Update entrypoint.sh.template

entrypoint.sh.template

@@ -1,11 +1,20 @@
 #!/bin/bash
 
 if [[ "$QUANTIZATION" == "false" ]]; then 
-  echo "WTF3-1!" 
-  text-generation-launcher --model-id $MODEL_NAME --num-shard 1 --port 8080 --trust-remote-code & 
+  text-generation-launcher --model-id $MODEL_NAME \
+    --num-shard 1 --port 8080 --trust-remote-code \
+    --max-concurrent-request $MAX_CONCURRENT_REQUESTS \
+    --max-top-n-tokens $MAX_INPUT_LENGTH \
+    --max-input-length $MAX_TOTAL_TOKENS \
+    & 
 else 
-  echo "WTF3-2!" 
-  text-generation-launcher --model-id $MODEL_NAME --num-shard 1 --port 8080 --trust-remote-code --quantize $QUANTIZATION & 
+  text-generation-launcher --model-id $MODEL_NAME \
+    --num-shard 1 --port 8080 --trust-remote-code \
+    --max-concurrent-request $MAX_CONCURRENT_REQUESTS \
+    --max-top-n-tokens $MAX_INPUT_LENGTH \
+    --max-input-length $MAX_TOTAL_TOKENS \
+    --quantize $QUANTIZATION \
+    & 
 fi
 
 # Wait for text-generation-inference to start