backups/app_v1.py

import os
import torch
import gradio as gr
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

# Environment variables
os.environ["TOKENIZERS_PARALLELISM"] = "0"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

# Global variables for model and tokenizer
model = None
tokenizer = None

def get_gpu_memory():
    return torch.cuda.memory_allocated() / 1024 / 1024  # Convert to MiB

class TorchTracemalloc:
    def __init__(self):
        self.begin = 0
        self.peak = 0

    def __enter__(self):
        torch.cuda.empty_cache()
        torch.cuda.reset_peak_memory_stats()
        torch.cuda.synchronize()
        self.begin = get_gpu_memory()
        return self

    def __exit__(self, *exc):
        torch.cuda.synchronize()
        self.peak = (
            torch.cuda.max_memory_allocated() / 1024 / 1024
        )  # Convert to MiB

    def consumed(self):
        return self.peak - self.begin

def load_model_and_tokenizer(model_name, dtype, kv_bits):
    global model, tokenizer
    if model is None or tokenizer is None:
        tokenizer = AutoTokenizer.from_pretrained(model_name)
        special_tokens = {"pad_token": "<PAD>"}
        tokenizer.add_special_tokens(special_tokens)

        config = AutoConfig.from_pretrained(model_name)
        if kv_bits != "unquantized":
            quantizer_path = f"codebooks/{model_name.split('/')[-1]}_{kv_bits}bit.xmad"
            setattr(config, "quantizer_path", quantizer_path)

        dtype = torch.__dict__.get(dtype, torch.float32)
        model = AutoModelForCausalLM.from_pretrained(model_name, config=config, torch_dtype=dtype, device_map="auto")

        if len(tokenizer) > model.get_input_embeddings().weight.shape[0]:
            model.resize_token_embeddings(len(tokenizer))

        tokenizer.padding_side = "left"
        model.config.pad_token_id = tokenizer.pad_token_id

    return model, tokenizer

# Initialize model and tokenizer
model, tokenizer = load_model_and_tokenizer("NousResearch/Hermes-2-Theta-Llama-3-8B", "fp16", "1")

def process_dialog(dialog, model, tokenizer, max_tokens, temperature):
    prompt = tokenizer.apply_chat_template(
        dialog, tokenize=False, add_generation_prompt=True
    )
    tokenized_input_prompt_ids = tokenizer(
        prompt, return_tensors="pt"
    ).input_ids.to(model.device)

    torch.cuda.empty_cache()
    torch.cuda.reset_peak_memory_stats()

    with TorchTracemalloc() as tt:
        start_time = time.time()
        with torch.no_grad():
            token_ids_for_each_answer = model.generate(
                tokenized_input_prompt_ids,
                max_new_tokens=max_tokens,
                temperature=temperature,
                do_sample=True,
                eos_token_id=tokenizer.eos_token_id,
                pad_token_id=tokenizer.pad_token_id,
            )
        torch.cuda.synchronize()
        end_time = time.time()

    response = token_ids_for_each_answer[0][
        tokenized_input_prompt_ids.shape[-1] :
    ]
    cleaned_response = tokenizer.decode(
        response,
        skip_special_tokens=True,
        clean_up_tokenization_spaces=True,
    )

    return cleaned_response

def respond(message, history, system_message, max_tokens, temperature):
    dialog = [{"role": "system", "content": system_message}]
    for val in history:
        if val[0]:
            dialog.append({"role": "user", "content": val[0]})
        if val[1]:
            dialog.append({"role": "assistant", "content": val[1]})
    dialog.append({"role": "user", "content": message})

    response = process_dialog(dialog, model, tokenizer, max_tokens, temperature)
    history.append((message, response))
    return history, history

# Initialize Gradio ChatInterface
demo = gr.ChatInterface(
    fn=respond,
    additional_inputs=[
        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
    ],
    theme="default",
    title="1bit llama3 by xMAD.ai",
    description="The first industrial level 1 bit quantization Llama3, we can achieve 800 tokens per second on NVIDIA V100 and 1200 on NVIDIA A100, 90% cost down of your cloud hosting cost",
    css=".scrollable { height: 400px; overflow-y: auto; padding: 10px; border: 1px solid #ccc; }"
)

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