Skip to content

Store pre_att_rms_out as BF16 #987

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

Open
Mikyx-1 wants to merge 1 commit into
base: dev
Choose a base branch
from
+8 −5
Open

Store pre_att_rms_out as BF16 #987

Show file tree
Hide file tree
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
6 changes: 4 additions & 2 deletions gemma/activations.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,7 +197,9 @@ struct AttentionActivations {
MatStorageT<KV_t> vit_K_T;
MatStorageT<KV_t> vit_V_T;

MatStorageT<float> pre_att_rms_out;
// BF16 because this is only ever the A operand of the QKV MatMuls, which
// would otherwise decompress it to BF16 on every call; see `pre_ffw_rms_out`.
MatStorageT<BF16> pre_att_rms_out;
MatStorageT<float> att_out; // attention output
MatStorageT<float> att_out_reps; // attention output for each thread.
MatStorageT<float> softmax_max; // see OnlineSoftmaxState
Expand Down Expand Up @@ -308,7 +310,7 @@ struct AttentionActivationsPtrs {
MatPtrT<KV_t> vit_V_T;

// Output of RMSNorm before attention, size batch_size x model_dim.
MatPtrT<float> pre_att_rms_out;
MatPtrT<BF16> pre_att_rms_out;
// Attention output computed from att * V, size batch_size x (q_heads *
// qkv_dim).
MatPtrT<float> att_out;
Expand Down
7 changes: 4 additions & 3 deletions gemma/attention_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,12 +43,13 @@ HWY_BEFORE_NAMESPACE();
namespace gcpp {
namespace HWY_NAMESPACE {

void FillRandom(MatPtrT<float>& mat, uint64_t seed) {
template <typename T>
void FillRandom(MatPtrT<T>& mat, uint64_t seed) {
hwy::RandomState rng(seed);
for (size_t r = 0; r < mat.Rows(); ++r) {
float* row = mat.Row(r);
T* row = mat.Row(r);
for (size_t c = 0; c < mat.Cols(); ++c) {
row[c] = static_cast<float>(RandomGaussian(rng));
row[c] = hwy::ConvertScalarTo<T>(RandomGaussian(rng));
}
}
}
Expand Down