handle inputs from Siglip/Siglip2 non-automapped encoder layers

src/transformers/models/siglip/modeling_siglip.py

@@ -679,6 +679,11 @@ def forward(
 
 
 class SiglipVisionTransformer(nn.Module):
+    _can_record_outputs = {
+        "hidden_states": SiglipEncoderLayer,
+        "attentions": SiglipAttention,
+    }
+
     def __init__(self, config: SiglipVisionConfig):
         super().__init__()
         self.config = config
@@ -691,6 +696,7 @@ def __init__(self, config: SiglipVisionConfig):
         if self.use_head:
             self.head = SiglipMultiheadAttentionPoolingHead(config)
 
+    @check_model_inputs(tie_last_hidden_states=False)
     @auto_docstring
     def forward(
         self,

src/transformers/models/siglip2/modeling_siglip2.py

@@ -384,6 +384,11 @@ def forward(
 
 
 class Siglip2VisionTransformer(nn.Module):
+    _can_record_outputs = {
+        "hidden_states": Siglip2EncoderLayer,
+        "attentions": Siglip2Attention,
+    }
+
     def __init__(self, config: Siglip2VisionConfig):
         super().__init__()
         self.config = config
@@ -396,6 +401,7 @@ def __init__(self, config: Siglip2VisionConfig):
         if self.use_head:
             self.head = Siglip2MultiheadAttentionPoolingHead(config)
 
+    @check_model_inputs(tie_last_hidden_states=False)
     @auto_docstring
     def forward(
         self,
@@ -811,7 +817,6 @@ def get_input_embeddings(self) -> nn.Module:
         return self.vision_model.embeddings.patch_embedding
 
     @check_model_inputs(tie_last_hidden_states=False)
-    @auto_docstring
     def forward(
         self,
         pixel_values: torch.FloatTensor,

src/transformers/models/siglip2/modular_siglip2.py

@@ -37,6 +37,7 @@
 
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask
 from ...utils import auto_docstring, filter_out_non_signature_kwargs
+from ...utils.generic import check_model_inputs
 
 
 class Siglip2TextConfig(SiglipTextConfig):
@@ -314,6 +315,7 @@ def forward(self, hidden_state: torch.Tensor, attention_mask: Optional[torch.Ten
 
 class Siglip2VisionModel(SiglipVisionModel):
     # Update: add `spatial_shapes` and `pixel_attention_mask`
+    @check_model_inputs(tie_last_hidden_states=False)
     def forward(
         self,
         pixel_values: torch.FloatTensor,

src/transformers/utils/generic.py

@@ -836,7 +836,7 @@ def wrapper(self, *args, **kwargs):
                     all_args[k] = v
 
             # _can_record_outputs is None by default
-            capture_flags = _CAN_RECORD_REGISTRY.get(str(self.__class__)) or {}  # there is a weak ref for executorch
+            capture_flags = _CAN_RECORD_REGISTRY.get(str(self.__class__)) or getattr(self, "_can_record_outputs", {})
             recordable_keys = {
                 f"output_{k}": all_args.get(
                     f"output_{k}",