Build uploaded using `kernels`.

.gitattributes

@@ -81,3 +81,4 @@ build/torch210-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_9a0d032.abi3.so filt
 build/torch29-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_9a0d032.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-cu128-aarch64-linux/_tinygrad_rms_cuda_9a0d032.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_9a0d032.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cu128-x86_64-windows/_tinygrad_rms_cuda_6e9aef6.pyd filter=lfs diff=lfs merge=lfs -text

build/torch210-cu128-x86_64-windows/__init__.py

@@ -0,0 +1,63 @@
+from typing import Optional, Tuple
+
+import torch
+
+from ._ops import ops
+
+
+def tinygrad_rms_norm(
+    x: torch.Tensor,
+    epsilon: float = 1e-6,
+    out: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """
+    Compute RMSNorm using tinygrad-style CUDA kernels.
+
+    RMSNorm(x) = x * (1 / sqrt(mean(x^2) + epsilon))
+
+    This implementation uses a two-kernel approach:
+    1. Compute 1/sqrt(mean(x^2) + epsilon) for each row
+    2. Multiply input by the computed factor
+
+    Args:
+        x: Input tensor of shape (..., hidden_size)
+        epsilon: Small constant for numerical stability
+        out: Optional pre-allocated output tensor
+
+    Returns:
+        Tuple of (output tensor, rms_inv tensor)
+    """
+    if out is None:
+        out = torch.empty_like(x)
+
+    hidden_size = x.size(-1)
+    num_rows = x.numel() // hidden_size
+    rms_inv = torch.empty(num_rows, dtype=x.dtype, device=x.device)
+
+    ops.tinygrad_rms_norm(out, rms_inv, x, epsilon)
+    return out, rms_inv
+
+
+def tinygrad_rms_norm_simple(
+    x: torch.Tensor,
+    epsilon: float = 1e-6,
+    out: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """
+    Compute RMSNorm using tinygrad-style CUDA kernels.
+
+    This is a simpler interface that only returns the normalized output.
+
+    Args:
+        x: Input tensor of shape (..., hidden_size)
+        epsilon: Small constant for numerical stability
+        out: Optional pre-allocated output tensor
+
+    Returns:
+        Normalized output tensor
+    """
+    if out is None:
+        out = torch.empty_like(x)
+
+    ops.tinygrad_rms_norm_inplace(out, x, epsilon)
+    return out

build/torch210-cu128-x86_64-windows/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tinygrad_rms_cuda_6e9aef6
+ops = torch.ops._tinygrad_rms_cuda_6e9aef6
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_6e9aef6::{op_name}"

build/torch210-cu128-x86_64-windows/_tinygrad_rms_cuda_6e9aef6.pyd

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebb4b0bd73ae5c954fc25a1fce249af4ac184a1bf3fb537e77ca36d21b3044ef
+size 320000

build/torch210-cu128-x86_64-windows/metadata.json

@@ -0,0 +1,21 @@
+{
+  "version": 1,
+  "license": "MIT",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "10.1",
+      "12.0+PTX",
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch210-cu128-x86_64-windows/tinygrad_rms/__init__.py

@@ -0,0 +1,26 @@
+import ctypes
+import sys
+
+import importlib
+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")))