create app.py

app.py

@@ -0,0 +1,233 @@
+import gradio as gr
+import json
+import requests
+import os
+from text_generation import Client, InferenceAPIClient
+
+#Streaming endpoint for OPENAI ChatGPT
+API_URL = "https://api.openai.com/v1/chat/completions"
+#Streaming endpoint for OPENCHATKIT
+os.environ['API_URL_TGTHR'] = 'https://openchat.ngrok.io'
+
+#Predict function for CHATGPT
+def predict_chatgpt(inputs, top_p_chatgpt, temperature_chatgpt, openai_api_key, chat_counter_chatgpt, chatbot_chatgpt=[], history=[]):  
+    #Define payload and header for chatgpt API
+    payload = {
+    "model": "gpt-3.5-turbo",
+    "messages": [{"role": "user", "content": f"{inputs}"}],
+    "temperature" : 1.0,
+    "top_p":1.0,
+    "n" : 1,
+    "stream": True,
+    "presence_penalty":0,
+    "frequency_penalty":0,
+    }
+
+    headers = {
+    "Content-Type": "application/json",
+    "Authorization": f"Bearer {openai_api_key}"
+    }
+    #debug
+    #print(f"chat_counter_chatgpt - {chat_counter_chatgpt}")
+    
+    #Handling the different roles for ChatGPT
+    if chat_counter_chatgpt != 0 :
+        messages=[]
+        for data in chatbot:
+          temp1 = {}
+          temp1["role"] = "user" 
+          temp1["content"] = data[0] 
+          temp2 = {}
+          temp2["role"] = "assistant" 
+          temp2["content"] = data[1]
+          messages.append(temp1)
+          messages.append(temp2)
+        temp3 = {}
+        temp3["role"] = "user" 
+        temp3["content"] = inputs
+        messages.append(temp3)
+        payload = {
+        "model": "gpt-3.5-turbo",
+        "messages": messages, #[{"role": "user", "content": f"{inputs}"}],
+        "temperature" : temperature, #1.0,
+        "top_p": top_p, #1.0,
+        "n" : 1,
+        "stream": True,
+        "presence_penalty":0,
+        "frequency_penalty":0,
+        }
+
+    chat_counter_chatgpt+=1
+    
+    history.append(inputs)
+    # make a POST request to the API endpoint using the requests.post method, passing in stream=True
+    response = requests.post(API_URL, headers=headers, json=payload, stream=True)
+    token_counter = 0 
+    partial_words = "" 
+
+    counter=0
+    for chunk in response.iter_lines():
+        #Skipping the first chunk
+        if counter == 0:
+          counter+=1
+          continue
+        # check whether each line is non-empty
+        if chunk.decode() :
+          chunk = chunk.decode()
+          # decode each line as response data is in bytes
+          if len(chunk) > 12 and "delta" in json.loads(chunk[6:])['choices'][0]:
+              partial_words = partial_words + json.loads(chunk[6:])['choices'][0]["delta"]["content"]
+              if token_counter == 0:
+                history.append(" " + partial_words)
+              else:
+                history[-1] = partial_words
+              chat = [(history[i], history[i + 1]) for i in range(0, len(history) - 1, 2) ]  # convert to tuples of list
+              token_counter+=1
+              yield chat, history, chat_counter_chatgpt  # this resembles {chatbot: chat, state: history}  
+
+#Predict function for OPENCHATKIT
+def predict_together(model: str,
+        inputs: str,
+        top_p: float,
+        temperature: float,
+        top_k: int,
+        repetition_penalty: float,
+        watermark: bool,
+        chatbot,
+        history,):
+  
+    client = Client(os.getenv("API_URL_TGTHR")) #get_client(model)
+    # debug
+    print(f"^^client is - {client}")
+    user_name, assistant_name = "<human>:", "<bot>:" #get_usernames(model)
+
+    history.append(inputs)
+
+    past = []
+    for data in chatbot:
+        user_data, model_data = data
+
+        if not user_data.startswith(user_name):
+            user_data = user_name + user_data
+        if not model_data.startswith("\n\n" + assistant_name):
+            model_data = "\n\n" + assistant_name + model_data
+
+        past.append(user_data + model_data + "\n\n")
+
+    if not inputs.startswith(user_name):
+        inputs = user_name + inputs
+
+    total_inputs = "".join(past) + inputs + "\n\n" + assistant_name
+    # truncate total_inputs
+    total_inputs = total_inputs[-1000:]
+
+    partial_words = ""
+
+    for i, response in enumerate(client.generate_stream(
+            total_inputs,
+            top_p=top_p,
+            top_k=top_k,
+            repetition_penalty=repetition_penalty,
+            watermark=watermark,
+            temperature=temperature,
+            max_new_tokens=500,
+            stop_sequences=[user_name.rstrip(), assistant_name.rstrip()],
+    )):
+        if response.token.special:
+            continue
+
+        partial_words = partial_words + response.token.text
+        if partial_words.endswith(user_name.rstrip()):
+            partial_words = partial_words.rstrip(user_name.rstrip())
+        if partial_words.endswith(assistant_name.rstrip()):
+            partial_words = partial_words.rstrip(assistant_name.rstrip())
+
+        if i == 0:
+            history.append(" " + partial_words)
+        else:
+            history[-1] = partial_words
+
+        chat = [
+            (history[i].strip(), history[i + 1].strip()) for i in range(0, len(history) - 1, 2)
+        ]
+        yield chat, history
+
+
+def reset_textbox():
+    return gr.update(value="")
+
+def reset_chat(chatbot, state):
+    # debug
+    print(f"^^chatbot value is - {chatbot}")
+    print(f"^^state value is - {state}")
+    return None, []
+
+
+title = """<h1 align="center">🚀ChatGPT & 🔥OpenChatKit Comparison Gradio Demo</h1>"""
+description = """Language models can be conditioned to act like dialogue agents through a conversational prompt that typically takes the form:
+```
+User: <utterance>
+Assistant: <utterance>
+User: <utterance>
+Assistant: <utterance>
+...
+```
+In this app, you can explore the outputs of multiple LLMs when prompted in similar ways.
+"""
+
+with gr.Blocks(css="""#col_container {width: 1000px; margin-left: auto; margin-right: auto;}
+                #chatgpt {height: 520px; overflow: auto;}
+                #chattogether {height: 520px; overflow: auto;} """ ) as demo:
+                #clear {width: 100px; height:50px; font-size:12px}""") as demo:
+    gr.HTML(title)
+    with gr.Row():
+      with gr.Column(scale=14):
+          with gr.Box():
+              with gr.Row():
+                  with gr.Column(scale=13):
+                    openai_api_key = gr.Textbox(type='password', label="Enter your OpenAI API key here for ChatGPT")
+                    inputs = gr.Textbox(placeholder="Hi there!", label="Type an input and press Enter" )
+                  with gr.Column(scale=1):
+                      #b1 = gr.Button(elem_id = 'run')
+                      b2 = gr.Button('Clear up Chatbots!', elem_id = 'clear').style(full_width=True)
+                  state_chatgpt = gr.State([])
+                  state_together = gr.State([])
+
+          with gr.Box():
+              with gr.Row():
+                  chatbot_chatgpt = gr.Chatbot(elem_id="chatgpt", label='ChatGPT API - OPENAI')
+                  chatbot_together = gr.Chatbot(elem_id="chattogether", label='OpenChatKit - Text Generation')
+          
+      with gr.Column(scale=2, elem_id='parameters'):
+          with gr.Box():
+              gr.HTML("Parameters for #OpenCHAtToolKit")
+              top_p = gr.Slider(minimum=-0, maximum=1.0,value=0.95, step=0.05,interactive=True, label="Top-p",)
+              temperature = gr.Slider(minimum=-0, maximum=5.0, value=0.5, step=0.1, interactive=True, label="Temperature", )
+              top_k = gr.Slider( minimum=1, maximum=50, value=4, step=1, interactive=True, label="Top-k",)
+              repetition_penalty = gr.Slider( minimum=0.1, maximum=3.0, value=1.03, step=0.01, interactive=True, label="Repetition Penalty",)
+              watermark = gr.Checkbox(value=True, label="Text watermarking")
+              model = gr.CheckboxGroup(value="Rallio67/joi2_20B_instruct_alpha",
+                  choices=["togethercomputer/GPT-NeoXT-Chat-Base-20B", "Rallio67/joi2_20B_instruct_alpha", "google/flan-t5-xxl", "google/flan-ul2", "bigscience/bloomz", "EleutherAI/gpt-neox-20b",],
+                  label="Model",visible=False,) 
+              temp_textbox_together = gr.Textbox(value=model.choices[0], visible=False) 
+
+          with gr.Box():
+              gr.HTML("Parameters for OpenAI's ChatGPT")
+              top_p_chatgpt = gr.Slider( minimum=-0, maximum=1.0, value=1.0, step=0.05, interactive=True, label="Top-p",)
+              temperature_chatgpt = gr.Slider( minimum=-0, maximum=5.0, value=1.0, step=0.1, interactive=True, label="Temperature",)
+              chat_counter_chatgpt = gr.Number(value=0, visible=False, precision=0)
+
+    inputs.submit(reset_textbox, [], [inputs])
+    
+    inputs.submit( predict_chatgpt, 
+                [inputs, top_p_chatgpt, temperature_chatgpt, openai_api_key, chat_counter_chatgpt, chatbot_chatgpt, state_chatgpt], 
+                [chatbot_chatgpt, state_chatgpt, chat_counter_chatgpt],)
+    inputs.submit( predict_together,
+                [temp_textbox_together, inputs, top_p, temperature, top_k, repetition_penalty, watermark, chatbot_together, state_together, ],
+                [chatbot_together, state_together],)
+
+    b2.click(reset_chat, [chatbot_chatgpt, state_chatgpt], [chatbot_chatgpt, state_chatgpt])
+    b2.click(reset_chat, [chatbot_together, state_together], [chatbot_together, state_together])
+    
+    gr.Markdown(description)
+demo.queue(concurrency_count=16).launch(height= 2500, debug=True)