[veomni] feat: add DeepSeek-V3 to MOE_PARAM_HANDERS

[Luosuu]

What does this PR do?

Add deepseek_v3 entry to MOE_PARAM_HANDERS mapping in verl/workers/engine/veomni/utils.py, reusing the existing _map_moe_params_qwen3_moe handler since DeepSeek-V3 shares the same MoE parameter mapping logic as Qwen3-MoE.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: https://github.com/verl-project/verl/pulls?q=is%3Apr+deepseek_v3+moe
• Format the PR title as [{modules}] {type}: {description}

Test

This is a configuration-level change (adding a key to a dictionary mapping). The mapping reuses an existing, tested handler function (_map_moe_params_qwen3_moe). No new logic is introduced.

API and Usage Example

No API changes. DeepSeek-V3 MoE models will now be automatically handled by the veomni engine during parameter mapping.

Design & Code Changes

• Added "deepseek_v3": _map_moe_params_qwen3_moe to the MOE_PARAM_HANDERS dict in verl/workers/engine/veomni/utils.py.

Checklist Before Submitting

• Read the Contribute Guide.
• Apply pre-commit checks.
• Add / Update the documentation. (N/A - no doc changes needed for this config addition)
• Add unit or end-to-end test(s). Not feasible: this is a single-line dictionary entry addition reusing an existing handler; no new logic to test.
• Once your PR is ready for CI, send a message in the ci-request channel.
• If your PR is related to the recipe submodule, please also update the reference. (N/A)

[Luosuu]

cc @pengwu22 @wuxibin89

[gemini-code-assist]

Code Review

This pull request adds support for the deepseek_v3 model by mapping it to the existing qwen3_moe parameter handler. However, feedback indicates that reusing this handler propagates an existing bug where parameter suffixes are redundantly appended, resulting in incorrect mapping keys like '.weight.weight'.

[pengwu22]

It’s verified compatible with veomni, right?

[Luosuu]

It’s verified compatible with veomni, right?

yes, verified by end-to-end moonlight training