fast_dynamic_batcher 1.1.0

Example Usage

To use dynamic batching in FastAPI, you have to first create an instance of the InferenceModel class. Initiate your ML model in its init method and use it in its infer method:

from typing import Any
from fast_dynamic_batcher.inference_template import InferenceModel


class DemoModel(InferenceModel):
   def __init__(self):
       super().__init__()
       # Initiate your ML model here

   def infer(self, inputs: list[Any]) -> list[Any]:
       # Run your inputs as a batch for your model
       ml_output = ... # Your inference outputs
       return ml_output

Subsequently, use your InferenceModel instance to initiate our DynBatcher:

from contextlib import asynccontextmanager

from anyio import CapacityLimiter
from anyio.lowlevel import RunVar

from fast_dynamic_batcher.dyn_batcher import DynBatcher


@asynccontextmanager
async def lifespan(app: FastAPI):
   RunVar("_default_thread_limiter").set(CapacityLimiter(16))
   global dyn_batcher
   dyn_batcher = DynBatcher(DemoModel, max_batch_size = 8, max_delay = 0.1)
   yield
   dyn_batcher.stop()

app = FastAPI(lifespan=lifespan)

@app.post("/predict/")
async def predict(
   input_model: YourInputPydanticModel
):
   return await dyn_batcher.process_batched(input_model)

The DynBatcher can be initiated in the FastAPI lifespans as a global variable. It can be further customized with the max_batch_size and max_delay variables. Subsequently, use it in your FastAPI endpoints by registering your inputs by calling its process_batched method.

Our dynamic batching algorithm will then wait for either the number of inputs to equal the max_batch_size, or until max_delay seconds have passed. In the latter case, a batch may contain between 1 and max_batch_size inputs. Once, either condition is met, a batch will be processed by calling the infer method of your InferenceModel instance.

Installation

The Fast Dynamic Batcher library can be installed with pip:

pip install fast_dynamic_batcher

Performance Tests

We tested the performance of our dynamic batching solution against a baseline without batching on a Colab instance with a T4 GPU as well as on a laptop with an Intel i7-1250U CPU. The experiments were conducted by using this testing script. The results are reported in the table below:

On GPUs, which benefit greatly from large batch sizes, we achieved a speed-up of almost 2.5x by creating dynamic batches of size 16. On CPUs, the gains are more modest with a speed-up of 1.3x.