Skip to content

Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types #6962

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
albertvillanova merged 2 commits into from
Aug 31, 2026
Merged

Pin layer_types so the tiny Gemma3 and Olmo3 models cover both attention types #6962

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
ffee169
Pin layer_types to one of each attention type for tiny Gemma3
albertvillanova Aug 28, 2026
f229067
Pin layer_types to one of each attention type for tiny Olmo3
albertvillanova Aug 28, 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
3 changes: 3 additions & 0 deletions scripts/generate_tiny_models/for_causal_lm/olmo3_for_causal_lm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,9 @@
num_key_value_heads=2,
num_hidden_layers=2,
intermediate_size=32,
# One of each attention type. Deriving the pattern from num_hidden_layers=2 gives two
# sliding layers (the reference is 3:1 over 32 layers), so the global RoPE path would never run.
layer_types=["sliding_attention", "full_attention"],
# Non-size fields kept aligned with the reference so the tiny config only differs in what we scale down.
max_position_embeddings=65536,
rms_norm_eps=1e-06,
Expand Down
4 changes: 3 additions & 1 deletion scripts/generate_tiny_models/for_conditional_generation/gemma3_for_conditional_generation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,9 @@
"hidden_size": 16,
"num_attention_heads": 4,
"num_key_value_heads": 2,
"layer_types": None, # Set it automatically from num_hidden_layers
# One of each attention type. Deriving the pattern from num_hidden_layers=2 gives two
# sliding layers (the reference is 5:1 over 34 layers), so the global RoPE path would never run.
"layer_types": ["sliding_attention", "full_attention"],
"intermediate_size": 32,
}
vision_config = {
Expand Down
Loading