fixed stream

app.py

@@ -75,6 +75,9 @@ def chat_fn(message, history, temperature, top_p, top_k):
         "<assistant>\n"
     )
 
+    # ★ ここが重要：累積して yield
+    output = ""
+
     for chunk in vlm_infer_stream(
         model=model_gpu,
         image_tensor=image_tensor,
@@ -84,13 +87,15 @@ def chat_fn(message, history, temperature, top_p, top_k):
         top_p=top_p if top_p > 0 else None,
         top_k=top_k if top_k > 0 else None,
     ):
-        yield chunk
+        output += chunk
+        yield output
 
     model_gpu.to("cpu")
     torch.cuda.empty_cache()
 
 
 
+
 # =====================================================
 # UI (ChatInterface, multimodal)
 # =====================================================