Gradio: cache examples in a lazy manner

This should avoid runtime errors on startup, by lack of gpu quota
requirements update: lazy cache support comes after 4.25
See https://www.gradio.app/changelog

Signed-off-by: Raphael Glon <oOraph@users.noreply.github.com>

app.py

@@ -21,7 +21,7 @@ if not torch.cuda.is_available():
     DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU. </p>"
 IS_COLAB = utils.is_google_colab() or os.getenv("IS_COLAB") == "1"
 HF_TOKEN = os.getenv("HF_TOKEN")
-CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES") == "1"
+CACHE_EXAMPLES = 'lazy' if torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES") == "1" else False
 MIN_IMAGE_SIZE = int(os.getenv("MIN_IMAGE_SIZE", "512"))
 MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "2048"))
 USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE") == "1"

requirements.txt

@@ -1,6 +1,6 @@
 accelerate==0.27.2
 diffusers==0.26.3
-gradio==4.20.0
+gradio>=4.25.0
 invisible-watermark==0.2.0
 Pillow==10.2.0
 spaces==0.24.0