Fix: Safe device selection for CPU/GPU compatibility

- clip_retrieval.py: Add _resolve_device() with CUDA detection & CPU fallback
- demo.py: Auto-select GPT engine (EngineFast for CUDA, Engine for CPU)
- engine.py: Normalize device_map to string for CLIP text encoder

Resolves: CUDA availability issues on HF Spaces CPU instances
Support: Works on both GPU and CPU tiers

File: code/cube3d/training/engine.py

code/cube3d/training/engine.py

@@ -78,11 +78,12 @@ class Engine:
             codebook = self.shape_model.bottleneck.block.get_codebook()
             codebook = self.gpt_model.shape_proj(codebook).detach()
         self.gpt_model.transformer.wte.weight.data[: codebook.shape[0]] = codebook
-        #import ipdb; ipdb.set_trace()
+
+        device_map = self.device.type if isinstance(self.device, torch.device) else self.device
         self.text_model = CLIPTextModelWithProjection.from_pretrained(
             self.cfg.text_model_pretrained_model_name_or_path,
             force_download=False,
-            device_map=self.device,
+            device_map=device_map,
             cache_dir=HF_CACHE_DIR,
         ).eval()
         print("------text_model device---------", self.text_model.device)