Skip to content

gemma : more consistent attention scaling for v2 and v3 #13951

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 3 commits into from
Jun 2, 2025
Merged

gemma : more consistent attention scaling for v2 and v3 #13951

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
36469ad
gemma : fix attn scale for 27B
ggerganov Jun 1, 2025
67c4346
cont : apply scale before attn
ggerganov Jun 1, 2025
fbc6df0
cont : consistent attention scaling
ggerganov Jun 2, 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
20 changes: 11 additions & 9 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -953,6 +953,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
case 46: type = LLM_TYPE_27B; break;
default: type = LLM_TYPE_UNKNOWN;
}

// ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
hparams.f_attention_scale = type == LLM_TYPE_27B
? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
: 1.0f / std::sqrt(float(hparams.n_embd_head_k));
} break;
case LLM_ARCH_GEMMA3:
{
Expand All @@ -973,6 +978,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
default: type = LLM_TYPE_UNKNOWN;
}

// ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
hparams.f_attention_scale = type == LLM_TYPE_27B
? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
: 1.0f / std::sqrt(float(hparams.n_embd_head_k));
Expand Down Expand Up @@ -8481,14 +8487,7 @@ struct llm_build_gemma2_iswa : public llm_graph_context {
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);

// ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
switch (model.type) {
case LLM_TYPE_2B:
case LLM_TYPE_9B:
case LLM_TYPE_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); break;
default: GGML_ABORT("fatal error");
};
cb(Qcur, "Qcur_scaled", il);
Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);

cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Expand Down Expand Up @@ -8629,9 +8628,12 @@ struct llm_build_gemma3_iswa : public llm_graph_context {
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);

// ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315
Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale);

cur = build_attn(inp_attn, gf,
model.layers[il].wo, NULL,
Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il);
Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il);
}

cur = build_norm(cur,
Expand Down
Loading