Uploaded using `kernel-builder`.
Browse files- build/torch212-metal-aarch64-darwin/{_ggml_norm_metal_1dd4e91.abi3.so → _ggml_norm_metal_f178548.abi3.so} +1 -1
- build/torch212-metal-aarch64-darwin/_ops.py +3 -3
- build/torch212-metal-aarch64-darwin/layers.py +0 -7
- build/torch212-metal-aarch64-darwin/metadata.json +5 -5
- build/torch213-metal-aarch64-darwin/{_ggml_norm_metal_1dd4e91.abi3.so → _ggml_norm_metal_f178548.abi3.so} +1 -1
- build/torch213-metal-aarch64-darwin/_ops.py +3 -3
- build/torch213-metal-aarch64-darwin/layers.py +0 -7
- build/torch213-metal-aarch64-darwin/metadata.json +5 -5
build/torch212-metal-aarch64-darwin/{_ggml_norm_metal_1dd4e91.abi3.so → _ggml_norm_metal_f178548.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 189424
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:df0d604dfe7c1e3f35956c8d04be5132775312cc231f49abffd89c87861051c7
|
| 3 |
size 189424
|
build/torch212-metal-aarch64-darwin/_ops.py
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
import torch
|
| 2 |
-
from . import
|
| 3 |
-
ops = torch.ops.
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
-
return f"
|
|
|
|
| 1 |
import torch
|
| 2 |
+
from . import _ggml_norm_metal_f178548
|
| 3 |
+
ops = torch.ops._ggml_norm_metal_f178548
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_ggml_norm_metal_f178548::{op_name}"
|
build/torch212-metal-aarch64-darwin/layers.py
CHANGED
|
@@ -40,13 +40,6 @@ class RMSNormZeroCentered(nn.Module):
|
|
| 40 |
"""
|
| 41 |
|
| 42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 43 |
-
# `kernels` grafts on the layer name, and a name is inherited by every subclass -- through
|
| 44 |
-
# modular, and through plain Python, where `kernel_layer_name` is just a class attribute. So
|
| 45 |
-
# a layer that normalises over *groups* of the last axis rather than all of it can arrive
|
| 46 |
-
# here (Qwen4Exp builds three of them) and this kernel would quietly answer the wrong
|
| 47 |
-
# question. Its own `_norm` is the one that knows, so grouped instances keep it.
|
| 48 |
-
if getattr(self, "group_size", None) is not None:
|
| 49 |
-
return (self._norm(x.float()) * (1.0 + self.weight.float())).type_as(x)
|
| 50 |
# `float()` and `type_as` are both no-ops when x is already f32, which is what a GGUF model
|
| 51 |
# runs in, so that path is the single dispatch it looks like.
|
| 52 |
return ops.rms_norm(x.float(), _folded_weight(self, lambda w: 1.0 + w), self.eps).type_as(x)
|
|
|
|
| 40 |
"""
|
| 41 |
|
| 42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
# `float()` and `type_as` are both no-ops when x is already f32, which is what a GGUF model
|
| 44 |
# runs in, so that path is the single dispatch it looks like.
|
| 45 |
return ops.rms_norm(x.float(), _folded_weight(self, lambda w: 1.0 + w), self.eps).type_as(x)
|
build/torch212-metal-aarch64-darwin/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "ggml-norm",
|
| 3 |
-
"id": "
|
| 4 |
"version": 1,
|
| 5 |
"license": "MIT",
|
| 6 |
"python-depends": [],
|
|
@@ -11,9 +11,9 @@
|
|
| 11 |
"algorithm": "sha256",
|
| 12 |
"files": {
|
| 13 |
"__init__.py": "Cmj03F1/1Zc1c1XRKOh64yMkFTydJP30N+eU7aOwOlA=",
|
| 14 |
-
"
|
| 15 |
-
"_ops.py": "
|
| 16 |
-
"layers.py": "
|
| 17 |
}
|
| 18 |
},
|
| 19 |
"provenance": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
"dirty": false
|
| 24 |
},
|
| 25 |
"kernel": {
|
| 26 |
-
"sha": "
|
| 27 |
"dirty": false
|
| 28 |
}
|
| 29 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "ggml-norm",
|
| 3 |
+
"id": "_ggml_norm_metal_f178548",
|
| 4 |
"version": 1,
|
| 5 |
"license": "MIT",
|
| 6 |
"python-depends": [],
|
|
|
|
| 11 |
"algorithm": "sha256",
|
| 12 |
"files": {
|
| 13 |
"__init__.py": "Cmj03F1/1Zc1c1XRKOh64yMkFTydJP30N+eU7aOwOlA=",
|
| 14 |
+
"_ggml_norm_metal_f178548.abi3.so": "3w1gTf58Hj81lWyNBL5RMndTEswjH0mr/9ich4YQUcc=",
|
| 15 |
+
"_ops.py": "vfjD1b82HX0GAHAocRISvyJjwpDbXGb4AIeRsPhs7G0=",
|
| 16 |
+
"layers.py": "vzJROAscDehDcNaBtUJE7hV6tN7qit0nVJ5TPuWlWJc="
|
| 17 |
}
|
| 18 |
},
|
| 19 |
"provenance": {
|
|
|
|
| 23 |
"dirty": false
|
| 24 |
},
|
| 25 |
"kernel": {
|
| 26 |
+
"sha": "f178548516bff02585ce5f9afcc367596b8eef45",
|
| 27 |
"dirty": false
|
| 28 |
}
|
| 29 |
}
|
build/torch213-metal-aarch64-darwin/{_ggml_norm_metal_1dd4e91.abi3.so → _ggml_norm_metal_f178548.abi3.so}
RENAMED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
size 189424
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:63db191a9f8f4d348399a54b126cba6ace92a29ceb059cbadd8507b2e2a7613b
|
| 3 |
size 189424
|
build/torch213-metal-aarch64-darwin/_ops.py
CHANGED
|
@@ -1,9 +1,9 @@
|
|
| 1 |
import torch
|
| 2 |
-
from . import
|
| 3 |
-
ops = torch.ops.
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
-
return f"
|
|
|
|
| 1 |
import torch
|
| 2 |
+
from . import _ggml_norm_metal_f178548
|
| 3 |
+
ops = torch.ops._ggml_norm_metal_f178548
|
| 4 |
|
| 5 |
def add_op_namespace_prefix(op_name: str):
|
| 6 |
"""
|
| 7 |
Prefix op by namespace.
|
| 8 |
"""
|
| 9 |
+
return f"_ggml_norm_metal_f178548::{op_name}"
|
build/torch213-metal-aarch64-darwin/layers.py
CHANGED
|
@@ -40,13 +40,6 @@ class RMSNormZeroCentered(nn.Module):
|
|
| 40 |
"""
|
| 41 |
|
| 42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 43 |
-
# `kernels` grafts on the layer name, and a name is inherited by every subclass -- through
|
| 44 |
-
# modular, and through plain Python, where `kernel_layer_name` is just a class attribute. So
|
| 45 |
-
# a layer that normalises over *groups* of the last axis rather than all of it can arrive
|
| 46 |
-
# here (Qwen4Exp builds three of them) and this kernel would quietly answer the wrong
|
| 47 |
-
# question. Its own `_norm` is the one that knows, so grouped instances keep it.
|
| 48 |
-
if getattr(self, "group_size", None) is not None:
|
| 49 |
-
return (self._norm(x.float()) * (1.0 + self.weight.float())).type_as(x)
|
| 50 |
# `float()` and `type_as` are both no-ops when x is already f32, which is what a GGUF model
|
| 51 |
# runs in, so that path is the single dispatch it looks like.
|
| 52 |
return ops.rms_norm(x.float(), _folded_weight(self, lambda w: 1.0 + w), self.eps).type_as(x)
|
|
|
|
| 40 |
"""
|
| 41 |
|
| 42 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
# `float()` and `type_as` are both no-ops when x is already f32, which is what a GGUF model
|
| 44 |
# runs in, so that path is the single dispatch it looks like.
|
| 45 |
return ops.rms_norm(x.float(), _folded_weight(self, lambda w: 1.0 + w), self.eps).type_as(x)
|
build/torch213-metal-aarch64-darwin/metadata.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
| 1 |
{
|
| 2 |
"name": "ggml-norm",
|
| 3 |
-
"id": "
|
| 4 |
"version": 1,
|
| 5 |
"license": "MIT",
|
| 6 |
"python-depends": [],
|
|
@@ -11,9 +11,9 @@
|
|
| 11 |
"algorithm": "sha256",
|
| 12 |
"files": {
|
| 13 |
"__init__.py": "Cmj03F1/1Zc1c1XRKOh64yMkFTydJP30N+eU7aOwOlA=",
|
| 14 |
-
"
|
| 15 |
-
"_ops.py": "
|
| 16 |
-
"layers.py": "
|
| 17 |
}
|
| 18 |
},
|
| 19 |
"provenance": {
|
|
@@ -23,7 +23,7 @@
|
|
| 23 |
"dirty": false
|
| 24 |
},
|
| 25 |
"kernel": {
|
| 26 |
-
"sha": "
|
| 27 |
"dirty": false
|
| 28 |
}
|
| 29 |
}
|
|
|
|
| 1 |
{
|
| 2 |
"name": "ggml-norm",
|
| 3 |
+
"id": "_ggml_norm_metal_f178548",
|
| 4 |
"version": 1,
|
| 5 |
"license": "MIT",
|
| 6 |
"python-depends": [],
|
|
|
|
| 11 |
"algorithm": "sha256",
|
| 12 |
"files": {
|
| 13 |
"__init__.py": "Cmj03F1/1Zc1c1XRKOh64yMkFTydJP30N+eU7aOwOlA=",
|
| 14 |
+
"_ggml_norm_metal_f178548.abi3.so": "Y9sZGp+PTTSDmaVLEmy6as6SopzrBZy63YUHsuKnYTs=",
|
| 15 |
+
"_ops.py": "vfjD1b82HX0GAHAocRISvyJjwpDbXGb4AIeRsPhs7G0=",
|
| 16 |
+
"layers.py": "vzJROAscDehDcNaBtUJE7hV6tN7qit0nVJ5TPuWlWJc="
|
| 17 |
}
|
| 18 |
},
|
| 19 |
"provenance": {
|
|
|
|
| 23 |
"dirty": false
|
| 24 |
},
|
| 25 |
"kernel": {
|
| 26 |
+
"sha": "f178548516bff02585ce5f9afcc367596b8eef45",
|
| 27 |
"dirty": false
|
| 28 |
}
|
| 29 |
}
|