Build uploaded using `kernels`.

.gitattributes

@@ -79,3 +79,4 @@ build/torch210-cxx11-cpu-x86_64-linux/_rmsnorm_ce2b5cc.abi3.so filter=lfs diff=l
 build/torch210-cxx11-xpu20253-x86_64-linux/_rmsnorm_ce2b5cc.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-cpu-x86_64-linux/_rmsnorm_ce2b5cc.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-xpu20252-x86_64-linux/_rmsnorm_ce2b5cc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-xpu20253-x86_64-windows/rmsnorm/_rmsnorm_4cd2f5b.pyd filter=lfs diff=lfs merge=lfs -text

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

@@ -0,0 +1,4 @@
+{
+  "version": 1,
+  "python-depends": []
+}

build/torch210-xpu20253-x86_64-windows/rmsnorm/__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/torch210-xpu20253-x86_64-windows/rmsnorm/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _rmsnorm_4cd2f5b
+ops = torch.ops._rmsnorm_4cd2f5b
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_rmsnorm_4cd2f5b::{op_name}"

build/torch210-xpu20253-x86_64-windows/rmsnorm/_rmsnorm_4cd2f5b.pyd

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13c739cbd5b54166522364d80f8565fddb2ec3bbb77ac3348b35077f70802bfd
+size 2363904

build/torch210-xpu20253-x86_64-windows/rmsnorm/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"]