fix: multiple tool calls in remote-vllm chat_completion #2161
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
bbrowning
requested review from
SLR722,
ashwinb,
dineshyv,
dltn,
ehhuang,
hardikjshah,
leseb,
raghotham,
sixianyi0721,
terrytangyuan,
vladimirivic and
yanxi0830
as code owners
facebook-github-bot
added
the
CLA Signed
ashwinb
reviewed
May 14, 2025
This fixes an issue in how we used the tool_call_buf from streaming tool calls in the remote-vllm provider where it would end up concatenating parameters from multiple different tool call results instead of aggregating the results from each tool call separately. It also fixes an issue found while digging into that where we were accidentally mixing the json string form of tool call parameters with the string representation of the python form, which mean we'd end up with single quotes in what should be double-quoted json strings. The following tests are now passing 100% for the remote-vllm provider, where some of the test_text_inference were failing before this change: ``` VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_text_inference.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/inference/test_vision_inference.py --vision-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" ``` Many of the agent tests are passing, although some are failing due to bugs in vLLM's pythonic tool parser for Llama models. See the PR at vllm-project/vllm#17917 and a gist at https://gist.github.com/bbrowning/b5007709015cb2aabd85e0bd08e6d60f for changes needed there, which will have to get made upstream in vLLM. Agent tests: ``` VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/agents/test_agents.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" ```` Signed-off-by: Ben Browning <bbrownin@redhat.com>
This updates test_agents.py a bit after testing with Llama 4 Scout and the remote-vllm provider. The main difference here is a bit more verbose prompting to encourage tool calls because Llama 4 Scout likes to reply that polyjuice is fictional and has no boiling point vs calling our custom tool unless it's prodded a bit. Also, the remote-vllm distribution doesn't use input/output shields by default so test_multi_tool_calls was adjusted to only expect the shield results if shields are in use and otherwise not check for shield usage. Note that it requires changes to the vLLM pythonic tool parser to pass these tests - those are listed at https://gist.github.com/bbrowning/4734240ce96b4264340caa9584e47c9e With this change, all of the agent tests pass with Llama 4 Scout and remote-vllm except one of the RAG tests, that looks to be an unrelated (and pre-existing) failure. ``` VLLM_URL="http://localhost:8000/v1" INFERENCE_MODEL="RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" LLAMA_STACK_CONFIG=remote-vllm python -m pytest -v tests/integration/agents/test_agents.py --text-model "RedHatAI/Llama-4-Scout-17B-16E-Instruct-FP8-dynamic" ``` Signed-off-by: Ben Browning <bbrownin@redhat.com>
This code was unnecessarily verbose when calculating has_input_shield and has_output_shield values, and was simplified to take advantage of Python's intrinsic treatment of empty lists and None values as false. Signed-off-by: Ben Browning <bbrownin@redhat.com>
bbrowning
force-pushed
the
remote-vllm-multi-tools
branch
from
44799a7 to
18c2494
Compare
|
Rebased on top of some conflicting vllm changes, and I re-ran all the remote-vllm unit tests as well as all test_text_inference, test_vision_inference, and test_agents with the commands outlined in the PR description, ensuring they all still pass (except that 1 agent test that was failing for remote-vllm previously, and needs a separate fix unrelated to this). |
|
Thanks Ben! |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
This fixes an issue in how we used the tool_call_buf from streaming tool calls in the remote-vllm provider where it would end up concatenating parameters from multiple different tool call results instead of aggregating the results from each tool call separately.
It also fixes an issue found while digging into that where we were accidentally mixing the json string form of tool call parameters with the string representation of the python form, which mean we'd end up with single quotes in what should be double-quoted json strings.
Closes #1120
Test Plan
The following tests are now passing 100% for the remote-vllm provider, where some of the test_text_inference were failing before this change:
All but one of the agent tests are passing (including the multi-tool one). See the PR at vllm-project/vllm#17917 and a gist at https://gist.github.com/bbrowning/4734240ce96b4264340caa9584e47c9e for changes needed there, which will have to get made upstream in vLLM.
Agent tests: