Build uploaded using `kernels`.

.gitattributes

@@ -40,3 +40,4 @@ build/torch29-xpu20252-x86_64-windows/rotary/_rotary_66b961a.pyd filter=lfs diff
 build/torch210-cu128-x86_64-windows/rotary/_rotary_9f63cc2.pyd filter=lfs diff=lfs merge=lfs -text
 build/torch210-xpu20253-x86_64-windows/rotary/_rotary_9f63cc2.pyd filter=lfs diff=lfs merge=lfs -text
 build/torch210-cu128-x86_64-windows/_rotary_cuda_07a01e5.pyd filter=lfs diff=lfs merge=lfs -text
+build/torch210-xpu20253-x86_64-windows/_rotary_xpu_07a01e5.pyd filter=lfs diff=lfs merge=lfs -text

build/torch210-xpu20253-x86_64-windows/__init__.py

@@ -0,0 +1,52 @@
+from typing import Optional, Tuple
+import torch
+
+from ._ops import ops
+
+
+def apply_rotary(
+    x1: torch.Tensor,
+    x2: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    out1: torch.Tensor,
+    out2: torch.Tensor,
+    conj: bool,
+) -> None:
+    ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
+
+
+def apply_rotary_transformers(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    cos: torch.Tensor,
+    sin: torch.Tensor,
+    unsqueeze_dim: int = 1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Rotary kernel implementation wrapper
+    Adapts rotary kernel 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/torch210-xpu20253-x86_64-windows/{rotary/_ops.py → _ops.py}

@@ -1,9 +1,9 @@
 import torch
-from . import _rotary_9f63cc2
-ops = torch.ops._rotary_9f63cc2
+from . import _rotary_xpu_07a01e5
+ops = torch.ops._rotary_xpu_07a01e5
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rotary_9f63cc2::{op_name}"
+    return f"_rotary_xpu_07a01e5::{op_name}"

build/torch210-xpu20253-x86_64-windows/{rotary/_rotary_9f63cc2.pyd → _rotary_xpu_07a01e5.pyd}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a8ece9a8c015b26093abe355961e81d14f6a26bf809cce3f465f1125d97039ea
-size 202240
+oid sha256:34579b75b36dea1b18446ac679fde1e89d16b25350e2cb8213acfc19434bba8b
+size 396288

build/torch210-xpu20253-x86_64-windows/metadata.json

@@ -1,4 +1,5 @@
 {
   "version": 1,
+  "license": "BSD-3-Clause",
   "python-depends": []
 }

build/torch210-xpu20253-x86_64-windows/rotary/__init__.py

@@ -1,52 +1,26 @@
-from typing import Optional, Tuple
-import torch
-
-from ._ops import ops
-
-
-def apply_rotary(
-    x1: torch.Tensor,
-    x2: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-    out1: torch.Tensor,
-    out2: torch.Tensor,
-    conj: bool,
-) -> None:
-    ops.apply_rotary(x1, x2, cos, sin, out1, out2, conj)
-
-
-def apply_rotary_transformers(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    cos: torch.Tensor,
-    sin: torch.Tensor,
-    unsqueeze_dim: int = 1,
-) -> Tuple[torch.Tensor, torch.Tensor]:
-    """
-    Rotary kernel implementation wrapper
-    Adapts rotary kernel 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"]
+import ctypes
+import sys
+
+import importlib
+from pathlib import Path
+from types import ModuleType
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))