Fix: Remove engine caching to prevent LDR generation state corruption

code/demo.py

@@ -51,44 +51,7 @@ def get_clip_retriever_cached():
     print(f"✅ CLIP retriever loaded ({retriever.features.shape[0]} designs)")
     return retriever
 
-@functools.lru_cache(maxsize=1)
-def get_gpt_engine_cached():
-    """Lazy load GPT engine (initialized only once, cached)"""
-    print("🔧 Initializing GPT engine (one-time setup)...")
-
-    # Use absolute path relative to this file (fixes HF Spaces deployment)
-    config_path = os.path.join(os.path.dirname(__file__), 'cube3d/configs/open_model_v0.5.yaml')
-    gpt_ckpt_path = None  # test mode doesn't use this
-
-    # Detect HuggingFace Spaces environment
-    is_hf_space = os.getenv("SPACE_ID") is not None
-
-    if is_hf_space:
-        from huggingface_hub import hf_hub_download
-        print(f"Loading GPT model from HuggingFace Model Hub...")
-        shape_ckpt_path = hf_hub_download(
-            repo_id="0xZohar/object-assembler-models",
-            filename="save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors",
-            cache_dir=HF_CACHE_DIR,
-            local_files_only=False  # Allow runtime download on first use
-        )
-        save_gpt_ckpt_path = shape_ckpt_path
-        print(f"✅ GPT model loaded from cache: {shape_ckpt_path}")
-    else:
-        shape_ckpt_path = 'model_weights/save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors'
-        save_gpt_ckpt_path = 'model_weights/save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors'
-
-    # GPU T4 small environment: Use EngineFast with CUDA (required for t2t method)
-    device = torch.device("cuda")
-    print("✅ Using EngineFast on GPU (T4 small)")
-
-    engine = EngineFast(
-        config_path, gpt_ckpt_path, shape_ckpt_path, save_gpt_ckpt_path,
-        device=device,
-        mode='test'
-    )
-    print(f"✅ GPT engine initialized on {device}")
-    return engine
+# Removed cached engine - creates fresh instance each time to prevent state corruption
 
 # 确保临时目录存在（远程服务器路径）
 TMP_DIR = "./tmp/ldr_processor_demo"
@@ -333,22 +296,42 @@ def generate_ldr_gpu(ldr_content, ldr_path):
         List of predicted LDR lines
     """
     print("🤖 Running GPT model to generate new assembly sequence...")
+    print("   Using CUDA graphs, this will take some time to warmup and capture the graph.")
 
     stride = 5
     rot_num = 24
     bert_shift = 1
     shift = 0
 
-    # Lazy load GPT engine (cached, initialized only once)
-    engine = get_gpt_engine_cached()
+    # Prepare checkpoint paths
+    config_path = os.path.join(os.path.dirname(__file__), 'cube3d/configs/open_model_v0.5.yaml')
+    gpt_ckpt_path = None  # test mode doesn't use this
 
-    device = engine.device
-    if device.type == "cuda":
-        print("   Using CUDA graphs (this will take some time to warmup)")
+    # Detect HuggingFace Spaces environment
+    is_hf_space = os.getenv("SPACE_ID") is not None
+
+    if is_hf_space:
+        from huggingface_hub import hf_hub_download
+        checkpoint_path = hf_hub_download(
+            repo_id="0xZohar/object-assembler-models",
+            filename="save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors",
+            cache_dir=HF_CACHE_DIR,
+            local_files_only=False
+        )
+        shape_ckpt_path = checkpoint_path
+        save_gpt_ckpt_path = checkpoint_path
     else:
-        print("   Running on CPU (slower). Set GPT_DEVICE=cuda or enable GPU/ZeroGPU for faster runs.")
+        shape_ckpt_path = 'model_weights/save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors'
+        save_gpt_ckpt_path = 'model_weights/save_shape_cars_whole_p_rot_scratch_4mask_randp.safetensors'
 
-    print("   Graph compiled, starting generation...")
+    # Create fresh engine instance (fixes state corruption from caching)
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    engine = EngineFast(
+        config_path, gpt_ckpt_path, shape_ckpt_path, save_gpt_ckpt_path,
+        device=device,
+        mode='test'
+    )
+    print("   Compiled the graph.")
 
     targets_source = torch.from_numpy(ldr_content[0]).to(device).unsqueeze(0)
     targets = targets_source.clone()