Skip to content

[GPU] Skip activation dynamic quantization for 4-bit FC on discrete GPU #36907

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
allnes wants to merge 2 commits into from
+16 −0
Closed

[GPU] Skip activation dynamic quantization for 4-bit FC on discrete GPU #36907

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b98cae3
[GPU] Skip activation dynamic quantization for 4-bit FC on discrete GPU
allnes Jul 15, 2026
4175d31
Merge branch 'master' into gpu-dq-skip-dgpu-int4
allnes Jul 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
16 changes: 16 additions & 0 deletions src/plugins/intel_gpu/src/plugin/transformations_pipeline.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1723,6 +1723,22 @@ void TransformationsPipeline::apply(std::shared_ptr<ov::Model> func) {
return true;
}
}

// 4-bit FC is memory-bound by the int4 weights, so on the discrete-GPU oneDNN
// path an i8 activation adds a per-token quantize + i8 GEMM without any win over
// f16. Keep f16 (all batches) for discrete GPU + 4-bit weights; iGPU/8-bit keep DQ.
{
const auto& dev_info = m_context->get_engine().get_device_info();
const bool is_wei_4bit = cldnn::one_of(root->get_input_element_type(1),
{ov::element::i4, ov::element::u4});
if (dev_info.dev_type == cldnn::device_type::discrete_gpu && is_wei_4bit) {
GPU_DEBUG_TRACE << root->get_friendly_name()
<< " dyn_quan is turned off: discrete GPU + 4-bit weights"
" (f16 FC is faster and batch-consistent)" << std::endl;
return true;

@isanghao isanghao Jul 16, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

we should use dynamic quantization for int4 weights because it shows performance gain.

}
}

uint64_t adj_group_size = dynamic_quantization_group_size;
const bool is_wei_i8u8 = cldnn::one_of(root->get_input_element_type(1), {ov::element::i8, ov::element::u8});
if (ov::intel_gpu::DynamicQuantizeFullyConnected::ShouldUseGs128(is_wei_i8u8, use_gs128_for_int8_per_token, adj_group_size, use_gs128_for_linear_attention)) {
Expand Down
Loading