feat:添加--mode启动参数

docker/Dockerfile

@@ -378,6 +378,7 @@ RUN printf '%s\n' \
     '  --llama-checkpoint-path "${LLAMA_CHECKPOINT_PATH}" \' \
     '  --decoder-checkpoint-path "${DECODER_CHECKPOINT_PATH}" \' \
     '  --decoder-config-name "${DECODER_CONFIG_NAME}" \' \
+    '  --mode "reduce-overhead" \' \
     '  ${DEVICE_ARGS} ${COMPILE_ARGS} ${HALF_ARGS}' \
     > /app/start_server.sh && chmod +x /app/start_server.sh
 

fish_speech/models/text2semantic/inference.py

@@ -541,6 +541,17 @@ def generate_long(
     assert 0 < top_p <= 1, "top_p must be in (0, 1]"
     assert 0 < temperature < 2, "temperature must be in (0, 2)"
 
+    logger.info(f"generate_long.param.device: {device}")
+    logger.info(f"generate_long.param.text: {text}")
+    logger.info(f"generate_long.param.max_new_tokens: {max_new_tokens}")
+    logger.info(f"generate_long.param.top_p: {top_p}")
+    logger.info(f"generate_long.param.top_k: {top_k}")
+    logger.info(f"generate_long.param.temperature: {temperature}")
+    logger.info(f"generate_long.param.compile: {compile}")
+    logger.info(f"generate_long.param.chunk_length: {chunk_length}")
+    logger.info(f"generate_long.param.prompt_text: {prompt_text}")
+    logger.info(f"generate_long.param.prompt_tokens: {prompt_tokens}")
+
     use_prompt = bool(prompt_text) and bool(prompt_tokens)
     if use_prompt and isinstance(prompt_text, str):
         prompt_text = [prompt_text]