Update app_v3.py

app_v3.py

@@ -23,7 +23,6 @@ from typing import Generator
 import gradio as gr
 from huggingface_hub import CommitScheduler, HfApi, logging
 from debug import log_params, scheduler
-import torch._dynamo
 logging.set_verbosity_debug()
 
 huggingface_token = os.getenv("HUGGINFACE_TOKEN")
@@ -47,13 +46,13 @@ pipe = FluxControlNetPipeline.from_pretrained(
 )
 pipe.to("cuda")
 
-torch._dynamo.config.suppress_errors = True
+# torch._dynamo.config.suppress_errors = True
 # For FLUX models, compiling VAE decode can also be beneficial if needed, though UNet is primary.
 # pipe.vae.decode = torch.compile(pipe.vae.decode, mode="reduce-overhead", fullgraph=True) # Uncomment if VAE compile helps
-try:
-    pipe.vae.decode = torch.compile(pipe.vae.decode, mode="default")
-except Exception as e:
-    print(f"Compile failed: {e}")
+# try:
+#     pipe.vae.decode = torch.compile(pipe.vae.decode, mode="default")
+# except Exception as e:
+#     print(f"Compile failed: {e}")
 
 # 2. Memory Efficient Attention (xFormers): Reduces memory usage and improves speed
 # Requires xformers library installation. Beneficial even with high VRAM.