fixed button bugs

app_rag.py

@@ -95,28 +95,32 @@ def _request_stop():
             event.set()
 
 
-def _on_stop_clicked():
+def _on_stop_clicked(use_rag: bool):
     _request_stop()
+    loaded = _is_rag_runtime_loaded() if use_rag else _RUNTIME_LOADED
+    stage = "Stop requested (RAG)" if use_rag else "Stop requested"
     return _format_status(
-        stage="Stop requested",
-        loaded=_RUNTIME_LOADED,
+        stage=stage,
+        loaded=loaded,
         device="unknown",
         loading_percent="--",
         error="Stop requested. Waiting for backend generation to halt.",
     )
 
 
-def _on_clear_clicked():
+def _on_clear_clicked(current_use_rag: bool):
     # Clear should also stop any in-flight generation to avoid concurrent
     # updates from the stream generator after UI has been reset.
     _request_stop()
+    loaded = _is_rag_runtime_loaded() if current_use_rag else _RUNTIME_LOADED
+    stage = "Idle (RAG mode)" if current_use_rag else "Idle"
     return (
         DEFAULT_INSTRUCTION,
         "",
-        False,
+        gr.update(),
         "",
-        _format_status(stage="Idle", loaded=_RUNTIME_LOADED, device="unknown", loading_percent="0%"),
-        _format_mode_indicator(False),
+        _format_status(stage=stage, loaded=loaded, device="unknown", loading_percent="0%"),
+        _format_mode_indicator(current_use_rag),
         "0.00s",
     )
 
@@ -187,7 +191,6 @@ def _load_demo_examples():
                 "instruction": instruction,
                 "input": user_input,
                 "max_new_tokens": max_new_tokens,
-                "use_rag": False,
             }
         )
 
@@ -208,11 +211,11 @@ def _load_demo_examples():
     return parsed, None
 
 
-def _apply_example(example: dict):
+def _apply_example(example: dict, current_use_rag: bool):
     max_tokens_update = (
         example["max_new_tokens"] if example.get("max_new_tokens") is not None else gr.update()
     )
-    use_rag = bool(example.get("use_rag", False))
+    use_rag = bool(example["use_rag"]) if "use_rag" in example else bool(current_use_rag)
     return example["instruction"], example["input"], max_tokens_update, use_rag, _format_mode_indicator(use_rag)
 
 
@@ -757,8 +760,8 @@ with gr.Blocks(title="Numen Scriptorium Demo") as demo:
                 for example in parsed_examples:
                     example_btn = gr.Button(example["label"], variant="secondary")
                     example_btn.click(
-                        fn=lambda ex=example: _apply_example(ex),
-                        inputs=None,
+                        fn=lambda current_mode, ex=example: _apply_example(ex, current_mode),
+                        inputs=[use_rag],
                         outputs=[instruction, user_input, max_new_tokens, use_rag, mode_panel],
                     )
 
@@ -783,11 +786,11 @@ with gr.Blocks(title="Numen Scriptorium Demo") as demo:
         outputs=[output, status_panel, elapsed_text],
     )
 
-    stop_btn.click(fn=_on_stop_clicked, inputs=None, outputs=[status_panel], cancels=[run_event])
+    stop_btn.click(fn=_on_stop_clicked, inputs=[use_rag], outputs=[status_panel], cancels=[run_event])
 
     clear_btn.click(
         fn=_on_clear_clicked,
-        inputs=None,
+        inputs=[use_rag],
         outputs=[instruction, user_input, use_rag, output, status_panel, mode_panel, elapsed_text],
         cancels=[run_event],
     )

src/numen_scriptorium/inference/qwen.py

@@ -110,6 +110,7 @@ def generate(
             temperature=temperature,
             top_p=top_p,
             eos_token_id=tokenizer.eos_token_id,
+            generator=generator,
         )
     text = tokenizer.decode(outputs[0], skip_special_tokens=True)
     if "回答：" in text:
@@ -157,6 +158,8 @@ def stream_generate(
         eos_token_id=tokenizer.eos_token_id,
         streamer=streamer,
     )
+    if generator is not None:
+        generate_kwargs["generator"] = generator
     if stop_event is not None:
         generate_kwargs["stopping_criteria"] = StoppingCriteriaList([_EventStoppingCriteria(stop_event)])