Update app.py

app.py

@@ -46,7 +46,6 @@ tokenizer_info = xgr.TokenizerInfo.from_huggingface(
 
 grammar_compiler = xgr.GrammarCompiler(tokenizer_info)
 compiled_grammar = grammar_compiler.compile_json_schema(Person)
-xgr_logits_processor = xgr.contrib.hf.LogitsProcessor(compiled_grammar)
 
 default_value = "A 57-year-old male presented with fever (38.9°C), chest pain, cough, and progressive dyspnea. The patient exhibited tachypnea (34 breaths/min) and tachycardia (134 bpm). Auscultation revealed decreased breath sounds in both lung bases, with crackles on the left. A chest X-ray revealed bilateral pleural opacities and enlargement of the cardiac silhouette ( A). Echocardiography showed moderate pericardial effusion affecting the entire cardiac silhouette. Pericardiocentesis yielded 250 mL of exudative fluid. A CT scan of the chest showed pneumonia in the left lower lobe, bilateral pleural effusion, and moderate pericardial effusion ( B). Thoracentesis was performed and yielded 1,050 mL of exudative fluid. Laboratory tests yielded the following data: white blood cell count, 11.78 × 109 cells/L (84.3% neutrophils, 4.3% lymphocytes, and 9.1% monocytes); platelet count, 512 × 109/L; serum C-reactive protein, 31.27 mg/dL; serum creatinine, 0.94 mg/dL; serum sodium, 133 mEq/L; and serum potassium, 3.72 mEq/L. Examination of the pleural fluid showed a pH of 7.16, a glucose level of 4.5 mg/dL, proteins at 49.1 g/L, and an LDH content of 1,385 U/L. A urinary pneumococcal antigen test was positive. Pleural fluid culture was positive for S. pneumoniae. The patient was treated for four weeks with amoxicillin-clavulanate (2.2 g/8 h, i.v.) plus levofloxacin (500 mg twice a day), together with a nonsteroidal anti-inflammatory drug (ibuprofen, 800 mg/day), after which there was nearly complete resolution of the alterations seen on the chest X-ray and CT scan."
 
@@ -198,6 +197,10 @@ def summarize(
             add_generation_prompt=True,
             enable_thinking=False, # only relevant for qwen
         )
+
+        # We cannot reset here because __call__ is not invoked when stop token is sampled.
+        # Therefore, each `generate()` call needs to instantiate an LogitsProcessor.
+        xgr_logits_processor = xgr.contrib.hf.LogitsProcessor(compiled_grammar)
     
         model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
         generated_ids = model.generate(