Build uploaded using `kernels`.

.gitattributes

@@ -94,3 +94,7 @@ build/torch210-cxx11-cpu-x86_64-linux/_rmsnorm_cpu_cec90b8.abi3.so filter=lfs di
 build/torch210-cxx11-xpu20253-x86_64-linux/_rmsnorm_xpu_cec90b8.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-cxx11-cpu-x86_64-linux/_rmsnorm_cpu_cec90b8.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-xpu20252-x86_64-linux/_rmsnorm_xpu_cec90b8.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cpu-x86_64-linux/_rmsnorm_cpu_1a02f6f.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-xpu20253-x86_64-linux/_rmsnorm_xpu_1a02f6f.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cpu-x86_64-linux/_rmsnorm_cpu_1a02f6f.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-xpu20253-x86_64-linux/_rmsnorm_xpu_1a02f6f.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-cxx11-cpu-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rmsnorm_cpu_cec90b8
-ops = torch.ops._rmsnorm_cpu_cec90b8
+from . import _rmsnorm_cpu_1a02f6f
+ops = torch.ops._rmsnorm_cpu_1a02f6f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rmsnorm_cpu_cec90b8::{op_name}"
+    return f"_rmsnorm_cpu_1a02f6f::{op_name}"

build/torch210-cxx11-cpu-x86_64-linux/{_rmsnorm_cpu_cec90b8.abi3.so → _rmsnorm_cpu_1a02f6f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:707454a1958381abde25f844df3a461c933df1c0f72a7f85916950df2d9efe77
+oid sha256:c102259696d99bbe9d4c686b4293195548faa4856123a358d44aab3d90148620
 size 2006072

build/torch210-cxx11-xpu20253-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rmsnorm_xpu_cec90b8
-ops = torch.ops._rmsnorm_xpu_cec90b8
+from . import _rmsnorm_xpu_1a02f6f
+ops = torch.ops._rmsnorm_xpu_1a02f6f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rmsnorm_xpu_cec90b8::{op_name}"
+    return f"_rmsnorm_xpu_1a02f6f::{op_name}"

build/torch210-cxx11-xpu20253-x86_64-linux/{_rmsnorm_xpu_cec90b8.abi3.so → _rmsnorm_xpu_1a02f6f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f189d83a1f1245edd0a35839df2686bc806d33a2a1f4c30fbeefe2dd58de55df
+oid sha256:9a87f0910ab215646183ecd9f4b2cbc5be6c72c3eee20d167f42f71c14629e65
 size 104793360

build/torch211-cxx11-cpu-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _rmsnorm_cpu_cec90b8
-ops = torch.ops._rmsnorm_cpu_cec90b8
+from . import _rmsnorm_cpu_1a02f6f
+ops = torch.ops._rmsnorm_cpu_1a02f6f
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_rmsnorm_cpu_cec90b8::{op_name}"
+    return f"_rmsnorm_cpu_1a02f6f::{op_name}"

build/torch211-cxx11-cpu-x86_64-linux/{_rmsnorm_cpu_cec90b8.abi3.so → _rmsnorm_cpu_1a02f6f.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:931baff3800b3e5bdfa223f3bc1149eb10013914437acc1ca88eec255f552c83
+oid sha256:439ac1a1bc4a6095844795cbccd7f2137c101bce3e3415bcebb3fd2b0dfcb97b
 size 2001976

build/torch211-cxx11-xpu20253-x86_64-linux/__init__.py

@@ -0,0 +1,27 @@
+from . import layers
+
+from ._ops import ops
+
+
+def apply_rms_norm(input, weight, eps):
+    # ops.apply_rms_norm returns [output, rstd]
+    return ops.apply_rms_norm(
+            input,
+            weight,
+            eps,
+    )[0]
+
+def apply_rms_norm_backward(grad_output, input, weight, output, rstd, eps, input_requires_grad=True, weight_requires_grad=True):
+    return ops.apply_rms_norm_backward(
+            grad_output,
+            input,
+            weight,
+            output,
+            rstd,
+            eps,
+            input_requires_grad,
+            weight_requires_grad
+    )
+
+__all__ = ["layers", "apply_rms_norm_forward", "apply_rms_norm_backward"]
+

build/torch211-cxx11-xpu20253-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _rmsnorm_xpu_1a02f6f
+ops = torch.ops._rmsnorm_xpu_1a02f6f
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_rmsnorm_xpu_1a02f6f::{op_name}"

build/torch211-cxx11-xpu20253-x86_64-linux/_rmsnorm_xpu_1a02f6f.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:153aa232ee4f342e92075140aa796e86ccd2f55f07d27bcad90890ed2fac57bf
+size 104793120

build/torch211-cxx11-xpu20253-x86_64-linux/layers.py

@@ -0,0 +1,59 @@
+import torch
+from ._ops import ops
+
+class RMSNormFunction(torch.autograd.Function):
+    @staticmethod
+    def forward(ctx, hidden_states, weight, variance_epsilon):
+        ctx.variance_epsilon = variance_epsilon
+        output, rstd = ops.apply_rms_norm(hidden_states, weight, variance_epsilon)
+        ctx.save_for_backward(hidden_states, weight, output, rstd)
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        hidden_states, weight, output, rstd = ctx.saved_tensors
+        grads = ops.apply_rms_norm_backward(
+            grad_output,
+            hidden_states,
+            weight,
+            output,
+            rstd,
+            ctx.variance_epsilon,
+            ctx.needs_input_grad[0],
+            ctx.needs_input_grad[1]
+        )
+        return grads[0], grads[1], None
+
+class RMSNorm(torch.nn.Module):
+    """
+    RMSNorm module that uses the optimized LigerRMSNormFunction.
+    
+    Args:
+        hidden_size (int): The size of the hidden dimension.
+        eps (float, optional): The epsilon value for numerical stability. Defaults to 1e-6.
+        offset (float, optional): Offset value to shift the weight tensor. Defaults to 0.0.
+        casting_mode (str, optional): The casting mode to use. Defaults to "llama".
+        in_place (bool, optional): Whether to modify dY in-place to store dX during backward. Defaults to True.
+    """
+    
+
+    weight: torch.Tensor
+    variance_epsilon: float
+    
+    def forward(self, hidden_states):
+        """
+        Apply RMS normalization to the input tensor.
+        
+        Args:
+            hidden_states (torch.Tensor): Input tensor of shape (B, T, H) or (BxT, H)
+            
+        Returns:
+            torch.Tensor: Normalized tensor of the same shape as input
+        """
+        return RMSNormFunction.apply(
+            hidden_states,
+            self.weight, 
+            self.variance_epsilon,
+        )
+    
+__all__ = ["RMSNorm"]

build/torch211-cxx11-xpu20253-x86_64-linux/metadata.json

@@ -0,0 +1,8 @@
+{
+  "version": 1,
+  "license": "Apache-2.0",
+  "python-depends": [],
+  "backend": {
+    "type": "xpu"
+  }
+}

build/torch211-cxx11-xpu20253-x86_64-linux/rmsnorm/__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")))