Upload app.py

app.py

@@ -354,16 +354,17 @@ def generate_text():
     prompt = data['prompt']
     max_new_tokens = data.get('max_new_tokens', 256)
     temperature = data.get('temperature', 0.0)
-    system_prompt = data.get('system_prompt', 'You are an expert assistant.')
+    system_prompt = data.get('system_prompt', 'You are an expert real-time translation assistant.')
     messages = [
         {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt}
     ]
+    # enable_thinking=False for ALL routes to prevent Qwen3 from leaking internal monologue
     encoded = tokenizer.apply_chat_template(
         messages,
         add_generation_prompt=True,
         tokenize=True,
-        enable_thinking=False if IS_DIFFUSION else None
+        enable_thinking=False
     )
     if IS_DIFFUSION:
         input_ids = torch.tensor([encoded], dtype=torch.long, device=device)
@@ -403,14 +404,14 @@ def generate_text_stream():
     prompt = data['prompt']
     max_new_tokens = data.get('max_new_tokens', 256)
     temperature = data.get('temperature', 0.0)
-    system_prompt = data.get('system_prompt', 'You are an expert assistant.')
+    system_prompt = data.get('system_prompt', 'You are an expert real-time translation assistant.')
     messages = [
         {"role": "system", "content": system_prompt},
         {"role": "user", "content": prompt}
     ]
     encoded = tokenizer.apply_chat_template(
         messages, add_generation_prompt=True, tokenize=True,
-        enable_thinking=False if IS_DIFFUSION else None
+        enable_thinking=False
     )
     if IS_DIFFUSION:
         input_ids = torch.tensor([encoded], dtype=torch.long, device=device)
@@ -463,7 +464,7 @@ def generate_text_sse():
     ]
     encoded = tokenizer.apply_chat_template(
         messages, add_generation_prompt=True, tokenize=True,
-        enable_thinking=False if IS_DIFFUSION else None
+        enable_thinking=False
     )
     input_ids = torch.tensor([encoded], dtype=torch.long, device=device)
 
@@ -480,7 +481,9 @@ def generate_text_sse():
                 temperature=temperature, cfg_scale=cfg_scale, remasking=remasking,
                 capture_interval=capture_interval
             ):
-                yield f"data: {json.dumps(state)}\n\n"
+                yield f"data: {json.dumps(state)}
+
+"
         else:
             streamer = TextIteratorStreamer(
                 tokenizer, skip_prompt=True, skip_special_tokens=True
@@ -493,14 +496,25 @@ def generate_text_sse():
                 do_sample=True if temperature > 0 else False,
                 pad_token_id=tokenizer.eos_token_id,
             )
+
             def _generate():
                 with torch.inference_mode():
                     model.generate(**generation_kwargs)
+
             thread = Thread(target=_generate)
             thread.start()
+
+            accumulated = []
             for text in streamer:
-                yield f"data: {json.dumps({'type': 'token', 'text': text})}\n\n"
-            yield f"data: {json.dumps({'type': 'final', 'text': '', 'total_steps': 1})}\n\n"
+                accumulated.append(text)
+                yield f"data: {json.dumps({'type': 'token', 'text': text})}
+
+"
+
+            full_text = "".join(accumulated)
+            yield f"data: {json.dumps({'type': 'final', 'text': full_text, 'total_steps': 1})}
+
+"
 
     return Response(
         stream(), mimetype='text/event-stream',