Update app.py

app.py

@@ -1,84 +1,109 @@
 import gradio as gr
-import transformers
-from torch import bfloat16
-# from dotenv import load_dotenv  # if you wanted to adapt this for a repo that uses auth
-from threading import Thread
-
-
-#HF_AUTH = os.getenv('HF_AUTH')
-model_id ="tatihden/gemma_mental_health_7b_it_en"
-
-bnb_config = transformers.BitsAndBytesConfig(
-    load_in_4bit=True,
-    bnb_4bit_quant_type='nf4',
-    bnb_4bit_use_double_quant=True,
-    bnb_4bit_compute_dtype=bfloat16
-)
-model_config = transformers.AutoConfig.from_pretrained(
-    model_id,
-    #use_auth_token=HF_AUTH
-)
-
-model = transformers.AutoModelForCausalLM.from_pretrained(
-    model_id,
-    trust_remote_code=True,
-    config=model_config,
-    quantization_config=bnb_config,
-    device_map='auto',
-    #use_auth_token=HF_AUTH
-)
-
-tokenizer = transformers.AutoTokenizer.from_pretrained(
-    model_id,
-    #use_auth_token=HF_AUTH
-)
-
-
-DESCRIPTION = """
-# CalmChat
-This is a streaming Chat Interface implementation of [gemma_mental_health_7b_it_en](https://huggingface.co/tatihden/gemma_mental_health_7b_it_en).
-Sometimes the model doesn't appropriately hit its stop token. Feel free to hit "stop" and "retry" if this happens to you. 
-"""
-
-system_prompt = "You are helpful mental health AI."
-
-def prompt_build(system_prompt, user_inp, hist):
-    prompt = f"""### System:\n{system_prompt}\n\n"""
-    
-    for pair in hist:
-        prompt += f"""### User:\n{pair[0]}\n\n### Assistant:\n{pair[1]}\n\n"""
-
-    prompt += f"""### User:\n{user_inp}\n\n### Assistant:"""
+from huggingface_hub import InferenceClient
+import random
+
+models = [
+    "https://huggingface.co/tatihden/gemma_mental_health_7b_it_en",
+    "https://huggingface.co/tatihden/gemma_mental_health_2b_en",
+    "https://huggingface.co/tatihden/gemma_mental_health_2b_it_en"
+]
+
+clients = []
+for model in models:
+    clients.append(InferenceClient(model))
+
+
+def format_prompt(message, history):
+    prompt = ""
+    if history:
+        for user_prompt, bot_response in history:
+            prompt += f"<start_of_turn>user{user_prompt}<end_of_turn>"
+            prompt += f"<start_of_turn>model{bot_response}"
+    prompt += f"<start_of_turn>user{message}<end_of_turn><start_of_turn>model"
     return prompt
 
-def chat(user_input, history):
 
-    prompt = prompt_build(system_prompt, user_input, history)
-    model_inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
-
-    streamer = transformers.TextIteratorStreamer(tokenizer, timeout=10., skip_prompt=True, skip_special_tokens=True)
+def chat_inf(system_prompt, prompt, history, client_choice, seed, temp, tokens, top_p, rep_p):
+    client = clients[int(client_choice) - 1]
+    if not history:
+        history = []
+        hist_len = 0
+    if history:
+        hist_len = len(history)
+        print(hist_len)
 
     generate_kwargs = dict(
-        model_inputs,
-        streamer=streamer,
-        max_new_tokens=2048,
+        temperature=temp,
+        max_new_tokens=tokens,
+        top_p=top_p,
+        repetition_penalty=rep_p,
         do_sample=True,
-        top_p=0.95,
-        temperature=0.8,
-        top_k=50
+        seed=seed,
     )
-    t = Thread(target=model.generate, kwargs=generate_kwargs)
-    t.start()
-
-    model_output = ""
-    for new_text in streamer:
-        model_output += new_text
-        yield model_output
-    return model_output
-
-
-with gr.Blocks() as demo:
-    gr.Markdown(DESCRIPTION)
-    chatbot = gr.ChatInterface(fn=chat)
-
-demo.queue().launch()
+    formatted_prompt = format_prompt(f"{system_prompt}, {prompt}", history)
+    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True,
+                                    return_full_text=False)
+    output = ""
+
+    for response in stream:
+        output += response.token.text
+        yield [(prompt, output)]
+    history.append((prompt, output))
+    yield history
+
+
+def clear_fn():
+    return None
+
+
+rand_val = random.randint(1, 1111111111111111)
+
+
+def check_rand(inp, val):
+    if inp is True:
+        return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=random.randint(1, 1111111111111111))
+    else:
+        return gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, value=int(val))
+
+
+with gr.Blocks(css=background-color: rgb(74 222 128),theme=gr.themes.Soft()) as app:
+    gr.HTML(
+        """<center><h1 style='font-size:xx-large;'>Google Gemma Models</h1></center>""")
+    with gr.Group():
+        with gr.Row():
+            client_choice = gr.Dropdown(label="Models", type='index', choices=[c for c in models], value=models[0],
+                                        interactive=True)
+    chat_b = gr.Chatbot(height=500)
+    with gr.Group():
+        with gr.Row():
+            with gr.Column(scale=1):
+                with gr.Group():
+                    rand = gr.Checkbox(label="Random Seed", value=True)
+                    seed = gr.Slider(label="Seed", minimum=1, maximum=1111111111111111, step=1, value=rand_val)
+                    tokens = gr.Slider(label="Max new tokens", value=6400, minimum=0, maximum=8000, step=64,
+                                       interactive=True, visible=True, info="The maximum number of tokens")
+            with gr.Column(scale=1):
+                with gr.Group():
+                    temp = gr.Slider(label="Temperature", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
+                    top_p = gr.Slider(label="Top-P", step=0.01, minimum=0.01, maximum=1.0, value=0.9)
+                    rep_p = gr.Slider(label="Repetition Penalty", step=0.1, minimum=0.1, maximum=2.0, value=1.0)
+
+    with gr.Group():
+        with gr.Row():
+            with gr.Column(scale=3):
+                sys_inp = gr.Textbox(label="System Prompt (optional)")
+                inp = gr.Textbox(label="Prompt")
+                with gr.Row():
+                    btn = gr.Button("Chat")
+                    stop_btn = gr.Button("Stop")
+                    clear_btn = gr.Button("Clear")
+
+    chat_sub = inp.submit(check_rand, [rand, seed], seed).then(chat_inf,
+                                                               [sys_inp, inp, chat_b, client_choice, seed, temp, tokens,
+                                                                top_p, rep_p], chat_b)
+    go = btn.click(check_rand, [rand, seed], seed).then(chat_inf,
+                                                        [sys_inp, inp, chat_b, client_choice, seed, temp, tokens, top_p,
+                                                         rep_p], chat_b)
+    stop_btn.click(None, None, None, cancels=[go, chat_sub])
+    clear_btn.click(clear_fn, None, [chat_b])
+app.queue(default_concurrency_limit=10).launch()