app.py

import gradio as gr

import os, sys
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig, pipeline
from transformers import LlamaTokenizer
import torch
import spaces
import psutil

# Define the model repository
REPO_NAME = 'schuler/experimental-JP47D20'
# REPO_NAME = 'schuler/experimental-JP47D21-KPhi-3-micro-4k-instruct'

# How to cache?
@spaces.GPU()
def load_model(repo_name):
    # tokenizer = AutoTokenizer.from_pretrained(repo_name, trust_remote_code=True)
    tokenizer = LlamaTokenizer.from_pretrained(repo_name, trust_remote_code=True)
    generator_conf = GenerationConfig.from_pretrained(repo_name)
    model = AutoModelForCausalLM.from_pretrained(repo_name, trust_remote_code=True, torch_dtype=torch.bfloat16, attn_implementation="eager")
    # model.to('cuda')
    return tokenizer, generator_conf, model

# tokenizer, generator_conf, model, generator = False, False, False, False 
# with gr.Blocks() as main_block:

tokenizer, generator_conf, model = load_model(REPO_NAME)
global_error = ''
try:
  generator = pipeline("text-generation", model=model, tokenizer=tokenizer)
except Exception as e:
  global_error =  f"Failed to load model: {str(e)}"

@spaces.GPU()
def local_generate(
    prompt,
    generation_config,
    max_new_tokens,
    do_sample=True,
    top_p=0.25,
    repetition_penalty=1.2,
    temperature=1.0
):
    response_output = generator(
        prompt,
        generation_config=generation_config,
        max_new_tokens=max_new_tokens,
        do_sample=do_sample,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        temperature=temperature
    )        
    generated_text = response_output[0]['generated_text']
    # Extract the assistant's response
    result = generated_text[len(prompt):]    
    return result

def respond(
    message,
    history: list[tuple[str, str]],
    system_message,
    max_tokens,
    temperature,
    top_p,
):
    result = 'none'
    try:
        # Build the conversation prompt
        prompt = ''
        messages = []
        if (len(system_message)>0):
            prompt = "<|assistant|>"+system_message+f"<|end|>\n"    
        for val in history:
            if val[0]:
                messages.append({"role": "user", "content": val[0]})
            if val[1]:
                messages.append({"role": "assistant", "content": val[1]})
    
        messages.append({"role": "user", "content": message})
    
        for hmessage in messages:
            role = "<|assistant|>" if hmessage['role'] == 'assistant' else "<|user|>"
            prompt += f"{role}{hmessage['content']}<|end|>"        
        prompt += f"<|assistant|>"

        tokens_cnt = 0
        tokens_inc = 64
        last_token_len = 1
        full_result = ''
        while ( (tokens_cnt < max_tokens) and (last_token_len > 0) ):
            # Generate the response
            result = local_generate(
                prompt,
                generation_config=generator_conf,
                max_new_tokens=tokens_inc,
                do_sample=True,
                top_p=top_p,
                repetition_penalty=1.2,
                temperature=temperature
            )
            full_result = full_result + result
            prompt = prompt + result
            tokens_cnt = tokens_cnt + tokens_inc
            last_token_len = len(result)
            yield full_result
        
    except Exception as error:
        exc_type, exc_obj, exc_tb = sys.exc_info()
        fname = os.path.split(exc_tb.tb_frame.f_code.co_filename)[1]
        result = str(error) +':'+ exc_type +':'+ fname +':'+ exc_tb.tb_lineno
        yield result

"""
For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
"""
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
embed_params = sum(p.numel() for p in model.model.embed_tokens.parameters())*2
non_embed_params = (trainable_params - embed_params) / 1e6
cpu_usage = psutil.cpu_percent(interval=1)
status_text = \
  f"This chat uses the {REPO_NAME} model with {model.get_memory_footprint() / 1e6:.2f} MB memory footprint. " + \
  f"Current CPU usage is {cpu_usage:.2f}% . '" + \
  f"Total number of non embedding trainable parameters: {non_embed_params:.2f} million. " + \
  f"You may ask questions such as 'What is biology?' or 'What is the human body?'"

# """
demo = gr.ChatInterface(
    respond,
    additional_inputs=[
        gr.Textbox(value="" + global_error, label="System message"),
        gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=1.0, step=0.1, label="Temperature"),
        gr.Slider(
            minimum=0.1,
            maximum=1.0,
            value=0.25,
            step=0.05,
            label="Top-p (nucleus sampling)",
        ),
    ],
    description=status_text
)
"""
with gr.Blocks() as demo:
    # Display the status text at the top
    gr.Markdown(status_text)
    # Create the ChatInterface
    chat = gr.ChatInterface(
        respond,
        additional_inputs=[
            gr.Textbox(value="" + global_error, label="System message"),
            gr.Slider(minimum=1, maximum=4096, value=1024, step=1, label="Max new tokens"),
            gr.Slider(minimum=0.1, maximum=4.0, value=1.0, step=0.1, label="Temperature"),
            gr.Slider(
                minimum=0.1,
                maximum=1.0,
                value=0.25,
                step=0.05,
                label="Top-p (nucleus sampling)",
            ),
        ],
    )
"""

if __name__ == "__main__":
    demo.launch()