Fix blank screen during generation, stream thinking tokens, Enter to send, lower GPU duration

app.py

@@ -123,6 +123,19 @@ HEADER = """
 </div>
 """
 
+HEAD = """
+<script>
+document.addEventListener("keydown", (e) => {
+  if (e.key !== "Enter" || e.shiftKey || e.isComposing) return;
+  const ta = document.querySelector("#twil-input textarea");
+  if (!ta || e.target !== ta) return;
+  e.preventDefault();
+  e.stopPropagation();
+  document.querySelector("#send-btn")?.click();
+}, true);
+</script>
+"""
+
 PLACEHOLDER = """
 <div class="twil-empty">
   <img class="twil-cube" src="/gradio_api/file=brand/webai-cube-256.webp" alt="" />
@@ -155,12 +168,18 @@ CSS = """
 
 * { box-shadow: none !important; text-shadow: none !important; }
 html, body, #root, .gradio-container, .gradio-container > .main,
-.gradio-container .contain, .fillable, .gradio-container .wrap,
+.gradio-container .contain, .fillable,
 .gradio-container .column, .gradio-container .row, .contain, .wrapper {
   background: var(--weba-canvas) !important;
   box-shadow: none !important;
   filter: none !important;
 }
+/* Never paint over content while a job is running: the status tracker
+   overlay must stay transparent or the chat looks blank mid-generation. */
+.gradio-container [data-testid="status-tracker"],
+.gradio-container .wrap.default {
+  background: transparent !important;
+}
 html, body, .gradio-container, .gradio-container > .main, .fillable,
 .contain, .app {
   height: 100% !important;
@@ -651,6 +670,9 @@ def _parse_think(raw: str) -> tuple[str, str, bool]:
     """Return (think, answer, think_open) from a possibly streaming reply."""
     open_idx = raw.find(THINK_OPEN)
     if open_idx == -1:
+        # A partially streamed opening tag is not an answer yet.
+        if raw.strip() and THINK_OPEN.startswith(raw.strip()):
+            return "", "", False
         return "", raw, False
     rest = raw[open_idx + len(THINK_OPEN) :]
     close_idx = rest.find(THINK_CLOSE)
@@ -662,9 +684,11 @@ def _parse_think(raw: str) -> tuple[str, str, bool]:
     return think, answer, False
 
 
-def _gpu_seconds(history, max_new_tokens=2048, *args, **kwargs):
-    tokens = int(max_new_tokens or 2048)
-    return min(180, max(45, 25 + tokens // 18))
+def _gpu_seconds(history, max_new_tokens=1024, *args, **kwargs):
+    # ZeroGPU adds its own startup overhead on top of this request, and
+    # anonymous visitors have a small quota — keep the ask modest.
+    tokens = int(max_new_tokens or 1024)
+    return min(90, max(25, 10 + tokens // 30))
 
 
 def _history_for_model(history: list) -> list[dict]:
@@ -685,15 +709,20 @@ def _history_for_model(history: list) -> list[dict]:
 
 def _stream_tokens(conversation: list[dict], max_new_tokens: int, temperature: float, top_p: float, enable_thinking: bool) -> Iterator[str]:
     if PREVIEW:
+        import time
+
         demo = (
             f"{THINK_OPEN}\n"
-            "Check the premises, then the conclusion. The argument is a classic Barbara syllogism.\n"
+            "Check the premises, then the conclusion. Premise 1 gives Rain -> Wet. "
+            "Premise 2 observes Wet. Inferring Rain from Wet affirms the consequent, "
+            "which is invalid. The argument is a classic Barbara syllogism otherwise.\n"
             f"{THINK_CLOSE}\n\n"
             "Answer: entailment."
         )
         acc = ""
         for ch in demo:
             acc += ch
+            time.sleep(0.02)
             yield acc
         return
 
@@ -801,6 +830,10 @@ def generate_reply(history, max_new_tokens, temperature, top_p, enable_thinking)
             answer_msg["content"] = raw.strip()
             if not any(m is answer_msg for m in history):
                 history.append(answer_msg)
+        elif any(m is answer_msg for m in history):
+            # A partial tag was mistaken for an answer earlier; retract it.
+            history.remove(answer_msg)
+            answer_msg["content"] = ""
         yield history
 
     if think_msg["content"]:
@@ -849,7 +882,7 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
                     )
                     send = gr.Button("↑", elem_id="send-btn", scale=0)
                 with gr.Accordion("Parameters", open=False, elem_id="params-box"):
-                    max_new_tokens = gr.Slider(256, 4096, value=2048, step=256, label="Max new tokens")
+                    max_new_tokens = gr.Slider(256, 4096, value=1024, step=256, label="Max new tokens")
                     temperature = gr.Slider(0, 1.5, value=0, step=0.05, label="Temperature (0 = greedy)")
                     top_p = gr.Slider(0.1, 1.0, value=0.95, step=0.05, label="Top-p")
                     enable_thinking = gr.Checkbox(value=True, label="Enable thinking")
@@ -876,7 +909,7 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
         [chatbot, max_new_tokens, temperature, top_p, enable_thinking],
         chatbot,
         concurrency_limit=1,
-        show_progress="minimal",
+        show_progress="hidden",
     )
     prompt.submit(
         queue_message,
@@ -888,7 +921,7 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
         [chatbot, max_new_tokens, temperature, top_p, enable_thinking],
         chatbot,
         concurrency_limit=1,
-        show_progress="minimal",
+        show_progress="hidden",
     )
     for btn, (_, template) in zip(chip_btns, CHIPS):
         btn.click(lambda t=template: t, outputs=prompt)
@@ -908,13 +941,14 @@ with gr.Blocks(fill_height=True, fill_width=True, elem_id="app-root") as demo:
             [chatbot, max_new_tokens, temperature, top_p, enable_thinking],
             chatbot,
             concurrency_limit=1,
-            show_progress="minimal",
+            show_progress="hidden",
         )
 
 if __name__ == "__main__":
     demo.launch(
         theme=THEME,
         css=CSS,
+        head=HEAD,
         allowed_paths=[str(ASSETS), str(ASSETS.parent)],
         ssr_mode=False,
     )