Mirrors
/
hf_text-generation-inference
mirror of https://github.com/huggingface/text-generation-inference.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
Large Language Model Text Generation Inference
7 Commits 181 Branches 50 Tags 606 MiB
Python 71%
Rust 22.2%
Cuda 4.7%
C++ 1%
Nix 0.4%
Other 0.6%
Go to file
Olivier Dehaene f11965c11d support deepspeed 2022-10-13 11:05:44 +02:00
k6 Add load testing 2022-10-11 10:36:51 +02:00
proto Refactored gRPC interface 2022-10-11 16:50:54 +02:00
router Use axum 2022-10-11 18:14:39 +02:00
server support deepspeed 2022-10-13 11:05:44 +02:00
.gitignore Add load testing 2022-10-11 10:36:51 +02:00
README.md Refactored gRPC interface 2022-10-11 16:50:54 +02:00

README.md

Text Generation Inference

A Rust and gRPC server for text generation inference.

Load Tests

See k6/load_test.js We send the default examples with a 1 second delay between each request.

Stages:

  • Ramp up to 50 concurrent requests per second in 1min
  • Ramp up from 50 to 100 concurrent requests per second in 2min
  • Ramp down to 0 concurrent requests per second in 1min
avg min med max p(90) p(95) RPS
Original code 8.9s 1s 9.12s 16.69s 13.7s 14.26s 5.9
ISO with original code 8.88s 959.53ms 8.89s 17.08s 13.34s 14.12s 5.94
New batching logic 5.44s 1.27s 5.28s 13.12s 7.78s 8.92s 9.08

Install

cd server
pip install .

cd router
cargo build --release

Run

python server/bloom_inference/main.py bigscience/bloom --num-gpus 8 --shard-directory /dev/shm/models

./router/target/release/router

TODO:

  • Improve model download
    • Store "shardable" layers separately and layer by layer
  • Add batching args to router CLI
  • Add docstrings + comments everywhere as the codebase is fairly complicated
  • Add tests
  • Add shutdown logic in router and server
  • Improve multi-processing logic in server
  • Improve error handling everywhere
  • Improve past key layer indexing?