Upload app.py

app.py

@@ -426,6 +426,9 @@ def is_research_paper(text, use_ai=True):
         else:
             return False, 0, f"Does not appear to be a research paper. {rule_reason}"
 
+# Rest of the code remains exactly the same...
+# (create_messages, convert_file, process_file, generate functions remain unchanged)
+
 def create_messages(review_fields, paper_text):
     messages = [
         {"role": "system", "content": SYSTEM_PROMPT_TEMPLATE.format(review_fields=review_fields)},
@@ -433,8 +436,7 @@ def create_messages(review_fields, paper_text):
     ]
     return messages
 
-# 降低 GPU 时长上限，避免超配额
-@spaces.GPU(duration=60)
+@spaces.GPU()
 def convert_file(filepath):
     full_text, images, out_metadata = convert_single_pdf(
             filepath, model_refs, metadata=metadata, max_pages=MAX_PAGES
@@ -476,20 +478,13 @@ def process_file(file):
     
     return paper_text, True
 
-# 将 GPU 时长从 190 -> 60；并做轻量限流（截断 + 减少新token）
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=190)
 def generate(paper_text, review_template):
     # Final check before generation
     is_paper, confidence, reason = is_research_paper(paper_text, use_ai=False)  # Quick check without AI
     if not is_paper:
         return f"⚠️ Cannot generate review: {reason}"
-
-    # 软截断，避免一次生成过长导致超时（可按需调整）
-    # 说明：Llama-8B 在 bfloat16 下 2k~4k new tokens 已较长，这里对输入做保守控制。
-    max_chars_in = 18000
-    if isinstance(paper_text, str) and len(paper_text) > max_chars_in:
-        paper_text = paper_text[:max_chars_in] + "\n\n...[truncated for inference]...\n"
-
+    
     messages = create_messages(review_template, paper_text)
     input_ids = tokenizer.apply_chat_template(
         messages,
@@ -505,7 +500,7 @@ def generate(paper_text, review_template):
         input_ids=input_ids,
         attention_mask=attention_mask,
         streamer=streamer,
-        max_new_tokens=2048,      # 从 4096 降到 2048（必要时继续下调）
+        max_new_tokens=4096,
         do_sample=True,
         temperature=0.6,
         top_p=0.9,
@@ -514,13 +509,11 @@ def generate(paper_text, review_template):
     thread = Thread(target=model.generate, kwargs=generation_kwargs)
     thread.start()
     generated_text = ""
-    try:
-        for new_text in streamer:
-            generated_text += new_text
-            yield generated_text.replace("<|eot_id|>", "")
-    except Exception as e:
-        yield (generated_text + f"\n\n\n⚠️ Generation interrupted: {type(e).__name__}: {e}").strip()
+    for new_text in streamer:
+        generated_text += new_text
+        yield generated_text.replace("<|eot_id|>", "")
 
+# UI code remains exactly the same...
 # ui
 title = """<h1 align="center">OpenReviewer</h1>
 <div align="center">Using <a href="https://huggingface.co/maxidl/Llama-OpenReviewer-8B" target="_blank"><code>Llama-OpenReviewer-8B</code></a> - Built with Llama</div>
@@ -592,7 +585,6 @@ with gr.Blocks(theme=theme) as demo:
         else:
             return gr.update(interactive=False), gr.update(value=f"❌ **Not a research paper**: {reason}", visible=True)
     
-    # 事件保持不变（API 名沿用 /handle_file_upload、/handle_text_change、/generate）
     file_input.upload(handle_file_upload, file_input, [paper_text_field, validation_status, generate_button])
     paper_text_field.change(handle_text_change, paper_text_field, [generate_button, validation_status])
     
@@ -610,11 +602,8 @@ with gr.Blocks(theme=theme) as demo:
         inputs=None, 
         outputs=generate_button
     )
-
+    
     demo.title = "OpenReviewer"
 
-# 限制并发，避免累计占用导致单次执行拉长
-demo.queue(concurrency_count=1)
-
 if __name__ == "__main__":
     demo.launch()