web_gradio.py

import os
import sys

import fire
import gradio as gr
import torch
import transformers
from peft import PeftModel
from transformers import GenerationConfig, AutoModelForCausalLM, AutoTokenizer

from utils.callbacks import Iteratorize, Stream
from utils.prompter import Prompter

if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

try:
    if torch.backends.mps.is_available():
        device = "mps"
except:  # noqa: E722
    pass


def main(
    load_8bit: bool = False,
    base_model: str = "",
    lora_weights: str = "", #tloen/alpaca-lora-7b",
    prompt_template: str = "",  # The prompt template to use, will default to alpaca.
    server_name: str = "0.0.0.0",  # Allows to listen on all interfaces by providing '0.
    share_gradio: bool = True,
    pad_token: str = None,
    bos_token: str = None, 
    eos_token: str = None,
    qwen_api: bool = False, 
):
    base_model = base_model or os.environ.get("BASE_MODEL", "")
    assert (
        base_model
    ), "Please specify a --base_model, e.g. --base_model='huggyllama/llama-7b'"

    prompter = Prompter(prompt_template)
    tokenizer = AutoTokenizer.from_pretrained(
        base_model, 
        trust_remote_code=True,
        pad_token=pad_token,
        bos_token=bos_token,
        eos_token=eos_token,
    )
    if device == "cuda":
        model = AutoModelForCausalLM.from_pretrained(
            base_model,
            load_in_8bit=load_8bit,
            torch_dtype=torch.float16,
            device_map="auto",
            trust_remote_code=True
        )
        if len(lora_weights) > 0:
            model = PeftModel.from_pretrained(
                model,
                lora_weights,
                torch_dtype=torch.float16,
            )
    elif device == "mps":
        model = AutoModelForCausalLM.from_pretrained(
            base_model,
            device_map={"": device},
            torch_dtype=torch.float16,
            trust_remote_code=True
        )
        if len(lora_weights) > 0:
            model = PeftModel.from_pretrained(
                model,
                lora_weights,
                device_map={"": device},
                torch_dtype=torch.float16,
            )
    else:
        model = AutoModelForCausalLM.from_pretrained(
            base_model, device_map={"": device}, low_cpu_mem_usage=True, trust_remote_code=True
        )
        if len(lora_weights) > 0:
            model = PeftModel.from_pretrained(
                model,
                lora_weights,
                device_map={"": device},
            )

    # unwind broken decapoda-research config
    #model.config.pad_token_id = tokenizer.pad_token_id = 0  # unk
    #model.config.bos_token_id = 1
    #model.config.eos_token_id = 2
    print("pad", tokenizer.pad_token, tokenizer.pad_token_id)
    print("bos", tokenizer.bos_token, tokenizer.bos_token_id)
    print("eos", tokenizer.eos_token, tokenizer.eos_token_id)

    #if not load_8bit:
    #    model.half()  # seems to fix bugs for some users.

    model.eval()
    if torch.__version__ >= "2" and sys.platform != "win32":
        model = torch.compile(model)

    def evaluate(
        instruction,
        max_new_tokens=128,
        stream_output=True,
        **kwargs,
    ):
        prompt = prompter.generate_prompt(instruction)
        inputs = tokenizer(prompt, return_tensors="pt")
        input_ids = inputs["input_ids"].to(device)
        generation_config = GenerationConfig.from_pretrained(
            base_model
        )
        '''
        generation_config = GenerationConfig(
            temperature=temperature,
            top_p=top_p,
            top_k=top_k,
            num_beams=num_beams,
            do_sample=True,
            repetition_penalty=1.3,
            **kwargs,
        )
        '''

        generate_params = {
            "input_ids": input_ids,
            "generation_config": generation_config,
            "return_dict_in_generate": True,
            "output_scores": True,
            "max_new_tokens": max_new_tokens,
        }

        if stream_output:
            # Stream the reply 1 token at a time.
            # This is based on the trick of using 'stopping_criteria' to create an iterator,
            # from https://github.com/oobabooga/text-generation-webui/blob/ad37f396fc8bcbab90e11ecf17c56c97bfbd4a9c/modules/text_generation.py#L216-L243.

            def generate_with_callback(callback=None, **kwargs):
                kwargs.setdefault(
                    "stopping_criteria", transformers.StoppingCriteriaList()
                )
                kwargs["stopping_criteria"].append(
                    Stream(callback_func=callback)
                )
                with torch.no_grad():
                    model.generate(**kwargs)

            def generate_with_streaming(**kwargs):
                return Iteratorize(
                    generate_with_callback, kwargs, callback=None
                )

            with generate_with_streaming(**generate_params) as generator:
                for output in generator:
                    # new_tokens = len(output) - len(input_ids[0])
                    decoded_output = tokenizer.decode(output)

                    if output[-1] in [tokenizer.eos_token_id]:
                        break

                    yield prompter.get_response(decoded_output)
            return  # early return for stream_output

        # Without streaming
        with torch.no_grad():
            generation_output = model.generate(
                input_ids=input_ids,
                generation_config=generation_config,
                return_dict_in_generate=True,
                output_scores=True,
                max_new_tokens=max_new_tokens,
            )
        s = generation_output.sequences[0]
        output = tokenizer.decode(s)
        print(output)
        yield prompter.get_response(output)

    gr.Interface(
        fn=evaluate,
        inputs=[
            gr.components.Textbox(
                lines=2,
                label="Instruction",
                placeholder="Tell me about alpacas.",
            ),
            gr.components.Slider(
                minimum=1, maximum=2000, step=1, value=128, label="Max tokens"
            ),
            gr.components.Checkbox(label="Stream output", value=True),
        ],
        outputs=[
            gr.components.Textbox(
                lines=5,
                label="Output",
            )
        ],
        title="🦙🌲 LLM utils",
    ).queue().launch(server_name="0.0.0.0", share=False, prevent_thread_lock=qwen_api)
    # Old testing code follows.

    if qwen_api and prompt_template == 'qwen':
        from qwen_openai_api import run_server
        run_server(tokenizer, model, "0.0.0.0", 8000)

    """
    # testing code for readme
    for instruction in [
        "Tell me about alpacas.",
        "Tell me about the president of Mexico in 2019.",
        "Tell me about the king of France in 2019.",
        "List all Canadian provinces in alphabetical order.",
        "Write a Python program that prints the first 10 Fibonacci numbers.",
        "Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'.",  # noqa: E501
        "Tell me five words that rhyme with 'shock'.",
        "Translate the sentence 'I have no mouth but I must scream' into Spanish.",
        "Count up from 1 to 500.",
    ]:
        print("Instruction:", instruction)
        print("Response:", evaluate(instruction))
        print()
    """


if __name__ == "__main__":
    fire.Fire(main)