fix: throw warnings if xformers or flash-attn cant be used

configuration_clip.py

@@ -8,6 +8,7 @@ import os
 from copy import deepcopy
 from typing import Any, Dict, List, Optional, Union
 
+import torch
 from transformers import PretrainedConfig, logging
 
 logger = logging.get_logger(__name__)
@@ -157,6 +158,7 @@ class JinaCLIPConfig(PretrainedConfig):
         use_vision_xformers: Optional[bool] = None,
         matryoshka_dimensions: Optional[List[int]] = None,
         truncate_dim: Optional[int] = None,
+        torch_dtype: Optional[Union[str, torch.dtype]] = None,
         **kwargs,
     ):
         # If `_config_dict` exist, we use them for the backward compatibility.
@@ -284,6 +286,18 @@ class JinaCLIPConfig(PretrainedConfig):
                     'projections with `add_projections=True`.'
                 )
 
+        if (
+            torch_dtype
+            and hasattr(torch, torch_dtype)
+            and type(getattr(torch, torch_dtype)) is torch.dtype
+        ):
+            self.torch_dtype = getattr(torch, torch_dtype)
+        else:
+            self.torch_dtype = torch_dtype
+
+        if not self.use_text_flash_attn or not torch.cuda.is_available():
+            self.torch_dtype = torch.float32
+
     @classmethod
     def from_text_vision_configs(
         cls,