Uploaded using `kernel-builder`.

build/torch210-cxx11-cu126-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8f83d7eac250567d574bca8b45b0df27ce6aeea3a50ddcba333b8ff143503d97
+oid sha256:a78e7fffe93acd3db377ab1850aa4927885a1ddc9ed699cd7a8044fd28e3803b
 size 8606368

build/torch210-cxx11-cu126-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch210-cxx11-cu128-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e245677fedf207341fd149c5a2b1af4a3417da8f26a1b59f78f4fb5a16cbbc32
+oid sha256:ba51c83bd4d18c329f5f48d279d5a4983fb66c79eb67e1408741a8d965ac4137
 size 11621120

build/torch210-cxx11-cu128-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch210-cxx11-cu128-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch210-cxx11-cu130-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4090b9a5be5fd27aabb5ecef49550a543f72c3221973e78cc0e918a2f057716a
+oid sha256:b0ddf3c6c7c6cb973af5af6b7be05b513a336abf969fdd9441c60c82ce0c1258
 size 9825896

build/torch210-cxx11-cu130-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch210-cxx11-cu130-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu126-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ab0eba678eee353ed44a90fc18ec09eba18d3d8420f274b9ccff36a5ef561f5c
+oid sha256:34d9d5e50c94aaa7d47f2aae8cf4f893d92dc24215ebbcba7c2d3ade54209bd2
 size 8606480

build/torch211-cxx11-cu126-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch211-cxx11-cu126-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu128-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:21f63f40573eb787e89fc886341e9028be0ed8b0b6c44a5e2bc8439f541cae20
+oid sha256:fac9605c94a04d948e421567004b5ec92485bec3410dde06e19a9f86309f6648
 size 11621232

build/torch211-cxx11-cu128-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch211-cxx11-cu128-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch211-cxx11-cu130-aarch64-linux/{_deformable_detr_cuda_86f75d9.abi3.so → _deformable_detr_cuda_c35b54a.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5404a3384eb16c51370320aff1795107eca9afa9b63a3a436c7b254facd42ddb
+oid sha256:56a6ec91cabb15aa60ec67fe35ddcf28f753ad31f40a065721e8e6e2137f7b11
 size 9891552

build/torch211-cxx11-cu130-aarch64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_cuda_86f75d9
-ops = torch.ops._deformable_detr_cuda_86f75d9
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_cuda_86f75d9::{op_name}"
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch211-cxx11-cu130-aarch64-linux/metadata.json

@@ -1,6 +1,6 @@
 {
   "name": "deformable-detr",
-  "id": "_deformable_detr_cuda_86f75d9",
+  "id": "_deformable_detr_cuda_c35b54a",
   "version": 1,
   "license": "Apache-2.0",
   "python-depends": [],

build/torch212-cxx11-cu126-aarch64-linux/__init__.py

@@ -0,0 +1,46 @@
+from typing import List
+import torch
+
+from ._ops import ops
+from . import layers
+
+
+def ms_deform_attn_backward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    grad_output: torch.Tensor,
+    im2col_step: int,
+) -> List[torch.Tensor]:
+    return ops.ms_deform_attn_backward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        grad_output,
+        im2col_step,
+    )
+
+
+def ms_deform_attn_forward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    im2col_step: int,
+) -> torch.Tensor:
+    return ops.ms_deform_attn_forward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        im2col_step,
+    )
+
+
+__all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]

build/torch212-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25570d02c333bf8a76d1d894778ad7eaeb7a87a8d2d63e4b13c0d4cb28187b9d
+size 8606640

build/torch212-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch212-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu126-aarch64-linux/layers.py

@@ -0,0 +1,84 @@
+from typing import List, Union, Tuple
+
+from torch import Tensor
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class MultiScaleDeformableAttentionFunction(Function):
+    @staticmethod
+    def forward(
+        context,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        context.im2col_step = im2col_step
+        output = ops.ms_deform_attn_forward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            context.im2col_step,
+        )
+        context.save_for_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(context, grad_output):
+        (
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        ) = context.saved_tensors
+        grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            grad_output,
+            context.im2col_step,
+        )
+
+        return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
+
+
+class MultiScaleDeformableAttention(nn.Module):
+    def forward(
+        self,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_spatial_shapes_list: List[Tuple],
+        level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        return MultiScaleDeformableAttentionFunction.apply(
+            value,
+            value_spatial_shapes,
+            level_start_index,
+            sampling_locations,
+            attention_weights,
+            im2col_step,
+        )
+
+
+__all__ = ["MultiScaleDeformableAttention"]

build/torch212-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,20 @@
+{
+  "name": "deformable-detr",
+  "id": "_deformable_detr_cuda_c35b54a",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0+PTX"
+    ]
+  }
+}

build/torch212-cxx11-cu130-aarch64-linux/__init__.py

@@ -0,0 +1,46 @@
+from typing import List
+import torch
+
+from ._ops import ops
+from . import layers
+
+
+def ms_deform_attn_backward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    grad_output: torch.Tensor,
+    im2col_step: int,
+) -> List[torch.Tensor]:
+    return ops.ms_deform_attn_backward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        grad_output,
+        im2col_step,
+    )
+
+
+def ms_deform_attn_forward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    im2col_step: int,
+) -> torch.Tensor:
+    return ops.ms_deform_attn_forward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        im2col_step,
+    )
+
+
+__all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]

build/torch212-cxx11-cu130-aarch64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a17b5cf9073ab3c77746a5716c919784d0bbd1cc045df5a632c7a73a4a4a80e3
+size 9891672

build/torch212-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch212-cxx11-cu130-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu130-aarch64-linux/layers.py

@@ -0,0 +1,84 @@
+from typing import List, Union, Tuple
+
+from torch import Tensor
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class MultiScaleDeformableAttentionFunction(Function):
+    @staticmethod
+    def forward(
+        context,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        context.im2col_step = im2col_step
+        output = ops.ms_deform_attn_forward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            context.im2col_step,
+        )
+        context.save_for_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(context, grad_output):
+        (
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        ) = context.saved_tensors
+        grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            grad_output,
+            context.im2col_step,
+        )
+
+        return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
+
+
+class MultiScaleDeformableAttention(nn.Module):
+    def forward(
+        self,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_spatial_shapes_list: List[Tuple],
+        level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        return MultiScaleDeformableAttentionFunction.apply(
+            value,
+            value_spatial_shapes,
+            level_start_index,
+            sampling_locations,
+            attention_weights,
+            im2col_step,
+        )
+
+
+__all__ = ["MultiScaleDeformableAttention"]

build/torch212-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "name": "deformable-detr",
+  "id": "_deformable_detr_cuda_c35b54a",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0+PTX",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch212-cxx11-cu132-aarch64-linux/__init__.py

@@ -0,0 +1,46 @@
+from typing import List
+import torch
+
+from ._ops import ops
+from . import layers
+
+
+def ms_deform_attn_backward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    grad_output: torch.Tensor,
+    im2col_step: int,
+) -> List[torch.Tensor]:
+    return ops.ms_deform_attn_backward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        grad_output,
+        im2col_step,
+    )
+
+
+def ms_deform_attn_forward(
+    value: torch.Tensor,
+    spatial_shapes: torch.Tensor,
+    level_start_index: torch.Tensor,
+    sampling_loc: torch.Tensor,
+    attn_weight: torch.Tensor,
+    im2col_step: int,
+) -> torch.Tensor:
+    return ops.ms_deform_attn_forward(
+        value,
+        spatial_shapes,
+        level_start_index,
+        sampling_loc,
+        attn_weight,
+        im2col_step,
+    )
+
+
+__all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]

build/torch212-cxx11-cu132-aarch64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:50b08835f3b62d9201ad56a558381af696ba4b216ae18a6a9e38eeb4a5385e5a
+size 9959808

build/torch212-cxx11-cu132-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_c35b54a
+ops = torch.ops._deformable_detr_cuda_c35b54a
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_c35b54a::{op_name}"

build/torch212-cxx11-cu132-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch212-cxx11-cu132-aarch64-linux/layers.py

@@ -0,0 +1,84 @@
+from typing import List, Union, Tuple
+
+from torch import Tensor
+from torch.autograd import Function
+from torch.autograd.function import once_differentiable
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class MultiScaleDeformableAttentionFunction(Function):
+    @staticmethod
+    def forward(
+        context,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        context.im2col_step = im2col_step
+        output = ops.ms_deform_attn_forward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            context.im2col_step,
+        )
+        context.save_for_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        )
+        return output
+
+    @staticmethod
+    @once_differentiable
+    def backward(context, grad_output):
+        (
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+        ) = context.saved_tensors
+        grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
+            value,
+            value_spatial_shapes,
+            value_level_start_index,
+            sampling_locations,
+            attention_weights,
+            grad_output,
+            context.im2col_step,
+        )
+
+        return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
+
+
+class MultiScaleDeformableAttention(nn.Module):
+    def forward(
+        self,
+        value: Tensor,
+        value_spatial_shapes: Tensor,
+        value_spatial_shapes_list: List[Tuple],
+        level_start_index: Tensor,
+        sampling_locations: Tensor,
+        attention_weights: Tensor,
+        im2col_step: int,
+    ):
+        return MultiScaleDeformableAttentionFunction.apply(
+            value,
+            value_spatial_shapes,
+            level_start_index,
+            sampling_locations,
+            attention_weights,
+            im2col_step,
+        )
+
+
+__all__ = ["MultiScaleDeformableAttention"]

build/torch212-cxx11-cu132-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "name": "deformable-detr",
+  "id": "_deformable_detr_cuda_c35b54a",
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0+PTX",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}