Upload app.py with huggingface_hub

app.py

@@ -137,8 +137,18 @@ class StopOnEvent(StoppingCriteria):
 
 def gpu_duration(message, history, system_prompt=GENERAL, thinking="none", max_new_tokens=512,
                  temperature=1.0, top_p=0.95, *args, **kwargs) -> int:
-    """Initial bounded reservation; tune after measuring live requests."""
-    return min(120, 30 + int(max(64, min(2048, int(max_new_tokens)))) // 25)
+    """Budget from live ~32 tokens/s plus a margin for input processing.
+
+    The 2026-10-26 smoke tests took 1.22, 1.90 and 7.94 seconds. The linear
+    output allowance adds ~40% to the longest test's rate; long inputs get
+    an extra prefill margin. This estimates a reservation, not a speed promise.
+    """
+    from math import ceil
+
+    limit = max(64, min(2048, int(max_new_tokens)))
+    batch, _ = build_inputs(plain_text(message).strip(), history, system_prompt, thinking)
+    input_tokens = int(batch["input_ids"].shape[-1])
+    return min(120, max(15, ceil(5 + limit * 0.045 + input_tokens * 0.003)))
 
 
 @spaces.GPU(duration=gpu_duration)
@@ -168,7 +178,7 @@ def chat(message: str, history: list, system_prompt: str = GENERAL, thinking: st
     stop = threading.Event()
     streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=False, timeout=5)
     outcome = {}
-    budget = gpu_duration(message, history, system_prompt, thinking, max_new_tokens, temperature, top_p) - 8
+    budget = gpu_duration(message, history, system_prompt, thinking, max_new_tokens, temperature, top_p) - 5
     started = time.perf_counter()
 
     def generate():
@@ -278,7 +288,7 @@ with gr.Blocks(title="Spark 한국어 대화 실험실") as demo:
         with gr.Column(scale=3):
             chatbot = gr.Chatbot(height=470, show_label=False, placeholder="예시를 고르거나 메시지를 입력하세요.")
             with gr.Row():
-                message = gr.Textbox(placeholder="한국어로 말을 걸어보세요", show_label=False, lines=2, scale=6)
+                message = gr.Textbox(label="메시지", placeholder="한국어로 말을 걸어보세요", show_label=False, lines=2, scale=6)
                 send = gr.Button("보내기", variant="primary", scale=1)
             with gr.Row():
                 stop_button = gr.Button("중지")