Skip to content

perf: run block-wise int4/int8 QMoE CPU experts on the MLAS QNBit GEMM kernels #32644

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
Tianlei Wu (tianleiwu) merged 11 commits into from
Sep 17, 2026
Merged

perf: run block-wise int4/int8 QMoE CPU experts on the MLAS QNBit GEMM kernels #32644

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
6f88804
perf: run block-wise int4/int8 QMoE CPU experts on the MLAS QNBit GEM…
ykhrustalev Sep 16, 2026
33b3684
refactor: tighten the QMoE CPU QNBit prepack and GEMM plumbing
ykhrustalev Sep 16, 2026
385cc3f
feat: QMoE accuracy_level attribute, zero points and a fallback warni…
ykhrustalev Sep 16, 2026
b036131
fix: dependent template keyword for Tensor::Data calls in the QMoE QN…
ykhrustalev Sep 16, 2026
c54c674
style: clang-format the QMoE QNBit packed-experts struct
ykhrustalev Sep 16, 2026
28ec438
fix: address QMoE QNBit review findings
ykhrustalev Sep 16, 2026
d026ecb
test: cover the 8-bit QNBit default and its forced-fp32 fallback
ykhrustalev Sep 16, 2026
61173ff
fix: document the 8-bit int8 default, SafeInt the shared-buffer size …
ykhrustalev Sep 16, 2026
acebeeb
fix: prefer int8 QNBit kernels for accuracy_level=4 and keep the per-…
ykhrustalev Sep 17, 2026
0d244bc
test: capture QMoE outputs for both element types in the block-wise h…
ykhrustalev Sep 17, 2026
bc5f446
test: pin the remaining QMoE QNBit policy cells
ykhrustalev Sep 17, 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
2 changes: 2 additions & 0 deletions docs/ContribOperators.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5675,6 +5675,8 @@ This version of the operator has been available since version 1 of the 'com.micr
#### Attributes

<dl>
<dt><tt>accuracy_level</tt> : int</dt>
<dd>Minimum accuracy level of the expert GEMMs on CPU, with the MatMulNBits meaning. For block-wise 4-bit experts, 0 (default) or 1 keeps fp32 activations and 4 allows int8 activations (int8 dot-product kernels). Block-wise 8-bit experts have no fp32 kernel and use int8 activations at every level. Other values are treated as 0.</dd>
<dt><tt>activation_alpha</tt> : float</dt>
<dd>Alpha parameter used in activation function.</dd>
<dt><tt>activation_beta</tt> : float</dt>
Expand Down
Loading
Loading