OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

[Refactor] Explicilty spell out the namespace for device() function (#153248)

Browse Source 
Summary: To prepare for the coming up header-only file change. The same files have been using a mixed style of using at::device() and device(). Given these .cpp files are not in the at namespace, it makes sense to spell them out explicitly.

Differential Revision: [D74577412](https://our.internmc.facebook.com/intern/diff/D74577412)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153248
Approved by: https://github.com/cyyever, https://github.com/albanD, https://github.com/janeyx99
This commit is contained in:
Bin Bao 2025-05-09 08:13:56 -07:00 committed by PyTorch MergeBot
parent 0ef5ba43a6
commit e8f7a97e2e
8 changed files with 24 additions and 24 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_deserialize.cpp
View File

@ -106,7 +106,7 @@ c10::intrusive_ptr<LinearPackedParamsBase> PackedLinearWeight::deserialize(
std::get<weight_scales_index>(serialized),
weight_zero_points,
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQInt8));
at::device(c10::kCPU).dtype(c10::kQInt8));
}
const at::Tensor loaded_weight_values =

4
aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear_unpack.cpp
View File

@ -46,7 +46,7 @@ LinearPackedSerializationType PackedLinearWeight::unpack() {
scales,
zero_points,
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQInt8));
at::device(c10::kCPU).dtype(c10::kQInt8));
}
int8_t* weight_ptr_int8 =
@ -100,7 +100,7 @@ LinearPackedSerializationType PackedLinearWeightQnnp::unpack() {
scales,
zero_points,
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQInt8));
at::device(c10::kCPU).dtype(c10::kQInt8));
}
int8_t* weight_ptr_int8 =

10
aten/src/ATen/native/quantized/cpu/LinearUnpackImpl.cpp
View File

@ -32,16 +32,16 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedLinearWeight::unpack() {
{N, K}, at::device(c10::kCPU).dtype(c10::kQInt8), w_scale[0], w_zp[0]);
} else if (q_scheme == c10::kPerChannelAffine) {
auto scales = at::from_blob(
w_scale.data(), w_scale.size(), device(c10::kCPU).dtype(c10::kFloat));
w_scale.data(), w_scale.size(), at::device(c10::kCPU).dtype(c10::kFloat));
auto zero_points = at::from_blob(
w_zp.data(), w_zp.size(), device(c10::kCPU).dtype(c10::kInt));
w_zp.data(), w_zp.size(), at::device(c10::kCPU).dtype(c10::kInt));
weight_origin = at::_empty_per_channel_affine_quantized(
{N, K},
scales.toType(c10::kDouble),
zero_points.toType(c10::kLong),
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQInt8));
at::device(c10::kCPU).dtype(c10::kQInt8));
}
int8_t* weight_ptr_int8 =
@ -81,7 +81,7 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedLinearWeightsQnnp::
auto scales = at::from_blob(
weight_scales_data,
w_scales.sizes()[0] - kPaddingChannels,
device(c10::kCPU).dtype(c10::kFloat));
at::device(c10::kCPU).dtype(c10::kFloat));
at::Tensor zero_points = at::empty(
w_zero_points.size() - kPaddingChannels, at::device(c10::kCPU).dtype(c10::kLong));
@ -93,7 +93,7 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedLinearWeightsQnnp::
scales,
zero_points.toType(c10::kLong),
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQInt8))
at::device(c10::kCPU).dtype(c10::kQInt8))
.contiguous();
} else {
TORCH_INTERNAL_ASSERT(false, "Unsupported quantization scheme.");

8
aten/src/ATen/native/quantized/cpu/qconv.cpp
View File

@ -448,7 +448,7 @@ at::Tensor PackedConvWeight<kSpatialDim>::apply_impl(
at::Tensor output = kSpatialDim == 2
? at::_empty_affine_quantized(
output_shape,
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(c10::kQUInt8)
.memory_format(c10::MemoryFormat::ChannelsLast),
output_scale,
@ -460,7 +460,7 @@ at::Tensor PackedConvWeight<kSpatialDim>::apply_impl(
output_shape[2],
output_shape[3],
output_shape[4],
device(c10::kCPU).dtype(c10::kQUInt8),
at::device(c10::kCPU).dtype(c10::kQUInt8),
output_scale,
output_zero_point);
at::Tensor buffer =
@ -1225,7 +1225,7 @@ at::Tensor PackedConvWeightsOnednn<kSpatialDim>::apply_impl(
ideep::dims dst_dims = ideep::dims({output_sizes.cbegin(), output_sizes.cend()});
at::Tensor output = at::_empty_affine_quantized(
dst_dims,
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(c10::kQUInt8)
.memory_format(kSpatialDim == 2 ?
c10::MemoryFormat::ChannelsLast :
@ -1593,7 +1593,7 @@ static at::Tensor _quantized_convolution_onednn(
accum.value() :
at::empty(
dst_dims,
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(fp32_output ? c10::kFloat : (bfloat16_output ? c10::kBFloat16 : c10::kByte))
.memory_format(kSpatialDim == 2 ?
c10::MemoryFormat::ChannelsLast :

12
aten/src/ATen/native/quantized/cpu/qconv_unpack_impl.cpp
View File

@ -37,7 +37,7 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedConvWeight<
unpacked_weights = kSpatialDim == 2
? at::_empty_affine_quantized(
{output_channels, C_per_G, kernel_h, kernel_w},
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(c10::kQInt8)
.memory_format(c10::MemoryFormat::ChannelsLast),
w_scale[0],
@ -50,7 +50,7 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedConvWeight<
kernel_d,
kernel_h,
kernel_w,
device(c10::kCPU).dtype(c10::kQInt8),
at::device(c10::kCPU).dtype(c10::kQInt8),
w_scale[0],
w_zp[0]);
} else if (q_scheme == c10::kPerChannelAffine) {
@ -58,16 +58,16 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedConvWeight<
!transpose(),
"Per Channel Quantization is currently disabled for transposed conv");
auto scales = at::from_blob(
w_scale.data(), w_scale.size(), device(c10::kCPU).dtype(c10::kFloat));
w_scale.data(), w_scale.size(), at::device(c10::kCPU).dtype(c10::kFloat));
auto zero_points = at::from_blob(
w_zp.data(), w_zp.size(), device(c10::kCPU).dtype(c10::kInt));
w_zp.data(), w_zp.size(), at::device(c10::kCPU).dtype(c10::kInt));
unpacked_weights = kSpatialDim == 2
? at::_empty_per_channel_affine_quantized(
{output_channels, C_per_G, kernel_h, kernel_w},
scales.toType(c10::kDouble),
zero_points.toType(c10::kLong),
0, /* The output channel axis is 0 */
device(c10::kCPU).dtype(c10::kQInt8),
at::device(c10::kCPU).dtype(c10::kQInt8),
c10::MemoryFormat::ChannelsLast)
: at::native::fbgemm_utils::
MakeEmptyPerChannelAffineQuantizedChannelsLast3dTensor(
@ -76,7 +76,7 @@ std::tuple<at::Tensor, std::optional<at::Tensor>> PackedConvWeight<
kernel_d,
kernel_h,
kernel_w,
device(c10::kCPU).dtype(c10::kQInt8),
at::device(c10::kCPU).dtype(c10::kQInt8),
scales.toType(c10::kDouble),
zero_points.toType(c10::kLong));
} else {

8
aten/src/ATen/native/quantized/cpu/qembeddingbag_unpack.cpp
View File

@ -44,9 +44,9 @@ at::Tensor PackedEmbeddingBagWeight::unpack() {
num_elem_per_byte};
auto scales = at::from_blob(
w_scale.data(), w_scale.size(), device(c10::kCPU).dtype(c10::kFloat));
w_scale.data(), w_scale.size(), at::device(c10::kCPU).dtype(c10::kFloat));
auto zero_points = at::from_blob(
w_zp.data(), w_zp.size(), device(c10::kCPU).dtype(c10::kFloat));
w_zp.data(), w_zp.size(), at::device(c10::kCPU).dtype(c10::kFloat));
auto output_columns = output_shape[1];
uint8_t* output_data = nullptr;
@ -58,7 +58,7 @@ at::Tensor PackedEmbeddingBagWeight::unpack() {
scales.toType(c10::kFloat),
zero_points.toType(c10::kFloat),
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQUInt8));
at::device(c10::kCPU).dtype(c10::kQUInt8));
output_data = static_cast<uint8_t*>(weight_origin.data_ptr());
} else {
// We create empty qtensor with the full output shape, and dtype set to
@ -69,7 +69,7 @@ at::Tensor PackedEmbeddingBagWeight::unpack() {
scales.toType(c10::kFloat),
zero_points.toType(c10::kFloat),
0, // The output channel axis is 0
device(c10::kCPU).dtype(c10::kQUInt4x2));
at::device(c10::kCPU).dtype(c10::kQUInt4x2));
output_data = static_cast<uint8_t*>(weight_origin.data_ptr());
}

2
aten/src/ATen/native/quantized/cpu/qlinear.cpp
View File

@ -1015,7 +1015,7 @@ static at::Tensor linear_int8_with_onednn_weight(
other.value() :
at::empty(
dst_dims,
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(fp32_output ? c10::kFloat : (bf16_output ? c10::kBFloat16 : c10::kByte))
);
if (output.numel() == 0) {

2
aten/src/ATen/native/quantized/cpu/qlinear_dynamic.cpp
View File

@ -652,7 +652,7 @@ static at::Tensor linear_dynamic_fp16_with_onednn_weight(
std::vector<int64_t> dst_dims = {M, N};
at::Tensor output = at::empty(
dst_dims,
device(c10::kCPU)
at::device(c10::kCPU)
.dtype(c10::kFloat)
);
if (output.numel() == 0) {