Skip to content

[CPU]Improve dynamic 4bit moe performance #27240

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
merged 4 commits into from
Nov 4, 2025
Merged

[CPU]Improve dynamic 4bit moe performance #27240

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
00e2d64
[CPU]Improve dynamic 4bit moe performance
xiangze-arm Sep 17, 2025
863aca8
Fix pre-commit
xiangze-arm Oct 22, 2025
c083ca4
Revert silu_and_mul change
xiangze-arm Nov 3, 2025
0119fc3
Merge branch 'main' into int4_moe
bigPYJ1151 Nov 3, 2025
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
33 changes: 12 additions & 21 deletions csrc/moe/dynamic_4bit_int_moe_cpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,30 +87,23 @@ torch::Tensor dynamic_4bit_int_moe_cpu(
const int64_t g_eff_13 = (group_size != -1) ? group_size : H;
const int64_t g_eff_2 = (group_size != -1) ? group_size : I;

// Per-expert outputs filled in parallel
std::vector<torch::Tensor> y_list(E);
y_list.resize(E);
auto X_all = x_c.index_select(/*dim=*/0, expert_tokens);
if (apply_router_weight_on_input) {
X_all = X_all.mul(expert_gates.unsqueeze(1));
}
auto Y_all = at::empty({offsets[E], H}, x_c.options());

at::parallel_for(0, E, 1, [&](int64_t e_begin, int64_t e_end) {
c10::InferenceMode guard;
for (int64_t e = e_begin; e < e_end; ++e) {
const int64_t te = counts[e];
if (te == 0) {
y_list[e] = at::empty({0, H}, x_c.options());
continue;
}

const int64_t start = offsets[e];

auto sel_tokens =
expert_tokens.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);
auto gates_e =
expert_gates.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);

auto x_e = x_c.index_select(/*dim=*/0, sel_tokens);

if (apply_router_weight_on_input) {
x_e = x_e.mul(gates_e.unsqueeze(1));
}
auto x_e = X_all.narrow(/*dim=*/0, /*start=*/start, /*length=*/te);

auto w13_e = w13_packed.select(/*dim=*/0, e);
auto w2_e = w2_packed.select(/*dim=*/0, e);
Expand All @@ -137,17 +130,15 @@ torch::Tensor dynamic_4bit_int_moe_cpu(
// W2
auto y = mm(act, w2_e, g_eff_2, /*in_features=*/I, /*out_features=*/H);

if (!apply_router_weight_on_input) {
y = y.mul(gates_e.unsqueeze(1));
}

// Store per-expert result
y_list[e] = y;
Y_all.narrow(/*dim=*/0, /*start=*/start, /*length=*/te).copy_(y);
}
});

// Concatenate all expert outputs to match expert_tokens order
auto Y_all = at::cat(y_list, /*dim=*/0);
if (!apply_router_weight_on_input) {
Y_all = Y_all.mul(expert_gates.unsqueeze(1));
}

auto out = at::zeros({T, H}, x.options());
out =
at::index_add(out, /*dim=*/0, /*index=*/expert_tokens, /*source=*/Y_all);
Expand Down