feat:添加workers

.env

@@ -1,4 +1,4 @@
 API_PORT=8080
 COMPILE=1
 HALF=1
-PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:64,expandable_segments:True

docker/Dockerfile

@@ -379,6 +379,7 @@ RUN printf '%s\n' \
     '  --llama-checkpoint-path "${LLAMA_CHECKPOINT_PATH}" \' \
     '  --decoder-checkpoint-path "${DECODER_CHECKPOINT_PATH}" \' \
     '  --decoder-config-name "${DECODER_CONFIG_NAME}" \' \
+    '  --workers 3 \' \
     '  ${DEVICE_ARGS} ${COMPILE_ARGS} ${HALF_ARGS}' \
     > /app/start_server.sh && chmod +x /app/start_server.sh