Skip to content

Normalize OpenPangu RoPE scaling defaults #28361

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
skyloevil wants to merge 5 commits into
base: main
Choose a base branch
from
+31 −11
Open

Normalize OpenPangu RoPE scaling defaults #28361

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
7eee77f
Load rope scaling from config in OpenPangu
skyloevil Nov 8, 2025
ceefa88
Normalize rope scaling in embedded attention
skyloevil Nov 9, 2025
94c7e95
Silence flake8 SIM108 warning
skyloevil Nov 9, 2025
c7634d2
Apply ruff format after pre-commit
skyloevil Nov 9, 2025
f0e2d61
Respect custom rope scaling types
skyloevil Nov 9, 2025
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
42 changes: 31 additions & 11 deletions vllm/model_executor/models/openpangu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,29 @@ def check_ffn_act_fn(act_fn: str):
)


def _normalize_rope_scaling_defaults(
rope_scaling_config: dict[str, Any] | None, max_position_embeddings: int
) -> dict[str, Any]:
"""Return a DeepSeek-YaRN compatible rope_scaling dict."""
rope_scaling = dict(rope_scaling_config) if rope_scaling_config is not None else {}
original_scaling_type = rope_scaling.get("type")

rope_scaling.setdefault("beta_fast", 32)
rope_scaling.setdefault("beta_slow", 1)
rope_scaling.setdefault("factor", 1)
rope_scaling.setdefault("mscale", 1.0)
rope_scaling.setdefault("mscale_all_dim", 1.0)
rope_scaling.setdefault("type", "yarn")
rope_scaling.setdefault("original_max_position_embeddings", max_position_embeddings)
if "rope_type" not in rope_scaling:
if original_scaling_type is not None:
rope_scaling["rope_type"] = original_scaling_type
else:
rope_scaling["rope_type"] = "deepseek_yarn"

return rope_scaling


class OpenPanguMLP(nn.Module):
def __init__(
self,
Expand Down Expand Up @@ -338,17 +361,10 @@ def __init__(
prefix=f"{prefix}.o_proj",
)

# TODO: remove hard coding
rope_scaling = {
"beta_fast": 32,
"beta_slow": 1,
"factor": 1,
"mscale": 1.0,
"mscale_all_dim": 1.0,
"original_max_position_embeddings": max_position_embeddings,
"type": "yarn",
"rope_type": "deepseek_yarn",
}
rope_scaling_config = getattr(config, "rope_scaling", None)
rope_scaling = _normalize_rope_scaling_defaults(
rope_scaling_config, max_position_embeddings
)
self.rotary_emb = get_rope(
qk_rope_head_dim,
rotary_dim=qk_rope_head_dim,
Expand Down Expand Up @@ -529,6 +545,10 @@ def _init_rotary_emb(
if is_gguf and config.model_type == "PanguEmbedded":
is_neox_style = False

rope_scaling = _normalize_rope_scaling_defaults(
rope_scaling, self.max_position_embeddings
)

self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
Expand Down