Update app.py

app.py

@@ -61,15 +61,16 @@ audio_pipe = pipeline(
 
 
 def process_text(text, count):
-    inputs = text_tokenizer(text, return_tensors="pt",
-                            return_attention_mask=False)
+    inputs = text_tokenizer.encode(text, return_tensors="pt")
+    input_embeds = tuned_phi2.get_submodule(
+            'model.embed_tokens')(inputs).to(device)
     prediction = text_tokenizer.batch_decode(
-        base_phi2_text.generate(
-            **inputs,
-            max_new_tokens=count,
-            bos_token_id=text_tokenizer.bos_token_id,
-            eos_token_id=text_tokenizer.eos_token_id,
-            pad_token_id=text_tokenizer.pad_token_id
+         tuned_phi2.generate(
+                inputs_embeds=inputs,
+                max_new_tokens=30,
+                bos_token_id=text_tokenizer.bos_token_id,
+                eos_token_id=text_tokenizer.eos_token_id,
+                pad_token_id=text_tokenizer.pad_token_id
         )
     )
     return prediction[0].rstrip('<|endoftext|>').rstrip("\n")
@@ -103,7 +104,7 @@ def generate_response(image, audio, text, count):
         q_tokens = text_tokenizer.encode(
             overall_input,
             return_tensors='pt').to(device)
-        question_token_embeddings = base_phi2_text.get_submodule(
+        question_token_embeddings = tuned_phi2.get_submodule(
             'model.embed_tokens')(q_tokens).to(device)
         inputs = torch.concat(
             (img_tokens.unsqueeze(0), question_token_embeddings),