Hugging Face's logo

Spaces:
m-ric
/
Aria
Running on Zero

App Files Community
andito HF staff commited on
Commit
d30dedf
•
1 Parent(s): 0fcfdf2

Update app.py

Browse files
Files changed (1) hide show
  1. app.py +4 -8
app.py CHANGED
@@ -13,11 +13,7 @@ processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM_converted_4")
13
 
14
  model = Idefics3ForConditionalGeneration.from_pretrained("HuggingFaceTB/SmolVLM_converted_4",
15
  torch_dtype=torch.bfloat16,
16
- #_attn_implementation="flash_attention_2"
17
- ).to("cuda")
18
-
19
- BAD_WORDS_IDS = processor.tokenizer(["<image>", "<fake_token_around_image>", "<row_", "apiro", "\u2500lrow_", "row_1"], add_special_tokens=False).input_ids
20
- EOS_WORDS_IDS = [processor.tokenizer.eos_token_id]
21
 
22
  @spaces.GPU
23
  def model_inference(
@@ -78,9 +74,9 @@ def model_inference(
78
 
79
 
80
  with gr.Blocks(fill_height=True) as demo:
81
- gr.Markdown("## IDEFICS3-Llama 🐶")
82
- gr.Markdown("Play with [HuggingFaceM4/Idefics3-8B-Llama3](https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3) in this demo. To get started, upload an image and text or try one of the examples.")
83
- gr.Markdown("**Disclaimer:** Idefics3 does not include an RLHF alignment stage, so it may not consistently follow prompts or handle complex tasks. However, this doesn't mean it is incapable of doing so. Adding a prefix to the assistant's response, such as Let's think step for a reasoning question or `<html>` for HTML code generation, can significantly improve the output in practice. You could also play with the parameters such as the temperature in non-greedy mode.")
84
  with gr.Column():
85
  image_input = gr.Image(label="Upload your Image", type="pil", scale=1)
86
  query_input = gr.Textbox(label="Prompt")
 
13
 
14
  model = Idefics3ForConditionalGeneration.from_pretrained("HuggingFaceTB/SmolVLM_converted_4",
15
  torch_dtype=torch.bfloat16,
16
+ _attn_implementation="flash_attention_2").to("cuda")
 
 
 
 
17
 
18
  @spaces.GPU
19
  def model_inference(
 
74
 
75
 
76
  with gr.Blocks(fill_height=True) as demo:
77
+ gr.Markdown("## SmolVLM 🐶")
78
+ gr.Markdown("Play with [HuggingFaceTB/SmolVLM](https://huggingface.co/HuggingFaceTB/SmolVLM) in this demo. To get started, upload an image and text or try one of the examples.")
79
+ gr.Markdown("**Disclaimer:** SmolVLM does not include an RLHF alignment stage, so it may not consistently follow prompts or handle complex tasks. However, this doesn't mean it is incapable of doing so. Adding a prefix to the assistant's response, such as Let's think step for a reasoning question or `<html>` for HTML code generation, can significantly improve the output in practice. You could also play with the parameters such as the temperature in non-greedy mode.")
80
  with gr.Column():
81
  image_input = gr.Image(label="Upload your Image", type="pil", scale=1)
82
  query_input = gr.Textbox(label="Prompt")