Update app.py

app.py

@@ -35,12 +35,17 @@ pipe = DiffusionPipeline.from_pretrained(
     ),
     tokenizer=tokenizer,
     feature_extractor=feature_extractor,
-    torch_dtype=torch.bfloat16,
-    use_safetensors=True,
-    device_map="balanced",          # automatically spreads submodules CPU/GPU
-    offload_folder="offload"    # where to spill CPU-offloaded weights
+    torch_dtype=torch.float16,   # load weights in half-precision
+    revision="fp16",             # if your repo provides fp16 weights
+    use_safetensors=True
 )
 
+# 4) Memory savings hooks (all on your single GPU + CPU offload)
+pipe.enable_attention_slicing()                     # slice big attention maps
+pipe.enable_vae_slicing()                           # slice VAE decode
+pipe.enable_xformers_memory_efficient_attention()   # if xformers is installed
+pipe.enable_model_cpu_offload()                     # offload idle submodules to CPU
+
 pipe = pipe.to(device)
 
 MAX_SEED = np.iinfo(np.int32).max