kernels-bot commited on
Commit
79a3cae
·
verified ·
1 Parent(s): 18186f3

Uploaded using `kernel-builder`.

Browse files
Files changed (44) hide show
  1. build/torch210-cxx11-cu126-x86_64-linux/_ops.py +3 -3
  2. build/torch210-cxx11-cu126-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  3. build/torch210-cxx11-cu126-x86_64-linux/metadata.json +1 -1
  4. build/torch210-cxx11-cu128-x86_64-linux/_ops.py +3 -3
  5. build/torch210-cxx11-cu128-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  6. build/torch210-cxx11-cu128-x86_64-linux/metadata.json +1 -1
  7. build/torch210-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  8. build/torch210-cxx11-cu130-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  9. build/torch210-cxx11-cu130-x86_64-linux/metadata.json +1 -1
  10. build/torch210-cxx11-xpu20253-x86_64-linux/_ops.py +3 -3
  11. build/{torch211-cxx11-xpu20253-x86_64-linux/_rotary_xpu_86f75d9.abi3.so → torch210-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so} +1 -1
  12. build/torch210-cxx11-xpu20253-x86_64-linux/metadata.json +1 -1
  13. build/torch211-cxx11-cu126-x86_64-linux/_ops.py +3 -3
  14. build/torch211-cxx11-cu126-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  15. build/torch211-cxx11-cu126-x86_64-linux/metadata.json +1 -1
  16. build/torch211-cxx11-cu128-x86_64-linux/_ops.py +3 -3
  17. build/torch211-cxx11-cu128-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  18. build/torch211-cxx11-cu128-x86_64-linux/metadata.json +1 -1
  19. build/torch211-cxx11-cu130-x86_64-linux/_ops.py +3 -3
  20. build/torch211-cxx11-cu130-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} +1 -1
  21. build/torch211-cxx11-cu130-x86_64-linux/metadata.json +1 -1
  22. build/torch211-cxx11-xpu20253-x86_64-linux/_ops.py +3 -3
  23. build/{torch210-cxx11-xpu20253-x86_64-linux/_rotary_xpu_86f75d9.abi3.so → torch211-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so} +1 -1
  24. build/torch211-cxx11-xpu20253-x86_64-linux/metadata.json +1 -1
  25. build/torch212-cxx11-cu126-x86_64-linux/__init__.py +52 -0
  26. build/torch212-cxx11-cu126-x86_64-linux/_ops.py +9 -0
  27. build/torch212-cxx11-cu126-x86_64-linux/_rotary_cuda_725ebfc.abi3.so +3 -0
  28. build/torch212-cxx11-cu126-x86_64-linux/metadata.json +20 -0
  29. build/torch212-cxx11-cu126-x86_64-linux/rotary/__init__.py +26 -0
  30. build/torch212-cxx11-cu130-x86_64-linux/__init__.py +52 -0
  31. build/torch212-cxx11-cu130-x86_64-linux/_ops.py +9 -0
  32. build/torch212-cxx11-cu130-x86_64-linux/_rotary_cuda_725ebfc.abi3.so +3 -0
  33. build/torch212-cxx11-cu130-x86_64-linux/metadata.json +21 -0
  34. build/torch212-cxx11-cu130-x86_64-linux/rotary/__init__.py +26 -0
  35. build/torch212-cxx11-cu132-x86_64-linux/__init__.py +52 -0
  36. build/torch212-cxx11-cu132-x86_64-linux/_ops.py +9 -0
  37. build/torch212-cxx11-cu132-x86_64-linux/_rotary_cuda_725ebfc.abi3.so +3 -0
  38. build/torch212-cxx11-cu132-x86_64-linux/metadata.json +21 -0
  39. build/torch212-cxx11-cu132-x86_64-linux/rotary/__init__.py +26 -0
  40. build/torch212-cxx11-xpu20253-x86_64-linux/__init__.py +52 -0
  41. build/torch212-cxx11-xpu20253-x86_64-linux/_ops.py +9 -0
  42. build/torch212-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so +3 -0
  43. build/torch212-cxx11-xpu20253-x86_64-linux/metadata.json +10 -0
  44. build/torch212-cxx11-xpu20253-x86_64-linux/rotary/__init__.py +26 -0
build/torch210-cxx11-cu126-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch210-cxx11-cu126-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b1d5394246163c10eb12e82bd4151ef5f26044ebef06ff6eb978204ce9569fe4
3
  size 8200568
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:39d341299aaec16caebbdebde5fa235e8d2fe93304f64f39d5d25aca27bb1e47
3
  size 8200568
build/torch210-cxx11-cu126-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch210-cxx11-cu128-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch210-cxx11-cu128-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2776016e55fced29f3f7da4592a64690abc7d13f13a1abeb97d0e5ac0ae79570
3
  size 11905904
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cc218cdb6a055456094a6f518d2f3b43fb1a0d516907cc8dedee3d6511b3217b
3
  size 11905904
build/torch210-cxx11-cu128-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch210-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch210-cxx11-cu130-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d48e50159c8070fad7389d57fd8a828ae09b0bf7c7823d9ce175e60a68d04321
3
  size 10307160
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:51653f32f84e840f9e7182eb129bf31e528c513d59802f8560909cfcc3707b85
3
  size 10307160
build/torch210-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch210-cxx11-xpu20253-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_xpu_86f75d9
3
- ops = torch.ops._rotary_xpu_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_xpu_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_xpu_725ebfc
3
+ ops = torch.ops._rotary_xpu_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_xpu_725ebfc::{op_name}"
build/{torch211-cxx11-xpu20253-x86_64-linux/_rotary_xpu_86f75d9.abi3.so → torch210-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2e78d1a36648ffeaa4835901d94e93de2d1439fedd504345d71984fc5cbfebd0
3
  size 2301504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1a6425022ede63a5cdaa95626f393fde741b34037d7bd588044d324c64a0588
3
  size 2301504
build/torch210-cxx11-xpu20253-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_xpu_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_xpu_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch211-cxx11-cu126-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch211-cxx11-cu126-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dcafe175d2f5a54a5133884109c478238dfe9000f178a317d27881b5b32e168d
3
  size 8193600
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7b43b3223ec25d835af8c4020d87eca05c60efc111211a2a7196c3d692e77179
3
  size 8193600
build/torch211-cxx11-cu126-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch211-cxx11-cu128-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch211-cxx11-cu128-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:04ea5f41dc222373c069accb827df58ab6156edba6e2684fd5ebd6b1ec2993ce
3
  size 11894840
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3037fcf467a2bddcc34cfa70d188ae2ed190a87c6b56818f20c0c39b9bc2342f
3
  size 11894840
build/torch211-cxx11-cu128-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch211-cxx11-cu130-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_cuda_86f75d9
3
- ops = torch.ops._rotary_cuda_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_cuda_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch211-cxx11-cu130-x86_64-linux/{_rotary_cuda_86f75d9.abi3.so → _rotary_cuda_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:81f997832b95bdd5bfd0d0ce46ef13f910345b0c47ae163431b8ecde7947482d
3
  size 10296096
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5d95c0078ab5f1e7462f6b607890d58044d40bd02a5ecd749c9e6cac33424faf
3
  size 10296096
build/torch211-cxx11-cu130-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_cuda_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch211-cxx11-xpu20253-x86_64-linux/_ops.py CHANGED
@@ -1,9 +1,9 @@
1
  import torch
2
- from . import _rotary_xpu_86f75d9
3
- ops = torch.ops._rotary_xpu_86f75d9
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
- return f"_rotary_xpu_86f75d9::{op_name}"
 
1
  import torch
2
+ from . import _rotary_xpu_725ebfc
3
+ ops = torch.ops._rotary_xpu_725ebfc
4
 
5
  def add_op_namespace_prefix(op_name: str):
6
  """
7
  Prefix op by namespace.
8
  """
9
+ return f"_rotary_xpu_725ebfc::{op_name}"
build/{torch210-cxx11-xpu20253-x86_64-linux/_rotary_xpu_86f75d9.abi3.so → torch211-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so} RENAMED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:31984c0927a018a70c49d356690a4cd08f39ec0109f6a5a440e93f05bd6da25e
3
  size 2301504
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8482e87f3cb8d30070a9367ea2d01398497f52fc7e2369db3d763118c3c92566
3
  size 2301504
build/torch211-cxx11-xpu20253-x86_64-linux/metadata.json CHANGED
@@ -1,6 +1,6 @@
1
  {
2
  "name": "rotary",
3
- "id": "_rotary_xpu_86f75d9",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
 
1
  {
2
  "name": "rotary",
3
+ "id": "_rotary_xpu_725ebfc",
4
  "version": 1,
5
  "license": "BSD-3-Clause",
6
  "python-depends": [],
build/torch212-cxx11-cu126-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Tuple
2
+ import torch
3
+
4
+ from ._ops import ops
5
+
6
+
7
+ def apply_rotary(
8
+ x1: torch.Tensor,
9
+ x2: torch.Tensor,
10
+ cos: torch.Tensor,
11
+ sin: torch.Tensor,
12
+ out1: torch.Tensor,
13
+ out2: torch.Tensor,
14
+ conj: bool,
15
+ ) -> None:
16
+ ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
17
+
18
+
19
+ def apply_rotary_transformers(
20
+ q: torch.Tensor,
21
+ k: torch.Tensor,
22
+ cos: torch.Tensor,
23
+ sin: torch.Tensor,
24
+ unsqueeze_dim: int = 1,
25
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
26
+ """
27
+ Rotary kernel implementation wrapper
28
+ Adapts rotary kernel implementation to match transformers apply_rotary_pos_emb signature
29
+ """
30
+ cos = cos.unsqueeze(unsqueeze_dim)
31
+ sin = sin.unsqueeze(unsqueeze_dim)
32
+
33
+ q_rotated = q.clone()
34
+ k_rotated = k.clone()
35
+
36
+ # Get half dimension for rotation
37
+ half_dim = q.shape[-1] // 2
38
+ q1 = q_rotated[..., :half_dim]
39
+ q2 = q_rotated[..., half_dim:]
40
+ k1 = k_rotated[..., :half_dim]
41
+ k2 = k_rotated[..., half_dim:]
42
+ if cos.shape[-1] != half_dim:
43
+ # Trim cos/sin to match half_dim
44
+ cos = cos[..., :half_dim]
45
+ sin = sin[..., :half_dim]
46
+
47
+ apply_rotary(q1, q2, cos, sin, q1, q2, False)
48
+ apply_rotary(k1, k2, cos, sin, k1, k2, False)
49
+ return q_rotated, k_rotated
50
+
51
+
52
+ __all__ = ["apply_rotary", "apply_rotary_transformers"]
build/torch212-cxx11-cu126-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch212-cxx11-cu126-x86_64-linux/_rotary_cuda_725ebfc.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b7e9bd09e74ab8c1a7468d01fc6c21185ef9fce899f24e020d09c17e8cfe4a08
3
+ size 8195384
build/torch212-cxx11-cu126-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,20 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
+ "version": 1,
5
+ "license": "BSD-3-Clause",
6
+ "python-depends": [],
7
+ "backend": {
8
+ "type": "cuda",
9
+ "archs": [
10
+ "7.0",
11
+ "7.2",
12
+ "7.5",
13
+ "8.0",
14
+ "8.6",
15
+ "8.7",
16
+ "8.9",
17
+ "9.0+PTX"
18
+ ]
19
+ }
20
+ }
build/torch212-cxx11-cu126-x86_64-linux/rotary/__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/torch212-cxx11-cu130-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Tuple
2
+ import torch
3
+
4
+ from ._ops import ops
5
+
6
+
7
+ def apply_rotary(
8
+ x1: torch.Tensor,
9
+ x2: torch.Tensor,
10
+ cos: torch.Tensor,
11
+ sin: torch.Tensor,
12
+ out1: torch.Tensor,
13
+ out2: torch.Tensor,
14
+ conj: bool,
15
+ ) -> None:
16
+ ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
17
+
18
+
19
+ def apply_rotary_transformers(
20
+ q: torch.Tensor,
21
+ k: torch.Tensor,
22
+ cos: torch.Tensor,
23
+ sin: torch.Tensor,
24
+ unsqueeze_dim: int = 1,
25
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
26
+ """
27
+ Rotary kernel implementation wrapper
28
+ Adapts rotary kernel implementation to match transformers apply_rotary_pos_emb signature
29
+ """
30
+ cos = cos.unsqueeze(unsqueeze_dim)
31
+ sin = sin.unsqueeze(unsqueeze_dim)
32
+
33
+ q_rotated = q.clone()
34
+ k_rotated = k.clone()
35
+
36
+ # Get half dimension for rotation
37
+ half_dim = q.shape[-1] // 2
38
+ q1 = q_rotated[..., :half_dim]
39
+ q2 = q_rotated[..., half_dim:]
40
+ k1 = k_rotated[..., :half_dim]
41
+ k2 = k_rotated[..., half_dim:]
42
+ if cos.shape[-1] != half_dim:
43
+ # Trim cos/sin to match half_dim
44
+ cos = cos[..., :half_dim]
45
+ sin = sin[..., :half_dim]
46
+
47
+ apply_rotary(q1, q2, cos, sin, q1, q2, False)
48
+ apply_rotary(k1, k2, cos, sin, k1, k2, False)
49
+ return q_rotated, k_rotated
50
+
51
+
52
+ __all__ = ["apply_rotary", "apply_rotary_transformers"]
build/torch212-cxx11-cu130-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch212-cxx11-cu130-x86_64-linux/_rotary_cuda_725ebfc.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:720b8c363ad2feaeda8476661c5627265524b8cd704a1b5f4c7946ef9e2d0c9c
3
+ size 10301968
build/torch212-cxx11-cu130-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
+ "version": 1,
5
+ "license": "BSD-3-Clause",
6
+ "python-depends": [],
7
+ "backend": {
8
+ "type": "cuda",
9
+ "archs": [
10
+ "10.0",
11
+ "11.0",
12
+ "12.0+PTX",
13
+ "7.5",
14
+ "8.0",
15
+ "8.6",
16
+ "8.7",
17
+ "8.9",
18
+ "9.0"
19
+ ]
20
+ }
21
+ }
build/torch212-cxx11-cu130-x86_64-linux/rotary/__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/torch212-cxx11-cu132-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Tuple
2
+ import torch
3
+
4
+ from ._ops import ops
5
+
6
+
7
+ def apply_rotary(
8
+ x1: torch.Tensor,
9
+ x2: torch.Tensor,
10
+ cos: torch.Tensor,
11
+ sin: torch.Tensor,
12
+ out1: torch.Tensor,
13
+ out2: torch.Tensor,
14
+ conj: bool,
15
+ ) -> None:
16
+ ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
17
+
18
+
19
+ def apply_rotary_transformers(
20
+ q: torch.Tensor,
21
+ k: torch.Tensor,
22
+ cos: torch.Tensor,
23
+ sin: torch.Tensor,
24
+ unsqueeze_dim: int = 1,
25
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
26
+ """
27
+ Rotary kernel implementation wrapper
28
+ Adapts rotary kernel implementation to match transformers apply_rotary_pos_emb signature
29
+ """
30
+ cos = cos.unsqueeze(unsqueeze_dim)
31
+ sin = sin.unsqueeze(unsqueeze_dim)
32
+
33
+ q_rotated = q.clone()
34
+ k_rotated = k.clone()
35
+
36
+ # Get half dimension for rotation
37
+ half_dim = q.shape[-1] // 2
38
+ q1 = q_rotated[..., :half_dim]
39
+ q2 = q_rotated[..., half_dim:]
40
+ k1 = k_rotated[..., :half_dim]
41
+ k2 = k_rotated[..., half_dim:]
42
+ if cos.shape[-1] != half_dim:
43
+ # Trim cos/sin to match half_dim
44
+ cos = cos[..., :half_dim]
45
+ sin = sin[..., :half_dim]
46
+
47
+ apply_rotary(q1, q2, cos, sin, q1, q2, False)
48
+ apply_rotary(k1, k2, cos, sin, k1, k2, False)
49
+ return q_rotated, k_rotated
50
+
51
+
52
+ __all__ = ["apply_rotary", "apply_rotary_transformers"]
build/torch212-cxx11-cu132-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _rotary_cuda_725ebfc
3
+ ops = torch.ops._rotary_cuda_725ebfc
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_rotary_cuda_725ebfc::{op_name}"
build/torch212-cxx11-cu132-x86_64-linux/_rotary_cuda_725ebfc.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:eb233b9be14a2d9c5bbb7b860973db6dd48aad4f5b760dd358880d079bff3b23
3
+ size 10490208
build/torch212-cxx11-cu132-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "rotary",
3
+ "id": "_rotary_cuda_725ebfc",
4
+ "version": 1,
5
+ "license": "BSD-3-Clause",
6
+ "python-depends": [],
7
+ "backend": {
8
+ "type": "cuda",
9
+ "archs": [
10
+ "10.0",
11
+ "11.0",
12
+ "12.0+PTX",
13
+ "7.5",
14
+ "8.0",
15
+ "8.6",
16
+ "8.7",
17
+ "8.9",
18
+ "9.0"
19
+ ]
20
+ }
21
+ }
build/torch212-cxx11-cu132-x86_64-linux/rotary/__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/torch212-cxx11-xpu20253-x86_64-linux/__init__.py ADDED
@@ -0,0 +1,52 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Tuple
2
+ import torch
3
+
4
+ from ._ops import ops
5
+
6
+
7
+ def apply_rotary(
8
+ x1: torch.Tensor,
9
+ x2: torch.Tensor,
10
+ cos: torch.Tensor,
11
+ sin: torch.Tensor,
12
+ out1: torch.Tensor,
13
+ out2: torch.Tensor,
14
+ conj: bool,
15
+ ) -> None:
16
+ ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
17
+
18
+
19
+ def apply_rotary_transformers(
20
+ q: torch.Tensor,
21
+ k: torch.Tensor,
22
+ cos: torch.Tensor,
23
+ sin: torch.Tensor,
24
+ unsqueeze_dim: int = 1,
25
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
26
+ """
27
+ Rotary kernel implementation wrapper
28
+ Adapts rotary kernel implementation to match transformers apply_rotary_pos_emb signature
29
+ """
30
+ cos = cos.unsqueeze(unsqueeze_dim)
31
+ sin = sin.unsqueeze(unsqueeze_dim)
32
+
33
+ q_rotated = q.clone()
34
+ k_rotated = k.clone()
35
+
36
+ # Get half dimension for rotation
37
+ half_dim = q.shape[-1] // 2
38
+ q1 = q_rotated[..., :half_dim]
39
+ q2 = q_rotated[..., half_dim:]
40
+ k1 = k_rotated[..., :half_dim]
41
+ k2 = k_rotated[..., half_dim:]
42
+ if cos.shape[-1] != half_dim:
43
+ # Trim cos/sin to match half_dim
44
+ cos = cos[..., :half_dim]
45
+ sin = sin[..., :half_dim]
46
+
47
+ apply_rotary(q1, q2, cos, sin, q1, q2, False)
48
+ apply_rotary(k1, k2, cos, sin, k1, k2, False)
49
+ return q_rotated, k_rotated
50
+
51
+
52
+ __all__ = ["apply_rotary", "apply_rotary_transformers"]
build/torch212-cxx11-xpu20253-x86_64-linux/_ops.py ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ from . import _rotary_xpu_725ebfc
3
+ ops = torch.ops._rotary_xpu_725ebfc
4
+
5
+ def add_op_namespace_prefix(op_name: str):
6
+ """
7
+ Prefix op by namespace.
8
+ """
9
+ return f"_rotary_xpu_725ebfc::{op_name}"
build/torch212-cxx11-xpu20253-x86_64-linux/_rotary_xpu_725ebfc.abi3.so ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:42a11c6606ad23c9cd2f3e8dff22be3a2349efefff46a1e42cb87f7785b7fcb6
3
+ size 2301512
build/torch212-cxx11-xpu20253-x86_64-linux/metadata.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "name": "rotary",
3
+ "id": "_rotary_xpu_725ebfc",
4
+ "version": 1,
5
+ "license": "BSD-3-Clause",
6
+ "python-depends": [],
7
+ "backend": {
8
+ "type": "xpu"
9
+ }
10
+ }
build/torch212-cxx11-xpu20253-x86_64-linux/rotary/__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")))