Uploaded using `kernel-builder`.

build/torch210-cxx11-cu126-x86_64-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 sha256:5cd54d93f125add18b9b0399072b6af3f0f2c4b13ac541a3dcf496a0ee0a3f30
+oid sha256:4770d1a90085f7daf05a938f72bf0c87bc88786418673c7a9fa23ab71b4c3cea
 size 8541080

build/torch210-cxx11-cu126-x86_64-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-x86_64-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-x86_64-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 sha256:6a5684a08a44bf76b3b316308d0ba25da74417ab1d123152b428a14a44681d83
+oid sha256:a4f249cec923d18ff5a0cbde52cf496020f7e35eb0cbd19b9fadd8487523542b
 size 11524560

build/torch210-cxx11-cu128-x86_64-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-x86_64-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-x86_64-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:f685d0d2da9804b005293a21a5d9d4304350307ae858c3d47c26a7daf2fb910a
+oid sha256:a227a15462af545f30b2a53abf128104c118aee20098f91b9e5b57a3c4ee61b7
 size 9809000

build/torch210-cxx11-cu130-x86_64-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-x86_64-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-x86_64-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 sha256:66b5846d32ffbd7daac9f2f03fe71034a3000aeb7dfe715620035d4aac7ab852
+oid sha256:1ee2e4be12cd8c45cd3c04352ebd46b754f9407aa9d8614e2398d5cad3969e96
 size 8541176

build/torch211-cxx11-cu126-x86_64-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-x86_64-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-x86_64-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:dc4718caad26b95601f98b49967e84cc0c3d9e31bb48d3aaef86fba81af2213e
+oid sha256:1ed7de8f0d91469452568abc6eab182cfc607a035ba3a22fd607378184409f35
 size 11528752

build/torch211-cxx11-cu128-x86_64-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-x86_64-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-x86_64-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:d559f99c3efee3c42da70211a901c8a1de5cb2a8c4cd933daf17ed94b9796dd8
+oid sha256:4d4283e89185a13455f1bed02d4e16fc79559006afc71b2a1441436bdffbd9ab
 size 9809096

build/torch211-cxx11-cu130-x86_64-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-x86_64-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-x86_64-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-x86_64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a67ddd7c9f58db967d4898a8ea2488600290efb4d43d643da119174b7f202b23
+size 8536248

build/torch212-cxx11-cu126-x86_64-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-x86_64-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-x86_64-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-x86_64-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-x86_64-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-x86_64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1395c3badf6a4992814e084253292f68d974b35ccc02603e189ae5adeec35780
+size 9804120

build/torch212-cxx11-cu130-x86_64-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-x86_64-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-x86_64-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-x86_64-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-x86_64-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-x86_64-linux/_deformable_detr_cuda_c35b54a.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4521cdb5743a43cef143f5d13285a8fb52b0e1c36d253a347b92284a22c9332e
+size 9900584

build/torch212-cxx11-cu132-x86_64-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-x86_64-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-x86_64-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-x86_64-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"
+    ]
+  }
+}