Spaces:
Running
on
Zero
Running
on
Zero
File size: 3,582 Bytes
abe7a5d 24306aa 04a03fc a5dc4f3 04a03fc abe7a5d ce7f04f de49e6a 6f080e6 ab9b2ec abe7a5d 64b2a00 abe7a5d 74c9a68 abe7a5d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 |
"""
File: app.py
Description: Chat with the vision language model Gemma3.
Author: Didier Guillevic
Date: 2025-03-16
"""
import spaces
from huggingface_hub import login, whoami
import os
token = os.getenv('HF_TOKEN')
login(token=token)
import gradio as gr
from transformers import AutoProcessor, Gemma3ForConditionalGeneration
from transformers import TextIteratorStreamer
from threading import Thread
import torch
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_id = "google/gemma-3-4b-it"
processor = AutoProcessor.from_pretrained(model_id, use_fast=True, padding_side="left")
model = Gemma3ForConditionalGeneration.from_pretrained(
model_id,
torch_dtype=torch.bfloat16
).to(device).eval()
@torch.inference_mode()
@spaces.GPU
def process(message, history):
"""Generate the model response in streaming mode given message and history
"""
print(f"{history=}")
# Get the user's text and list of images
user_text = message.get("text", "")
user_images = message.get("files", []) # List of images
# Build the message list including history
messages = []
combined_user_input = [] # Combine images and text if found in same turn.
for user_turn, bot_turn in history:
if isinstance(user_turn, tuple): # Image input
image_content = [{"type": "image", "url": image_url} for image_url in user_turn]
combined_user_input.extend(image_content)
elif isinstance(user_turn, str): # Text input
combined_user_input.append({"type":"text", "text": user_turn})
if combined_user_input and bot_turn:
messages.append({'role': 'user', 'content': combined_user_input})
messages.append({'role': 'assistant', 'content': [{"type": "text", "text": bot_turn}]})
combined_user_input = [] # reset the combined user input.
# Build the user message's content from the provided message
user_content = []
if user_text:
user_content.append({"type": "text", "text": user_text})
for image in user_images:
user_content.append({"type": "image", "url": image})
messages.append({'role': 'user', 'content': user_content})
# Generate model's response
inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt"
).to(model.device, dtype=torch.bfloat16)
streamer = TextIteratorStreamer(
processor, skip_prompt=True, skip_special_tokens=True)
generation_kwargs = dict(
inputs,
streamer=streamer,
max_new_tokens=1_024,
do_sample=False
)
thread = Thread(target=model.generate, kwargs=generation_kwargs)
thread.start()
partial_message = ""
for new_text in streamer:
partial_message += new_text
yield partial_message
#
# User interface
#
with gr.Blocks() as demo:
chat_interface = gr.ChatInterface(
fn=process,
title="Multimedia Chat",
description="Chat with text or text+image.",
multimodal=True,
examples=[
"How can we rationalize quantum entanglement?",
"Peux-tu expliquer le terme 'quantum spin'?",
{'files': ['./sample_ID.jpeg',], 'text': 'Describe this image in a few words.'},
{
'files': ['./sample_ID.jpeg',],
'text': (
'Could you extract the information present in the image '
'and present it as a bulleted list?')
},
]
)
demo.launch()
|