ADd buggy PEFT

app.py

@@ -31,6 +31,7 @@ MODEL_PATHS = {
     "grade_more": "hallisky/lora-grade-highschool-llama-3-8b", 
     "grade_less": "hallisky/lora-grade-elementary-llama-3-8b",
 }
+FIRST_MODEL = list(MODEL_PATHS.keys())[0]
 
 DESCRIPTION = """\
 # Authorship Obfuscation
@@ -47,8 +48,20 @@ if not torch.cuda.is_available():
 if torch.cuda.is_available():
     device = "cuda"
     model_id = "meta-llama/Meta-Llama-3-8B"
-    model = AutoModelForCausalLM.from_pretrained(model_id).to(device) # device_map="auto" requires accelerate
-    tokenizer = AutoTokenizer.from_pretrained(model_id)
+
+    tokenizer = AutoTokenizer.from_pretrained(model_id, add_bos_token=True, add_eos_token=False, padding_side="left")
+    tokenizer.add_special_tokens({'pad_token': '<padding_token>'})
+
+    base_model = AutoModelForCausalLM.from_pretrained(model_id).to(device) # device_map="auto" requires accelerate
+    base_model.resize_token_embeddings(len(tokenizer)) # Resize to add pad token. Value doesn't matter
+    # Load in the first model
+    model = PeftModel.from_pretrained(base_model, MODEL_PATHS[FIRST_MODEL], adapter_name=FIRST_MODEL).to(device)
+    # Load in the rest of the models
+    for cur_adapter in MODEL_PATHS.keys():
+        if cur_adapter != FIRST_MODEL:
+            model.load_adapter(MODEL_PATHS[cur_adapter], adapter_name=cur_adapter)
+
+    model.eval()
 
 # Global variable to store the latest obfuscation result
 user_id = str(uuid4())  # Generate a unique session-specific user ID
@@ -98,13 +111,15 @@ def greet(input_text, length, function_words, grade_level, sarcasm, formality, v
         f"Expository: {expository}"
     )
 
+    converted_text = convert_data_to_format(input_text)
+    response = converted_text
     with torch.no_grad():
         outputs = model.generate(
-            input_ids=tokenizer(input_text, return_tensors="pt").input_ids.to(device),
+            input_ids=tokenizer(converted_text, return_tensors="pt").input_ids.to(device),
             max_length=100,
             num_return_sequences=1,
         )
-        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    #     response = tokenizer.decode(outputs[0], skip_special_tokens=True)
         
     # Save the new obfuscation result and reset feedback
     latest_obfuscation = {
@@ -123,6 +138,7 @@ def greet(input_text, length, function_words, grade_level, sarcasm, formality, v
             "narrative": narrative,
             "expository": expository
         },
+        "input": input_text,
         "output": response,
         "feedback_rating": "No Feedback Selected",
         "feedback_text": ""