Triton Server CUDA_ERROR_ILLEGAL_ADDRESS: an illegal memory access was encountered

Triton Inference Server Image: nvcr.io/nvidia/tritonserver:20.03.1-py3
GPU: T4
OS: Ubuntu 18.04
Driver: 440.33.01
CUDA version: 10.1, cudnn 7.6.5
Tensorflow: 2.2

Hi, I am using Triton inference server for doing dynamic batching based inference.
Following is my config.pbtxt file -

name: “effdet”
platform: “tensorflow_savedmodel”
max_batch_size: 64

input {
name: “image_arrays:0”
data_type: TYPE_UINT8
dims: -1
dims: -1
dims: -1
}

output {
name: “detections:0”
data_type: TYPE_FP32
dims: 100
dims: 7
}

instance_group [
{
count: 1
kind: KIND_GPU
}
]

dynamic_batching {
preferred_batch_size: [8, 64]
max_queue_delay_microseconds: 100
}

default_model_filename: “model.savedmodel”

I am successfully able to run the inferrence using a single instance of my effdet model.
However changing number of model instances to anything other than 1, i.e

instance_group [
{
count: 2 #[or 3,4 etc]
kind: KIND_GPU
}
]
leads to an error CUDA_ERROR_ILLEGAL_ADDRESS: an illegal memory access was encountered.

Refering to the screenshot, can someone please assist me with a solution.

Thank you,
Shubham