feat: add device_map argument to try avoid CUDA init RuntimeError, bump lisa-on-cuda to version 1.4.6

lisa_on_cuda/utils/app_helpers.py

@@ -224,7 +224,9 @@ def get_model(args_to_parse, internal_logger: logging = None, inference_decorato
     ) if inference_decorator else prepare_model_vision_tower(
         _model, args_to_parse, torch_dtype
     )
-    vision_tower.to(device=args_to_parse.local_rank)
+    # set device to device_map try to avoid CUDA init RuntimeError on ZeroGPU huggingface hardware
+    device = device_map if device_map else args_to_parse.local_rank
+    vision_tower.to(device=device)
     internal_logger.debug("vision tower loaded, prepare clip image processor...")
     _clip_image_processor = CLIPImageProcessor.from_pretrained(_model.config.vision_tower)
     internal_logger.debug("clip image processor done.")

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "lisa-on-cuda"
-version = "1.4.5"
+version = "1.4.6"
 description = "LISA (Reasoning Segmentation via Large Language Model) on cuda, now with huggingface ZeroGPU support!"
 authors = ["alessandro trinca tornidor <[email protected]>"]
 license = "Apache 2.0"
@@ -8,7 +8,7 @@ readme = "README.md"
 
 [metadata]
 name = "lisa-on-cuda"
-version = "1.4.5"
+version = "1.4.6"
 
 [tool.poetry.urls]
 Source = "https://huggingface.co/spaces/aletrn/lisa-on-cuda/"