# What does this PR do?
his PR allows users to customize the template used for chunks when
inserted into the context. Additionally, this enables metadata injection
into the context of an LLM for RAG. This makes a naive and crude
assumption that each chunk should include the metadata, this is
obviously redundant when multiple chunks are returned from the same
document. In order to remove any sort of duplication of chunks, we'd
have to make much more significant changes so this is a reasonable first
step that unblocks users requesting this enhancement in
https://github.com/meta-llama/llama-stack/issues/1767.
In the future, this can be extended to support citations.
List of Changes:
- `llama_stack/apis/tools/rag_tool.py`
- Added `chunk_template` field in `RAGQueryConfig`.
- Added `field_validator` to validate the `chunk_template` field in
`RAGQueryConfig`.
- Ensured the `chunk_template` field includes placeholders `{index}` and
`{chunk.content}`.
- Updated the `query` method to use the `chunk_template` for formatting
chunk text content.
- `llama_stack/providers/inline/tool_runtime/rag/memory.py`
- Modified the `insert` method to pass `doc.metadata` for chunk
creation.
- Enhanced the `query` method to format results using `chunk_template`
and exclude unnecessary metadata fields like `token_count`.
- `llama_stack/providers/utils/memory/vector_store.py`
- Updated `make_overlapped_chunks` to include metadata serialization and
token count for both content and metadata.
- Added error handling for metadata serialization issues.
- `pyproject.toml`
- Added `pydantic.field_validator` as a recognized `classmethod`
decorator in the linting configuration.
- `tests/integration/tool_runtime/test_rag_tool.py`
- Refactored test assertions to separate `assert_valid_chunk_response`
and `assert_valid_text_response`.
- Added integration tests to validate `chunk_template` functionality
with and without metadata inclusion.
- Included a test case to ensure `chunk_template` validation errors are
raised appropriately.
- `tests/unit/rag/test_vector_store.py`
- Added unit tests for `make_overlapped_chunks`, verifying chunk
creation with overlapping tokens and metadata integrity.
- Added tests to handle metadata serialization errors, ensuring proper
exception handling.
- `docs/_static/llama-stack-spec.html`
- Added a new `chunk_template` field of type `string` with a default
template for formatting retrieved chunks in RAGQueryConfig.
- Updated the `required` fields to include `chunk_template`.
- `docs/_static/llama-stack-spec.yaml`
- Introduced `chunk_template` field with a default value for
RAGQueryConfig.
- Updated the required configuration list to include `chunk_template`.
- `docs/source/building_applications/rag.md`
- Documented the `chunk_template` configuration, explaining how to
customize metadata formatting in RAG queries.
- Added examples demonstrating the usage of the `chunk_template` field
in RAG tool queries.
- Highlighted default values for `RAG` agent configurations.
# Resolves https://github.com/meta-llama/llama-stack/issues/1767
## Test Plan
Updated both `test_vector_store.py` and `test_rag_tool.py` and tested
end-to-end with a script.
I also tested the quickstart to enable this and specified this metadata:
```python
document = RAGDocument(
document_id="document_1",
content=source,
mime_type="text/html",
metadata={"author": "Paul Graham", "title": "How to do great work"},
)
```
Which produced the output below:

This highlights the usefulness of the additional metadata. Notice how
the metadata is redundant for different chunks of the same document. I
think we can update that in a subsequent PR.
# Documentation
I've added a brief comment about this in the documentation to outline
this to users and updated the API documentation.
---------
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
# What does this PR do?
In the Responses API, we convert incoming response requests to chat
completion requests. When streaming the resulting chunks of those chat
completion requests, inference providers that use OpenAI clients will
often return a `type=None` value in the tool call parts of the response.
This causes issues when we try to dump and load that response into our
pydantic model, because type cannot be None in the Responses API model
we're loading these into.
So, strip the "type" field, if present, off those chat completion tool
call results before dumping and loading them as our typed pydantic
models, which will apply our default value for that type field.
## Test Plan
This was found via manual testing of the Responses API with codex, where
I was getting errors in some tool call situations. I added a unit test
to simulate this scenario and verify the fix, as well as manual codex
testing to verify the fix.
Signed-off-by: Ben Browning <bbrownin@redhat.com>
note: the openai provider exposes the litellm specific model names to
the user. this change is compatible with that. the litellm names should
be deprecated.
# 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?
This is a combination of what was previously 3 separate PRs - #2069,
#2075, and #2083. It turns out all 3 of those are needed to land a
working function calling Responses implementation. The web search
builtin tool was already working, but this wires in support for custom
function calling.
I ended up combining all three into one PR because they all had lots of
merge conflicts, both with each other but also with #1806 that just
landed. And, because landing any of them individually would have only
left a partially working implementation merged.
The new things added here are:
* Storing of input items from previous responses and restoring of those
input items when adding previous responses to the conversation state
* Handling of multiple input item messages roles, not just "user"
messages.
* Support for custom tools passed into the Responses API to enable
function calling outside of just the builtin websearch tool.
Closes#2074Closes#2080
## Test Plan
### Unit Tests
Several new unit tests were added, and they all pass. Ran via:
```
python -m pytest -s -v tests/unit/providers/agents/meta_reference/test_openai_responses.py
```
### Responses API Verification Tests
I ran our verification run.yaml against multiple providers to ensure we
were getting a decent pass rate. Specifically, I ensured the new custom
tool verification test passed across multiple providers and that the
multi-turn examples passed across at least some of the providers (some
providers struggle with the multi-turn workflows still).
Running the stack setup for verification testing:
```
llama stack run --image-type venv tests/verifications/openai-api-verification-run.yaml
```
Together, passing 100% as an example:
```
pytest -s -v 'tests/verifications/openai_api/test_responses.py' --provider=together-llama-stack
```
## Documentation
We will need to start documenting the OpenAI APIs, but for now the
Responses stuff is still rapidly evolving so delaying that.
---------
Signed-off-by: Derek Higgins <derekh@redhat.com>
Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Derek Higgins <derekh@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
The ollama provider was using an older variant of the code to convert
incoming parameters from the OpenAI API completions and chat completion
endpoints into requests that get sent to the backend provider over its
own OpenAI client. This updates it to use the common
`prepare_openai_completion_params` method used elsewhere, which takes
care of removing stray `None` values even for nested structures.
Without this, some other parameters, even if they have values of `None`,
make their way to ollama and actually influence its inference output as
opposed to when those parameters are not sent at all.
## Test Plan
This passes tests/integration/inference/test_openai_completion.py and
fixes the issue found in #2098, which was tested via manual curl
requests crafted a particular way.
Closes#2098
Signed-off-by: Ben Browning <bbrownin@redhat.com>
```
$ INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct \
CHROMADB_URL=http://localhost:8000 \
llama stack build --image-type conda --image-name llama \
--providers vector_io=remote::chromadb,inference=remote::ollama \
--run
...
File ".../llama_stack/providers/remote/vector_io/chroma/chroma.py", line 31, in <module>
ChromaClientType = chromadb.AsyncHttpClient | chromadb.PersistentClient
TypeError: unsupported operand type(s) for |: 'function' and 'function'
```
issue: AsyncHttpClient and PersistentClient are functions that return
AsyncClientAPI and ClientAPI types, respectively. | cannot be used to
construct a type from functions.
previously the code was Union[AsyncHttpClient, PersistentClient], which
did not trigger an error
# What does this PR do?
Closes#2135
# What does this PR do?
This PR fixes the behavior of the `/tool-runtime/rag-tool/query`
endpoint when invoked with an empty `vector_db_ids` parameter.
As of now, it simply returns an empty result, which leads to a
misleading error message from the server and makes it difficult and
time-consuming to detect the problem with the input parameter.
The proposed fix is to return an indicative error message in this case.
## Test Plan
Running the following script:
```
agent = Agent(
client,
model=MODEL_ID,
instructions=SYSTEM_PROMPT,
tools=[
dict(
name="builtin::rag/knowledge_search",
args={
"vector_db_ids": [],
},
)
],
)
response = agent.create_turn(
messages=[
{
"role": "user",
"content": "How to install OpenShift?",
}
],
session_id=agent.create_session(f"rag-session")
)
```
results in the following error message in the non-patched version:
```
{"type": "function", "name": "knowledge_search", "parameters": {"query": "installing OpenShift"}}400: Invalid value: Tool call result (id: 494b8020-90bb-449b-aa76-10960d6b2cc2, name: knowledge_search) does not have any content
```
and in the following one in the patched version:
```
{"type": "function", "name": "knowledge_search", "parameters": {"query": "installing OpenShift"}}400: Invalid value: No vector DBs were provided to the RAG tool. Please provide at least one DB.
```
# What does this PR do?
Adds the API to query metrics from telemetry.
## Test Plan
llama stack run ~/.llama/distributions/fireworks/fireworks-run.yaml
---------
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# 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>