[CI][LoRA] Preserve QKV fuser split sizes through LoRA wrappers

[AndreasKaratzas]

This PR fixes the LoRA regression in CI:

• https://buildkite.com/vllm/ci/builds/76671/list?jid=019f3973-c928-4da0-997d-33258ddbe1c5&tab=output

The regression was probably introduced by:

• Make the Transformers modeling backend as fast as native vLLM #47187

After transformers QKV fuser it rewrites separate q_proj, k_proj, and v_proj calls into one fused qkv_proj call, then splits the fused output with qkv_proj.split_sizes. It sets split_sizes on the generated QKVParallelLinear, but the LoRA wrapper did not preserve that dynamic attribute. When test_ilama_lora initialized the Transformers backend with LoRA enabled, the rewritten forward referenced MergedQKVParallelLinearWithLoRA.split_sizes and failed during engine startup.

[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.

[hmellor]

I'll look at this ASAP

[hmellor]

The cause is correct, but I think we should use a different solution. I shouldn't have patched an attribute onto QKVParallelLinear as it is brittle in cases like these. I would push an alternative but the PR was made from an org fork so I can't. I'll make a separate PR

[hmellor]

blocking in the meantime

[hmellor]

Here is my PR #47832