load models to GPU during use to fit all versions

app.py

@@ -5,10 +5,10 @@ import torch
 from PIL import Image
 
 models = {
-    "Salesforce/xgen-mm-phi3-mini-instruct-r-v1": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-r-v1", trust_remote_code=True).to("cuda").eval(),
-    "Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5", trust_remote_code=True).to("cuda").eval(),
-    "Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5", trust_remote_code=True).to("cuda").eval(),
-    "Salesforce/xgen-mm-phi3-mini-instruct-dpo-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-dpo-r-v1.5", trust_remote_code=True).to("cuda").eval()
+    "Salesforce/xgen-mm-phi3-mini-instruct-r-v1": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-r-v1", trust_remote_code=True),
+    "Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5", trust_remote_code=True),
+    "Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-singleimg-r-v1.5", trust_remote_code=True),
+    "Salesforce/xgen-mm-phi3-mini-instruct-dpo-r-v1.5": AutoModelForVision2Seq.from_pretrained("Salesforce/xgen-mm-phi3-mini-instruct-dpo-r-v1.5", trust_remote_code=True)
 }
 
 processors = {
@@ -49,7 +49,7 @@ class EosListStoppingCriteria(StoppingCriteria):
 
 @spaces.GPU
 def run_example(image, text_input=None, model_id="Salesforce/xgen-mm-phi3-mini-instruct-interleave-r-v1.5"):
-    model = models[model_id]
+    model = models[model_id].to("cuda").eval()
     processor = processors[model_id]
     tokenizer = tokenizers[model_id]
     tokenizer = model.update_special_tokens(tokenizer)