Skip to content

Deduplicate packed matrix splitting #983

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
copybara-service merged 1 commit into from
Aug 19, 2026
+25 −37
Merged

Deduplicate packed matrix splitting #983

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
62 changes: 25 additions & 37 deletions gemma/weights.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -69,9 +69,10 @@ void LayerWeightsPtrs::InitAttWeights(std::vector<MatOwner>& mat_owners,
HWY_ASSERT(attn_vec_einsum_w.Rows() == heads * model_dim);
HWY_ASSERT(attn_vec_einsum_w.Cols() == qkv_dim);

const MatPadding padding = (att_weights.GetType() == Type::kQ4_0)
? MatPadding::kPacked
: MatPadding::kOdd;
const MatPadding padding =
(att_weights.GetType() == Type::kQ4_0)
? MatPadding::kPacked
: MatPadding::kOdd;
{
std::lock_guard<std::mutex> lock(g_mat_owners_mutex);
mat_owners.push_back(MatOwner());
Expand All @@ -84,8 +85,6 @@ void LayerWeightsPtrs::InitAttWeights(std::vector<MatOwner>& mat_owners,
const size_t dst_row_bytes = Q4_0Stream::PackedEnd(cols);
HWY_ASSERT(dst_row_bytes == heads * src_row_bytes);



uint8_t* dst_ptr = att_weights.RowBytes(0);
const uint8_t* src_ptr = attn_vec_einsum_w.RowBytes(0);

Expand All @@ -110,6 +109,24 @@ void LayerWeightsPtrs::InitAttWeights(std::vector<MatOwner>& mat_owners,
att_weights.SetScale(attn_vec_einsum_w.Scale());
}

static void SplitPackedMatrix(MatPtr& parent, size_t split_row, MatPtr& w1,
MatPtr& w2) {
const size_t stride = parent.Stride();
uint8_t* base_ptr = parent.RowBytes(0);
w1.SetPtr(base_ptr, stride);

size_t split_bytes = 0;
if (parent.GetType() == Type::kQ4_0) {
split_bytes = Q4_0Stream::PackedEnd(split_row * stride);
} else if (parent.GetType() == Type::kNUQ) {
split_bytes = NuqStream::PackedEnd(split_row * stride);
} else {
w2.SetPtr(parent.RowBytes(split_row), stride);
return;
}
w2.SetPtr(base_ptr + split_bytes, stride);
}

// For FFN. Fast, only updates pointers.
void LayerWeightsPtrs::SplitW1() {
// Used for Gemma layers; FFWVit uses different tensors.
Expand All @@ -132,22 +149,8 @@ void LayerWeightsPtrs::SplitW1() {
HWY_ASSERT(gating_einsum_w1.Cols() == gating_einsum_w.Cols());
HWY_ASSERT(gating_einsum_w2.Cols() == gating_einsum_w.Cols());

if (gating_einsum_w.GetType() == Type::kQ4_0) {
const size_t stride = gating_einsum_w.Stride();
uint8_t* base_ptr = gating_einsum_w.RowBytes(0);
gating_einsum_w1.SetPtr(base_ptr, stride);
gating_einsum_w2.SetPtr(base_ptr + Q4_0Stream::PackedEnd(ff_hidden_dim * stride), stride);
} else if (gating_einsum_w.GetType() == Type::kNUQ) {
const size_t stride = gating_einsum_w.Stride();
uint8_t* base_ptr = gating_einsum_w.RowBytes(0);
gating_einsum_w1.SetPtr(base_ptr, stride);
gating_einsum_w2.SetPtr(
base_ptr + NuqStream::PackedEnd(ff_hidden_dim * stride), stride);
} else {
const size_t stride = gating_einsum_w.Stride();
gating_einsum_w1.SetPtr(gating_einsum_w.RowBytes(0), stride);
gating_einsum_w2.SetPtr(gating_einsum_w.RowBytes(ff_hidden_dim), stride);
}
SplitPackedMatrix(gating_einsum_w, ff_hidden_dim, gating_einsum_w1,
gating_einsum_w2);
gating_einsum_w1.SetType(gating_einsum_w.GetType());
gating_einsum_w2.SetType(gating_einsum_w.GetType());
gating_einsum_w1.SetScale(gating_einsum_w.Scale());
Expand Down Expand Up @@ -192,22 +195,7 @@ void LayerWeightsPtrs::SplitAttW1() {
HWY_ASSERT(qkv_einsum_w1.Cols() == qkv_einsum_w.Cols());
HWY_ASSERT(qkv_einsum_w2.Cols() == qkv_einsum_w.Cols());

if (qkv_einsum_w.GetType() == Type::kQ4_0) {
const size_t stride = qkv_einsum_w.Stride();
uint8_t* base_ptr = qkv_einsum_w.RowBytes(0);
qkv_einsum_w1.SetPtr(base_ptr, stride);
qkv_einsum_w2.SetPtr(base_ptr + Q4_0Stream::PackedEnd(w1_rows * stride), stride);
} else if (qkv_einsum_w.GetType() == Type::kNUQ) {
const size_t stride = qkv_einsum_w.Stride();
uint8_t* base_ptr = qkv_einsum_w.RowBytes(0);
qkv_einsum_w1.SetPtr(base_ptr, stride);
qkv_einsum_w2.SetPtr(base_ptr + NuqStream::PackedEnd(w1_rows * stride),
stride);
} else {
const size_t stride = qkv_einsum_w.Stride();
qkv_einsum_w1.SetPtr(qkv_einsum_w.RowBytes(0), stride);
qkv_einsum_w2.SetPtr(qkv_einsum_w.RowBytes(w1_rows), stride);
}
SplitPackedMatrix(qkv_einsum_w, w1_rows, qkv_einsum_w1, qkv_einsum_w2);
qkv_einsum_w1.SetType(qkv_einsum_w.GetType());
qkv_einsum_w2.SetType(qkv_einsum_w.GetType());
qkv_einsum_w1.SetScale(qkv_einsum_w.Scale());
Expand Down
Loading