Skip to content

[Feature] Support glm yarn scaling rope #7893

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

Open
Sunny-bot1 wants to merge 1 commit into
base: develop
Choose a base branch
from
Open

[Feature] Support glm yarn scaling rope #7893

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
25 changes: 22 additions & 3 deletions fastdeploy/model_executor/layers/rotary_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -268,7 +268,7 @@ def forward(
return query, key


class GptOssScalingRotaryEmbedding:
class YarnScalingRotaryEmbedding:
def __init__(
self,
rotary_dim,
Expand Down Expand Up @@ -345,10 +345,29 @@ def get_rope_impl(
rotary_emb_layer = QwenRotaryEmbedding(rotary_dim, base, partial_rotary_factor)
rotary_emb = rotary_emb_layer(position_ids)
elif architecture.startswith("Glm"):
rotary_emb_layer = GlmRotaryEmbedding(rotary_dim, base, partial_rotary_factor)
rope_scaling = getattr(model_config, "rope_scaling", None)
if (
rope_scaling is not None
and isinstance(rope_scaling, dict)
and rope_scaling.get("rope_type", rope_scaling.get("type", "")) == "yarn"
and "factor" in rope_scaling
):
yarn_rotary_dim = int(rotary_dim * partial_rotary_factor) if partial_rotary_factor < 1.0 else rotary_dim
rotary_emb_layer = YarnScalingRotaryEmbedding(
rotary_dim=yarn_rotary_dim,
base=base,
original_max_position_embeddings=rope_scaling["original_max_position_embeddings"],
scale=rope_scaling["factor"],
mscale=rope_scaling.get("mscale", 1.0),
beta_fast=rope_scaling.get("beta_fast", 32),
Copy link
Copy Markdown

@PaddlePaddle-bot PaddlePaddle-bot May 22, 2026

Choose a reason for hiding this comment

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

🔴 Bug rope_scaling["original_max_position_embeddings"] 直接字典访问，存在 KeyError 风险。

当前条件只验证了 "factor" in rope_scaling，但未检查 "original_max_position_embeddings" 是否存在。若用户的 rope_scaling 配置中缺少该字段，运行时会直接抛出 KeyError 导致服务崩溃。

建议修复（使用 .get() + 显式报错）：

original_max_pos = rope_scaling.get("original_max_position_embeddings")
if original_max_pos is None:
    raise ValueError(
        "rope_scaling must contain 'original_max_position_embeddings' when rope_type is 'yarn'"
    )
rotary_emb_layer = YarnScalingRotaryEmbedding(
    rotary_dim=yarn_rotary_dim,
    base=base,
    original_max_position_embeddings=original_max_pos,
    ...
)

beta_slow=rope_scaling.get("beta_slow", 1),
use_neox_rotary_style=False,
)
else:
rotary_emb_layer = GlmRotaryEmbedding(rotary_dim, base, partial_rotary_factor)
rotary_emb = rotary_emb_layer(position_ids)
elif architecture.startswith("GptOss"):
rotary_emb_layer = GptOssScalingRotaryEmbedding(
rotary_emb_layer = YarnScalingRotaryEmbedding(
rotary_dim=model_config.head_dim,
base=model_config.rope_theta,
original_max_position_embeddings=model_config.rope_scaling["original_max_position_embeddings"],
Expand Down
Loading