Update app.py

app.py

@@ -421,7 +421,7 @@ def handsome_chat_completions():
         return jsonify({"error": "Invalid request data"}), 400
 
     model_name = data['model']
-
+    
     api_key = select_key(model_name)
 
     if not api_key:
@@ -434,7 +434,7 @@ def handsome_chat_completions():
                 )
             }
         ), 429
-
+    
     if model_name == "deepseek-reasoner":
         for param in ["temperature", "top_p", "presence_penalty", "frequency_penalty", "logprobs", "top_logprobs"]:
             if param in data:
@@ -444,7 +444,7 @@ def handsome_chat_completions():
         "Authorization": f"Bearer {api_key}",
         "Content-Type": "application/json"
     }
-
+    
     try:
         start_time = time.time()
         response = requests.post(
@@ -479,7 +479,6 @@ def handsome_chat_completions():
                 prompt_tokens = 0
                 completion_tokens = 0
                 response_content = ""
-                formatted_chunks = [] # Collect formatted chunks for final response
                 for line in full_response_content.splitlines():
                     if line.startswith("data:"):
                         line = line[5:].strip()
@@ -496,24 +495,17 @@ def handsome_chat_completions():
                                     "usage"
                                 ]["completion_tokens"]
 
-                            if "choices" in response_json and len(response_json["choices"]) > 0:
-                                delta = response_json["choices"][0].get("delta", {})
-                                if "reasoning_content" in delta and delta["reasoning_content"]:
-                                    reasoning_lines = delta["reasoning_content"].splitlines()
-                                    formatted_reasoning = "\n".join(f"> {line}" for line in reasoning_lines)
-                                    response_content += formatted_reasoning
-                                    formatted_chunks.append({
-                                        "choices": [
-                                            {"delta": {"content": formatted_reasoning}, "index": 0, "finish_reason": None}
-                                        ]
-                                    })
-                                if "content" in delta and delta["content"]:
-                                    response_content += delta["content"]
-                                    formatted_chunks.append({
-                                        "choices": [
-                                            {"delta": {"content": delta["content"]}, "index": 0, "finish_reason": None}
-                                        ]
-                                    })
+                            if (
+                                "choices" in response_json and
+                                len(response_json["choices"]) > 0 and
+                                "delta" in response_json["choices"][0] and
+                                "content" in response_json[
+                                    "choices"
+                                ][0]["delta"]
+                            ):
+                                response_content += response_json[
+                                    "choices"
+                                ][0]["delta"]["content"]
 
                             if (
                                 "usage" in response_json and
@@ -574,18 +566,10 @@ def handsome_chat_completions():
                     request_timestamps.append(time.time())
                     token_counts.append(prompt_tokens+completion_tokens)
 
-                # Generate the final stream response
-                def generate_formatted_stream():
-                    for chunk in formatted_chunks:
-                        yield f"data: {json.dumps(chunk)}\n\n".encode('utf-8')
-                    yield "data: [DONE]\n\n".encode('utf-8')
-
-
-                return Response(
-                    stream_with_context(generate_formatted_stream()),
-                    content_type="text/event-stream"
-                )
-
+            return Response(
+                stream_with_context(generate()),
+                content_type=response.headers['Content-Type']
+            )
         else:
             response.raise_for_status()
             end_time = time.time()