File size: 7,005 Bytes
e352103 ad382c8 e352103 0a651e1 e352103 0a1a1a0 e352103 ad382c8 e352103 0a651e1 ad382c8 e352103 9ad1205 e352103 b926faa e352103 0029ec4 b926faa 0029ec4 e352103 0029ec4 e352103 0029ec4 e352103 0029ec4 e352103 3ada93a ad382c8 5af142a e352103 74ee333 e352103 b926faa e352103 7660e3c ad382c8 3ada93a ad382c8 63f23ec ad382c8 e352103 b926faa e352103 109d07a |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 |
import gradio as gr
from transformers import AutoProcessor, AutoModelForVision2Seq
import re
import time
from PIL import Image
import torch
import spaces
import subprocess
#subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM-Instruct")
model = AutoModelForVision2Seq.from_pretrained("HuggingFaceTB/SmolVLM-Instruct",
torch_dtype=torch.bfloat16,
#_attn_implementation="flash_attention_2"
).to("cuda")
@spaces.GPU
def model_inference(
images, text, assistant_prefix, decoding_strategy, temperature, max_new_tokens,
repetition_penalty, top_p
):
if text == "" and not images:
gr.Error("Please input a query and optionally image(s).")
if text == "" and images:
gr.Error("Please input a text query along the image(s).")
if isinstance(images, Image.Image):
images = [images]
resulting_messages = [
{
"role": "user",
"content": [{"type": "image"}] + [
{"type": "text", "text": text}
]
}
]
if assistant_prefix:
text = f"{assistant_prefix} {text}"
prompt = processor.apply_chat_template(resulting_messages, add_generation_prompt=True)
inputs = processor(text=prompt, images=[images], return_tensors="pt")
inputs = {k: v.to("cuda") for k, v in inputs.items()}
generation_args = {
"max_new_tokens": max_new_tokens,
"repetition_penalty": repetition_penalty,
}
assert decoding_strategy in [
"Greedy",
"Top P Sampling",
]
if decoding_strategy == "Greedy":
generation_args["do_sample"] = False
elif decoding_strategy == "Top P Sampling":
generation_args["temperature"] = temperature
generation_args["do_sample"] = True
generation_args["top_p"] = top_p
generation_args.update(inputs)
# Generate
generated_ids = model.generate(**generation_args)
generated_texts = processor.batch_decode(generated_ids[:, generation_args["input_ids"].size(1):], skip_special_tokens=True)
return generated_texts[0]
with gr.Blocks(fill_height=True) as demo:
gr.Markdown("## SmolVLM: Small yet Mighty 💫")
gr.Markdown("Play with [HuggingFaceTB/SmolVLM-Instruct](https://huggingface.co/HuggingFaceTB/SmolVLM-Instruct) in this demo. To get started, upload an image and text or try one of the examples.")
with gr.Column():
image_input = gr.Image(label="Upload your Image", type="pil", scale=1)
query_input = gr.Textbox(label="Prompt")
assistant_prefix = gr.Textbox(label="Assistant Prefix", placeholder="Let's think step by step.")
submit_btn = gr.Button("Submit")
output = gr.Textbox(label="Output")
with gr.Accordion(label="Examples and Advanced Generation Parameters", open=False):
examples=[
["example_images/rococo.jpg", "What art era is this?", None, "Greedy", 0.4, 512, 1.2, 0.8],
["example_images/examples_wat_arun.jpg", "Give me travel tips for the area around this monument.", None, "Greedy", 0.4, 512, 1.2, 0.8],
["example_images/examples_invoice.png", "What is the due date and the invoice date?", None, "Greedy", 0.4, 512, 1.2, 0.8],
["example_images/s2w_example.png", "What is this UI about?", None, "Greedy", 0.4, 512, 1.2, 0.8],
["example_images/examples_weather_events.png", "Where do the severe droughts happen according to this diagram?", None, "Greedy", 0.4, 512, 1.2, 0.8],
]
# Hyper-parameters for generation
max_new_tokens = gr.Slider(
minimum=8,
maximum=1024,
value=512,
step=1,
interactive=True,
label="Maximum number of new tokens to generate",
)
repetition_penalty = gr.Slider(
minimum=0.01,
maximum=5.0,
value=1.2,
step=0.01,
interactive=True,
label="Repetition penalty",
info="1.0 is equivalent to no penalty",
)
temperature = gr.Slider(
minimum=0.0,
maximum=5.0,
value=0.4,
step=0.1,
interactive=True,
label="Sampling temperature",
info="Higher values will produce more diverse outputs.",
)
top_p = gr.Slider(
minimum=0.01,
maximum=0.99,
value=0.8,
step=0.01,
interactive=True,
label="Top P",
info="Higher values is equivalent to sampling more low-probability tokens.",
)
decoding_strategy = gr.Radio(
[
"Greedy",
"Top P Sampling",
],
value="Greedy",
label="Decoding strategy",
interactive=True,
info="Higher values is equivalent to sampling more low-probability tokens.",
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(
visible=(
selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
)
),
inputs=decoding_strategy,
outputs=temperature,
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(
visible=(
selection in ["contrastive_sampling", "beam_sampling", "Top P Sampling", "sampling_top_k"]
)
),
inputs=decoding_strategy,
outputs=repetition_penalty,
)
decoding_strategy.change(
fn=lambda selection: gr.Slider(visible=(selection in ["Top P Sampling"])),
inputs=decoding_strategy,
outputs=top_p,
)
gr.Examples(
examples = examples,
inputs=[image_input, query_input, assistant_prefix, decoding_strategy, temperature,
max_new_tokens, repetition_penalty, top_p],
outputs=output,
fn=model_inference, cache_examples=False
)
submit_btn.click(model_inference, inputs = [image_input, query_input, assistant_prefix, decoding_strategy, temperature,
max_new_tokens, repetition_penalty, top_p], outputs=output)
demo.launch(debug=True) |