app.py

# Imports
import os
import sys
import gradio as gr
import spaces
import torch
import librosa
from PIL import Image
from decord import VideoReader, cpu
from transformers import AutoModel, AutoTokenizer, AutoProcessor

# Variables
DEVICE = "auto"
if DEVICE == "auto":
    DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[SYSTEM] | Using {DEVICE} type compute device.")

DEFAULT_INPUT = "Describe in one paragraph."
MAX_FRAMES = 64

model_name = "openbmb/MiniCPM-o-2_6"

repo = AutoModel.from_pretrained(model_name, trust_remote_code=True, attn_implementation="sdpa", torch_dtype=torch.bfloat16).to(DEVICE)
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)

css = '''
.gradio-container{max-width: 560px !important}
h1{text-align:center}
footer {
    visibility: hidden
}
'''

def encode_video(video_path):
    def uniform_sample(idxs, n):
        gap = len(idxs) / n
        return [idxs[int(i*gap + gap/2)] for i in range(n)]

    vr = VideoReader(video_path, ctx=cpu(0))
    fps = round(vr.get_avg_fps())
    idxs = list(range(0, len(vr), fps))
    if len(idxs) > MAX_FRAMES:
        idxs = uniform_sample(idxs, MAX_FRAMES)
    frames = vr.get_batch(idxs).asnumpy()
    return [Image.fromarray(f.astype("uint8")) for f in frames]

@spaces.GPU(duration=60)
def generate(image, video, audio, instruction=DEFAULT_INPUT, sampling=False, temperature=0.7, top_p=0.8, top_k=100, repetition_penalty=1.05, max_tokens=512):

    print(image)
    print(video)
    print(audio)
    print(instruction)

    if image is not None:
        img = Image.fromarray(image.astype("uint8"), "RGB")
        inputs = [{"role": "user", "content": [img, instruction]}]
    elif video is not None:
        vid = encode_video(video)
        inputs = [{"role": "user", "content": vid + [instruction]}]
    elif audio is not None:
        aud, _ = librosa.load(audio, sr=16000, mono=True)
        inputs = [{"role": "user", "content": [aud, instruction]}]
    else:
        return "No input provided."

    params = {
        "msgs": inputs,
        "tokenizer": tokenizer,
        #"processor": processor,
        "sampling": sampling,
        "temperature": temperature,
        "top_p": top_p,
        "top_k": top_k,
        "repetition_penalty": repetition_penalty,
        "max_new_tokens": max_tokens,
    }

    output = repo.chat(**params)

    print(output)

    return output

def cloud():
    print("[CLOUD] | Space maintained.")

# Initialize
with gr.Blocks(css=css) as main:
    with gr.Column():
        input = gr.Image(label="Image")
        input_2 = gr.Video(label="Video")
        input_3 = gr.Audio(label="Audio", type="filepath")
        instruction = gr.Textbox(lines=1, value=DEFAULT_INPUT, label="Instruction")
        sampling = gr.Checkbox(value=False, label="Sampling")
        temperature = gr.Slider(minimum=0.01, maximum=1.99, step=0.01, value=0.7, label="Temperature")
        top_p = gr.Slider(minimum=0, maximum=1, step=0.01, value=0.8, label="Top P")
        top_k = gr.Slider(minimum=0, maximum=1000, step=1, value=100, label="Top K")
        repetition_penalty = gr.Slider(minimum=0.01, maximum=1.99, step=0.01, value=1.05, label="Repetition Penalty")
        max_tokens = gr.Slider(minimum=1, maximum=4096, step=1, value=512, label="Max Tokens")
        submit = gr.Button("▶")
        maintain = gr.Button("☁️")
        
    with gr.Column():
        output = gr.Textbox(lines=1, value="", label="Output")

    submit.click(fn=generate, inputs=[input, input_2, input_3, instruction, sampling, temperature, top_p, top_k, repetition_penalty, max_tokens], outputs=[output], queue=False)
    maintain.click(cloud, inputs=[], outputs=[], queue=False)

main.launch(show_api=True)