Update configuration_deepseekocr.py

configuration_deepseekocr.py

@@ -5,90 +5,64 @@
 
 from transformers.configuration_utils import PretrainedConfig
 from transformers.utils import logging
+from .configuration_deepseek_v2 import DeepseekV2Config
 
 logger = logging.get_logger(__name__)
 
 DEEPSEEK_OCR_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
 
-
-class DeepseekOCRConfig(PretrainedConfig):
+class DeepseekOCRConfig(DeepseekV2Config):
     """
-    Configuration class for the Deepseek-OCR model.
-    Allows Hugging Face Auto classes to recognize this custom architecture.
+    Config for Deepseek-OCR.
+
+    It extends DeepseekV2Config (the language model config) and adds
+    OCR / vision-specific fields: projector_config, vision_config,
+    candidate_resolutions, etc.
 
-    Example:
-    ```python
-    from transformers import AutoConfig
-    config = AutoConfig.from_pretrained("specsGuy/Deepseek-ocr", trust_remote_code=True)
-    print(config.model_type)  # deepseekocr
-    ```
+    This lets DeepseekOCRModel (which subclasses DeepseekV2Model)
+    see ALL the attributes it expects (hidden_act, attention_bias, etc.)
+    while still letting us keep multimodal metadata.
     """
 
-    model_type = "deepseekocr"  # ✅ REQUIRED for AutoConfig detection
+    model_type = "deepseekocr"
 
     def __init__(
         self,
-        vocab_size=129280,
-        hidden_size=1280,
-        intermediate_size=6848,
-        num_hidden_layers=12,
-        num_attention_heads=10,
-        num_key_value_heads=10,
-        max_position_embeddings=8192,
-        moe_intermediate_size=896,
-        n_group=1,
-        n_routed_experts=64,
-        n_shared_experts=2,
-        num_experts_per_tok=6,
-        first_k_dense_replace=1,
-        topk_group=1,
-        topk_method="greedy",
-        use_mla=False,
-        bos_token_id=0,
-        eos_token_id=1,
-        torch_dtype="bfloat16",
+        # OCR / vision specific
+        candidate_resolutions=None,
+        global_view_pos="head",
+        tile_tag="2D",
         projector_config=None,
         vision_config=None,
         language_config=None,
+        torch_dtype="bfloat16",
         **kwargs,
     ):
-        # Core architecture settings
-        self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.num_key_value_heads = num_key_value_heads
-        self.max_position_embeddings = max_position_embeddings
-        self.moe_intermediate_size = moe_intermediate_size
-        self.attention_dropout = kwargs.get("attention_dropout", 0.0)
-        self.attention_bias = kwargs.get("attention_bias", False)
+        """
+        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 language_config is not None and isinstance(language_config, dict):
+            base = dict(language_config)
+            base.update(kwargs)
+            kwargs = base
 
-        # Mixture of Experts & attention params
-        self.n_group = n_group
-        self.n_routed_experts = n_routed_experts
-        self.n_shared_experts = n_shared_experts
-        self.num_experts_per_tok = num_experts_per_tok
-        self.first_k_dense_replace = first_k_dense_replace
-        self.topk_group = topk_group
-        self.topk_method = topk_method
-        self.use_mla = use_mla
+        # Let DeepseekV2Config set all LM-related fields:
+        super().__init__(torch_dtype=torch_dtype, **kwargs)
 
-        # Tokens & dtype
-        self.bos_token_id = bos_token_id
-        self.eos_token_id = eos_token_id
-        self.torch_dtype = torch_dtype
+        # OCR-specific / multimodal metadata
+        self.candidate_resolutions = candidate_resolutions or [[1024, 1024]]
+        self.global_view_pos = global_view_pos
+        self.tile_tag = tile_tag
 
-        # Subconfigs (for multimodal alignment)
+        # Keep the sub-configs so modeling_deepseekocr can read them if needed
         self.projector_config = projector_config
         self.vision_config = vision_config
         self.language_config = language_config
 
-        super().__init__(
-            bos_token_id=bos_token_id,
-            eos_token_id=eos_token_id,
-            **kwargs,
-        )
+        logger.info("✅ DeepseekOCRConfig initialized (inherits DeepseekV2Config).")
 
-        logger.info("✅ Initialized DeepseekOCRConfig successfully.")