Skip to content

Move Gemma4 per-layer embeddings to embedding model (Alternative A) #296

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
justinchuby merged 4 commits into from
May 18, 2026
Merged

Move Gemma4 per-layer embeddings to embedding model (Alternative A) #296

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6bbb2cd
Move Gemma4 per-layer embeddings to embedding model (Alternative A)
justinchuby May 8, 2026
7d30095
Fix lint: lowercase local variable name in preprocess_weights
justinchuby May 8, 2026
3ef5750
Address review: ORT version comment, drop input_ids from decoder mock…
justinchuby May 8, 2026
3881873
Merge branch 'main' into gemma4-perlayer-fused
justinchuby May 9, 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
8 changes: 4 additions & 4 deletions src/mobius/integrations/ort_genai/auto_export_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1055,6 +1055,7 @@ class FakeConfig:
decoder = _mock_model_with_inputs(
[
"inputs_embeds",
"per_layer_inputs",
"input_ids",
"attention_mask",
"position_ids",
Comment thread
justinchuby marked this conversation as resolved.
Expand Down Expand Up @@ -1108,8 +1109,8 @@ def test_gemma4_vision_inputs(self, tmp_path):
assert "image_grid_thw" not in vision_inputs
assert data["model"]["vision"]["spatial_merge_size"] == 2

def test_gemma4_decoder_has_input_ids_and_inputs_embeds(self, tmp_path):
"""Gemma4 decoder has both inputs_embeds and input_ids."""
def test_gemma4_decoder_has_per_layer_inputs_and_inputs_embeds(self, tmp_path):
"""Gemma4 decoder has inputs_embeds and per_layer_inputs."""
pkg = self._make_gemma4_pkg()
path = _write_genai_config(
pkg.config,
Expand All @@ -1128,7 +1129,7 @@ def test_gemma4_decoder_has_input_ids_and_inputs_embeds(self, tmp_path):
data = json.load(f)
decoder_inputs = data["model"]["decoder"]["inputs"]
assert "inputs_embeds" in decoder_inputs
assert "input_ids" in decoder_inputs
assert "per_layer_inputs" in decoder_inputs
Comment thread
justinchuby marked this conversation as resolved.
# KV cache templates are present
assert decoder_inputs["past_key_names"] == "past_key_values.%d.key"

Expand Down Expand Up @@ -1342,7 +1343,6 @@ def test_gemma4_genai_config_from_real_model(self, tmp_path):
# Decoder inputs introspected from graph
decoder_inputs = data["model"]["decoder"]["inputs"]
assert "inputs_embeds" in decoder_inputs
assert "input_ids" in decoder_inputs
assert "attention_mask" in decoder_inputs
assert "position_ids" in decoder_inputs
assert decoder_inputs["past_key_names"] == ("past_key_values.%d.key")
Expand Down
Loading
Loading