Optimizing

app.py

@@ -16,6 +16,10 @@ def warmup():
     Run once on Space startup, on CPU only.
     Downloads models & builds pipeline into cache.
     """
+    torch.backends.cuda.matmul.allow_tf32 = True
+    torch.backends.cudnn.allow_tf32 = True
+    torch.set_float32_matmul_precision("high")
+    
     print("🔥 Warmup starting: downloading models & building pipeline on CPU...")
     pipe = load_pipeline()
     print("🔥 Warmup done. Pipeline ready on CPU.")

diffqrcoder_wrapper.py

@@ -1,6 +1,6 @@
 # diffqrcoder_wrapper.py
 import torch
-from diffusers import ControlNetModel, DDIMScheduler
+from diffusers import ControlNetModel, DPMSolverMultistepScheduler
 from PIL import Image
 import qrcode
 from huggingface_hub import hf_hub_download
@@ -71,12 +71,17 @@ def load_pipeline():
         use_auth_token=True,  # uses the Space's HF token
     )
 
-    pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+    pipe.scheduler = DPMSolverMultistepScheduler.from_config(pipe.scheduler.config)
 
     # Memory helpers – cheaper attention
     try:
-        pipe.enable_attention_slicing()
-        # Optional: pipe.enable_xformers_memory_efficient_attention()
+        # pipe.enable_attention_slicing()
+        try:
+            pipe.enable_xformers_memory_efficient_attention()
+            print("✅ xFormers attention enabled.")
+        except Exception as e:
+            print("⚠️ xFormers not available:", repr(e))
+
     except Exception as e:
         print("⚠️ Could not enable attention optimizations:", repr(e))
 

requirements.txt

@@ -12,4 +12,5 @@ opencv-python==4.11.0.86
 accelerate==1.3.0
 safetensors
 Pillow
-qrcode
+qrcode
+xformers