app.py

import os
import json
import subprocess
from threading import Thread

import torch
import spaces
import gradio as gr
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig, TextIteratorStreamer

subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)

MODEL_ID = os.environ.get("MODEL_ID")
CHAT_TEMPLATE = os.environ.get("CHAT_TEMPLATE")
MODEL_NAME = MODEL_ID.split("/")[-1]
CONTEXT_LENGTH = int(os.environ.get("CONTEXT_LENGTH"))
COLOR = os.environ.get("COLOR")
EMOJI = os.environ.get("EMOJI")
DESCRIPTION = os.environ.get("DESCRIPTION")


@spaces.GPU()
def predict(message, history, system_prompt, temperature, max_new_tokens, top_k, repetition_penalty, top_p):
    # Format history with a given chat template
    if CHAT_TEMPLATE == "Auto":
        stop_tokens = [tokenizer.eos_token_id]
        instruction = []
        for user, assistant in history:
            instruction.extend([{"role": "user", "content": user}, {"role": "assistant", "content": assistant}])
        instruction.append({"role": "user", "content": message})
    elif CHAT_TEMPLATE == "ChatML":
        stop_tokens = ["<|endoftext|>", "<|im_end|>"]
        instruction = '<|im_start|>system\n' + system_prompt + '\n<|im_end|>\n'
        for user, assistant in history:
            instruction += '<|im_start|>user\n' + user + '\n<|im_end|>\n<|im_start|>assistant\n' + assistant
        instruction += '\n<|im_start|>user\n' + message + '\n<|im_end|>\n<|im_start|>assistant\n'
    elif CHAT_TEMPLATE == "Mistral Instruct":
        stop_tokens = ["</s>", "[INST]", "[INST] ", "<s>", "[/INST]", "[/INST] "]
        instruction = '<s>[INST] ' + system_prompt
        for user, assistant in history:
            instruction += user + ' [/INST] ' + assistant + '</s>[INST]'
        instruction += ' ' + message + ' [/INST]'
    else:
        raise Exception("Incorrect chat template, select 'ChatML' or 'Mistral Instruct'")
    print(instruction)
    
    streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True)
    enc = tokenizer([instruction], return_tensors="pt", padding=True, truncation=True)
    input_ids, attention_mask = enc.input_ids, enc.attention_mask

    if input_ids.shape[1] > CONTEXT_LENGTH:
        input_ids = input_ids[:, -CONTEXT_LENGTH:]

    generate_kwargs = dict(
        {"input_ids": input_ids.to(device), "attention_mask": attention_mask.to(device)},
        streamer=streamer,
        do_sample=True,
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_k=top_k,
        repetition_penalty=repetition_penalty,
        top_p=top_p
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()
    outputs = []
    for new_token in streamer:
        outputs.append(new_token)
        if new_token in stop_tokens:
            break
        yield "".join(outputs)


# Load model
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
quantization_config = BitsAndBytesConfig(
    load_in_4bit=False,
    bnb_4bit_compute_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    device_map="auto",
    quantization_config=quantization_config,
    attn_implementation="flash_attention_2",
)

# Create Gradio interface
gr.ChatInterface(
    predict,
    title=EMOJI + " " + MODEL_NAME,
    description=DESCRIPTION,
    additional_inputs_accordion=gr.Accordion(label="⚙️ Parameters", open=False),
    additional_inputs=[
        gr.Textbox("", label="System prompt"),
        gr.Slider(0, 1, 0.8, label="Temperature"),
        gr.Slider(128, 4096, 1024, label="Max new tokens"),
        gr.Slider(1, 80, 40, label="Top K sampling"),
        gr.Slider(0, 2, 1.1, label="Repetition penalty"),
        gr.Slider(0, 1, 0.95, label="Top P sampling"),
    ],
    theme=gr.themes.Soft(primary_hue=COLOR),
).queue().launch()