mirror of
https://github.com/meta-llama/llama-stack.git
synced 2025-12-03 09:53:45 +00:00
f34f22f8c7
# 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. |
||
|---|---|---|
| .. | ||
| agents | ||
| datasets | ||
| eval | ||
| fixtures | ||
| inference | ||
| inspect | ||
| post_training | ||
| providers | ||
| safety | ||
| scoring | ||
| telemetry | ||
| test_cases | ||
| tool_runtime | ||
| tools | ||
| vector_io | ||
| __init__.py | ||
| conftest.py | ||
| metadata.py | ||
| README.md | ||
| report.py | ||
Llama Stack Integration Tests
We use pytest for parameterizing and running tests. You can see all options with:
cd tests/integration
# this will show a long list of options, look for "Custom options:"
pytest --help
Here are the most important options:
--stack-config: specify the stack config to use. You have three ways to point to a stack:- a URL which points to a Llama Stack distribution server
- a template (e.g.,
fireworks,together) or a path to a run.yaml file - a comma-separated list of api=provider pairs, e.g.
inference=fireworks,safety=llama-guard,agents=meta-reference. This is most useful for testing a single API surface.
--env: set environment variables, e.g. --env KEY=value. this is a utility option to set environment variables required by various providers.
Model parameters can be influenced by the following options:
--text-model: comma-separated list of text models.--vision-model: comma-separated list of vision models.--embedding-model: comma-separated list of embedding models.--safety-shield: comma-separated list of safety shields.--judge-model: comma-separated list of judge models.--embedding-dimension: output dimensionality of the embedding model to use for testing. Default: 384
Each of these are comma-separated lists and can be used to generate multiple parameter combinations. Note that tests will be skipped if no model is specified.
Experimental, under development, options:
--record-responses: record new API responses instead of using cached ones--report: path where the test report should be written, e.g. --report=/path/to/report.md
Examples
Run all text inference tests with the together distribution:
pytest -s -v tests/integration/inference/test_text_inference.py \
--stack-config=together \
--text-model=meta-llama/Llama-3.1-8B-Instruct
Run all text inference tests with the together distribution and meta-llama/Llama-3.1-8B-Instruct:
pytest -s -v tests/integration/inference/test_text_inference.py \
--stack-config=together \
--text-model=meta-llama/Llama-3.1-8B-Instruct
Running all inference tests for a number of models:
TEXT_MODELS=meta-llama/Llama-3.1-8B-Instruct,meta-llama/Llama-3.1-70B-Instruct
VISION_MODELS=meta-llama/Llama-3.2-11B-Vision-Instruct
EMBEDDING_MODELS=all-MiniLM-L6-v2
export TOGETHER_API_KEY=<together_api_key>
pytest -s -v tests/integration/inference/ \
--stack-config=together \
--text-model=$TEXT_MODELS \
--vision-model=$VISION_MODELS \
--embedding-model=$EMBEDDING_MODELS
Same thing but instead of using the distribution, use an adhoc stack with just one provider (fireworks for inference):
export FIREWORKS_API_KEY=<fireworks_api_key>
pytest -s -v tests/integration/inference/ \
--stack-config=inference=fireworks \
--text-model=$TEXT_MODELS \
--vision-model=$VISION_MODELS \
--embedding-model=$EMBEDDING_MODELS
Running Vector IO tests for a number of embedding models:
EMBEDDING_MODELS=all-MiniLM-L6-v2
pytest -s -v tests/integration/vector_io/ \
--stack-config=inference=sentence-transformers,vector_io=sqlite-vec \
--embedding-model=$EMBEDDING_MODELS