torch.cuda.OutOfMemoryError: CUDA out of memory.

code
import json
import sagemaker
import boto3
from sagemaker.huggingface import HuggingFaceModel, get_huggingface_llm_image_uri

try:
role = sagemaker.get_execution_role()
except ValueError:
iam = boto3.client('iam')
role = iam.get_role(RoleName='sagemaker_execution_role')['Role']['Arn']

Hub Model configuration. https://huggingface.co/models

hub = {
'HF_MODEL_ID':'deepseek-ai/DeepSeek-Coder-V2-Base',
'SM_NUM_GPUS': json.dumps(1)
}

create Hugging Face Model Class

huggingface_model = HuggingFaceModel(
image_uri=get_huggingface_llm_image_uri("huggingface",version="3.0.1"),
env=hub,
role=role,
)

deploy model to SageMaker Inference

predictor = huggingface_model.deploy(
initial_instance_count=1,
instance_type="ml.g5.2xlarge",
container_startup_health_check_timeout=300,
)

send request

predictor.predict({
"inputs": "Hi, what can you help me with?",
})