debug

app.py

@@ -333,10 +333,11 @@ def construction_layout():
     # quantizer = quantizer.to("cuda")
     # tokenizer = tokenizer.to("cuda")
     model.lm = model.lm.to("cuda")
+    print(model.lm.device)
     return model, quantizer, tokenizer, params_dict["width"], params_dict["height"], device
 
 @torch.no_grad()   
-@spaces.GPU(enable_queue=True, duration=60) 
+@spaces.GPU(duration=60) 
 def evaluate_v1(inputs, model, quantizer, tokenizer, width, height, device, do_sample=False, temperature=1.0, top_p=1.0, top_k=50):
     json_example = inputs
     input_intension = '{"wholecaption":"' + json_example["wholecaption"] + '","layout":[{"layer":'
@@ -344,14 +345,17 @@ def evaluate_v1(inputs, model, quantizer, tokenizer, width, height, device, do_s
     inputs = tokenizer(
         input_intension, return_tensors="pt"
     ).to(model.lm.device)
+    print(inputs.device)
     print("tokenizer2")
     
     stopping_criteria = StoppingCriteriaList()
     stopping_criteria.append(StopAtSpecificTokenCriteria(token_id_list=[128000]))
 
     print("lm1")
+    print(model.lm.device)
     outputs = model.lm.generate(**inputs, use_cache=True, max_length=8000, stopping_criteria=stopping_criteria, do_sample=do_sample, temperature=temperature, top_p=top_p, top_k=top_k)
     print("lm2")
+    
     inputs_length = inputs['input_ids'].shape[1] 
     outputs = outputs[:, inputs_length:]
     
@@ -427,7 +431,7 @@ def construction():
     
     return pipeline, transp_vae
 
-@spaces.GPU(enable_queue=True, duration=60)
+@spaces.GPU(duration=60)
 def test_one_sample(validation_box, validation_prompt, true_gs, inference_steps, pipeline, generator, transp_vae):
     print(validation_box)
     output, rgba_output, _, _ = pipeline(
@@ -474,7 +478,17 @@ def svg_test_one_sample(validation_prompt, validation_box_str, seed, true_gs, in
     svg_file_path = './image.svg'
     os.makedirs(os.path.dirname(svg_file_path), exist_ok=True)
     with open(svg_file_path, 'w', encoding='utf-8') as f:
-        f.write(svg_img)       
+        f.write(svg_img)    
+        
+    if not isinstance(result_images, list):
+        raise TypeError("result_images 必须是一个列表")
+    else:
+        print(len(result_images))
+    
+    if not os.path.exists(svg_file_path):
+        raise FileNotFoundError(f"文件 {svg_file_path} 未创建")
+    if os.path.getsize(svg_file_path) == 0:
+        raise ValueError(f"文件 {svg_file_path} 内容为空")
     
     return result_images, svg_file_path
     

modeling_crello.py

@@ -196,6 +196,7 @@ class CrelloModel(PreTrainedModel):
     self,
     labels: torch.LongTensor,
   ):
+    print("inside Crello")
     batch_size = labels.shape[0]
     full_labels = labels.detach().clone()
 
@@ -219,10 +220,12 @@ class CrelloModel(PreTrainedModel):
           pad_idx.append(k + 1)
     assert len(pad_idx) == batch_size, (len(pad_idx), batch_size)
     
+    print("inside Crello, lm1")
     output = self.lm( inputs_embeds=input_embs,
                       # input_ids=labels,
                       labels=full_labels,
                       output_hidden_states=True)
+    print("inside Crello, lm2")
     
     return output, full_labels, input_embs_norm