Build uploaded using `kernels`.

.gitattributes

@@ -95,3 +95,10 @@ build/torch211-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filt
 build/torch211-cxx11-cu128-aarch64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch211-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-cxx11-cu129-aarch64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu126-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu128-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu126-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu128-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch29-cxx11-cu129-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch210-cxx11-cu126-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_9a0d032
-ops = torch.ops._tinygrad_rms_cuda_9a0d032
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_9a0d032::{op_name}"
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch210-cxx11-cu126-x86_64-linux/{_tinygrad_rms_cuda_9a0d032.abi3.so → _tinygrad_rms_cuda_6eb00dc.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:cc0991e99b3bc0752fb673d4f8a402d567115cf7b81298cf34d554c8b7613922
+oid sha256:b6dc5ab336157c3e6c577ecc8dca3790d4ac9273ae4b38ebd45e7dee0c6c20a5
 size 2123424

build/torch210-cxx11-cu126-x86_64-linux/tinygrad_rms/__init__.py

@@ -1,10 +1,10 @@
 import ctypes
+import importlib.util
 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

build/torch210-cxx11-cu128-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_9a0d032
-ops = torch.ops._tinygrad_rms_cuda_9a0d032
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_9a0d032::{op_name}"
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch210-cxx11-cu128-x86_64-linux/{_tinygrad_rms_cuda_9a0d032.abi3.so → _tinygrad_rms_cuda_6eb00dc.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8180078f139f93c7de5a02d7f26b1d545f32d1def25dfbba8e77d38c37a727cf
+oid sha256:645b392a5deadd70a5405bbceeb2161c9b852cb0a303b826e8ee633ee4443883
 size 2244040

build/torch210-cxx11-cu128-x86_64-linux/tinygrad_rms/__init__.py

@@ -1,10 +1,10 @@
 import ctypes
+import importlib.util
 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

build/torch210-cxx11-cu130-x86_64-linux/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_9a0d032
-ops = torch.ops._tinygrad_rms_cuda_9a0d032
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_9a0d032::{op_name}"
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch210-cxx11-cu130-x86_64-linux/{_tinygrad_rms_cuda_9a0d032.abi3.so → _tinygrad_rms_cuda_6eb00dc.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:613b67088db70e70b9b629dd61276368b3a7d8b871a6870116fc98b897ac3425
+oid sha256:3bd783d80e077f2983793e5fdc8a0f5478836851435a8b77f00b186fea64c8f4
 size 2245848

build/torch210-cxx11-cu130-x86_64-linux/tinygrad_rms/__init__.py

@@ -1,10 +1,10 @@
 import ctypes
+import importlib.util
 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

build/torch211-cxx11-cu126-x86_64-linux/__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/torch211-cxx11-cu126-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch211-cxx11-cu126-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6394aee38bd05dd68b98ab6e87ff95748c1662dad1b85178c475cd76de7cbf4
+size 2112256

build/torch211-cxx11-cu126-x86_64-linux/metadata.json

@@ -0,0 +1,18 @@
+{
+  "version": 1,
+  "license": "MIT",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "7.0",
+      "7.2",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0+PTX"
+    ]
+  }
+}

build/torch211-cxx11-cu126-x86_64-linux/tinygrad_rms/__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-x86_64-linux/__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/torch211-cxx11-cu128-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch211-cxx11-cu128-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2528a713f5fddd2d8ec0ab7dad1b08e59a693bcc5ea3d4c060a6c25760bc4956
+size 2232880

build/torch211-cxx11-cu128-x86_64-linux/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/torch211-cxx11-cu128-x86_64-linux/tinygrad_rms/__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-x86_64-linux/__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/torch211-cxx11-cu130-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch211-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1987caa31a62d13612871c2554f23facb2727ff58e5d274fc98ec39a105d24aa
+size 2234688

build/torch211-cxx11-cu130-x86_64-linux/metadata.json

@@ -0,0 +1,19 @@
+{
+  "version": 1,
+  "license": "MIT",
+  "python-depends": [],
+  "backend": {
+    "type": "cuda",
+    "archs": [
+      "10.0",
+      "11.0",
+      "12.0+PTX",
+      "7.5",
+      "8.0",
+      "8.6",
+      "8.7",
+      "8.9",
+      "9.0"
+    ]
+  }
+}

build/torch211-cxx11-cu130-x86_64-linux/tinygrad_rms/__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-x86_64-linux/__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/torch29-cxx11-cu129-x86_64-linux/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _tinygrad_rms_cuda_6eb00dc
+ops = torch.ops._tinygrad_rms_cuda_6eb00dc
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_6eb00dc::{op_name}"

build/torch29-cxx11-cu129-x86_64-linux/_tinygrad_rms_cuda_6eb00dc.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d7bdd7d0c61c7df6af0a414d49acfc4d61c5c23fc4bc8e9b0d9a5e633ac6a88b
+size 2262864

build/torch29-cxx11-cu129-x86_64-linux/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/torch29-cxx11-cu129-x86_64-linux/tinygrad_rms/__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")))