entrypoint.sh

#!/bin/bash

echo "MODEL_NAME: $MODEL_NAME"
echo "QUANTIZE: $QUANTIZATION"
echo "GRADIO_PORT: $GRADIO_PORT"

MODEL_NAME=$(echo $MODEL_NAME | sed 's/^[[:space:]]*//g;s/[[:space:]]*$//g')
QUANTIZATION=$(echo $QUANTIZATION | sed 's/^[[:space:]]*//g;s/[[:space:]]*$//g')
GRADIO_PORT=$(echo $GRADIO_PORT | sed 's/^[[:space:]]*//g;s/[[:space:]]*$//g')

if [[ "$QUANTIZATION" == "$QUANTIZATION" ]]; then
  echo "WTF1-1!"
else
  echo "WTF1-2!!"
fi

if [[ "$QUANTIZATION" == "false" ]]; then
  echo "WTF2-1!"
else
  echo "WTF2-2!!"
fi

if [[ "$QUANTIZATION" == false ]]; then
  echo "WTF3-1!"
  text-generation-launcher --model-id $MODEL_NAME --num-shard 1 --port 8080 --trust-remote-code &
else
  echo "WTF3-2!"
  text-generation-launcher --model-id $MODEL_NAME --num-shard 1 --port 8080 --trust-remote-code --quantize $QUANTIZATION &
fi

# Wait for text-generation-inference to start
curl --retry 60 --retry-delay 10 --retry-connrefused http://127.0.0.1:8080/health

# Start the gradio
python app/main.py --port $GRADIO_PORT &

# Wait for any process to exit
wait -n

# Exit with status of process that exited first
exit $?