Disable FSDP for single GPU inference to ensure GPU usage

app.py

@@ -110,8 +110,11 @@ if torch.cuda.is_available():
     print(f"✅ CUDA available! Device: {torch.cuda.get_device_name(0)}")
     print(f"   CUDA Version: {torch.version.cuda}")
     print(f"   PyTorch Version: {torch.__version__}")
+    print(f"   GPU Count: {torch.cuda.device_count()}")
+    print(f"   Current Device: {torch.cuda.current_device()}")
 else:
     print("⚠️  CUDA not available. Make sure GPU hardware is selected in Space settings.")
+    print(f"   PyTorch Version: {torch.__version__}")
 
 def generate_image(prompt, aspect_ratio, cfg, seed, checkpoint_file, config_path):
     """Generate image from text prompt."""

sample.py

@@ -205,7 +205,8 @@ def main(args: argparse.Namespace) -> None:
                 args.context_length = args.local_attn_window - 1
 
     # Override some settings for sampling
-    args.fsdp = 1  # sampling using FSDP if available.
+    # Disable FSDP for single GPU inference (FSDP can cause CPU fallback)
+    args.fsdp = 0  # Disable FSDP for single GPU - use regular GPU inference
     if args.use_pretrained_lm is not None:
         args.text = args.use_pretrained_lm