Skip to content

README.md

Latest commit

 

History

History
77 lines (63 loc) · 3.51 KB

README.md

File metadata and controls

77 lines (63 loc) · 3.51 KB
Raw

Auto Batching Proxy

It will automatically batch inference requests from independent users in a single batch request for efficiency, so that for users the interface looks like individual requests, but internally it is handled as a batch request, essentially it provide a REST API wrapper around some inference service like https://github.com/huggingface/text-embeddings-inference

Proxy server is configured with following parameters:

Max Wait Time - maximal time user request can wait for other requests to be accumulated in a batch
Max Batch Size - maximal number of requests that can be accumulated in a batch.

Setup Inference Service

First, try running inference service in a container with --model-id nomic-ai/nomic-embed-text-v1.5

docker run --rm -it -p 8080:80 --pull always \
    ghcr.io/huggingface/text-embeddings-inference:cpu-latest \
--model-id nomic-ai/nomic-embed-text-v1.5

if it fails to start, then try with some other alternatives. Currently, code is functional for
--model-id sentence-transformers/all-MiniLM-L6-v2 & --model-id sentence-transformers/all-mpnet-base-v2 Check /screenshots for some of the tried models.

docker run --rm -it -p 8080:80 --pull always \
  ghcr.io/huggingface/text-embeddings-inference:cpu-latest \
  --model-id sentence-transformers/all-MiniLM-L6-v2

Note: Backend does not support a batch size > 8 but our proxy will respect this config param & will not send requests (as well as max inputs, which is 32 for all-MiniLM-L6-v2) more than supported batch size.

Setup Proxy Service

RUST_LOG=INFO cargo run -- --max-batch-size 50 --max-wait-time-ms 3000

Unit tests
Relevant unit tests are provided inside /src source code files

Integration tests
Check the /tests folder, code is covered with various scenarios.

Run all tests via cargo test. Currently, tests are verified to be passed against
--model-id sentence-transformers/all-MiniLM-L6-v2 & --model-id sentence-transformers/all-mpnet-base-v2 & they also explain how/why which part of code was written for which particular use case.

Use the following simple CURL commands for quick testing

  • for inference
curl -X POST http://localhost:8080/embed \
  -H "Content-Type: application/json" \
  -d '{"inputs": ["Hello world"]}'
  • for proxy
curl -X POST http://localhost:3000/embed \
  -H "Content-Type: application/json" \
  -d '{"inputs": ["Hello", "World"]}'

to verify proxy is working for multiple concurrent requests

cd scripts 
./proxy_concurrent_calls.sh

Benchmark test results:
Following output is taken from

$ RUSTFLAGS="-A dead_code" cargo test test_compare_single_input_inference_service_vs_auto_batching_proxy_with_x_separate_requests -- --nocapture
  • -A shortcut for allow (to suppress warnings generated for unused functions, even though they are actually used in tests)
  • --nocapture will recover display output

Full output: timing_summary.png