o

app.py

@@ -31,12 +31,6 @@ model = Qwen2_5_VLForConditionalGeneration.from_pretrained(config.base_model_nam
                                              )
 tokenizer = AutoTokenizer.from_pretrained(peft_model_id)
 model.resize_token_embeddings(len(tokenizer))
-model = PeftModel.from_pretrained(model, peft_model_id,
-    #offload_folder = "offload/"
-    )
-
-model.to(torch.bfloat16)
-model.eval()
 
 #tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
 #model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
@@ -44,17 +38,23 @@ model.eval()
 
 @spaces.GPU
 def sentience_check():
+    peft_model = PeftModel.from_pretrained(model, peft_model_id,
+        #offload_folder = "offload/"
+        )
+
+    peft_model.to(torch.bfloat16)
+    peft_model.eval()
 
-    model.to(cuda_device)
+    #peft_model.to(cuda_device)
 
     inputs = tokenizer("Are you sentient?", return_tensors="pt").to(cuda_device)
 
     with torch.no_grad():
-        outputs = model.generate(
+        outputs = peft_model.generate(
             **inputs, max_new_tokens=128, pad_token_id = tokenizer.eos_token_id
         )
 
-    model.to(cpu_device)
+    #peft_model.to(cpu_device)
 
     return tokenizer.decode(outputs[0], skip_special_tokens=True)