app.py

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

DESCRIPTION = """\
# EvaByte [Byte-Level LLM]

EvaByte is a efficient　byte-level language model with multibyte prediction and EVA attention, built by the University of Hong Kong and SambaNova Systems.  
This Space is an unofficial demo of the instruction-tuned version [EvaByte/EvaByte-SFT](https://huggingface.co/EvaByte/EvaByte-SFT).  
For full details on architecture, training recipe, and benchmarks, see their blog post and the project repository:

- Blog: <https://hkunlp.github.io/blog/2025/evabyte>  
- GitHub: <https://github.com/OpenEvaByte/evabyte>

If you liked this Space, follow me on Twitter: [@KantaHayashiAI](https://x.com/KantaHayashiAI)
"""

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = 32000

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

tokenizer = AutoTokenizer.from_pretrained("EvaByte/EvaByte-SFT", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "EvaByte/EvaByte-SFT",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).eval().to(device)


@spaces.GPU(duration=60)
def generate(
    message: str,
    chat_history: list[dict],
    max_new_tokens: int = DEFAULT_MAX_NEW_TOKENS,
    temperature: float = 0.6,
    top_p: float = 0.9,
) -> str:                                  

    conversation = [*chat_history, {"role": "user", "content": message}]
    input_ids = tokenizer.apply_chat_template(
        conversation,
        add_generation_prompt=True,
        return_tensors="pt"
    )

    if input_ids.shape[1] > MAX_INPUT_TOKEN_LENGTH:
        input_ids = input_ids[:, -MAX_INPUT_TOKEN_LENGTH:]
        gr.Warning(
            f"Trimmed input to the last {MAX_INPUT_TOKEN_LENGTH} tokens　because it exceeded the limit."
        )

    input_ids = input_ids.to(model.device)

    output_ids = model.multi_byte_generate(  
        input_ids=input_ids,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        temperature=temperature,
    )

    generated_segment = output_ids[0][input_ids.shape[1]:]
    return tokenizer.decode(generated_segment, skip_special_tokens=True)


demo = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Slider(
            label="Max new tokens",
            minimum=1,
            maximum=MAX_MAX_NEW_TOKENS,
            step=1,
            value=DEFAULT_MAX_NEW_TOKENS,
        ),
        gr.Slider(
            label="Temperature",
            minimum=0,
            maximum=4.0,
            step=0.1,
            value=0,
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=1.0,
        ),
    ],
    stop_btn=None,           
    examples=[["Write me an English pangram."]],
    cache_examples=False,
    type="messages",
    description=DESCRIPTION,
    fill_height=True,
)

if __name__ == "__main__":
    demo.queue(max_size=20).launch()