Register MultiScaleDeformableAttention in mmcv ATTENTION registry

The stub class needs to be registered with @ATTENTION .register_module()
so that BaseTransformerLayer can find it when building the model.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

setup_mmcv_shim.py

@@ -164,8 +164,35 @@ class ConcatCell:
     def __init__(self, *args, **kwargs): raise NotImplementedError
 ''')
 
-# Create multi_scale_deform_attn stub
+# Create multi_scale_deform_attn stub with registry registration
 with open(os.path.join(ops_dir, 'multi_scale_deform_attn.py'), 'w') as f:
-    f.write('class MultiScaleDeformableAttention: pass\n')
+    f.write('''import warnings
+import torch
+import torch.nn as nn
+from mmcv.cnn.bricks.registry import ATTENTION
+from mmcv.runner import BaseModule
+
+@ATTENTION.register_module()
+class MultiScaleDeformableAttention(BaseModule):
+    def __init__(self, embed_dims=256, num_heads=8, num_levels=4, num_points=4,
+                 im2col_step=64, dropout=0.1, batch_first=False, norm_cfg=None,
+                 init_cfg=None, **kwargs):
+        super().__init__(init_cfg)
+        self.embed_dims = embed_dims
+        self.num_heads = num_heads
+        self.num_levels = num_levels
+        self.num_points = num_points
+        self.batch_first = batch_first
+        self.sampling_offsets = nn.Linear(embed_dims, num_heads * num_levels * num_points * 2)
+        self.attention_weights = nn.Linear(embed_dims, num_heads * num_levels * num_points)
+        self.value_proj = nn.Linear(embed_dims, embed_dims)
+        self.output_proj = nn.Linear(embed_dims, embed_dims)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, query, key=None, value=None, identity=None,
+                query_pos=None, key_padding_mask=None, reference_points=None,
+                spatial_shapes=None, level_start_index=None, **kwargs):
+        raise NotImplementedError("MultiScaleDeformableAttention CPU shim - forward not implemented")
+''')
 
 print('mmcv.ops shim created successfully')