app.py

import os
from threading import Thread
from typing import Iterator

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

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


DESCRIPTION = """\
# Gemma 3 270m IT 💎💬

Try this mini model by Google.

[🪪 **Model card**](https://huggingface.co/google/gemma-3-270m-it)
"""

MAX_MAX_NEW_TOKENS = 2048
DEFAULT_MAX_NEW_TOKENS = 1024
MAX_INPUT_TOKEN_LENGTH = int(os.getenv("MAX_INPUT_TOKEN_LENGTH", "4096"))

#device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Pick attention backend based on device availability
if torch.cuda.is_available():
    device = "cuda"
    attn_impl = "flash_attention_2"   # or "flash" depending on the library
    torch_dtype = torch.bfloat16      # or torch.float16
else:
    device = "cpu"
    attn_impl = "eager"
    torch_dtype = torch.bfloat16      # or float32, bfloat16 supported on CPUs with AVX512-BF16 or AMX (e.g., Intel Ice Lake / Sapphire Rapids, some newer AMD). But many ops may still fall back to float32.

# model_id = "google/gemma-3-270m-it"
model_id = "unsloth/gemma-3-270m-it"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True,)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    attn_implementation=attn_impl,
    trust_remote_code=True,
)
model.config.sliding_window = 4096
model.eval()


@spaces.GPU(duration=90)
def generate(
    message: str,
    chat_history: list[tuple[str, str]],
    system_message: str = "",
    max_new_tokens: int = 1024,
    temperature: float = 0.001,
    top_p: float = 1.0,
    top_k: int = 50,
    repetition_penalty: float = 1.0,
) -> Iterator[str]:
    conversation = [{"role": "system", "content": system_message}]
    for user, assistant in chat_history:
        conversation.extend(
            [
                {"role": "user", "content": user},
                {"role": "assistant", "content": assistant},
            ]
        )
    conversation.append({"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 from conversation as it was longer than {MAX_INPUT_TOKEN_LENGTH} tokens.")
    input_ids = input_ids.to(model.device)

    streamer = TextIteratorStreamer(tokenizer, timeout=20.0, skip_prompt=True, skip_special_tokens=True)
    generate_kwargs = dict(
        {"input_ids": input_ids},
        streamer=streamer,
        max_new_tokens=max_new_tokens,
        do_sample=True,
        top_p=top_p,
        top_k=top_k,
        temperature=temperature,
        num_beams=1,
        repetition_penalty=repetition_penalty,
        disable_compile=True,  #  https://ai.google.dev/gemma/docs/core/huggingface_text_full_finetune#test_model_inference
    )
    t = Thread(target=model.generate, kwargs=generate_kwargs)
    t.start()

    outputs = []
    for text in streamer:
        outputs.append(text)
        yield "".join(outputs)


chat_interface = gr.ChatInterface(
    fn=generate,
    additional_inputs=[
        gr.Textbox(
            value="",
            label="System message",
            render=False,
        ),        
        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=1.0,  # default from https://huggingface.co/docs/transformers/en/main_classes/text_generation
        ),
        gr.Slider(
            label="Top-p (nucleus sampling)",
            minimum=0.05,
            maximum=1.0,
            step=0.05,
            value=0.95,  # from https://huggingface.co/google/gemma-3-270m-it/blob/main/generation_config.json
        ),
        gr.Slider(
            label="Top-k",
            minimum=1,
            maximum=1000,
            step=1,
            value=64,  # from https://huggingface.co/google/gemma-3-270m-it/blob/main/generation_config.json
        ),
        gr.Slider(
            label="Repetition penalty",
            minimum=1.0,
            maximum=2.0,
            step=0.05,
            value=1.0,  # default from https://huggingface.co/docs/transformers/en/main_classes/text_generation
        ),
    ],
    stop_btn=None,
    examples = [
        ["Hi! How are you?"],
        ["Pros and cons of a long-term relationship. Bullet list with max 3 pros and 3 cons, concise."],
        ["How many hours does it take a man to eat a helicopter?"],
        ["How do you open a JSON file in Python?"],
        ["Make a bullet list of pros and cons of living in San Francisco. Maximum 2 pros and 2 cons."],
        ["Invent a short story with animals about the value of friendship."],
        ["Can you briefly explain what the Python programming language is?"],
        ["Write a 100-word article on 'Benefits of Open-Source in AI Research'."],
    ],
    cache_examples=False,
)

with gr.Blocks(css="style.css", fill_height=True, theme="soft") as demo:
    gr.Markdown(DESCRIPTION)
    gr.DuplicateButton(value="Duplicate Space for private use", elem_id="duplicate-button")
    chat_interface.render()

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