fix(deepseek-v4): preserve grouped wo_a LoRA semantics

[zhoutong-hai]

Problem

DeepSeek-V4 wo_a is not a single dense projection. It stores G independent group weights in a flattened column-parallel matrix, and each input group must use only the matching output group.

A generic LoRALinear wrapper creates two problems:

1. native forward reads self.wo_a.weight, but the wrapper owns the base module under to_wrap;
2. applying the flattened LoRA projection to every input group produces a G × G cross-group result, while the base operator uses only the group diagonal.

Fix

Add a V4-specific grouped projection helper that:

• unwraps the base weight from the adapter wrapper;
• computes the native grouped base projection;
• evaluates the Bridge adapter through its public adapter_forward contract;
• reshapes the flattened delta and selects only matching input/output groups;
• preserves the base-only path when the adapter is disabled; and
• fails closed on incompatible base, input, or adapter-output layouts.

The helper is used only by V4 wo_a; other LoRA projections retain their generic wrappers.

Validation

• Model-free numerical test compares against the exact effective grouped weight.
• Gradients reach both LoRA factors.
• Disabled-adapter and incompatible-layout cases are covered.
• tests/fast/plugins/test_deepseek_v4_lora.py: 3 passed.
• Ruff and diff checks pass.

Ordering

This PR should merge before #2772. That PR enables and wraps wo_a, and therefore depends on this grouped execution path.

Related rollout-engine execution PR: sgl-project/sglang#36579. V4 target registration and dimensions: sgl-project/sglang#36578.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.