Skip to content

convert: fold gemma4 audio projection scaling into the weight reparam #723

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
jakmro merged 2 commits into from
Jun 14, 2026
Merged

convert: fold gemma4 audio projection scaling into the weight reparam #723

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
9ecfb17
convert: fold gemma4 audio projection scaling into the weight reparam
ncylich Jun 14, 2026
0939e3b
convert: drop dead embed_vision_embedding entry from gemma4 scaling
ncylich Jun 14, 2026
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
2 changes: 1 addition & 1 deletion python/cactus/convert/cactus_adapters/tensor_io.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -292,7 +292,7 @@ def save_tensor_with_header(tensor, output_path, precision='INT8', transpose=Fal
data = data * GEMMA4_WEIGHT_SCALE
elif 'router_scale' in filename:
data = data / GEMMA4_WEIGHT_SCALE
elif filename in ('token_embeddings.weights', 'output_weight.weights', 'embed_vision_embedding.weights', 'embed_vision_proj.weights'):
elif filename in ('token_embeddings.weights', 'output_weight.weights', 'embed_vision_proj.weights', 'embed_audio_proj.weights'):
data = data / GEMMA4_WEIGHT_SCALE
elif filename == 'output_norm.weights':
data = data * GEMMA4_WEIGHT_SCALE
Expand Down
1 change: 0 additions & 1 deletion python/cactus/convert/model_adapters/naming.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,6 @@ class NameMatch:
GEMMA4_DIV_BASENAMES = {
"token_embeddings",
"output_weight",
"embed_vision_embedding",
}


Expand Down
4 changes: 0 additions & 4 deletions python/cactus/transpile/model_adapters.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1094,10 +1094,6 @@ def _gemma4_compute_native_like_audio_features(
if not callable(getattr(embed_audio, "forward", None)):
raise TypeError("Gemma4 audio embedder is not callable")
projected = embed_audio(inputs_embeds=audio_encodings)
# Native Cactus Gemma4 scales audio soft tokens before merging them into
# the language stream. HF leaves this implicit in its fp16 path, but the
# quantized Cactus decoder expects the native-scaled representation.
projected = projected * (1.0 / 16.0)
# The transpiled bundle is shape-specialized from representative media.
# Native Gemma4 audio preprocessing emits an unpadded feature tensor for
# that media, so the post-subsampling sequence is already the real token
Expand Down
Loading