Skip to content

Model Configuration

You set per model settings by adding a serving.properties file in the root of your model directory (or .zip). These apply for all engines and modes.

An example serving.properties can be found here.

Main properties

In serving.properties, you can set the following properties. Model properties are accessible to Translator and python handler functions.

  • engine: Which Engine to use, values include MXNet, PyTorch, TensorFlow, ONNX, DeepSpeed, etc.
  • load_on_devices: A ; delimited devices list, which the model to be loaded on, default to load on all devices.
  • translatorFactory: Specify the TranslatorFactory.
  • job_queue_size: Specify the job queue size at model level, this will override global job_queue_size, default is 1000.
  • batch_size: the dynamic batch size, default is 1.
  • max_batch_delay - the maximum delay for batch aggregation in millis, default value is 100 milliseconds.
  • max_idle_time - the maximum idle time in seconds before the worker thread is scaled down, default is 60 seconds.
  • log_request_metric: Enable model metrics (inference, pre-process and post-process latency) logging.
  • metrics_aggregation: Number of model metrics to aggregate, default is 1000.
  • minWorkers: Minimum number of workers, default is 1.
  • maxWorkers: Maximum number of workers, default is #CPU/OMP_NUM_THREAD for CPU, GPU default is 2, inferentia default is 2 (PyTorch engine), 1 (Python engine) .
  • gpu.minWorkers: Minimum number of workers for GPU.
  • gpu.maxWorkers: Maximum number of workers for GPU.
  • cpu.minWorkers: Minimum number of workers for CPU.
  • cpu.maxWorkers: Maximum number of workers for CPU.
  • required_memory_mb: Specify the required memory (CPU and GPU) in MB to load the model.
  • gpu.required_memory_mb: Specify the required GPU memory in MB to load the model.
  • reserved_memory_mb: Reserve memory in MB to avoid system out of memory.
  • gpu.reserved_memory_mb: Reserve GPU memory in MB to avoid system out of memory.

Option Properties

In serving.properties, you can also set options (prefixed with option) and properties. The options will be passed to Model.load(Path modelPath, String prefix, Map<String, ?> options) API. It allows you to set engine specific configurations. Here are some of the available option properties:

# set model file name prefix if different from folder name
option.modeName=resnet18_v1

# PyTorch options
option.mapLocation=true
option.extraFiles=foo.txt,bar.txt

# ONNXRuntime options
option.interOpNumThreads=2
option.intraOpNumThreads=2
option.executionMode=SEQUENTIAL
option.optLevel=BASIC_OPT
option.memoryPatternOptimization=true
option.cpuArenaAllocator=true
option.disablePerSessionThreads=true
option.customOpLibrary=myops.so
option.disablePerSessionThreads=true
option.ortDevice=TensorRT/ROCM/CoreML

# Python model options
# Mark model as failure after python process crashing 10 times
retry_threshold=10

option.pythonExecutable=python3
option.entryPoint=deepspeed.py
option.handler=hanlde
option.predict_timeout=120
option.model_loading_timeout=10
option.parallel_loading=true
option.tensor_parallel_degree=2
option.enable_venv=true
option.rolling_batch=auto
#option.rolling_batch=lmi-dist
option.max_rolling_batch_size=64

# max output size in bytes, default to 60M
option.max_output_size=67108864

Most of the options can also be overridden by an environment variable with the OPTION_ prefix. For example:

# to enable rolling batch with only environment variable:
export OPTION_ROLLING_BATCH=auto

Basic Model Configurations

You can set number of workers for each model: https://github.com/deepjavalibrary/djl-serving/blob/master/serving/src/test/resources/identity/serving.properties#L4-L8

For example, set minimum workers and maximum workers for your model:

minWorkers=32
maxWorkers=64

Or you can configure minimum workers and maximum workers differently for GPU and CPU:

gpu.minWorkers=2
gpu.maxWorkers=3
cpu.minWorkers=2
cpu.maxWorkers=4

job queue size, batch size, max batch delay, max worker idle time can be configured at per model level, this will override global settings:

job_queue_size=10
batch_size=2
max_batch_delay=1
max_idle_time=120

You can configure which device to load the model on, default is *:

load_on_devices=gpu4;gpu5
# or simply:
load_on_devices=4;5

Python model configuration

number of workers

For Python engine, we recommend set minWorkers and maxWorkers to be the same since python worker scale up and down is expensive.

You may also need to consider OMP_NUM_THREAD when setting number workers. OMP_NUM_THREAD is default to 1, you can unset OMP_NUM_THREAD by setting NO_OMP_NUM_THREADS=true. If OMP_NUM_THREAD is unset, the maxWorkers will be default to 2 (larger maxWorkers with non 1 OMP_NUM_THREAD can cause thread contention, and reduce throughput).

Set minimum workers and maximum workers for your model:

minWorkers=32
maxWorkers=64
# idle time in seconds before the worker thread is scaled down
max_idle_time=120

Or set minimum workers and maximum workers differently for GPU and CPU:

gpu.minWorkers=2
gpu.maxWorkers=3
cpu.minWorkers=2
cpu.maxWorkers=4

Note: Loading model in Python mode is pretty heavy. We recommend to set minWorker and maxWorker to be the same value to avoid unnecessary load and unload.

job queue size

Or override global job_queue_size:

job_queue_size=10

dynamic batching

To enable dynamic batching:

batch_size=2
max_batch_delay=1

rolling batch

To enable rolling batch for Python engine:

# lmi-dist requires running mpi mode
engine=MPI
option.rolling_batch=auto
# use FlashAttention
#option.rolling_batch=lmi-dist
#option.rolling_batch=scheduler
option.max_rolling_batch_size=64

Appendix

How to download uncompressed model from S3

To enable fast model downloading, you can store your model artifacts (weights) in a S3 bucket, and only keep the model code and metadata in the model.tar.gz (.zip) file. DJL can leverage s5cmd to download uncompressed files from S3 with extremely fast speed.

To enable s5cmd downloading, you can configure serving.properties as the following:

option.model_id=s3://YOUR_BUCKET/...

How to resolve python package conflict between models

If you want to deploy multiple python models, but their dependencies has conflict, you can enable python virtual environments for your model:

option.enable_venv=true