Update app.py

app.py

@@ -44,6 +44,12 @@ base_model_id = "hunyuanvideo-community/HunyuanVideo"
 
 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
 
+offload_config=OffloadConfig(
+    high_cpu_memory=True,
+    parameters_level=True,
+    compiler_transformer=False,
+)
+
 def init_predictor():
     global pipe
     text_encoder = LlamaModel.from_pretrained(
@@ -63,26 +69,21 @@ def init_predictor():
     quantize_(transformer, float8_weight_only(), device="cpu")
     transformer.to("cpu")
     torch.cuda.empty_cache()
-    offload_config=OffloadConfig(
-            high_cpu_memory=True,
-            parameters_level=True,
-            compiler_transformer=False,
-    )
     pipe = SkyreelsVideoPipeline.from_pretrained(
         base_model_id,
         transformer=transformer,
         text_encoder=text_encoder,
         torch_dtype=torch.bfloat16,
     ).to("cpu")
-    Offload.offload(
-            pipeline=pipe,
-            config=offload_config,
-    )
     
 negative_prompt = "Aerial view, aerial view, overexposed, low quality, deformation, a poor composition, bad hands, bad teeth, bad eyes, bad limbs, distortion"
 
 @spaces.GPU(duration=60)
 def generate(segment, image, prompt, size, guidance_scale, num_inference_steps, frames, seed, progress=gr.Progress(track_tqdm=True) ):
+    Offload.offload(
+        pipeline=pipe,
+        config=offload_config,
+    )
     random.seed(time.time())
     seed = int(random.randrange(4294967294))
     if segment==1: