Kernels:

kernels-community
/

tinygrad-rms

Trusted publisher

Kernel card Files Files and versions

xet

Community

kernels-bot commited on 27 days ago

Commit

728d66d

verified ·

1 Parent(s): 8a449d4

Uploaded using `kernel-builder`.

Browse files

Files changed (34) hide show

.gitattributes +9 -0
build/torch210-cxx11-cu126-aarch64-linux/_ops.py +3 -3
build/torch210-cxx11-cu126-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch210-cxx11-cu126-aarch64-linux/metadata.json +1 -1
build/torch210-cxx11-cu128-aarch64-linux/_ops.py +3 -3
build/torch210-cxx11-cu128-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch210-cxx11-cu128-aarch64-linux/metadata.json +1 -1
build/torch210-cxx11-cu130-aarch64-linux/_ops.py +3 -3
build/torch210-cxx11-cu130-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch210-cxx11-cu130-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu126-aarch64-linux/_ops.py +3 -3
build/torch211-cxx11-cu126-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch211-cxx11-cu126-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu128-aarch64-linux/_ops.py +3 -3
build/torch211-cxx11-cu128-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch211-cxx11-cu128-aarch64-linux/metadata.json +1 -1
build/torch211-cxx11-cu130-aarch64-linux/_ops.py +3 -3
build/torch211-cxx11-cu130-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} +1 -1
build/torch211-cxx11-cu130-aarch64-linux/metadata.json +1 -1
build/torch212-cxx11-cu126-aarch64-linux/__init__.py +63 -0
build/torch212-cxx11-cu126-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so +3 -0
build/torch212-cxx11-cu126-aarch64-linux/metadata.json +20 -0
build/torch212-cxx11-cu126-aarch64-linux/tinygrad_rms/__init__.py +26 -0
build/torch212-cxx11-cu130-aarch64-linux/__init__.py +63 -0
build/torch212-cxx11-cu130-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so +3 -0
build/torch212-cxx11-cu130-aarch64-linux/metadata.json +21 -0
build/torch212-cxx11-cu130-aarch64-linux/tinygrad_rms/__init__.py +26 -0
build/torch212-cxx11-cu132-aarch64-linux/__init__.py +63 -0
build/torch212-cxx11-cu132-aarch64-linux/_ops.py +9 -0
build/torch212-cxx11-cu132-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so +3 -0
build/torch212-cxx11-cu132-aarch64-linux/metadata.json +21 -0
build/torch212-cxx11-cu132-aarch64-linux/tinygrad_rms/__init__.py +26 -0

.gitattributes CHANGED Viewed

@@ -153,3 +153,12 @@ build/torch211-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filte
 build/torch212-cxx11-cu126-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu132-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text

 build/torch212-cxx11-cu126-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu130-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch212-cxx11-cu132-x86_64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu128-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch210-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu128-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch212-cxx11-cu132-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so filter=lfs diff=lfs merge=lfs -text

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch210-cxx11-cu126-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42834aa3e5e5e6d535ef20c2b0f62d98bae5445b8a2e06f2ed64ca6c7e829cf1
 size 2241168

 version https://git-lfs.github.com/spec/v1
+oid sha256:dd58eeef09d84e1f89625e9672a1e04704427c491ba0cadee0f1cf3bc7997563
 size 2241168

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch210-cxx11-cu128-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9df3b35b26c7971d2ebc0b6f98bb68e48a3839da9d9ef95e92458f8959313984
 size 2372360

 version https://git-lfs.github.com/spec/v1
+oid sha256:ba98bcb8da1711936256ddfd6faa502cc57a2850b9243fc7b3f2f79f2fc52560
 size 2372360

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch210-cxx11-cu130-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:1034f93746c54cd2ee8c8d1edda36f2347e7005cd47bef7d592e18f2a421a761
 size 2309232

 version https://git-lfs.github.com/spec/v1
+oid sha256:f777c857654ce0461994a1e278a26a8689b99312aeb992cfc3cbeb89b46a80ec
 size 2309232

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch211-cxx11-cu126-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6f68845038c29b83914010eeaab4459c6a29abc01dee8dfa0c5be10a53648ac2
 size 2237384

 version https://git-lfs.github.com/spec/v1
+oid sha256:b48d91db6f10cdabda9d5b13d8f6cf671a40afa0fc199fc5d9bd590483f4e5aa
 size 2237384

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch211-cxx11-cu128-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e70de3256587d30f0e85f254d45397eea6e995576aa03811360da270d10ae938
 size 2368568

 version https://git-lfs.github.com/spec/v1
+oid sha256:daeb102f08b78565d2dba7a48dcb7f6f3954cbcd819dd27e9d35eb22794494ee
 size 2368568

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

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

@@ -1,9 +1,9 @@
 import torch
-from . import _tinygrad_rms_cuda_86f75d9
-ops = torch.ops._tinygrad_rms_cuda_86f75d9
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_tinygrad_rms_cuda_86f75d9::{op_name}"

 import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch211-cxx11-cu130-aarch64-linux/{_tinygrad_rms_cuda_86f75d9.abi3.so → _tinygrad_rms_cuda_1917536.abi3.so} RENAMED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bff6659403c5a3c2ddcea3c59c5fac0e699434047c826f51de4fee845928e0f7
 size 2305440

 version https://git-lfs.github.com/spec/v1
+oid sha256:ffe2d6b8ceba031df62a81a3c49387594caf1d1e22b8fa50760fa205dc583957
 size 2305440

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

@@ -1,6 +1,6 @@
 {
   "name": "tinygrad-rms",
-  "id": "_tinygrad_rms_cuda_86f75d9",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

 {
   "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
   "version": 1,
   "license": "MIT",
   "python-depends": [],

build/torch212-cxx11-cu126-aarch64-linux/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu126-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch212-cxx11-cu126-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f59a3df54046a3fad9b1667e41944f989aaaa39924368c8787d3e3e086643e
+size 2240464

build/torch212-cxx11-cu126-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,20 @@

+{
+  "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
+  "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/torch212-cxx11-cu126-aarch64-linux/tinygrad_rms/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu130-aarch64-linux/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu130-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch212-cxx11-cu130-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:472028577a8a503e030bda19ec467097167c8460cd6faf9e90d813da66ddee62
+size 2308496

build/torch212-cxx11-cu130-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,21 @@

+{
+  "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
+  "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/torch212-cxx11-cu130-aarch64-linux/tinygrad_rms/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu132-aarch64-linux/__init__.py ADDED Viewed

	@@ -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/torch212-cxx11-cu132-aarch64-linux/_ops.py ADDED Viewed

	@@ -0,0 +1,9 @@

+import torch
+from . import _tinygrad_rms_cuda_1917536
+ops = torch.ops._tinygrad_rms_cuda_1917536
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_tinygrad_rms_cuda_1917536::{op_name}"

build/torch212-cxx11-cu132-aarch64-linux/_tinygrad_rms_cuda_1917536.abi3.so ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:c05d85f7ed88f1f1295872091abba69837f302c288a82fc1835af4f01ae009ea
+size 2376632

build/torch212-cxx11-cu132-aarch64-linux/metadata.json ADDED Viewed

	@@ -0,0 +1,21 @@

+{
+  "name": "tinygrad-rms",
+  "id": "_tinygrad_rms_cuda_1917536",
+  "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/torch212-cxx11-cu132-aarch64-linux/tinygrad_rms/__init__.py ADDED Viewed

	@@ -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")))