Build uploaded using `kernels`.

.gitattributes

@@ -33,3 +33,10 @@ 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/torch210-cxx11-cu126-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu128-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu130-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu126-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu128-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu130-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu129-aarch64-linux/_layer_norm_cuda_143103b.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-cxx11-cu126-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch210-cxx11-cu126-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e4ad4d5e8dee889d696c0b60ac2aa7c24586412b5605a33722f45a13a077b88d
+size 711710472

build/torch210-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch210-cxx11-cu126-aarch64-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/torch210-cxx11-cu126-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch210-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,13 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch210-cxx11-cu128-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch210-cxx11-cu128-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af5417fb89bb13567875514abfde1d4a087bb0295de1f53aaaf20edb0eb562c1
+size 1231083200

build/torch210-cxx11-cu128-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch210-cxx11-cu128-aarch64-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/torch210-cxx11-cu128-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch210-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,15 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch210-cxx11-cu130-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch210-cxx11-cu130-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c61421662fd4b22a3eb3d75fa5d9fcd3775ad8d8bbcf9c638b89e427111927b3
+size 1235993064

build/torch210-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch210-cxx11-cu130-aarch64-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/torch210-cxx11-cu130-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch210-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,15 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu126-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch211-cxx11-cu126-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d1b35f4d42743357fe2fbd9feb0ce20d28506154325e0745c97aa99d829bda8
+size 711706784

build/torch211-cxx11-cu126-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch211-cxx11-cu126-aarch64-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/torch211-cxx11-cu126-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch211-cxx11-cu126-aarch64-linux/metadata.json

@@ -0,0 +1,13 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu128-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch211-cxx11-cu128-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5e69e28fd3ca606b2519b7d64c292a8008db36f8b049d2bde0bf22c2fee0849
+size 1231079512

build/torch211-cxx11-cu128-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch211-cxx11-cu128-aarch64-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/torch211-cxx11-cu128-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch211-cxx11-cu128-aarch64-linux/metadata.json

@@ -0,0 +1,15 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu130-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch211-cxx11-cu130-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3147d8b7ea6ae20c72894191e79992c1825505e22b12b9fdf6df9d666504ab82
+size 1235989376

build/torch211-cxx11-cu130-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch211-cxx11-cu130-aarch64-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/torch211-cxx11-cu130-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch211-cxx11-cu130-aarch64-linux/metadata.json

@@ -0,0 +1,15 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch29-cxx11-cu129-aarch64-linux/__init__.py

@@ -0,0 +1,26 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+from . import layers
+
+def dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_fwd(input, gamma, beta, rowscale, colscale, x0_subset, z_subset, dropout_p, epsilon, rowscale_const, z_numrows, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_bwd(dz, dx, x, mu, rsigma, gamma, rowscale, colscale, x0_subset, z_subset, dropout_p, rowscale_const, x0_numrows, has_residual, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_fwd(input, gamma0, beta0, gamma1, beta1, dropout_p, epsilon, gen, residual_in_fp32, is_rms_norm)
+
+def dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm):
+    return ops.dropout_add_ln_parallel_residual_bwd(dz0, dz1, dx, x, mu, rsigma, gamma0, gamma1, dropout_p, has_x1, has_residual, is_rms_norm)
+
+__all__ = [
+    "layers",
+    "dropout_add_ln_fwd",
+    "dropout_add_ln_bwd",
+    "dropout_add_ln_parallel_residual_fwd",
+    "dropout_add_ln_parallel_residual_bwd",
+]

build/torch29-cxx11-cu129-aarch64-linux/_layer_norm_cuda_143103b.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5edebe8b0f52b1c169e80f3502abb83cfc4cf2438b5771387e8dfd9254eef1a8
+size 1282721000

build/torch29-cxx11-cu129-aarch64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _layer_norm_cuda_143103b
+ops = torch.ops._layer_norm_cuda_143103b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_layer_norm_cuda_143103b::{op_name}"

build/torch29-cxx11-cu129-aarch64-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-cu129-aarch64-linux/layers.py

@@ -0,0 +1,51 @@
+import torch
+import torch.nn as nn
+
+from ._ops import ops
+
+
+class LayerNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = False,
+        )
+        return output[0].view(hidden_states.shape)
+
+class LlamaRMSNorm(nn.Module):
+    weight: torch.Tensor
+    variance_epsilon: float
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        output = ops.dropout_add_ln_fwd(
+            hidden_states.view(-1, hidden_states.shape[-1]),
+            gamma = self.weight,
+            beta = None,
+            rowscale = None,
+            colscale = None,
+            x0_subset = None,
+            z_subset = None,
+            dropout_p = 0,
+            epsilon = self.variance_epsilon,
+            rowscale_const = 1.0,
+            z_numrows = hidden_states.shape[1],
+            gen = None,
+            residual_in_fp32 = False,
+            is_rms_norm = True,
+        )
+        return output[0].view(hidden_states.shape)

build/torch29-cxx11-cu129-aarch64-linux/metadata.json

@@ -0,0 +1,15 @@
+{
+  "version": 1,
+  "license": "BSD-3-Clause",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "12.0",
+      "8.0",
+      "8.9",
+      "9.0"
+    ]
+  }
+}