Update configuration_deepseekocr.py

configuration_deepseekocr.py

@@ -2,8 +2,6 @@
 # ------------------------------------------------------------
 # Configuration class for the Deepseek-OCR model
 # ------------------------------------------------------------
-
-from transformers.configuration_utils import PretrainedConfig
 from transformers.utils import logging
 from .configuration_deepseek_v2 import DeepseekV2Config
 
@@ -15,13 +13,9 @@ class DeepseekOCRConfig(DeepseekV2Config):
     """
     Config for Deepseek-OCR.
 
-    It extends DeepseekV2Config (the language model config) and adds
-    OCR / vision-specific fields: projector_config, vision_config,
-    candidate_resolutions, etc.
-
-    This lets DeepseekOCRModel (which subclasses DeepseekV2Model)
-    see ALL the attributes it expects (hidden_act, attention_bias, etc.)
-    while still letting us keep multimodal metadata.
+    Inherits all language-model fields from DeepseekV2Config
+    (hidden_size, hidden_act, attention_bias, etc.) and adds
+    OCR / vision specific metadata.
     """
 
     model_type = "deepseekocr"
@@ -35,34 +29,30 @@ class DeepseekOCRConfig(DeepseekV2Config):
         projector_config=None,
         vision_config=None,
         language_config=None,
-        torch_dtype="bfloat16",
         **kwargs,
     ):
-        """
-        Parameters that are *not* listed here (hidden_size, num_attention_heads,
-        num_hidden_layers, hidden_act, attention_bias, etc.) are passed via
-        **kwargs and handled by DeepseekV2Config.__init__.
-        """
-
-        # If a nested language_config is provided (like in your config.json),
-        # use it as a base and let top-level kwargs override it.
+        # If a nested language_config dict is provided in config.json,
+        # merge it into kwargs so DeepseekV2Config sees all LM params.
         if language_config is not None and isinstance(language_config, dict):
-            base = dict(language_config)
-            base.update(kwargs)
+            base = dict(language_config)  # copy
+            base.update(kwargs)           # top-level overrides nested
             kwargs = base
 
-        # Let DeepseekV2Config set all LM-related fields:
-        super().__init__(torch_dtype=torch_dtype, **kwargs)
+        # Let DeepseekV2Config handle all core model parameters.
+        # NOTE: we do NOT pass torch_dtype explicitly here, it will be
+        # picked from kwargs if present, so no "multiple values" error.
+        super().__init__(**kwargs)
 
-        # OCR-specific / multimodal metadata
+        # Store OCR-specific attributes
         self.candidate_resolutions = candidate_resolutions or [[1024, 1024]]
         self.global_view_pos = global_view_pos
         self.tile_tag = tile_tag
 
-        # Keep the sub-configs so modeling_deepseekocr can read them if needed
+        # Keep sub-configs around for the modeling code
         self.projector_config = projector_config
         self.vision_config = vision_config
         self.language_config = language_config
 
         logger.info("✅ DeepseekOCRConfig initialized (inherits DeepseekV2Config).")
 
+