Normalize OpenPangu RoPE scaling defaults

[skyloevil]

Summary

• Remove hard-coded RoPE scaling defaults and load values from model config.
• When instantiating OpenPanguEmbeddedAttention/OpenPanguMLAAttention, clone config.rope_scaling if it exists, otherwise start from {} and use default.
• Backfill all DeepSeek YaRN defaults (beta_fast, beta_slow, factor, mscale, mscale_all_dim, type, original_max_position_embeddings) .

Verification

• config.json add rope_scaling_config:

"rope_scaling": {
  "type": "yarn",
  "beta_fast": 48,
  "factor": 1.25,
  "mscale": 0.9,
  "mscale_all_dim": 0.5,
  "original_max_position_embeddings": 8192
}

• CUDA_VISIBLE_DEVICES=0 vllm serve FreedomIntelligence/openPangu-Embedded-7B --port 8818 --trust_remote_code --served-model-name openPangu-Embedded-7B – logs show every OpenPanguEmbeddedAttention layer printing the resolved rope_scaling_config, confirming it was read from the config (see screenshot).

• curl http://localhost:8818/v1/chat/completions … – inference succeeds with expected response, proving the model loads and serves normally (see screenshot).

• After verification passed, the debug log was removed.

[gemini-code-assist]

Code Review

This pull request aims to normalize the RoPE scaling defaults for OpenPangu models by reading from config.rope_scaling and backfilling any missing YaRN parameters. My main concern is that the implementation seems to be in OpenPanguMLAAttention, while the verification logs and model configuration for FreedomIntelligence/openPangu-Embedded-7B indicate that OpenPanguEmbeddedAttention is being used. This suggests the fix might be in the wrong place and would not be applied as intended. I've left a critical comment detailing this issue. Please verify the correct location for this logic.

[skyloevil]

/gemini review

[gemini-code-assist]

Code Review

This pull request introduces a helper function _normalize_rope_scaling_defaults to centralize the logic for handling RoPE scaling parameters for OpenPangu models. This change effectively removes hardcoded default values from OpenPanguMLAAttention and ensures that both OpenPanguMLAAttention and OpenPanguEmbeddedAttention consistently apply these defaults. The implementation is clean, robust, and improves the model's configurability by allowing rope_scaling parameters to be specified in the model's configuration file. The changes are well-verified and represent a good improvement in code maintainability and consistency.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[skyloevil]

/gemini review

[skyloevil]

@codex review

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Can't wait for the next one!

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[gemini-code-assist]

Code Review

This pull request refactors the RoPE scaling configuration for OpenPangu models by introducing a normalization function, _normalize_rope_scaling_defaults. This change removes hardcoded values and allows for using RoPE scaling parameters from the model's configuration, with sensible defaults. The implementation is applied to both OpenPanguMLAAttention and OpenPanguEmbeddedAttention.

My review identifies a potential logic issue in the new normalization function that could lead to inconsistent values for type and rope_type in the resulting configuration dictionary. I've provided a suggestion to refactor the logic for better clarity and correctness, ensuring that the configuration remains consistent regardless of which keys are provided by the user.