Fix: load model on CPU, move to GPU inside @spaces.GPU

app.py

@@ -8,14 +8,12 @@ import torch
 from diffusers import AutoencoderKLWan, WanPipeline
 from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
 
-# ── Load model once at startup ──────────────────────────────────────────
-# 1.3B model — fast and fits easily in A10G (24GB)
+# ── Load model on CPU at startup (ZeroGPU moves to GPU automatically) ───
 MODEL_ID = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
 
 vae = AutoencoderKLWan.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
 pipe = WanPipeline.from_pretrained(MODEL_ID, vae=vae, torch_dtype=torch.bfloat16)
 pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=3.0)
-pipe.to("cuda")
 
 # ── Resolution presets ──────────────────────────────────────────────────
 RESOLUTIONS = {
@@ -29,7 +27,9 @@ RESOLUTIONS = {
 def generate(prompt, negative_prompt, resolution, steps, guidance, seed, progress=gr.Progress(track_tqdm=True)):
     width, height = RESOLUTIONS.get(resolution, (480, 832))
 
-    generator = torch.Generator("cuda").manual_seed(seed) if seed >= 0 else None
+    pipe.to("cuda")
+
+    generator = torch.Generator("cuda").manual_seed(int(seed)) if seed >= 0 else None
 
     output = pipe(
         prompt=prompt,