dropout_layer_norm kernels build: torch2.9 cu12.8 cp312 sm80/90/100/120 (odyssey-v2.8.3-fused-1)

.gitattributes

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu128-aarch64-linux/dropout_layer_norm/dropout_layer_norm.cpython-312-aarch64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu128-x86_64-linux/dropout_layer_norm/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,22 @@
+---
+tags:
+  - kernel
+---
+
+# dropout_layer_norm
+
+Pre-built `dropout_layer_norm` CUDA extension from `csrc/layer_norm` of the
+[odysseyml/flash-attention](https://github.com/odysseyml/flash-attention) fork
+(tag `odyssey-v2.8.3-fused-1`), repackaged in HF kernels build-variant layout.
+
+Provides fused (dropout + residual add + LayerNorm/RMSNorm):
+`dropout_add_ln_fwd`, `dropout_add_ln_bwd`,
+`dropout_add_ln_parallel_residual_fwd`, `dropout_add_ln_parallel_residual_bwd`.
+
+Built for torch 2.9 / cu12.8 / cp312, sm_80/90/100/120 (x86_64 + aarch64).
+
+```python
+from kernels import get_kernel
+
+dropout_layer_norm = get_kernel("odyssey-systems/dropout-layer-norm")
+```

build/torch29-cxx11-cu128-aarch64-linux/dropout_layer_norm/__init__.py

@@ -0,0 +1,21 @@
+"""Pre-built CUDA extension from the flash-attention csrc tree (odysseyml fork).
+
+Packaged in HF kernels build-variant layout so it loads via
+``kernels.get_kernel("odyssey-systems/dropout_layer_norm")``.
+The shared library was compiled for sm_80/90/100/120
+(see https://github.com/odysseyml/flash-attention).
+"""
+
+import importlib.util
+from pathlib import Path
+
+# 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)
+
+globals().update({_name: getattr(_ext, _name) for _name in dir(_ext) if not _name.startswith("_")})
+
+del _name, _ext, _spec, _so_path, importlib, Path

build/torch29-cxx11-cu128-aarch64-linux/dropout_layer_norm/dropout_layer_norm.cpython-312-aarch64-linux-gnu.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cb431ebaceb7b12f0a5ad50eeca97d1aeaec16d13bbdedf807cae8a5dc18c2cb
+size 1258734424

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

@@ -0,0 +1,21 @@
+"""Pre-built CUDA extension from the flash-attention csrc tree (odysseyml fork).
+
+Packaged in HF kernels build-variant layout so it loads via
+``kernels.get_kernel("odyssey-systems/dropout_layer_norm")``.
+The shared library was compiled for sm_80/90/100/120
+(see https://github.com/odysseyml/flash-attention).
+"""
+
+import importlib.util
+from pathlib import Path
+
+# 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)
+
+globals().update({_name: getattr(_ext, _name) for _name in dir(_ext) if not _name.startswith("_")})
+
+del _name, _ext, _spec, _so_path, importlib, Path

build/torch29-cxx11-cu128-x86_64-linux/dropout_layer_norm/dropout_layer_norm.cpython-312-x86_64-linux-gnu.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ae8a09f42c4c44a01120b1e881f9966dada8eb4acdb5263297ff8b21305bbc5
+size 1259466744