feat: binary mm input compatible with stream mode. (#615)

Author: xiao-yu-chen

xllm/api_service/call.cpp

@@ -35,17 +35,17 @@ void Call::init() {
     x_request_time_ =
         *controller_->http_request().GetHeader("x-request-timems");
   }
-}
 
-bool Call::get_binary_payload(std::string& payload) {
-  payload.clear();
+  init_request_payload();
+}
 
+void Call::init_request_payload() {
   const auto infer_content_len =
       controller_->http_request().GetHeader(kInferContentLength);
   const auto content_len =
       controller_->http_request().GetHeader(kContentLength);
 
-  if (infer_content_len == nullptr || content_len == nullptr) return false;
+  if (infer_content_len == nullptr || content_len == nullptr) return;
 
   auto infer_len = std::stoul(*infer_content_len);
   auto len = std::stoul(*content_len);
@@ -54,12 +54,11 @@ bool Call::get_binary_payload(std::string& payload) {
     LOG(ERROR) << " content length is invalid:"
                << " infer content len is " << infer_len
                << " , content length is " << len;
-    return false;
+    return;
   }
 
   controller_->request_attachment().copy_to(
-      &payload, len - infer_len, infer_len);
-  return true;
+      &request_payload_, len - infer_len, infer_len);
 }
 
 }  // namespace xllm

xllm/api_service/call.h

@@ -29,7 +29,8 @@ class Call {
   std::string get_x_request_id() { return x_request_id_; }
   std::string get_x_request_time() { return x_request_time_; }
 
-  bool get_binary_payload(std::string& payload);
+  std::string& get_request_payload() { return request_payload_; }
+  void init_request_payload();
 
   virtual bool is_disconnected() const = 0;
 
@@ -41,6 +42,8 @@ class Call {
 
   std::string x_request_id_;
   std::string x_request_time_;
+
+  std::string request_payload_;
 };
 
 }  // namespace xllm

xllm/api_service/chat_service_impl.cpp

@@ -751,8 +751,7 @@ void MMChatServiceImpl::process_async_impl(std::shared_ptr<MMChatCall> call) {
   auto saved_streaming = request_params.streaming;
   auto saved_request_id = request_params.request_id;
 
-  std::string payload;
-  call->get_binary_payload(payload);
+  auto& payload = call->get_request_payload();
 
   // schedule the request
   master_->handle_request(

xllm/api_service/embedding_service_impl.cpp

@@ -163,8 +163,7 @@ void MMEmbeddingServiceImpl::process_async_impl(
   }
   auto request_id = request_params.request_id;
 
-  std::string payload;
-  call->get_binary_payload(payload);
+  auto& payload = call->get_request_payload();
 
   // schedule the request
   master_->handle_request(