feat: enable HF_TRANSFER, improve model loading logs, and increase GPU duration to 120s

app.py

@@ -1,4 +1,5 @@
 import os
+os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 import uuid
 import torch
 from diffusers import ErnieImagePipeline
@@ -10,17 +11,22 @@ import spaces
 torch.set_float32_matmul_precision("high")
 
 # Initialize Pipeline
-pipe = ErnieImagePipeline.from_pretrained(
-    "Baidu/ERNIE-Image-Turbo",
-    torch_dtype=torch.bfloat16,
-)
-
-if torch.cuda.is_available():
+print("Loading model Baidu/ERNIE-Image-Turbo... this may take a few minutes!", flush=True)
+try:
+    pipe = ErnieImagePipeline.from_pretrained(
+        "Baidu/ERNIE-Image-Turbo",
+        torch_dtype=torch.bfloat16,
+    )
+    print("Model loaded successfully. Moving to CUDA...", flush=True)
     pipe = pipe.to("cuda")
+    print("Model is on CUDA. Initializing Server...", flush=True)
+except Exception as e:
+    print(f"Error during model loading: {e}", flush=True)
+    raise
 
 app = Server()
 
-@spaces.GPU
+@spaces.GPU(duration=120)
 @app.api()
 def generate_image(prompt: str, width: int = 1024, height: int = 1024, guidance_scale: float = 1.0, num_inference_steps: int = 8, use_pe: bool = True) -> FileData:
     """Generate an image using ERNIE-Image-Turbo."""

requirements.txt

@@ -5,3 +5,4 @@ transformers
 accelerate
 safetensors
 git+https://github.com/huggingface/diffusers.git
+hf_transfer