0.42 Pharia added

app.py

@@ -8,7 +8,7 @@ import os
 import traceback
 
 from threading import Thread
-from random import shuffle
+from random import shuffle, choice
 
 logging.basicConfig(level=logging.DEBUG)
 
@@ -17,7 +17,6 @@ SPACER = '\n' + '*' * 40 + '\n'
 HF_TOKEN = os.environ.get("HF_TOKEN", None)
 login(token=HF_TOKEN)
 
-
 system_prompts = {
     "English": "You are a helpful chatbot that answers user input in a concise and witty way.",
     "German": "Du bist ein hilfreicher Chatbot, der Usereingaben knapp und originell beantwortet.",
@@ -25,12 +24,19 @@ system_prompts = {
     "Spanish": "Eres un chatbot servicial que responde a las entradas de los usuarios de forma concisa y original."
 }
 
-model_info = [{"id": "NousResearch/Meta-Llama-3.1-8B-Instruct",
+htmL_info = "<center><h1>Pharia Battle Royale</h1><p>Let the games begin: Try a prompt in a language you like. Set the parameters and vote for the best answers. After casting your vote, the bots reveal their identity.</p></center>"
+
+model_info = [{"id": "Aleph-Alpha/Pharia-1-LLM-7B-control-hf",
+                "name": "Pharia 1 LLM 7B control hf"}]
+
+challenger_models = [{"id": "NousResearch/Meta-Llama-3.1-8B-Instruct",
                     "name": "Meta Llama 3.1 8B Instruct"},
                     {"id": "mistralai/Mistral-7B-Instruct-v0.3",
                     "name": "Mistral 7B Instruct v0.3"}]
+
+model_info.append(choice(challenger_models))
 shuffle(model_info)
-logging.debug('Models shuffled')
+logging.debug(f'models shuffled. model[0]: {model_info[0]['name']}, model[1]: {model_info[1]['name']}')
 
 device = "cuda" 
 
@@ -50,15 +56,17 @@ try:
         device_map="auto",
         trust_remote_code=True,
     )
-    
+
 except Exception as e:
     logging.error(f'{SPACER} Error: {e}, Traceback {traceback.format_exc()}')
 
 
-def apply_chat_template(messages, add_generation_prompt=False):
-    """Adds chat template for Pharia. Expects a list of messages. 
+def apply_pharia_template(messages, add_generation_prompt=False):
+    """Chat template not defined in Pharia model configs. 
+    Adds chat template for Pharia. Expects a list of messages. 
     add_generation_prompt:bool extends tmplate for generation.
     """
+
     pharia_template = """<|begin_of_text|>"""
     role_map = {
         "system": "<|start_header_id|>system<|end_header_id|>\n",
@@ -99,21 +107,29 @@ def generate_both(system_prompt, input_text, chatbot_a, chatbot_b, max_new_token
         new_messages_a = system_prompt_list + chat_history_a + input_text_list
         new_messages_b = system_prompt_list + chat_history_b + input_text_list
 
-        input_ids_a = tokenizer_a.apply_chat_template(
-            new_messages_a,
-            add_generation_prompt=True,
-            dtype=torch.float16,
-            return_tensors="pt"
-        ).to(device)
-
-        input_ids_b = tokenizer_b.apply_chat_template(
-            new_messages_b,
-            add_generation_prompt=True,
-            dtype=torch.float16,
-            return_tensors="pt"
-        ).to(device)
-
-        logging.debug(f'model_a.device: {model_a.device}, model_b.device: {model_b.device}')
+        if "Pharia" in model_info[0]['id']:
+            formatted_conversation = apply_pharia_template(messages=new_messages_a, add_generation_prompt=True)
+            input_ids_a = tokenizer_a(formatted_conversation, return_tensors="pt").to(device)
+
+        else: 
+            input_ids_a = tokenizer_a.apply_chat_template(
+                new_messages_a,
+                add_generation_prompt=True,
+                dtype=torch.float16,
+                return_tensors="pt"
+            ).to(device)
+
+        if "Pharia" in model_info[1]['id']:
+            formatted_conversation = apply_pharia_template(messages=new_messages_a, add_generation_prompt=True)
+            input_ids_a = tokenizer_a(formatted_conversation, return_tensors="pt").to(device)
+
+        else:
+            input_ids_b = tokenizer_b.apply_chat_template(
+                new_messages_b,
+                add_generation_prompt=True,
+                dtype=torch.float16,
+                return_tensors="pt"
+            ).to(device)
 
         generation_kwargs_a = dict(
             input_ids=input_ids_a,
@@ -200,7 +216,7 @@ def reveal_bot(selection, chatbot_a, chatbot_b):
 with gr.Blocks() as demo:
     try:
         with gr.Column():
-            gr.HTML("<center><h1>Bot Battle Royale</h1><p>Try a prompt in a language you like. Set the parameters and vote for the best answers. Finally, the bots reveal their identity.</p></center>")
+            gr.HTML(htmL_info)
             with gr.Row(variant="compact"):
                 with gr.Column(scale=0):
                     language_dropdown = gr.Dropdown(