fix-attention-implementation-argument (#15)

- fix: attn arg (c16d5ecc6f527a6cfe77371ce05fecc89ff8b32a)

config.json

@@ -56,5 +56,6 @@
   "vocab_size": 151936,
   "truncate_dim": null,
   "task_names": ["retrieval", "text-matching", "code"],
-  "matryoshka_dims": [128, 256, 512, 1024]
+  "matryoshka_dims": [128, 256, 512, 1024],
+  "_attn_implementation": "flash_attention_2"
 }

modeling_jina_embeddings_v4.py

@@ -519,10 +519,7 @@ class JinaEmbeddingsV4Model(Qwen2_5_VLForConditionalGeneration):
         """
         if "torch_dtype" not in kwargs:
             kwargs["torch_dtype"] = "auto"
-            
-        if torch.cuda.is_available() and "attn_implementation" not in kwargs:
-            kwargs["attn_implementation"] = "flash_attention_2"
-        
+                    
         kwargs["key_mapping"] = super()._checkpoint_conversion_mapping
 
         base_model = super().from_pretrained(