Update app.py

app.py

@@ -1,10 +1,9 @@
-import spaces
 import gradio as gr
 import torch
 from diffusers import DiffusionPipeline
 from diffusers.quantizers import PipelineQuantizationConfig
 import imageio
-
+import spaces
 
 # Checkpoint ID
 ckpt_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
@@ -20,25 +19,23 @@ quant_config = PipelineQuantizationConfig(
     components_to_quantize=["transformer", "text_encoder"]
 )
 
-# Load pipeline with quantization
+# Load pipeline with quantization, without immediately moving to CUDA to avoid ConstantVariable error
 pipe = DiffusionPipeline.from_pretrained(
     ckpt_id,
     quantization_config=quant_config,
     torch_dtype=torch.bfloat16
-).to("cuda")
+)
 
-# Optimize memory and performance
+# Enable CPU offload and compile after offload
 pipe.enable_model_cpu_offload()
 torch._dynamo.config.recompile_limit = 1000
 torch._dynamo.config.capture_dynamic_output_shape_ops = True
-pipe.transformer.compile()
 
-# Duration function using progress context safely
+# Duration function
 def get_duration(prompt, height, width, 
                  negative_prompt, duration_seconds,
                  guidance_scale, steps,
-                 seed, randomize_seed,
-                 progress=None):  # Added default None for progress
+                 seed, randomize_seed):
     if steps > 4 and duration_seconds > 2:
         return 90
     elif steps > 4 or duration_seconds > 2:
@@ -46,13 +43,14 @@ def get_duration(prompt, height, width,
     else:
         return 60
 
-# Gradio inference function with safe GPU duration
-@spaces.GPU(duration=50)
-def generate_video(prompt, seed, steps, duration_seconds,   progress=gr.Progress(track_tqdm=True)):
+# Gradio inference function with spaces GPU decorator
+@spaces.GPU(duration=30)
+def generate_video(prompt, seed, steps, duration_seconds):
     generator = torch.manual_seed(seed) if seed else None
     fps = 8
     num_frames = duration_seconds * fps if duration_seconds else 16
 
+    # Run pipeline on default device with automatic offload
     video_frames = pipe(
         prompt=prompt,
         num_frames=num_frames,