Skip to content

update gemma4 rope #90

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
Jintao-Huang merged 7 commits into from
May 22, 2026
Merged

update gemma4 rope #90

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
45e547a
refactor partial_rotary_factor
Jintao-Huang May 22, 2026
59e1886
refactor rope partial_rotary_factor
Jintao-Huang May 22, 2026
f6028ed
fix
Jintao-Huang May 22, 2026
33f821d
fix
Jintao-Huang May 22, 2026
d50b85e
fix
Jintao-Huang May 22, 2026
59c9622
update
Jintao-Huang May 22, 2026
a71b91a
update
Jintao-Huang May 22, 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: 2 additions & 0 deletions src/mcore_bridge/config/model_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -287,6 +287,8 @@ def __post_init__(self):
self.rope_scaling = json_parse_to_dict(self.rope_scaling)
if 'type' in self.rope_scaling and 'rope_type' not in self.rope_scaling:
self.rope_scaling['rope_type'] = self.rope_scaling['type']
if self.multi_latent_attention and self.partial_rotary_factor is None:
self.partial_rotary_factor = self.qk_pos_emb_head_dim / self.kv_channels
Comment thread
Jintao-Huang marked this conversation as resolved.
Outdated

if self.add_bias_linear:
self.add_qkv_bias = True
Expand Down
5 changes: 2 additions & 3 deletions src/mcore_bridge/model/mm_gpts/gemma4.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -551,7 +551,7 @@ def _set_inv_freq(self):
rope_scaling = self.config.rope_scaling
self.config.rope_scaling = rope_scaling['sliding_attention']
new_inv_freq, attention_scaling = get_rope_inv_freq(self.config)
assert attention_scaling == 1, 'not support'
self.config.attention_scaling = attention_scaling
self.rotary_pos_emb.inv_freq = new_inv_freq.to(self.rotary_pos_emb.inv_freq.device)
# full
self.full_rotary_pos_emb = copy.copy(self.rotary_pos_emb)
Expand All @@ -561,9 +561,8 @@ def _set_inv_freq(self):
kwargs['head_dim_key'] = 'global_head_dim'
new_inv_freq, attention_scaling = get_rope_inv_freq(
self.config, text_config=self.config.hf_config.text_config, **kwargs)
assert attention_scaling == 1, 'not support'
self.full_rotary_pos_emb.inv_freq = new_inv_freq
self.config.attention_scaling = attention_scaling
self.config.full_attention_scaling = attention_scaling
Comment thread
Jintao-Huang marked this conversation as resolved.
Copy link
Copy Markdown

@gemini-code-assist gemini-code-assist Bot May 22, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The attribute full_attention_scaling is being dynamically added to the config object at runtime. It should be explicitly defined in the ModelConfig class in src/mcore_bridge/config/model_config.py to ensure consistency, proper documentation, and to avoid potential AttributeError if accessed before this initialization step.


self.config.rope_scaling = rope_scaling

Expand Down
2 changes: 1 addition & 1 deletion src/mcore_bridge/model/rope.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ def _get_dummy_config(config):
rope_scaling=config.rope_scaling,
rope_theta=config.rotary_base,
max_position_embeddings=config.max_position_embeddings,
head_dim=config.qk_pos_emb_head_dim if config.multi_latent_attention else config.kv_channels,
head_dim=config.kv_channels,
hidden_size=config.hidden_size,
num_attention_heads=config.num_attention_heads,
)
Expand Down
Loading