This gets the fireworks provider passing 100% of our OpenAI API
verification tests when run against a Llama Stack server using the
fireworks provider. Testing against Fireworks directly, without Llama
Stack in the middle, has a lower pass rate.
The main changes are are in how we divert Llama model OpenAI chat
completion requests to the Llama Stack chat completion API (vs
OpenAI), which applies all the client-side formatting necessary to get
tool calls working properly on Fireworks.
A side-effect of this work is any provider using the
OpenAIChatCompletionToLlamaStackMixin (renamed from
OpenAIChatCompletioonUnsupportedMixin) will also get a better
conversion from OpenAI to Llama Stack, including streaming and
non-stream responses.
A small change was required to
`llama_stack/models/llama/llama3/tool_utils.py` to get tests to 100%
because code there was incorrectly assuming any JSON response with a
`name` key was a tool call response. One of our verification tests
produces JSON keys with a `name` key that is not a tool call response,
so I tightened up the logic there to require both a `name` and
`parameters` key in the JSON response before it gets considered a
potential tool call. The `parameters` key was required by the code
anyway, but it wasn't explicitly checking for its existence.
Lastly, this adds some new verification test configs so we can see the
results of using OpenAI APIs against SaaS services directly compared
to hitting Llama Stack with a remote provider pointing at that SaaS
service.
You can run these verification tests like:
```
llama stack run \
--image-type venv \
tests/verifications/openai-api-verification-run.yaml
python tests/verifications/generate_report.py \
--run-tests \
--provider together fireworks openai \
together-llama-stack \
fireworks-llama-stack \
openai-llama-stack
```
Signed-off-by: Ben Browning <bbrownin@redhat.com>
# What does this PR do?
## Test Plan
(myenv) ➜ llama-stack python tests/verifications/generate_report.py
--providers fireworks,together,openai --run-tests
# What does this PR do?
- provider and their models now live in config.yaml
- better distinguish different cases within a test
- add model key to surface provider's model_id
- include example command to rerun single test case
## Test Plan
<img width="1173" alt="image"
src="https://github.com/user-attachments/assets/b414baf0-c768-451f-8c3b-c2905cf36fac"
/>
