Fix sliders, still has A100 issue with multiple PEFT

app.py

@@ -3,7 +3,6 @@ import gradio as gr
 import torch
 from peft import PeftModel, PeftConfig
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from IPython import embed
 
 import json
 from datetime import datetime
@@ -25,13 +24,19 @@ def convert_data_to_format(text):
 
 MODEL_PATHS = {
     "length_more": "hallisky/lora-length-long-llama-3-8b",
-    "length_less": "hallisky/lora-length-long-llama-3-8b",
+    "length_less": "hallisky/lora-length-short-llama-3-8b",
     "function_more": "hallisky/lora-function-more-llama-3-8b",
     "function_less": "hallisky/lora-function-less-llama-3-8b",
     "grade_more": "hallisky/lora-grade-highschool-llama-3-8b", 
     "grade_less": "hallisky/lora-grade-elementary-llama-3-8b",
+    "formality_more": "hallisky/lora-formality-formal-llama-3-8b", 
+    "formality_less": "hallisky/lora-formality-informal-llama-3-8b", 
+    "sarcasm_more": "hallisky/lora-sarcasm-more-llama-3-8b", 
+    "sarcasm_less": "hallisky/lora-sarcasm-less-llama-3-8b", 
+    "voice_passive": "hallisky/lora-voice-passive-llama-3-8b", 
+    "voice_active": "hallisky/lora-voice-active-llama-3-8b", 
 }
-FIRST_MODEL = list(MODEL_PATHS.keys())[0]
+FIRST_MODEL = list(MODEL_PATHS.keys())[5]
 
 DESCRIPTION = """\
 # Authorship Obfuscation
@@ -57,9 +62,9 @@ if torch.cuda.is_available():
     # 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)
+    # 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()
 
@@ -96,6 +101,19 @@ def greet(input_text, length, function_words, grade_level, sarcasm, formality, v
     global latest_obfuscation, user_id
     current_time = datetime.now().isoformat()
     
+    sliders_dict = {}
+    cur_keys = []
+    cur_keys.append(("length_more" if length > 0 else (None if length == 0 else "length_less"), length))
+    cur_keys.append(("function_more" if function_words > 0 else (None if function_words == 0 else "function_less"), function_words))
+    cur_keys.append(("grade_more" if grade_level > 0 else (None if grade_level == 0 else "grade_less"), grade_level))
+    cur_keys.append(("sarcasm_more" if sarcasm > 0 else (None if sarcasm == 0 else "sarcasm_less"), sarcasm))
+    cur_keys.append(("formality_more" if formality > 0 else (None if formality == 0 else "formality_less"), formality))
+    cur_keys.append(("voice_active" if voice > 0 else (None if voice == 0 else "voice_passive"), voice))
+
+    for cur_key in cur_keys:
+        if cur_key[0] is not None:
+            sliders_dict[cur_key[0]] = cur_key[1]
+
     response = (
         f"Hello!\n"
         f"Input Text: {input_text}\n"
@@ -112,14 +130,13 @@ def greet(input_text, length, function_words, grade_level, sarcasm, formality, v
     )
 
     converted_text = convert_data_to_format(input_text)
-    response = converted_text
-    with torch.no_grad():
-        outputs = model.generate(
-            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)
+
+    # Convert the list of strings in data to a list of model inputs
+    inputs = tokenizer(converted_text, return_tensors="pt", max_length=2048, truncation=True).to(device)
+    input_length = inputs.input_ids.shape[1]
+    with torch.no_grad(): 
+        outputs = model.generate(**inputs, max_length=100, top_p = 0.95)
+    response = tokenizer.decode(outputs[0, input_length:], skip_special_tokens=True)
         
     # Save the new obfuscation result and reset feedback
     latest_obfuscation = {