Upload model

config.json

@@ -2,6 +2,10 @@
   "architectures": [
     "LSPDetrModel"
   ],
+  "auto_map": {
+    "AutoConfig": "configuration.LSPDetrConfig",
+    "AutoModelForObjectDetection": "modeling.LSPDetrModel"
+  },
   "backbone": "microsoft/swinv2-tiny-patch4-window16-256",
   "backbone_config": null,
   "backbone_kwargs": {

modeling.py

@@ -1,5 +1,6 @@
 import math
 from functools import lru_cache
+from unittest.mock import patch
 
 import torch
 import torch.nn.functional as F
@@ -11,13 +12,28 @@ from torch.nn.attention.flex_attention import (
     create_block_mask,
     flex_attention,
 )
-from torch.nn.utils.parametrizations import orthogonal
+from torch.nn.utils.parametrizations import _Orthogonal, orthogonal
 from transformers.modeling_utils import PreTrainedModel
 from transformers.utils.backbone_utils import load_backbone
 
 from .configuration import LSPDetrConfig, STAConfig
 
 
+def _meta_safe_right_inverse(self, x: torch.Tensor) -> torch.Tensor:
+    if x.device.type == "meta":
+        # Skip validation; assume OK (unsafe but necessary for meta init)
+        return x
+
+    return _right_inverse(self, x)
+
+
+_right_inverse = _Orthogonal.right_inverse
+patch(
+    "torch.nn.utils.parametrizations._Orthogonal.right_inverse",
+    _meta_safe_right_inverse,
+).start()
+
+
 flex_attention = torch.compile(flex_attention, dynamic=True)
 
 
@@ -514,3 +530,7 @@ class LSPDetrModel(PreTrainedModel):
         )
 
         return self.decode_head(tgt, ref_points, features, h, w)
+
+
+P = orthogonal(nn.Linear(10, 10, bias=False, device="cpu"), orthogonal_map="cayley")
+P = orthogonal(nn.Linear(10, 10, bias=False, device="cpu"), orthogonal_map="cayley")