better default value

Author: tikikun

cpp/tensorrt_llm/nitro/controllers/tensorrtllm.cc

@@ -256,7 +256,6 @@ void tensorrtllm::chat_completion(
     samplingConfig.topP = std::vector{completion.top_p};
     samplingConfig.minLength = std::vector{outputLen};
     samplingConfig.repetitionPenalty = std::vector{completion.frequency_penalty};
-
     // Input preparation
 
     std::thread infThread(inferenceThread, inferState, inputIdsHost, callback, this,samplingConfig,inputLen,outputLen);

cpp/tensorrt_llm/nitro/models/chat_completion_request.h

@@ -21,11 +21,11 @@ inline inferences::ChatCompletionRequest fromRequest(const HttpRequest& req) {
   inferences::ChatCompletionRequest completion;
   if (jsonBody) {
     completion.stream = (*jsonBody).get("stream", false).asBool();
-    completion.max_tokens = (*jsonBody).get("max_tokens", 500).asInt();
+    completion.max_tokens = (*jsonBody).get("max_tokens", 2048).asInt();
     completion.top_p = (*jsonBody).get("top_p", 0.95).asFloat();
-    completion.temperature = (*jsonBody).get("temperature", 0.8).asFloat();
+    completion.temperature = (*jsonBody).get("temperature", 0.0f).asFloat();
     completion.frequency_penalty =
-        (*jsonBody).get("frequency_penalty", 0).asFloat();
+        (*jsonBody).get("frequency_penalty", 1.3).asFloat();
     completion.presence_penalty =
         (*jsonBody).get("presence_penalty", 0).asFloat();
     completion.messages = (*jsonBody)["messages"];