Make it compatible with ZeroGPU and add `trigger_mode`

app.py

@@ -1,3 +1,4 @@
+import spaces
 from diffusers import (
     StableDiffusionXLPipeline,
     EulerDiscreteScheduler,
@@ -55,6 +56,7 @@ unet.load_state_dict(load_file(hf_hub_download(REPO, CHECKPOINT), device="cuda")
 pipe = StableDiffusionXLPipeline.from_pretrained(
     BASE, unet=unet, torch_dtype=torch.float16, variant="fp16", safety_checker=False
 ).to("cuda")
+unet = unet.to(dtype=torch.float16)
 
 if USE_TAESD:
     pipe.vae = AutoencoderTiny.from_pretrained(
@@ -115,6 +117,7 @@ if SFAST_COMPILE:
     pipe = compile(pipe, config)
 
 
+@spaces.GPU
 def predict(prompt, seed=1231231):
     generator = torch.manual_seed(seed)
     last_time = time.time()
@@ -202,7 +205,7 @@ pipe("A girl smiling", num_inference_steps=2, guidance_scale=0).images[0].save("
         generate_bt.click(
             fn=predict, inputs=inputs, outputs=outputs, show_progress=False
         )
-        prompt.input(fn=predict, inputs=inputs, outputs=outputs, show_progress=False)
+        prompt.input(fn=predict, inputs=inputs, outputs=outputs, trigger_mode="always_last", show_progress=False)
         seed.change(fn=predict, inputs=inputs, outputs=outputs, show_progress=False)
 
 demo.queue()