Add grafted torch2.9+cu130 build variant

.gitattributes

@@ -111,3 +111,4 @@ build/torch211-cxx11-cu130-aarch64-linux/dropout_layer_norm.cpython-312-aarch64-
 build/torch211-cxx11-cu130-x86_64-linux/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-aarch64-linux/dropout_layer_norm.cpython-312-aarch64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-x86_64-linux/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu130-x86_64-linux/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text

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

@@ -0,0 +1,38 @@
+"""torch 2.9 / cu12.8 build variant grafted from the odysseyml flash-attention fork.
+
+Wraps the pre-built ``dropout_layer_norm`` extension from the
+``odyssey-fused-kernels`` wheel (tag ``odyssey-v2.8.3-fused-1``, built for
+sm_80/90/100/120) because upstream kernels-community/layer-norm only provides
+a cu129 build for torch 2.9.
+
+Unlike the kernels-community builds (which drop ``residual`` from the fused
+op signatures), this build keeps flash-attention's original signature with
+``residual`` as the second argument. Consumers can check ``SUPPORTS_RESIDUAL``.
+"""
+
+import importlib.util
+from pathlib import Path
+
+# True: kernels with flash-attention's original signature: fwd takes
+# (x0, residual, gamma, beta, ...) and bwd is supported.
+SUPPORTS_RESIDUAL = True
+
+# The extension's PyInit_* symbol is derived from the module name passed to
+# the loader, so the spec name must exactly equal the .so module name.
+_so_path = next(Path(__file__).parent.glob("dropout_layer_norm*.so"))
+_spec = importlib.util.spec_from_file_location("dropout_layer_norm", _so_path)
+_ext = importlib.util.module_from_spec(_spec)
+_spec.loader.exec_module(_ext)
+
+dropout_add_ln_fwd = _ext.dropout_add_ln_fwd
+dropout_add_ln_bwd = _ext.dropout_add_ln_bwd
+dropout_add_ln_parallel_residual_fwd = _ext.dropout_add_ln_parallel_residual_fwd
+dropout_add_ln_parallel_residual_bwd = _ext.dropout_add_ln_parallel_residual_bwd
+
+__all__ = [
+    "SUPPORTS_RESIDUAL",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch29-cxx11-cu130-x86_64-linux/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e60775caf40e0db1dc8de8a64b030ed3d3341a09ab7fe92c37f48981087d544d
+size 1020829008

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

@@ -0,0 +1,26 @@
+import ctypes
+import importlib.util
+import sys
+from pathlib import Path
+from types import ModuleType
+
+
+def _import_from_path(file_path: Path) -> ModuleType:
+    # We cannot use the module name as-is, after adding it to `sys.modules`,
+    # it would also be used for other imports. So, we make a module name that
+    # depends on the path for it to be unique using the hex-encoded hash of
+    # the path.
+    path_hash = "{:x}".format(ctypes.c_size_t(hash(file_path.absolute())).value)
+    module_name = path_hash
+    spec = importlib.util.spec_from_file_location(module_name, file_path)
+    if spec is None:
+        raise ImportError(f"Cannot load spec for {module_name} from {file_path}")
+    module = importlib.util.module_from_spec(spec)
+    if module is None:
+        raise ImportError(f"Cannot load module {module_name} from spec")
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)  # type: ignore
+    return module
+
+
+globals().update(vars(_import_from_path(Path(__file__).parent.parent / "__init__.py")))

build/torch29-cxx11-cu130-x86_64-linux/metadata.json

@@ -0,0 +1,14 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "8.0",
+      "9.0",
+      "10.0",
+      "12.0"
+    ]
+  }
+}