import gradio as gr
import json
import requests
import os
from text_generation import Client, InferenceAPIClient
# Load pre-trained model and tokenizer - for THUDM model
from transformers import AutoModel, AutoTokenizer
tokenizer_glm = AutoTokenizer.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True)
model_glm = AutoModel.from_pretrained("THUDM/chatglm-6b", trust_remote_code=True).half().cuda()
model_glm = model_glm.eval()
# Load pre-trained model and tokenizer for Chinese to English translator
from transformers import M2M100ForConditionalGeneration, M2M100Tokenizer
model_chtoen = M2M100ForConditionalGeneration.from_pretrained("facebook/m2m100_418M")
tokenizer_chtoen = M2M100Tokenizer.from_pretrained("facebook/m2m100_418M")
#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_chatgpt:
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_chatgpt, #1.0,
"top_p": top_p_chatgpt, #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) > 13 and "content" in json.loads(chunk[6:])['choices'][0]["delta"]:
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}
# Define function to generate model predictions and update the history
def predict_glm(input, history=[]):
response, history = model_glm.chat(tokenizer_glm, input, history)
# translate Chinese to English
history = [(query, translate_Chinese_English(response)) for query, response in history]
return history, history #[history] + updates
def translate_Chinese_English(chinese_text):
# translate Chinese to English
tokenizer_chtoen.src_lang = "zh"
encoded_zh = tokenizer_chtoen(chinese_text, return_tensors="pt")
generated_tokens = model_chtoen.generate(**encoded_zh, forced_bos_token_id=tokenizer_chtoen.get_lang_id("en"))
trans_eng_text = tokenizer_chtoen.batch_decode(generated_tokens, skip_special_tokens=True)
return trans_eng_text[0]
# Define function to generate model predictions and update the history
def predict_glm_stream(input, history=[]): #, top_p, temperature):
response, history = model_glm.chat(tokenizer_glm, input, history)
print(f"outside for loop resonse is ^^- {response}")
print(f"outside for loop history is ^^- {history}")
top_p = 1.0
temperature = 1.0
for response, history in model.stream_chat(tokenizer_glm, input, history, top_p=1.0, temperature=1.0): #max_length=max_length,
print(f"In for loop resonse is ^^- {response}")
print(f"In for loop history is ^^- {history}")
# translate Chinese to English
history = [(query, translate_Chinese_English(response)) for query, response in history]
print(f"In for loop translated history is ^^- {history}")
yield history, history #[history] + updates
"""
def predict(input, max_length, top_p, temperature, history=None):
if history is None:
history = []
for response, history in model.stream_chat(tokenizer, input, history, max_length=max_length, top_p=top_p,
temperature=temperature):
updates = []
for query, response in history:
updates.append(gr.update(visible=True, value="user：" + query)) #用户
updates.append(gr.update(visible=True, value="ChatGLM-6B：" + response))
if len(updates) < MAX_BOXES:
updates = updates + [gr.Textbox.update(visible=False)] * (MAX_BOXES - len(updates))
yield [history] + updates
"""
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 = """🔥🔥Comparison: ChatGPT & OpenChatKit
🚀A Gradio Streaming Demo
Official Demo: OpenChatKit feedback app"""
title = """🔥🔥Comparison: ChatGPT & Open Sourced CHatGLM-6B
🚀A Gradio Chatbot Demo
"""
description = """Language models can be conditioned to act like dialogue agents through a conversational prompt that typically takes the form:
```
User:
Assistant:
User:
Assistant:
...
```
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;}
#chatglm {height: 520px; overflow: auto;} """ ) as demo:
#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('🏃Run', elem_id = 'run').style(full_width=True)
b2 = gr.Button('🔄Clear up Chatbots!', elem_id = 'clear').style(full_width=True)
state_chatgpt = gr.State([])
#state_together = gr.State([])
state_glm = 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')
chatbot_glm = gr.Chatbot(elem_id="chatglm", label='THUDM-ChatGLM6B')
with gr.Column(scale=2, elem_id='parameters'):
with gr.Box():
gr.HTML("Parameters for #OpenCHAtKit", visible=False)
top_p = gr.Slider(minimum=-0, maximum=1.0,value=0.25, step=0.05,interactive=True, label="Top-p", visible=False)
temperature = gr.Slider(minimum=-0, maximum=5.0, value=0.6, step=0.1, interactive=True, label="Temperature", visible=False)
top_k = gr.Slider( minimum=1, maximum=50, value=50, step=1, interactive=True, label="Top-k", visible=False)
repetition_penalty = gr.Slider( minimum=0.1, maximum=3.0, value=1.01, step=0.01, interactive=True, label="Repetition Penalty", visible=False)
watermark = gr.Checkbox(value=True, label="Text watermarking", visible=False)
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_glm,
# [inputs, state_glm, ],
# [chatbot_glm, state_glm],)
#b1.click( predict_glm,
# [inputs, state_glm, ],
# [chatbot_glm, state_glm],)
inputs.submit( predict_glm_stream,
[inputs, state_glm, ],
[chatbot_glm, state_glm],)
b1.click( predict_glm_stream,
[inputs, state_glm, ],
[chatbot_glm, state_glm],)
b1.click( 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],)
b2.click(reset_chat, [chatbot_chatgpt, state_chatgpt], [chatbot_chatgpt, state_chatgpt])
#b2.click(reset_chat, [chatbot_together, state_together], [chatbot_together, state_together])
b2.click(reset_chat, [chatbot_glm, state_glm], [chatbot_glm, state_glm])
gr.HTML('''
Duplicate the Space and run securely with your OpenAI API Key''')
gr.Markdown(description)
demo.queue(concurrency_count=16).launch(height= 2500, debug=True)