Replace **dtype** with **torch_dtype** in Model Loading

The model card for Kosmos 2.5 has been updated to address an issue where the dtype parameter in the from_pretrained method was not functioning as intended. The code has been revised to use torch_dtype instead, aligning with the standard PyTorch API convention in Hugging Face's Transformers library. This change ensures proper data type specification during model loading and resolves issues with the previous dtype parameter.
The updated code is as follows:
```
model = Kosmos2_5ForConditionalGeneration.from_pretrained(repo, device_map=device, torch_dtype=dtype)
```
This update does not alter the model's functionality but ensures compatibility with the Hugging Face Transformers API.

README.md

@@ -33,7 +33,7 @@ from transformers import AutoProcessor, Kosmos2_5ForConditionalGeneration, infer
 repo = "microsoft/kosmos-2.5"
 device = "cuda:0"
 dtype = torch.bfloat16
-model = Kosmos2_5ForConditionalGeneration.from_pretrained(repo, device_map=device, dtype=dtype)
+model = Kosmos2_5ForConditionalGeneration.from_pretrained(repo, device_map=device, torch_dtype=dtype)
 processor = AutoProcessor.from_pretrained(repo)
 
 # sample image
@@ -71,7 +71,7 @@ from transformers import AutoProcessor, Kosmos2_5ForConditionalGeneration, infer
 repo = "microsoft/kosmos-2.5"
 device = "cuda:0"
 dtype = torch.bfloat16
-model = Kosmos2_5ForConditionalGeneration.from_pretrained(repo, device_map=device, dtype=dtype)
+model = Kosmos2_5ForConditionalGeneration.from_pretrained(repo, device_map=device, torch_dtype=dtype)
 processor = AutoProcessor.from_pretrained(repo)
 
 # sample image