Skip to content

[Gemma] Fix weights loader assertion crash in kReadBF16 mode. #939

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
Jun 19, 2026
+4 −1
Merged

[Gemma] Fix weights loader assertion crash in kReadBF16 mode. #939

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
5 changes: 4 additions & 1 deletion gemma/weights.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -515,6 +515,7 @@ struct TensorToRead {
// only for kReadBF16
bool keep_type = false;
Type prev_type;
size_t prev_packed_bytes = 0;
};

// Allocates multiple in parallel and binds to NUMA nodes.
Expand All @@ -533,6 +534,7 @@ static void AllocateAndBindAll(std::vector<TensorToRead>& tensors,
MatPtr& mat = *tensor.mat;

tensor.prev_type = mat.GetType();
tensor.prev_packed_bytes = mat.PackedBytes();
// We only care about MatMul inputs; skip F32 or small tensors.
if (tensor.prev_type == Type::kF32 || mat.Rows() < 1024) {
tensor.keep_type = true;
Expand Down Expand Up @@ -596,7 +598,8 @@ static void ReadAllToBF16(const std::vector<TensorToRead>& tensors,
// Validate blob size matches allocated buffer before any read.
// MapAll (line ~557) and MakeBatches (line ~645) both assert this;
// this path was the only one missing the check.
HWY_ASSERT_M(tensor.range.bytes == mat.PackedBytes(), mat.Name());
HWY_ASSERT_M(tensor.range.bytes == tensor.prev_packed_bytes,
mat.Name());

if (tensor.keep_type) {
HWY_ASSERT(reader.file().Read(
Expand Down
Loading