Reduce ORT memory: disable prepacking, basic optimization, 1 thread

app.py

@@ -50,12 +50,15 @@ def _load_all():
     print("[init] Downloading VAE...")
     vae_path = _download(repo, "vae_full.pt", MODELS_DIR, token)
 
-    print("[init] Creating ONNX Runtime session...")
+    print("[init] Creating ONNX Runtime session (mmap + low memory)...")
     t0 = time.time()
     opts = ort.SessionOptions()
-    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
-    opts.inter_op_num_threads = 2
+    opts.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_BASIC
+    opts.inter_op_num_threads = 1
     opts.intra_op_num_threads = 2
+    opts.enable_mem_pattern = True
+    opts.enable_mem_reuse = True
+    opts.add_session_config_entry("session.disable_prepacking", "1")
     dit_session = ort.InferenceSession(onnx_path, opts, providers=["CPUExecutionProvider"])
     print(f"[init] DiT session ready in {time.time() - t0:.0f}s")