Build uploaded using `kernels`.

build/torch28-cxx11-cu126-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch28-cxx11-cu126-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch28-cxx11-cu126-x86_64-linux/rotary/{_rotary_98ffc18.abi3.so → _rotary_dba7d1e.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3408f6e81f8f95fede157d02fdb2202bad3b740137d284f4929eb9b1fbd12bbb
-size 6817200
+oid sha256:afcfe44855a3e4bd60283a3c0e9b838a81177942d4997e48ff58e0542f168790
+size 5462056

build/torch28-cxx11-cu128-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch28-cxx11-cu128-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch28-cxx11-cu128-x86_64-linux/rotary/_rotary_98ffc18.abi3.so

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:99b283e0cc88386e746276ce9866dcd8d59c93bbff49ce02474357cca897d2a0
-size 10526400

build/{torch29-cxx11-cu126-x86_64-linux/rotary/_rotary_98ffc18.abi3.so → torch28-cxx11-cu128-x86_64-linux/rotary/_rotary_dba7d1e.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:10a3d2123227f7c24f5c56e5529de3d9d4873c741cfc1256a57d585021bb83e2
-size 6817088
+oid sha256:9092efc80c61670f0e9634e7b96a59c45c35cac470f85da5ba1a8ea184a15750
+size 7893120

build/torch28-cxx11-cu129-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch28-cxx11-cu129-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch28-cxx11-cu129-x86_64-linux/rotary/_rotary_98ffc18.abi3.so

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:50a460190d5b7e5f66016f23ce23989a66fa6e0bee967bf8b6ce9472706d55a6
-size 10586824

build/{torch29-cxx11-cu130-x86_64-linux/rotary/_rotary_98ffc18.abi3.so → torch28-cxx11-cu129-x86_64-linux/rotary/_rotary_dba7d1e.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5da8f52af54c4178e3494b3ec55c1c08fbf06ee4319c80b3de1c24b025faefb4
-size 9608112
+oid sha256:001611c5afea6f19fa26720bd93cca352fcb0eb9b50fb9786b03e80d0eefd1db
+size 7949960

build/torch28-cxx11-xpu20251-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch28-cxx11-xpu20251-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch28-cxx11-xpu20251-x86_64-linux/rotary/{_rotary_98ffc18.abi3.so → _rotary_dba7d1e.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ce70491bea7a387cea7ce474b8393dddc139645add8af8863d80c9ff3017a463
+oid sha256:036b2a03719657cb135cb93629f9da21c4d9d9f9537ff2792d151b32c2488c59
 size 2324536

build/torch29-cxx11-cu126-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch29-cxx11-cu126-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch29-cxx11-cu126-x86_64-linux/rotary/_rotary_dba7d1e.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:747406a42e14287a3cab577e65abf26cee64ec6324e2692be5ade3d22a1be2b3
+size 5461944

build/torch29-cxx11-cu128-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch29-cxx11-cu128-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch29-cxx11-cu128-x86_64-linux/rotary/_rotary_98ffc18.abi3.so

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8d39655db220a21ec886c5a896359755e247410999679e9bea86f51ee539b0b3
-size 10526504

build/torch29-cxx11-cu128-x86_64-linux/rotary/_rotary_dba7d1e.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9316296204902f06352d27d5702e11b54dc7580adec50fdde4be15d8351de03a
+size 7897320

build/torch29-cxx11-cu130-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch29-cxx11-cu130-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch29-cxx11-cu130-x86_64-linux/rotary/_rotary_dba7d1e.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c55687a6b3ec5f6fce8841387ca647eb2e57a4b387ea09e12c672a95040ea5c4
+size 7659376

build/torch29-cxx11-xpu20252-x86_64-linux/rotary/__init__.py

@@ -1,4 +1,4 @@
-from typing import Tuple
+from typing import Optional, Tuple
 import torch
 
 from ._ops import ops
@@ -12,8 +12,42 @@ def apply_rotary(
     out1: torch.Tensor,
     out2: torch.Tensor,
     conj: bool,
-):
+) -> None:
     ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
 
 
-__all__ = ["apply_rotary"]
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    position_ids: Optional[torch.Tensor] = None,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernels implementation to match transformers apply_rotary_pos_emb signature
+    """
+    cos = cos.unsqueeze(unsqueeze_dim)
+    sin = sin.unsqueeze(unsqueeze_dim)
+
+    q_rotated = q.clone()
+    k_rotated = k.clone()
+
+    # Get half dimension for rotation
+    half_dim = q.shape[-1] // 2
+    q1 = q_rotated[..., :half_dim]
+    q2 = q_rotated[..., half_dim:]
+    k1 = k_rotated[..., :half_dim]
+    k2 = k_rotated[..., half_dim:]
+    if cos.shape[-1] != half_dim:
+        # Trim cos/sin to match half_dim
+        cos = cos[..., :half_dim]
+        sin = sin[..., :half_dim]
+
+    apply_rotary(q1, q2, cos, sin, q1, q2, False)
+    apply_rotary(k1, k2, cos, sin, k1, k2, False)
+    return q_rotated, k_rotated
+
+
+__all__ = ["apply_rotary", "apply_rotary_transformers"]

build/torch29-cxx11-xpu20252-x86_64-linux/rotary/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_98ffc18
-ops = torch.ops._rotary_98ffc18
+from . import _rotary_dba7d1e
+ops = torch.ops._rotary_dba7d1e
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_98ffc18::{op_name}"
+    return f"_rotary_dba7d1e::{op_name}"

build/torch29-cxx11-xpu20252-x86_64-linux/rotary/{_rotary_98ffc18.abi3.so → _rotary_dba7d1e.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:32f216d78d17221dcd1233ee7b72c051ad6ce373f0d1c8a20d9cee5beda2b6a1
+oid sha256:70148df322edfbbcd55554ade33f184dcae0003a2f2b75e89d91f2a955dee7ca
 size 2273032