Fixes (DeepSeek, Florence-2) (#156)

* fix DeepSeek-VL2 quant loading

* fix florence LM only

* bump version

mlx_vlm/models/deepseek_vl_v2/deepseek_vl_v2.py

@@ -483,12 +483,12 @@ def from_pretrained(path_or_hf_repo: str):
     @staticmethod
     def sanitize(weights):
         def transform_key(key):
-            if "language" in key:
+            if "language" in key and "language_model" not in key:
                 if ".model" in key:
                     key = key.replace("language.model", "language_model.model")
                 if ".lm_head" in key:
                     key = key.replace("language", "language_model")
-            if "vision" in key:
+            if "vision" in key and "vision_tower" not in key:
                 key = key.replace("vision", "vision.vision_tower")
             if "view_seperator" in key:
                 key = key.replace("view_seperator", "view_separator")

mlx_vlm/models/florence2/florence2.py

@@ -317,6 +317,9 @@ def __call__(
             inputs_embeds, attention_mask = self._merge_input_ids_with_image_features(
                 image_features, inputs_embeds
             )
+        else:
+            inputs_embeds = None
+            attention_mask = None
 
         # Handle decoder input IDs
         if labels is not None and decoder_input_ids is None:

mlx_vlm/version.py

@@ -1 +1 @@
-__version__ = "0.1.5"
+__version__ = "0.1.6"