Skip to content

llama : fix Gemma-2 Query scaling factors #8473

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 2 commits into from
Jul 14, 2024
Merged

llama : fix Gemma-2 Query scaling factors #8473

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
df78f19
9B - query_pre_attn_scalar = 256 not 224
danielhanchen Jul 11, 2024
acc877f
llama : fix Gemma-2 Query scaling factor
ggerganov Jul 13, 2024
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: 0 additions & 5 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2473,11 +2473,6 @@ def set_gguf_parameters(self):
)
self.gguf_writer.add_sliding_window(self.hparams["sliding_window"])

# sanity check
attn_scalar = self.hparams["query_pre_attn_scalar"]
if attn_scalar != hparams["hidden_size"] / hparams["num_attention_heads"]:
raise ValueError("query_pre_attn_scalar must be equal to n_embd / n_head")

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused

Expand Down
7 changes: 6 additions & 1 deletion src/llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11679,7 +11679,12 @@ struct llm_build_context {
ext_factor, attn_factor, beta_fast, beta_slow);
cb(Qcur, "Qcur", il);

Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head)));
// ref: https://github.com/google/gemma_pytorch/commit/03e657582d17cb5a8617ebf333c1c16f3694670e
switch (model.type) {
case e_model::MODEL_9B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); break;
case e_model::MODEL_27B: Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd / n_head))); break;
default: GGML_ASSERT(false);
};
cb(Qcur, "Qcur_scaled", il);

Kcur = ggml_rope_ext(
Expand Down
Loading