementation Notes
- Agent names are optional string parameters with no additional
validation
- Names are not required to be unique - multiple agents can have the
same name
- The agent_id remains the unique identifier for an agent
---------
Co-authored-by: raghotham <raghotham@gmail.com>
# What does this PR do?
Add NVIDIA platform docs that serve as a starting point for Llama Stack
users and explains all supported microservices.
[//]: # (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.*]
[//]: # (## Documentation)
---------
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.com>
# What does this PR do?
This PR handles the case where a Customization Job's status is
`unknown`. Since we don't map `unknown` to a valid `JobStatus`, the
PostTraining provider throws an exception when fetching/listing a job.
[//]: # (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.*]
`./scripts/unit-tests.sh
tests/unit/providers/nvidia/test_supervised_fine_tuning.py` succeeds
[//]: # (## Documentation)
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.com>
# What does this PR do?
Adds `meta/llama-3.2-1b-instruct` to list of models that NeMo Customizer
can fine-tune. This is the model our example notebooks typically use for
fine-tuning.
[//]: # (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.*]
[//]: # (## Documentation)
Co-authored-by: Jash Gulabrai <jgulabrai@nvidia.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?
ollama's CLI supports running models via commands such as 'ollama run
llama3.2' this syntax does not work with the INFERENCE_MODEL llamastack
var as currently specifying a tag such as 'latest' is required
this commit will check to see if the 'latest' model is available and use
that model if a user passes a model name without a tag but the 'latest'
is available in ollama
## Test Plan
Behavior pre-code change
```bash
$ INFERENCE_MODEL=llama3.2 llama stack build --template ollama --image-type venv --run
...
INFO 2025-04-08 13:42:42,842 llama_stack.providers.remote.inference.ollama.ollama:80 inference: checking
connectivity to Ollama at `http://beanlab1.bss.redhat.com:11434`...
Traceback (most recent call last):
File "<frozen runpy>", line 198, in _run_module_as_main
File "<frozen runpy>", line 88, in _run_code
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/server/server.py", line 502, in <module>
main()
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/server/server.py", line 401, in main
impls = asyncio.run(construct_stack(config))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/lib64/python3.12/asyncio/runners.py", line 195, in run
return runner.run(main)
^^^^^^^^^^^^^^^^
File "/usr/lib64/python3.12/asyncio/runners.py", line 118, in run
return self._loop.run_until_complete(task)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/lib64/python3.12/asyncio/base_events.py", line 691, in run_until_complete
return future.result()
^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/stack.py", line 222, in construct_stack
await register_resources(run_config, impls)
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/stack.py", line 99, in register_resources
await method(**obj.model_dump())
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/providers/utils/telemetry/trace_protocol.py", line 102, in async_wrapper
result = await method(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/routers/routing_tables.py", line 294, in register_model
registered_model = await self.register_object(model)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/routers/routing_tables.py", line 228, in register_object
registered_obj = await register_object_with_provider(obj, p)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/distribution/routers/routing_tables.py", line 77, in register_object_with_provider
return await p.register_model(obj)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/providers/utils/telemetry/trace_protocol.py", line 102, in async_wrapper
result = await method(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/nathan/ai/llama-stack/repos/llama-stack/llama_stack/providers/remote/inference/ollama/ollama.py", line 315, in register_model
raise ValueError(
ValueError: Model 'llama3.2' is not available in Ollama. Available models: llama3.2:latest
++ error_handler 108
++ echo 'Error occurred in script at line: 108'
Error occurred in script at line: 108
++ exit 1
```
Behavior post-code change
```bash
$ INFERENCE_MODEL=llama3.2 llama stack build --template ollama --image-type venv --run
...
INFO 2025-04-08 13:58:17,365 llama_stack.providers.remote.inference.ollama.ollama:80 inference: checking
connectivity to Ollama at `http://beanlab1.bss.redhat.com:11434`...
WARNING 2025-04-08 13:58:18,190 llama_stack.providers.remote.inference.ollama.ollama:317 inference: Imprecise provider
resource id was used but 'latest' is available in Ollama - using 'llama3.2:latest'
INFO 2025-04-08 13:58:18,191 llama_stack.providers.remote.inference.ollama.ollama:308 inference: Pulling embedding
model `all-minilm:latest` if necessary...
INFO 2025-04-08 13:58:18,799 __main__:478 server: Listening on ['::', '0.0.0.0']:8321
INFO: Started server process [28378]
INFO: Waiting for application startup.
INFO 2025-04-08 13:58:18,803 __main__:148 server: Starting up
INFO: Application startup complete.
INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit)
...
```
## Documentation
Did not document this anywhere but happy to do so if there is an
appropriate place
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>
# What does this PR do?
Now a separate thread is started to execute training jobs. Training
requests now return job ID before the job completes. (Which fixes API
timeouts for any jobs that take longer than a minute.)
Note: the scheduler code is meant to be spun out in the future into a
common provider service that can be reused for different APIs and
providers. It is also expected to back the /jobs API proposed here:
https://github.com/meta-llama/llama-stack/discussions/1238
Hence its somewhat generalized form which is expected to simplify its
adoption elsewhere in the future.
Note: this patch doesn't attempt to implement missing APIs (e.g. cancel
or job removal). This work will belong to follow-up PRs.
[//]: # (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.*]
Added unit tests for the scheduler module. For the API coverage, did
manual testing and was able to run a training cycle on GPU. The initial
call returned job ID before the training completed, as (now) expected.
Artifacts are returned as expected.
```
JobArtifactsResponse(checkpoints=[{'identifier': 'meta-llama/Llama-3.2-3B-Instruct-sft-0', 'created_at': '2025-03-07T22:45:19.892714', 'epoch': 0, 'post_training_job_id': 'test-job2ee77104-2fd3-4a4e-84cf-f83f8b8f1f50', 'path': '/home/ec2-user/.llama/checkpoints/meta-llama/Llama-3.2-3B-Instruct-sft-0', 'training_metrics': None}], job_uuid='test-job2ee77104-2fd3-4a4e-84cf-f83f8b8f1f50')
```
The integration test is currently disabled for the provider. I will look
into how it can be enabled in a different PR / issue context.
[//]: # (## Documentation)
Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.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.
