natas_hw
/
openvino
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Allow signed int8 for WOQ (#24620) 

### Details:
 - Add i8 data type to list of supported dequantization types for woq

### Tickets:
 - *CVS-131044*
This commit is contained in:
Vladimir Paramuzov 2024-05-23 11:07:09 +04:00 committed by GitHub
parent 9ec1c4a1fe
commit 23d487bdfb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 6 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
src/plugins/intel_gpu/src/kernel_selector/kernels/fully_connected/fully_connected_kernel_bf_tiled.cpp
View File

@ -41,6 +41,8 @@ ParamsKey FullyConnected_bf_tiled::GetSupportedKey() const {
k.EnableOutputDataType(Datatype::UINT8);
k.EnableInputWeightsType(WeightsType::UINT4);
k.EnableInputWeightsType(WeightsType::INT4);
k.EnableInputWeightsType(WeightsType::UINT8);
k.EnableInputWeightsType(WeightsType::INT8);
k.EnableInputWeightsType(WeightsType::F16);
k.EnableInputWeightsType(WeightsType::F32);
k.EnableInputLayout(DataLayout::bf);

7
src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp
View File

@ -26,6 +26,7 @@
#include "low_precision/recurrent_cell.hpp"
#include "low_precision/strided_slice.hpp"
#include "openvino/core/deprecated.hpp"
#include "openvino/core/type/element_type.hpp"
#include "openvino/core/validation_util.hpp"
#include "openvino/op/constant.hpp"
#include "openvino/op/convolution.hpp"
@ -193,6 +194,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
using const_node_ptr = const std::shared_ptr<const ov::Node>;
const auto& defaultPrecisions = ov::pass::low_precision::precision_set::get_int8_support();
const ov::element::TypeVector supported_woq_types = {ov::element::u8, ov::element::i8, ov::element::u4, ov::element::i4};
bool enableInt8;
bool unroll_loop = config.get_property(ov::intel_gpu::enable_loop_unrolling);
{
@ -287,8 +289,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
// Disable subtract folding only for the dGPUs to meet the requirements of oneDNN:
// it expects to have the same data type for weights and zero points (apply it only for u8 data type, since other compression
// types are not supported by oneDNN)
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(ov::element::TypeVector{ov::element::u8, ov::element::u4, ov::element::i4},
!device_info.supports_immad);
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(supported_woq_types, !device_info.supports_immad);
// Need to check if transformations work correctly for mixed models with both compression and quantization at the same time.
if (!is_model_quantized)
@ -778,7 +779,7 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
if (device_info.supports_immad) {
// For OneDNN, ZP should not be folded for FC. But still, ZP should be folded for Gather.
// Therefore, run MarkDequantizationSubgraph again to fold ZP constant.
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(ov::element::TypeVector{ov::element::u8, ov::element::u4, ov::element::i4}, true);
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(supported_woq_types, true);
manager.register_pass<ov::pass::ConstantFolding>();
}
manager.register_pass<ov::pass::ConvertGatherToGatherCompressed>();