drbh commited on
Commit
8e5f9d7
·
unverified ·
0 Parent(s):

Migrated from kernels-community/deformable-detr

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +2 -0
  2. README.md +22 -0
  3. benchmarks/benchmark.py +250 -0
  4. build/torch210-cu128-x86_64-windows/__init__.py +46 -0
  5. build/torch210-cu128-x86_64-windows/_deformable_detr_cuda_d8a6191.pyd +3 -0
  6. build/torch210-cu128-x86_64-windows/_ops.py +9 -0
  7. build/torch210-cu128-x86_64-windows/deformable_detr/__init__.py +26 -0
  8. build/torch210-cu128-x86_64-windows/layers.py +84 -0
  9. build/torch210-cu128-x86_64-windows/metadata.json +21 -0
  10. build/torch210-cxx11-cu126-aarch64-linux/__init__.py +46 -0
  11. build/torch210-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  12. build/torch210-cxx11-cu126-aarch64-linux/_ops.py +9 -0
  13. build/torch210-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py +26 -0
  14. build/torch210-cxx11-cu126-aarch64-linux/layers.py +84 -0
  15. build/torch210-cxx11-cu126-aarch64-linux/metadata.json +18 -0
  16. build/torch210-cxx11-cu126-x86_64-linux/__init__.py +46 -0
  17. build/torch210-cxx11-cu126-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  18. build/torch210-cxx11-cu126-x86_64-linux/_ops.py +9 -0
  19. build/torch210-cxx11-cu126-x86_64-linux/deformable_detr/__init__.py +26 -0
  20. build/torch210-cxx11-cu126-x86_64-linux/layers.py +84 -0
  21. build/torch210-cxx11-cu126-x86_64-linux/metadata.json +18 -0
  22. build/torch210-cxx11-cu128-aarch64-linux/__init__.py +46 -0
  23. build/torch210-cxx11-cu128-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  24. build/torch210-cxx11-cu128-aarch64-linux/_ops.py +9 -0
  25. build/torch210-cxx11-cu128-aarch64-linux/deformable_detr/__init__.py +26 -0
  26. build/torch210-cxx11-cu128-aarch64-linux/layers.py +84 -0
  27. build/torch210-cxx11-cu128-aarch64-linux/metadata.json +21 -0
  28. build/torch210-cxx11-cu128-x86_64-linux/__init__.py +46 -0
  29. build/torch210-cxx11-cu128-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  30. build/torch210-cxx11-cu128-x86_64-linux/_ops.py +9 -0
  31. build/torch210-cxx11-cu128-x86_64-linux/deformable_detr/__init__.py +26 -0
  32. build/torch210-cxx11-cu128-x86_64-linux/layers.py +84 -0
  33. build/torch210-cxx11-cu128-x86_64-linux/metadata.json +21 -0
  34. build/torch210-cxx11-cu130-aarch64-linux/__init__.py +46 -0
  35. build/torch210-cxx11-cu130-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  36. build/torch210-cxx11-cu130-aarch64-linux/_ops.py +9 -0
  37. build/torch210-cxx11-cu130-aarch64-linux/deformable_detr/__init__.py +26 -0
  38. build/torch210-cxx11-cu130-aarch64-linux/layers.py +84 -0
  39. build/torch210-cxx11-cu130-aarch64-linux/metadata.json +19 -0
  40. build/torch210-cxx11-cu130-x86_64-linux/__init__.py +46 -0
  41. build/torch210-cxx11-cu130-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  42. build/torch210-cxx11-cu130-x86_64-linux/_ops.py +9 -0
  43. build/torch210-cxx11-cu130-x86_64-linux/deformable_detr/__init__.py +26 -0
  44. build/torch210-cxx11-cu130-x86_64-linux/layers.py +84 -0
  45. build/torch210-cxx11-cu130-x86_64-linux/metadata.json +19 -0
  46. build/torch211-cxx11-cu126-aarch64-linux/__init__.py +46 -0
  47. build/torch211-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so +3 -0
  48. build/torch211-cxx11-cu126-aarch64-linux/_ops.py +9 -0
  49. build/torch211-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py +26 -0
  50. build/torch211-cxx11-cu126-aarch64-linux/layers.py +84 -0
.gitattributes ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ *.so filter=lfs diff=lfs merge=lfs -text
2
+ build/torch210-cu128-x86_64-windows/_deformable_detr_cuda_d8a6191.pyd filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - kernels
5
+ ---
6
+
7
+ ![Status](https://hubwebhook.dholtz.com/shield?repo=kernels-community/deformable-detr)
8
+
9
+ ## deformable-detr
10
+
11
+ Kernel source: https://github.com/huggingface/kernels-community/tree/main/deformable-detr
12
+
13
+ ### Performance
14
+
15
+ <img class="dark:hidden border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_light_animation.svg" />
16
+ <img class="hidden dark:block border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_dark_animation.svg" />
17
+
18
+ <img class="dark:hidden border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_light_latency.svg" />
19
+ <img class="hidden dark:block border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_dark_latency.svg" />
20
+
21
+ <img class="dark:hidden border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_light_throughput.svg" />
22
+ <img class="hidden dark:block border border-gray-200 dark:border-gray-700 rounded-lg" src="media/benches_dark_throughput.svg" />
benchmarks/benchmark.py ADDED
@@ -0,0 +1,250 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn.functional as F
3
+
4
+ from kernels.benchmark import Benchmark
5
+
6
+
7
+ def ms_deform_attn_reference(
8
+ value: torch.Tensor,
9
+ spatial_shapes: torch.Tensor,
10
+ level_start_index: torch.Tensor,
11
+ sampling_locations: torch.Tensor,
12
+ attention_weights: torch.Tensor,
13
+ ) -> torch.Tensor:
14
+ batch, _, num_heads, channels = value.shape
15
+ _, num_query, _, num_levels, num_points, _ = sampling_locations.shape
16
+
17
+ # Split value by levels
18
+ value_list = []
19
+ for level_id in range(num_levels):
20
+ H, W = spatial_shapes[level_id]
21
+ start_idx = level_start_index[level_id]
22
+ end_idx = (
23
+ level_start_index[level_id + 1]
24
+ if level_id < num_levels - 1
25
+ else value.shape[1]
26
+ )
27
+ # (batch, H*W, num_heads, channels) -> (batch, num_heads, channels, H, W)
28
+ value_level = value[:, start_idx:end_idx, :, :].view(
29
+ batch, H, W, num_heads, channels
30
+ )
31
+ value_level = value_level.permute(0, 3, 4, 1, 2).contiguous()
32
+ value_list.append(value_level)
33
+
34
+ # Sample from each level
35
+ output = torch.zeros(
36
+ batch, num_query, num_heads, channels, device=value.device, dtype=value.dtype
37
+ )
38
+
39
+ for level_id in range(num_levels):
40
+ H, W = spatial_shapes[level_id]
41
+ value_level = value_list[level_id] # (batch, num_heads, channels, H, W)
42
+
43
+ # Get sampling locations for this level: (batch, num_query, num_heads, num_points, 2)
44
+ sampling_loc_level = sampling_locations[:, :, :, level_id, :, :]
45
+
46
+ # Convert from [0, 1] to [-1, 1] for grid_sample
47
+ grid = (
48
+ 2.0 * sampling_loc_level - 1.0
49
+ ) # (batch, num_query, num_heads, num_points, 2)
50
+
51
+ # Reshape for grid_sample: need (batch * num_heads, channels, H, W) and (batch * num_heads, num_query, num_points, 2)
52
+ value_level = value_level.view(batch * num_heads, channels, H.item(), W.item())
53
+ grid = grid.permute(
54
+ 0, 2, 1, 3, 4
55
+ ).contiguous() # (batch, num_heads, num_query, num_points, 2)
56
+ grid = grid.view(batch * num_heads, num_query, num_points, 2)
57
+
58
+ # Sample: output is (batch * num_heads, channels, num_query, num_points)
59
+ sampled = F.grid_sample(
60
+ value_level,
61
+ grid,
62
+ mode="bilinear",
63
+ padding_mode="zeros",
64
+ align_corners=False,
65
+ )
66
+
67
+ # Reshape back: (batch, num_heads, channels, num_query, num_points)
68
+ sampled = sampled.view(batch, num_heads, channels, num_query, num_points)
69
+ # -> (batch, num_query, num_heads, num_points, channels)
70
+ sampled = sampled.permute(0, 3, 1, 4, 2).contiguous()
71
+
72
+ # Get attention weights for this level: (batch, num_query, num_heads, num_points)
73
+ attn_level = attention_weights[:, :, :, level_id, :]
74
+
75
+ # Weighted sum over points: (batch, num_query, num_heads, channels)
76
+ output += (sampled * attn_level.unsqueeze(-1)).sum(dim=3)
77
+
78
+ # Reshape to (batch, num_query, num_heads * channels)
79
+ output = output.view(batch, num_query, num_heads * channels)
80
+ return output
81
+
82
+
83
+ class MSDeformAttnBenchmark(Benchmark):
84
+ seed: int = 42
85
+
86
+ def setup(self):
87
+ batch = 2
88
+ num_heads = 8
89
+ channels = 32 # embed_dim = num_heads * channels = 256
90
+ num_levels = 4
91
+ num_query = 300
92
+ num_points = 4
93
+ im2col_step = 64
94
+
95
+ # Spatial shapes for 4 levels: 64x64, 32x32, 16x16, 8x8
96
+ spatial_shapes = torch.tensor(
97
+ [[64, 64], [32, 32], [16, 16], [8, 8]],
98
+ dtype=torch.int64,
99
+ device=self.device,
100
+ )
101
+ # Calculate spatial_size = sum of H*W for all levels
102
+ spatial_size = (64 * 64) + (32 * 32) + (16 * 16) + (8 * 8) # 5440
103
+
104
+ # Level start indices
105
+ level_start_index = torch.tensor(
106
+ [0, 64 * 64, 64 * 64 + 32 * 32, 64 * 64 + 32 * 32 + 16 * 16],
107
+ dtype=torch.int64,
108
+ device=self.device,
109
+ )
110
+
111
+ self.value = torch.randn(
112
+ batch,
113
+ spatial_size,
114
+ num_heads,
115
+ channels,
116
+ device=self.device,
117
+ dtype=torch.float32,
118
+ )
119
+ self.spatial_shapes = spatial_shapes
120
+ self.level_start_index = level_start_index
121
+ self.sampling_loc = torch.rand(
122
+ batch,
123
+ num_query,
124
+ num_heads,
125
+ num_levels,
126
+ num_points,
127
+ 2,
128
+ device=self.device,
129
+ dtype=torch.float32,
130
+ )
131
+ self.attn_weight = torch.rand(
132
+ batch,
133
+ num_query,
134
+ num_heads,
135
+ num_levels,
136
+ num_points,
137
+ device=self.device,
138
+ dtype=torch.float32,
139
+ )
140
+ # Normalize attention weights
141
+ self.attn_weight = self.attn_weight / self.attn_weight.sum(-1, keepdim=True)
142
+ self.im2col_step = im2col_step
143
+
144
+ self.out = torch.empty(
145
+ batch,
146
+ num_query,
147
+ num_heads * channels,
148
+ device=self.device,
149
+ dtype=torch.float32,
150
+ )
151
+
152
+ def benchmark_forward(self):
153
+ self.out = self.kernel.ms_deform_attn_forward(
154
+ self.value,
155
+ self.spatial_shapes,
156
+ self.level_start_index,
157
+ self.sampling_loc,
158
+ self.attn_weight,
159
+ self.im2col_step,
160
+ )
161
+
162
+ def verify_forward(self) -> torch.Tensor:
163
+ return ms_deform_attn_reference(
164
+ self.value,
165
+ self.spatial_shapes,
166
+ self.level_start_index,
167
+ self.sampling_loc,
168
+ self.attn_weight,
169
+ )
170
+
171
+ def setup_large(self):
172
+ batch = 8
173
+ num_heads = 8
174
+ channels = 32
175
+ num_levels = 4
176
+ num_query = 900
177
+ num_points = 4
178
+ im2col_step = 64
179
+
180
+ spatial_shapes = torch.tensor(
181
+ [[64, 64], [32, 32], [16, 16], [8, 8]],
182
+ dtype=torch.int64,
183
+ device=self.device,
184
+ )
185
+ spatial_size = (64 * 64) + (32 * 32) + (16 * 16) + (8 * 8)
186
+
187
+ level_start_index = torch.tensor(
188
+ [0, 64 * 64, 64 * 64 + 32 * 32, 64 * 64 + 32 * 32 + 16 * 16],
189
+ dtype=torch.int64,
190
+ device=self.device,
191
+ )
192
+
193
+ self.value = torch.randn(
194
+ batch,
195
+ spatial_size,
196
+ num_heads,
197
+ channels,
198
+ device=self.device,
199
+ dtype=torch.float32,
200
+ )
201
+ self.spatial_shapes = spatial_shapes
202
+ self.level_start_index = level_start_index
203
+ self.sampling_loc = torch.rand(
204
+ batch,
205
+ num_query,
206
+ num_heads,
207
+ num_levels,
208
+ num_points,
209
+ 2,
210
+ device=self.device,
211
+ dtype=torch.float32,
212
+ )
213
+ self.attn_weight = torch.rand(
214
+ batch,
215
+ num_query,
216
+ num_heads,
217
+ num_levels,
218
+ num_points,
219
+ device=self.device,
220
+ dtype=torch.float32,
221
+ )
222
+ self.attn_weight = self.attn_weight / self.attn_weight.sum(-1, keepdim=True)
223
+ self.im2col_step = im2col_step
224
+
225
+ self.out = torch.empty(
226
+ batch,
227
+ num_query,
228
+ num_heads * channels,
229
+ device=self.device,
230
+ dtype=torch.float32,
231
+ )
232
+
233
+ def benchmark_large(self):
234
+ self.out = self.kernel.ms_deform_attn_forward(
235
+ self.value,
236
+ self.spatial_shapes,
237
+ self.level_start_index,
238
+ self.sampling_loc,
239
+ self.attn_weight,
240
+ self.im2col_step,
241
+ )
242
+
243
+ def verify_large(self) -> torch.Tensor:
244
+ return ms_deform_attn_reference(
245
+ self.value,
246
+ self.spatial_shapes,
247
+ self.level_start_index,
248
+ self.sampling_loc,
249
+ self.attn_weight,
250
+ )
build/torch210-cu128-x86_64-windows/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cu128-x86_64-windows/_deformable_detr_cuda_d8a6191.pyd ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f94825a148a77c630ae4f24f75f65e96dd6e3379653643aa2e81420ba61a9db3
3
+ size 9546240
build/torch210-cu128-x86_64-windows/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_d8a6191
3
+ ops = torch.ops._deformable_detr_cuda_d8a6191
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_d8a6191::{op_name}"
build/torch210-cu128-x86_64-windows/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import sys
3
+
4
+ import importlib
5
+ from pathlib import Path
6
+ from types import ModuleType
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cu128-x86_64-windows/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cu128-x86_64-windows/metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "10.0",
9
+ "10.1",
10
+ "12.0+PTX",
11
+ "7.0",
12
+ "7.2",
13
+ "7.5",
14
+ "8.0",
15
+ "8.6",
16
+ "8.7",
17
+ "8.9",
18
+ "9.0"
19
+ ]
20
+ }
21
+ }
build/torch210-cxx11-cu126-aarch64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:521f9ff226174c047c6d384b8e39a92a427c06496ad87465c554c2b51239e317
3
+ size 8606368
build/torch210-cxx11-cu126-aarch64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu126-aarch64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu126-aarch64-linux/metadata.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "7.0",
9
+ "7.2",
10
+ "7.5",
11
+ "8.0",
12
+ "8.6",
13
+ "8.7",
14
+ "8.9",
15
+ "9.0+PTX"
16
+ ]
17
+ }
18
+ }
build/torch210-cxx11-cu126-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu126-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ddc6b5a40afe614d97965a4f35113df59852130e21699792bb883dcdd8b1228f
3
+ size 8541080
build/torch210-cxx11-cu126-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu126-x86_64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu126-x86_64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu126-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "7.0",
9
+ "7.2",
10
+ "7.5",
11
+ "8.0",
12
+ "8.6",
13
+ "8.7",
14
+ "8.9",
15
+ "9.0+PTX"
16
+ ]
17
+ }
18
+ }
build/torch210-cxx11-cu128-aarch64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu128-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6de7e1d1805eb6d51f7588654319610b0a8fb55755ab5b6f2f2c42792ba060a9
3
+ size 11621120
build/torch210-cxx11-cu128-aarch64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu128-aarch64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu128-aarch64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu128-aarch64-linux/metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "10.0",
9
+ "10.1",
10
+ "12.0+PTX",
11
+ "7.0",
12
+ "7.2",
13
+ "7.5",
14
+ "8.0",
15
+ "8.6",
16
+ "8.7",
17
+ "8.9",
18
+ "9.0"
19
+ ]
20
+ }
21
+ }
build/torch210-cxx11-cu128-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu128-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c23af683cd86bb8f8a4426617c869403ea4fdba88b78935c676705e227523716
3
+ size 11524560
build/torch210-cxx11-cu128-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu128-x86_64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu128-x86_64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu128-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "10.0",
9
+ "10.1",
10
+ "12.0+PTX",
11
+ "7.0",
12
+ "7.2",
13
+ "7.5",
14
+ "8.0",
15
+ "8.6",
16
+ "8.7",
17
+ "8.9",
18
+ "9.0"
19
+ ]
20
+ }
21
+ }
build/torch210-cxx11-cu130-aarch64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu130-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e8bd9d5c6f5fdc673181ca67ec66a6f16d9c77a13fef13628b197aaa6d0ed98f
3
+ size 9890792
build/torch210-cxx11-cu130-aarch64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu130-aarch64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu130-aarch64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu130-aarch64-linux/metadata.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "10.0",
9
+ "11.0",
10
+ "12.0+PTX",
11
+ "7.5",
12
+ "8.0",
13
+ "8.6",
14
+ "8.7",
15
+ "8.9",
16
+ "9.0"
17
+ ]
18
+ }
19
+ }
build/torch210-cxx11-cu130-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch210-cxx11-cu130-x86_64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7c26502008f0002400df7ed708d8014210ad43655701340ce9089bc342bf9e4b
3
+ size 9808368
build/torch210-cxx11-cu130-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch210-cxx11-cu130-x86_64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch210-cxx11-cu130-x86_64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]
build/torch210-cxx11-cu130-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "version": 1,
3
+ "license": "Apache-2.0",
4
+ "python-depends": [],
5
+ "backend": {
6
+ "type": "cuda",
7
+ "archs": [
8
+ "10.0",
9
+ "11.0",
10
+ "12.0+PTX",
11
+ "7.5",
12
+ "8.0",
13
+ "8.6",
14
+ "8.7",
15
+ "8.9",
16
+ "9.0"
17
+ ]
18
+ }
19
+ }
build/torch211-cxx11-cu126-aarch64-linux/__init__.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List
2
+ import torch
3
+
4
+ from ._ops import ops
5
+ from . import layers
6
+
7
+
8
+ def ms_deform_attn_backward(
9
+ value: torch.Tensor,
10
+ spatial_shapes: torch.Tensor,
11
+ level_start_index: torch.Tensor,
12
+ sampling_loc: torch.Tensor,
13
+ attn_weight: torch.Tensor,
14
+ grad_output: torch.Tensor,
15
+ im2col_step: int,
16
+ ) -> List[torch.Tensor]:
17
+ return ops.ms_deform_attn_backward(
18
+ value,
19
+ spatial_shapes,
20
+ level_start_index,
21
+ sampling_loc,
22
+ attn_weight,
23
+ grad_output,
24
+ im2col_step,
25
+ )
26
+
27
+
28
+ def ms_deform_attn_forward(
29
+ value: torch.Tensor,
30
+ spatial_shapes: torch.Tensor,
31
+ level_start_index: torch.Tensor,
32
+ sampling_loc: torch.Tensor,
33
+ attn_weight: torch.Tensor,
34
+ im2col_step: int,
35
+ ) -> torch.Tensor:
36
+ return ops.ms_deform_attn_forward(
37
+ value,
38
+ spatial_shapes,
39
+ level_start_index,
40
+ sampling_loc,
41
+ attn_weight,
42
+ im2col_step,
43
+ )
44
+
45
+
46
+ __all__ = ["layers", "ms_deform_attn_forward", "ms_deform_attn_backward"]
build/torch211-cxx11-cu126-aarch64-linux/_deformable_detr_cuda_52e302f.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:422cd51f413bc85e60b606ab8ead8850331a3b5e288d8a0e4923b10a68c0f4e5
3
+ size 8606480
build/torch211-cxx11-cu126-aarch64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _deformable_detr_cuda_52e302f
3
+ ops = torch.ops._deformable_detr_cuda_52e302f
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_deformable_detr_cuda_52e302f::{op_name}"
build/torch211-cxx11-cu126-aarch64-linux/deformable_detr/__init__.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import ctypes
2
+ import importlib.util
3
+ import sys
4
+ from pathlib import Path
5
+ from types import ModuleType
6
+
7
+
8
+ def _import_from_path(file_path: Path) -> ModuleType:
9
+ # We cannot use the module name as-is, after adding it to `sys.modules`,
10
+ # it would also be used for other imports. So, we make a module name that
11
+ # depends on the path for it to be unique using the hex-encoded hash of
12
+ # the path.
13
+ path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
14
+ module_name = path_hash
15
+ spec = importlib.util.spec_from_file_location(module_name, file_path)
16
+ if spec is None:
17
+ raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
18
+ module = importlib.util.module_from_spec(spec)
19
+ if module is None:
20
+ raise ImportError(f"Cannot load module {module_name} from spec")
21
+ sys.modules[module_name] = module
22
+ spec.loader.exec_module(module) # type: ignore
23
+ return module
24
+
25
+
26
+ globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))
build/torch211-cxx11-cu126-aarch64-linux/layers.py ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Union, Tuple
2
+
3
+ from torch import Tensor
4
+ from torch.autograd import Function
5
+ from torch.autograd.function import once_differentiable
6
+ import torch.nn as nn
7
+
8
+ from ._ops import ops
9
+
10
+
11
+ class MultiScaleDeformableAttentionFunction(Function):
12
+ @staticmethod
13
+ def forward(
14
+ context,
15
+ value: Tensor,
16
+ value_spatial_shapes: Tensor,
17
+ value_level_start_index: Tensor,
18
+ sampling_locations: Tensor,
19
+ attention_weights: Tensor,
20
+ im2col_step: int,
21
+ ):
22
+ context.im2col_step = im2col_step
23
+ output = ops.ms_deform_attn_forward(
24
+ value,
25
+ value_spatial_shapes,
26
+ value_level_start_index,
27
+ sampling_locations,
28
+ attention_weights,
29
+ context.im2col_step,
30
+ )
31
+ context.save_for_backward(
32
+ value,
33
+ value_spatial_shapes,
34
+ value_level_start_index,
35
+ sampling_locations,
36
+ attention_weights,
37
+ )
38
+ return output
39
+
40
+ @staticmethod
41
+ @once_differentiable
42
+ def backward(context, grad_output):
43
+ (
44
+ value,
45
+ value_spatial_shapes,
46
+ value_level_start_index,
47
+ sampling_locations,
48
+ attention_weights,
49
+ ) = context.saved_tensors
50
+ grad_value, grad_sampling_loc, grad_attn_weight = ops.ms_deform_attn_backward(
51
+ value,
52
+ value_spatial_shapes,
53
+ value_level_start_index,
54
+ sampling_locations,
55
+ attention_weights,
56
+ grad_output,
57
+ context.im2col_step,
58
+ )
59
+
60
+ return grad_value, None, None, grad_sampling_loc, grad_attn_weight, None
61
+
62
+
63
+ class MultiScaleDeformableAttention(nn.Module):
64
+ def forward(
65
+ self,
66
+ value: Tensor,
67
+ value_spatial_shapes: Tensor,
68
+ value_spatial_shapes_list: List[Tuple],
69
+ level_start_index: Tensor,
70
+ sampling_locations: Tensor,
71
+ attention_weights: Tensor,
72
+ im2col_step: int,
73
+ ):
74
+ return MultiScaleDeformableAttentionFunction.apply(
75
+ value,
76
+ value_spatial_shapes,
77
+ level_start_index,
78
+ sampling_locations,
79
+ attention_weights,
80
+ im2col_step,
81
+ )
82
+
83
+
84
+ __all__ = ["MultiScaleDeformableAttention"]