Skip to content

MatmulNBits prepacking scales fix #27412

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

Merged
hariharans29 merged 11 commits into from
Feb 24, 2026
Merged

MatmulNBits prepacking scales fix #27412

Show file tree
Hide file tree
Changes from 6 commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
e579d41
Fix incorrect scales element count calculation while pre-packing scal…
hariharans29 Feb 21, 2026
34ce775
Add test
hariharans29 Feb 21, 2026
9d40b7d
Add test
hariharans29 Feb 23, 2026
51526a8
Add MatmulNBits unit test and remove shared lib test
hariharans29 Feb 23, 2026
c4e3e49
Add comment
hariharans29 Feb 23, 2026
6f5da27
Add test for fp16
hariharans29 Feb 23, 2026
c101131
Update onnxruntime/test/contrib_ops/matmul_4bits_test.cc
hariharans29 Feb 23, 2026
972a403
Experiment
hariharans29 Feb 23, 2026
26114e7
More fixes
hariharans29 Feb 23, 2026
8b9f0bb
Experiment - 2
hariharans29 Feb 23, 2026
2559293
Update tolerance
hariharans29 Feb 23, 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
6 changes: 3 additions & 3 deletions onnxruntime/contrib_ops/cpu/quantization/matmul_nbits.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -359,9 +359,9 @@ Status MatMulNBits<MLFloat16>::PrePack(const Tensor& tensor, int input_idx, /*ou
OpKernel::Info().TryGetConstantInput(InputIndex::scales, &scales);
if (scales && MlasQNBitGemmScalesPacked(K_, nbits_, block_size_, compute_type_, has_zp_input_)) {
auto sptr = scales->Data<MLFloat16>();
auto tensor_size = static_cast<size_t>(tensor.Shape().Size());
auto ptr = IAllocator::MakeUniquePtr<float>(alloc, tensor_size, true);
MlasConvertHalfToFloatBuffer(sptr, ptr.get(), tensor_size);
auto scales_size = static_cast<size_t>(scales->Shape().Size());
auto ptr = IAllocator::MakeUniquePtr<float>(alloc, scales_size, true);
MlasConvertHalfToFloatBuffer(sptr, ptr.get(), scales_size);
scales_fp32_ = std::move(ptr);
}

Expand Down
5 changes: 5 additions & 0 deletions onnxruntime/test/contrib_ops/matmul_4bits_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -495,6 +495,11 @@ TEST(MatMulNBits, Float16_4b_Accuracy4) {
TestMatMulNBitsTyped<MLFloat16, 100, 288, 93, 32, 4>();
TestMatMulNBitsTyped<MLFloat16, 100, 288, 93, 128, 4>();
TestMatMulNBitsTyped<MLFloat16, 100, 288, 1234, 16, 4>();

// See PR #27412 for details on the following test case,
// which is added to cover a specific failure case in the past.
TestMatMulNBitsTyped<MLFloat16, 369, 6144, 2048, 32, 4>();

Comment thread
hariharans29 marked this conversation as resolved.
}

TEST(MatMulNBits, LegacyShape_4b) {
Expand Down
Loading