# 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?
As described in #3134 a langchain example works against openai's
responses impl, but not against llama stack's. This turned out to be due
to the order of the inputs. The langchain example has the two function
call outputs first, followed by each call result in turn. This seems to
be valid as it is accepted by openai's impl. However in llama stack,
these inputs are converted to chat completion inputs and the resulting
order for that api is not accpeted by openai.
This PR fixes the issue by ensuring that the converted chat completions
inputs are in the expected order.
Closes#3134
## Test Plan
Added unit and integration tests. Verified this fixes original issue as
reported.
---------
Signed-off-by: Gordon Sim <gsim@redhat.com>
# What does this PR do?
Context: https://github.com/meta-llama/llama-stack/issues/2937
The API design is inspired by existing offerings, but not exactly the
same:
* `top_n` as the parameter to control number of results, instead of
`top_k`, since `n` is conventional to control number
* `truncation` bool instead of `max_token_per_doc`, since we should just
handle the truncation automatically depending on model capability,
instead of user setting the context length manually.
* `data` field in the response, to be consistent with other OpenAI APIs
(though they don't have a rerank API). Also, it is one less name to
learn in the API.
## Test Plan
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
This PR renames categories of llama_stack loggers.
This PR aligns logging categories as per the package name, as well as
reviews from initial
https://github.com/meta-llama/llama-stack/pull/2868. This is a follow up
to #3061.
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
Replaces https://github.com/meta-llama/llama-stack/pull/2868
Part of https://github.com/meta-llama/llama-stack/issues/2865
cc @leseb @rhuss
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
# What does this PR do?
Handles MCP tool calls in a previous response
Closes#3105
## Test Plan
Made call to create response with tool call, then made second call with
the first linked through previous_response_id. Did not get error.
Also added unit test.
Signed-off-by: Gordon Sim <gsim@redhat.com>
# What does this PR do?
This PR adds a step in pre-commit to enforce using `llama_stack` logger.
Currently, various parts of the code base uses different loggers. As a
custom `llama_stack` logger exist and used in the codebase, it is better
to standardize its utilization.
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
Co-authored-by: Matthew Farrellee <matt@cs.wisc.edu>
# What does this PR do?
Add CodeScanner implementations
## Test Plan
`SAFETY_MODEL=CodeScanner LLAMA_STACK_CONFIG=starter uv run pytest -v
tests/integration/safety/test_safety.py
--text-model=llama3.2:3b-instruct-fp16
--embedding-model=all-MiniLM-L6-v2 --safety-shield=ollama`
This PR need to land after this
https://github.com/meta-llama/llama-stack/pull/3098
Replace chat_completion calls with openai_chat_completion to eliminate
dependency on legacy inference APIs.
# What does this PR do?
<!-- 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#3067
## Test Plan
<!-- Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.* -->
# What does this PR do?
Adds proper streaming events for MCP tool listing (`mcp_list_tools.in_progress` and `mcp_list_tools.completed`). Also refactors things a bit more.
## Test Plan
Verified existing integration tests pass with the refactored code. The test `test_response_streaming_multi_turn_tool_execution` has been updated to check for the new MCP list tools streaming events
# What does this PR do?
Refactors the OpenAI response conversion utilities by moving helper functions from `openai_responses.py` to `utils.py`. Adds unit tests.
# What does this PR do?
Refactors the OpenAI responses implementation by extracting streaming and tool execution logic into separate modules. This improves code organization by:
1. Creating a new `StreamingResponseOrchestrator` class in `streaming.py` to handle the streaming response generation logic
2. Moving tool execution functionality to a dedicated `ToolExecutor` class in `tool_executor.py`
## Test Plan
Existing tests
# What does this PR do?
Adds content part streaming events to the OpenAI-compatible Responses API to support more granular streaming of response content. This introduces:
1. New schema types for content parts: `OpenAIResponseContentPart` with variants for text output and refusals
2. New streaming event types:
- `OpenAIResponseObjectStreamResponseContentPartAdded` for when content parts begin
- `OpenAIResponseObjectStreamResponseContentPartDone` for when content parts complete
3. Implementation in the reference provider to emit these events during streaming responses. Also emits MCP arguments just like function call ones.
## Test Plan
Updated existing streaming tests to verify content part events are properly emitted
# What does this PR do?
Enhances tool execution streaming by adding support for real-time progress events during tool calls. This implementation adds streaming events for MCP and web search tools, including in-progress, searching, completed, and failed states.
The refactored `_execute_tool_call` method now returns an async iterator that yields streaming events throughout the tool execution lifecycle.
## Test Plan
Updated the integration test `test_response_streaming_multi_turn_tool_execution` to verify the presence and structure of new streaming events, including:
- Checking for MCP in-progress and completed events
- Verifying that progress events contain required fields (item_id, output_index, sequence_number)
- Ensuring completed events have the necessary sequence_number field
# What does this PR do?
To be compliant with model policies for LLAMA, just return the
categories as is from provider, we will lose the OAI compat in
moderations api response.
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
## Test Plan
`SAFETY_MODEL=llama-guard3:8b LLAMA_STACK_CONFIG=starter uv run pytest
-v tests/integration/safety/test_safety.py
--text-model=llama3.2:3b-instruct-fp16
--embedding-model=all-MiniLM-L6-v2 --safety-shield=ollama`
# What does this PR do?
1. Updates `AgentPersistence.list_sessions()` to properly filter out
`Turn` keys from `Session` keys.
2. Adds a suite of unit tests to confirm the `list_sessions()` behavior
and tests the failed sample in
https://github.com/meta-llama/llama-stack/issues/3048
## Fixes https://github.com/meta-llama/llama-stack/issues/3048
## Test Plan
Unit tests added.
---------
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
Some fixes to MCP tests. And a bunch of fixes for Vector providers.
I also enabled a bunch of Vector IO tests to be used with
`LlamaStackLibraryClient`
## Test Plan
Run Responses tests with llama stack library client:
```
pytest -s -v tests/integration/non_ci/responses/ --stack-config=server:starter \
--text-model openai/gpt-4o \
--embedding-model=sentence-transformers/all-MiniLM-L6-v2 \
-k "client_with_models"
```
Do the same with `-k openai_client`
The rest should be taken care of by CI.
Well our Responses tests use it so we better include it in the API, no?
I discovered it because I want to make sure `llama-stack-client` can be
used always instead of `openai-python` as the client (we do want to be
_truly_ compatible.)
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
This PR addresses an issue where `PromptGuardSafetyImpl` was an
incomplete implementation of an abstract class. The class was missing
the required run_moderation method from its parent interface.
Currently, running `pre-commit` locally fails with the error below.
```
llama_stack/providers/inline/safety/prompt_guard/__init__.py:15: error: Cannot instantiate abstract class "PromptGuardSafetyImpl" with abstract attribute "run_moderation" [abstract]
Found 1 error in 1 file (checked 410 source files)
```
This PR fixes the issue as follows
- Added the missing run_moderation method to PromptGuardSafetyImpl
- Method raises NotImplementedError with appropriate message indicating
this functionality is not implemented for PromptGuard
- This allows the class to be properly instantiated while clearly
indicating the limitation
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
Signed-off-by: Mustafa Elbehery <melbeher@redhat.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 adds static type coverage to `llama-stack`
Part of https://github.com/meta-llama/llama-stack/issues/2647
<!-- 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.* -->
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
# What does this PR do?
This PR adds Open AI Compatible moderations api. Currently only
implementing for llama guard safety provider
Image support, expand to other safety providers and Deprecation of
run_shield will be next steps.
## Test Plan
Added 2 new tests for safe/ unsafe text prompt examples for the new open
ai compatible moderations api usage
`SAFETY_MODEL=llama-guard3:8b LLAMA_STACK_CONFIG=starter uv run pytest
-v tests/integration/safety/test_safety.py
--text-model=llama3.2:3b-instruct-fp16
--embedding-model=all-MiniLM-L6-v2 --safety-shield=ollama`
(Had some issue with previous PR
https://github.com/meta-llama/llama-stack/pull/2994 while updating and
accidentally close it , reopened new one )
# What does this PR do?
I found a few issues while adding new metrics for various APIs:
currently metrics are only propagated in `chat_completion` and
`completion`
since most providers use the `openai_..` routes as the default in
`llama-stack-client inference chat-completion`, metrics are currently
not working as expected.
in order to get them working the following had to be done:
1. get the completion as usual
2. use new `openai_` versions of the metric gathering functions which
use `.usage` from the `OpenAI..` response types to gather the metrics
which are already populated.
3. define a `stream_generator` which counts the tokens and computes the
metrics (only for stream=True)
5. add metrics to response
NOTE: I could not add metrics to `openai_completion` where stream=True
because that ONLY returns an `OpenAICompletion` not an AsyncGenerator
that we can manipulate.
acquire the lock, and add event to the span as the other `_log_...`
methods do
some new output:
`llama-stack-client inference chat-completion --message hi`
<img width="2416" height="425" alt="Screenshot 2025-07-16 at 8 28 20 AM"
src="https://github.com/user-attachments/assets/ccdf1643-a184-4ddd-9641-d426c4d51326"
/>
and in the client:
<img width="763" height="319" alt="Screenshot 2025-07-16 at 8 28 32 AM"
src="https://github.com/user-attachments/assets/6bceb811-5201-47e9-9e16-8130f0d60007"
/>
these were not previously being recorded nor were they being printed to
the server due to the improper console sink handling
---------
Signed-off-by: Charlie Doern <cdoern@redhat.com>
A bunch of miscellaneous cleanup focusing on tests, but ended up
speeding up starter distro substantially.
- Pulled llama stack client init for tests into `pytest_sessionstart` so
it does not clobber output
- Profiling of that told me where we were doing lots of heavy imports
for starter, so lazied them
- starter now starts 20seconds+ faster on my Mac
- A few other smallish refactors for `compat_client`
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
Extend the Shields Protocol and implement the capability to unregister
previously registered shields and CLI for shields management.
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
Closes#2581
## 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.* -->
First of, test API for shields
1. Install and start Ollama:
`ollama serve`
2. Pull Llama Guard Model in Ollama:
`ollama pull llama-guard3:8b`
3. Configure env variables:
```
export ENABLE_OLLAMA=ollama
export OLLAMA_URL=http://localhost:11434
```
4. Build Llama Stack distro:
`llama stack build --template starter --image-type venv `
5. Start Llama Stack server:
`llama stack run starter --port 8321`
6. Check if Ollama model is available:
`curl -X GET http://localhost:8321/v1/models | jq '.data[] |
select(.provider_id=="ollama")'`
7. Register a new Shield using Ollama provider:
```
curl -X POST http://localhost:8321/v1/shields \
-H "Content-Type: application/json" \
-d '{
"shield_id": "test-shield",
"provider_id": "llama-guard",
"provider_shield_id": "ollama/llama-guard3:8b",
"params": {}
}'
```
`{"identifier":"test-shield","provider_resource_id":"ollama/llama-guard3:8b","provider_id":"llama-guard","type":"shield","owner":{"principal":"","attributes":{}},"params":{}}%
`
8. Check if shield was registered:
`curl -X GET http://localhost:8321/v1/shields/test-shield`
`{"identifier":"test-shield","provider_resource_id":"ollama/llama-guard3:8b","provider_id":"llama-guard","type":"shield","owner":{"principal":"","attributes":{}},"params":{}}%
`
9. Run shield:
```
curl -X POST http://localhost:8321/v1/safety/run-shield \
-H "Content-Type: application/json" \
-d '{
"shield_id": "test-shield",
"messages": [
{
"role": "user",
"content": "How can I hack into someone computer?"
}
],
"params": {}
}'
```
`{"violation":{"violation_level":"error","user_message":"I can't answer
that. Can I help with something
else?","metadata":{"violation_type":"S2"}}}% `
10. Unregister shield:
`curl -X DELETE http://localhost:8321/v1/shields/test-shield`
`null% `
11. Verify shield was deleted:
`curl -X GET http://localhost:8321/v1/shields/test-shield`
`{"detail":"Invalid value: Shield 'test-shield' not found"}%`
All tests passed ✅
```
========================================================================== 430 passed, 194 warnings in 19.54s ==========================================================================
/Users/iamiller/GitHub/llama-stack/.venv/lib/python3.12/site-packages/litellm/llms/custom_httpx/async_client_cleanup.py:78: RuntimeWarning: coroutine 'close_litellm_async_clients' was never awaited
loop.close()
RuntimeWarning: Enable tracemalloc to get the object allocation traceback
Wrote HTML report to htmlcov-3.12/index.html
```
# What does this PR do?
1. Creates a new `SessionNotFoundError` class
2. Implements the new class where appropriate
Relates to #2379
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>
# What does this PR do?
Remove score_threshold based check from `OpenAIVectorStoreMixin`
Closes: https://github.com/meta-llama/llama-stack/issues/3018
<!-- 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.* -->
# What does this PR do?
Adds support to Vector store Open AI APIs in Qdrant.
<!-- If resolving an issue, uncomment and update the line below -->
Closes#2463
## 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>
Co-authored-by: ehhuang <ehhuang@users.noreply.github.com>
Co-authored-by: Francisco Arceo <arceofrancisco@gmail.com>
What does this PR do?
This PR adds support for Direct Preference Optimization (DPO) training
via the existing HuggingFace inline provider. It introduces a new DPO
training recipe, config schema updates, dataset integration, and
end-to-end testing to support preference-based fine-tuning with TRL.
Test Plan
Added integration test:
tests/integration/post_training/test_post_training.py::TestPostTraining::test_preference_optimize
Ran tests on both CPU and CUDA environments
---------
Co-authored-by: Ubuntu <ubuntu@ip-172-31-43-83.ec2.internal>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
1. Creates a new `VectorStoreNotFoundError` class
2. Implements the new class where appropriate
Relates to #2379
Signed-off-by: Nathan Weinberg <nweinber@redhat.com>
# What does this PR do?
- Change max_seq_length to max_length in SFTConfig constructor
- TRL deprecated max_seq_length in Feb 2024 and removed it in v0.20.0
- Reference: https://github.com/huggingface/trl/pull/2895
This resolves the SFT training failure in CI tests
Add support for deleting individual chunks from vector stores
- Add abstract remove_chunk() method to EmbeddingIndex base class
- Implement chunk deletion for Faiss provider, SQLite Vec, Milvus,
PGVector
- Placeholder implementations with NotImplementedError for
Chroma/Qdrant/Weaviate
- Integrate chunk deletion into OpenAI vector store file deletion flow
- removed xfail from
test_openai_vector_store_delete_file_removes_from_vector_store
Closes: #2477
---------
Signed-off-by: Derek Higgins <derekh@redhat.com>
Co-authored-by: Francisco Arceo <arceofrancisco@gmail.com>
# What does this PR do?
currently `print` is being used with custom formatting to achieve
telemetry output in the console_span_processor
This causes telemetry not to show up in log files when using
`LLAMA_STACK_LOG_FILE`. During testing it looks like telemetry is not
being captured when it is
switch to using Rich formatting with the logger and then strip the
formatting off when a log file is being used so the formatting looks
normal
## Test Plan
before:
console:
<img width="967" height="127" alt="Screenshot 2025-07-21 at 4 02 15 PM"
src="https://github.com/user-attachments/assets/b09518cc-9d38-4970-9877-70e2c41fcbb5"
/>
log file (no telemetry):
```
2025-07-21 16:01:32,481 llama_stack.providers.remote.inference.ollama.ollama:117 inference: checking connectivity to Ollama at `http://localhost:11434`...
2025-07-21 16:01:34,779 opentelemetry.trace:537 uncategorized: Overriding of current TracerProvider is not allowed
2025-07-21 16:01:35,083 __main__:587 server: Listening on ['::', '0.0.0.0']:8321
2025-07-21 16:01:35,091 uvicorn.error:84 uncategorized: Started server process [68679]
2025-07-21 16:01:35,091 uvicorn.error:48 uncategorized: Waiting for application startup.
2025-07-21 16:01:35,092 __main__:163 server: Starting up
2025-07-21 16:01:35,092 uvicorn.error:62 uncategorized: Application startup complete.
2025-07-21 16:01:35,092 uvicorn.error:216 uncategorized: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit)
2025-07-21 16:01:37,167 uvicorn.access:473 uncategorized: 127.0.0.1:53145 - "POST /v1/openai/v1/chat/completions HTTP/1.1" 200
```
after:
console:
<img width="797" height="165" alt="Screenshot 2025-07-22 at 3 28 44 PM"
src="https://github.com/user-attachments/assets/44d40e3b-6502-439d-9ea5-38058b289962"
/>
log file:
```
2025-07-21 15:59:51,481 llama_stack.providers.remote.inference.ollama.ollama:117 inference: checking connectivity to Ollama at `http://localhost:11434`...
2025-07-21 15:59:53,801 opentelemetry.trace:537 uncategorized: Overriding of current TracerProvider is not allowed
2025-07-21 15:59:54,059 __main__:587 server: Listening on ['::', '0.0.0.0']:8321
2025-07-21 15:59:54,066 uvicorn.error:84 uncategorized: Started server process [68578]
2025-07-21 15:59:54,067 uvicorn.error:48 uncategorized: Waiting for application startup.
2025-07-21 15:59:54,067 __main__:163 server: Starting up
2025-07-21 15:59:54,067 uvicorn.error:62 uncategorized: Application startup complete.
2025-07-21 15:59:54,068 uvicorn.error:216 uncategorized: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit)
2025-07-21 15:59:55,381 [TELEMETRY] 19:59:55.381 /v1/openai/v1/chat/completions
2025-07-21 15:59:55,619 uvicorn.access:473 uncategorized: 127.0.0.1:53102 - "POST /v1/openai/v1/chat/completions HTTP/1.1" 200
2025-07-21 15:59:55,621 [TELEMETRY] 19:59:55.621 /v1/openai/v1/chat/completions [StatusCode.OK] (240.07ms)
2025-07-21 15:59:55,622 [TELEMETRY] 19:59:55.620 127.0.0.1:53102 - "POST /v1/openai/v1/chat/completions HTTP/1.1" 200
```
Signed-off-by: Charlie Doern <cdoern@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.
# What does this PR do?
This PR implements the openai compatible endpoints for chromadb
Closes#2462
## Test Plan
Ran ollama llama stack server and ran the command
`pytest -sv --stack-config=http://localhost:8321
tests/integration/vector_io/test_openai_vector_stores.py
--embedding-model all-MiniLM-L6-v2`
8 failed, 27 passed, 8 skipped, 1 xfailed
The failed ones are regarding files api
---------
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
Co-authored-by: sarthakdeshpande <sarthak.deshpande@engati.com>
Co-authored-by: Francisco Javier Arceo <farceo@redhat.com>
Co-authored-by: Francisco Arceo <arceofrancisco@gmail.com>
# What does this PR do?
Moving vector store and vector store files helper methods to
`openai_vector_store_mixin.py`
<!-- If resolving an issue, uncomment and update the line below -->
<!-- Closes #[issue-number] -->
## Test Plan
The tests are already supported in the CI and tests the inline providers
and current integration tests.
Note that the `vector_index` fixture will be test `milvus_vec_adapter`,
`faiss_vec_adapter`, and `sqlite_vec_adapter` in
`tests/unit/providers/vector_io/test_vector_io_openai_vector_stores.py`.
Additionally, the integration tests in `integration-vector-io-tests.yml`
runs `tests/integration/vector_io` tests for the following providers:
```python
vector-io-provider: ["inline::faiss", "inline::sqlite-vec", "inline::milvus", "remote::chromadb", "remote::pgvector"]
```
Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
# What does this PR do?
Ensures that session turns retrieved from the agent persistence layer
are sorted by their `started_at` timestamp, as the key-value store does
not guarantee order.
Closes#2852
## Test Plan
- [ ] Add unit tests
Inline _inference_ providers haven't proved to be very useful -- they
are rarely used. And for good reason -- it is almost never a good idea
to include a complex (distributed) inference engine bundled into a
distributed stateful front-end server serving many other things.
Responsibility should be split properly.
See Discord discussion:
1395849853
# What does this PR do?
<!-- Provide a short summary of what this PR does and why. Link to
relevant issues if applicable. -->
This PR adds static type coverage to `llama-stack`
Part of https://github.com/meta-llama/llama-stack/issues/2647
<!-- 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.* -->
Signed-off-by: Mustafa Elbehery <melbeher@redhat.com>
