Build uploaded using `kernels`.

build/torch210-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/{torch29-cxx11-cu128-aarch64-linux/deformable_detr/_deformable_detr_320b408.abi3.so → torch210-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_a06632f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b54665ded7f4ed9bdedb765f3010a52ede13f0b9107aae15ae268f08dd171d21
-size 9980808
+oid sha256:4c9e12ad380ed55dd3c3e986d204dd48ac7fbe62bdcb71bb46feb2ecbef7c7d9
+size 8606368

build/{torch29-cxx11-cu126-aarch64-linux/deformable_detr → torch210-cxx11-cu126-aarch64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_320b408
-ops = torch.ops._deformable_detr_320b408
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_320b408::{op_name}"
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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/torch210-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,18 @@
+{
+  "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/torch210-cxx11-cu128-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/torch210-cxx11-cu128-aarch64-linux/_deformable_detr_cuda_a06632f.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0c15e71846d739201493eafdc6618817c3a9546fb521c0ed47edd15114ff61b3
+size 11621120

build/{torch29-cxx11-cu128-aarch64-linux/deformable_detr → torch210-cxx11-cu128-aarch64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_320b408
-ops = torch.ops._deformable_detr_320b408
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_320b408::{op_name}"
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch210-cxx11-cu128-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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/torch210-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch210-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/{torch29-cxx11-cu130-aarch64-linux/deformable_detr/_deformable_detr_320b408.abi3.so → torch210-cxx11-cu130-aarch64-linux/_deformable_detr_cuda_a06632f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6cc1595c80314a757e2d418bf2cf278ea2e1f8fadee0cd5440a753924ef87f0a
-size 9103096
+oid sha256:09114a32d0cf7f5d66931afb73956d78c68262abe63c6698d09571b903d6b0dd
+size 9890792

build/{torch29-cxx11-cu130-aarch64-linux/deformable_detr → torch210-cxx11-cu130-aarch64-linux}/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _deformable_detr_320b408
-ops = torch.ops._deformable_detr_320b408
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_deformable_detr_320b408::{op_name}"
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch210-cxx11-cu130-aarch64-linux/deformable_detr/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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/torch210-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,19 @@
+{
+  "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/torch29-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/torch29-cxx11-cu126-aarch64-linux/{deformable_detr/_deformable_detr_320b408.abi3.so → _deformable_detr_cuda_a06632f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e471a8e5692e1fb09dc5750fd1c76031fdb4e166082a70c9c248aa7b3a2388ca
-size 6966520
+oid sha256:e699c1c6b8c41b718c233089e16c5683fa782dd1b87e61e954941ab5f47d8341
+size 8604944

build/torch29-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch29-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py

@@ -1,46 +1,26 @@
-from typing import List
-import torch
+import ctypes
+import sys
 
-from ._ops import ops
-from . import layers
+import importlib
+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
 
-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"]
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch29-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/torch29-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,18 @@
+{
+  "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/torch29-cxx11-cu128-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/torch29-cxx11-cu128-aarch64-linux/_deformable_detr_cuda_a06632f.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee686cad364dc6a654fe2b43da0bf46fc1df3f3156dc1865a7e4811d811686e8
+size 11619232

build/torch29-cxx11-cu128-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch29-cxx11-cu128-aarch64-linux/deformable_detr/__init__.py

@@ -1,46 +1,26 @@
-from typing import List
-import torch
+import ctypes
+import sys
 
-from ._ops import ops
-from . import layers
+import importlib
+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
 
-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"]
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch29-cxx11-cu128-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/torch29-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch29-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/torch29-cxx11-cu130-aarch64-linux/_deformable_detr_cuda_a06632f.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0ab087e15a52004500698e23cb707eaec57fa16de384ea5388e0a396318f088
+size 9888904

build/torch29-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _deformable_detr_cuda_a06632f
+ops = torch.ops._deformable_detr_cuda_a06632f
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_deformable_detr_cuda_a06632f::{op_name}"

build/torch29-cxx11-cu130-aarch64-linux/deformable_detr/__init__.py

@@ -1,46 +1,26 @@
-from typing import List
-import torch
+import ctypes
+import sys
 
-from ._ops import ops
-from . import layers
+import importlib
+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
 
-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"]
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch29-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/torch29-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,19 @@
+{
+  "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"
+    ]
+  }
+}