IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[GPU] Revert transformation changes (#21749)

Browse Source 
* Revert "[GPU] Apply is_non_decompression_multiply() callback only for compressed models (#21719)"

This reverts commit 032ac898e2.

* [GPU] Apply MarkDequantizationSubgraph callback function only to compressed models
This commit is contained in:
Sergey Shlyapnikov 2023-12-20 07:46:09 +04:00 committed by GitHub
parent 8d704f6400
commit 51eab3c931
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 23 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

51
src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp
View File

@ -191,8 +191,9 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
bool enableInt8;
bool unroll_loop = config.get_property(ov::intel_gpu::enable_loop_unrolling);
{
ov::pass::Manager initial_transformations_manager;
initial_transformations_manager.set_per_pass_validation(false);
ov::pass::Manager manager;
auto pass_config = manager.get_pass_config();
manager.set_per_pass_validation(false);
// Temporary solution, global rt info cleanup is needed
for (auto& node : func->get_ops()) {
@ -200,9 +201,15 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
ov::disable_keep_const_precision(node);
}
enableInt8 = config.get_property(ov::intel_gpu::enable_lp_transformations) && ov::pass::low_precision::LowPrecision::isFunctionQuantized(func);
initial_transformations_manager.register_pass<ov::pass::InitNodeInfo>();
initial_transformations_manager.register_pass<EinsumDecomposition>();
auto is_model_quantized = ov::pass::low_precision::LowPrecision::isFunctionQuantized(func);
enableInt8 = config.get_property(ov::intel_gpu::enable_lp_transformations) && is_model_quantized;
if (enableInt8) {
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(
std::vector<ov::element::Type>{ ov::element::i8, ov::element::u8, ov::element::i4, ov::element::u4 });
}
manager.register_pass<ov::pass::InitNodeInfo>();
manager.register_pass<EinsumDecomposition>();
precisions_map fp_convert_precision_map = {
{ov::element::f64, ov::element::f32}
@ -251,19 +258,19 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
}
type_to_fuse_map empty_fuse_map = {};
initial_transformations_manager.register_pass<ov::pass::Validate>();
manager.register_pass<ov::pass::Validate>();
// fuse softmax, MVN patterns, so that they will not be marked as precision sensitive in ConvertPrecision
initial_transformations_manager.register_pass<ov::pass::SoftmaxFusion>();
initial_transformations_manager.register_pass<ov::pass::MVNFusion>();
manager.register_pass<ov::pass::SoftmaxFusion>();
manager.register_pass<ov::pass::MVNFusion>();
// decompose MVNs that sre not supported in GPU, so that they will be marked as precision sensitive in ConvertPrecision
initial_transformations_manager.register_pass<ov::pass::MVN6Decomposition>();
manager.register_pass<ov::pass::MVN6Decomposition>();
// Run these broadcast optimizations earlier to ensure that those are executed before NopElimination/ConstantFolding
initial_transformations_manager.register_pass<ov::pass::BroadcastElementwiseFusion>();
initial_transformations_manager.register_pass<ov::pass::BroadcastTransition>();
manager.register_pass<ov::pass::BroadcastElementwiseFusion>();
manager.register_pass<ov::pass::BroadcastTransition>();
initial_transformations_manager.register_pass<ov::pass::KeepConstantsPrecisionAndAddConverts>();
initial_transformations_manager.get_pass_config()->set_callback<ov::pass::KeepConstantsPrecisionAndAddConverts>(
manager.register_pass<ov::pass::KeepConstantsPrecisionAndAddConverts>();
pass_config->set_callback<ov::pass::KeepConstantsPrecisionAndAddConverts>(
[](const_node_ptr& node) -> bool {
auto next_node = node->get_output_target_inputs(0).begin()->get_node();
if (is_type<ov::op::v0::Convert>(next_node)) {
@ -272,22 +279,10 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
return !is_type<ov::op::v0::MatMul>(next_node);
});
initial_transformations_manager.register_pass<ov::pass::MarkDequantizationSubgraph>(ov::element::TypeVector{ov::element::u8,
ov::element::u4,
ov::element::i4}, true);
// Ignore nodes that are not related to FullyConnected and allow ConstantFolding to be applied to them
initial_transformations_manager.get_pass_config()->set_callback<ov::pass::MarkDequantizationSubgraph>(is_non_supported_decompression_op);
initial_transformations_manager.run_passes(func);
ov::pass::Manager manager;
auto pass_config = manager.get_pass_config();
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(ov::element::TypeVector{ov::element::u8, ov::element::u4, ov::element::i4}, true);
// Need to check if transfomrations work correctly for mixed models with both compression and quantization at the same time.
if (enableInt8) {
manager.register_pass<ov::pass::MarkDequantizationSubgraph>(
std::vector<ov::element::Type>{ ov::element::i8, ov::element::u8, ov::element::i4, ov::element::u4 });
}
if (!is_model_quantized)
pass_config->set_callback<ov::pass::MarkDequantizationSubgraph>(is_non_supported_decompression_op);
manager.register_pass<ov::intel_gpu::MoveConvertAfterGather>();