Update app.py

app.py

@@ -150,16 +150,17 @@ def to_wanted_embs(image_outputs, input_ids, attention_mask, cache_position=None
 
 @spaces.GPU()
 def generate_pali(user_emb):
-    prompt = 'caption en'
-    model_inputs = processor(text=prompt, images=torch.zeros(1, 3, 224, 224), return_tensors="pt")
-    # we need to get im_embs taken in here.
-    input_len = model_inputs["input_ids"].shape[-1]
-    input_embeds = to_wanted_embs(user_emb.squeeze()[None, None, :].repeat(1, 256, 1), 
-                        model_inputs["input_ids"].to(device), 
-                        model_inputs["attention_mask"].to(device))
-
-    generation = pali.generate(max_new_tokens=100, do_sample=True, top_p=.94, temperature=1.2, inputs_embeds=input_embeds)
-    decoded = processor.decode(generation[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
+    with torch.no_grad():
+        prompt = 'caption en'
+        model_inputs = processor(text=prompt, images=torch.zeros(1, 3, 224, 224), return_tensors="pt")
+        # we need to get im_embs taken in here.
+        input_len = model_inputs["input_ids"].shape[-1]
+        input_embeds = to_wanted_embs(user_emb.squeeze()[None, None, :].repeat(1, 256, 1), 
+                            model_inputs["input_ids"].to(device), 
+                            model_inputs["attention_mask"].to(device))
+    
+        generation = pali.generate(max_new_tokens=100, do_sample=True, top_p=.94, temperature=1.2, inputs_embeds=input_embeds)
+        decoded = processor.decode(generation[0], skip_special_tokens=True, clean_up_tokenization_spaces=True)
     return decoded