SIGN IN SIGN UP

[Test][Feature] Add Qwen3.5-27B multi-LoRA MTP eager serving E2E (#14794)

### What this PR does / why we need it?
This PR adds an end-to-end (E2E) test for Qwen3.5-27B serving with
multi-LoRA and built-in Multi-Token Prediction (MTP) in eager mode
(without ACL graph). It covers the serving path that preloads two named
LoRA modules and ensures that the LoRA routing works correctly by
verifying the chat completions output of both adapters.

PR CI is Hugging Face offline (`HF_HUB_OFFLINE=1`) and uses ModelScope.
There is no ModelScope-hosted Qwen3.5-27B PEFT adapter, so the test
synthesizes two rank-8 adapters with the 27B MLP shapes. Local
`MIX_LORA_A` / `MIX_LORA_B` still override that path.

### Does this PR introduce _any_ user-facing change?
No.

### How was this patch tested?
Added
`tests/e2e/pull_request/four_card/lora/test_qwen35_27b_multi_lora_mtp_eager.py`
which starts `RemoteOpenAIServer` with TP=4, `--enforce-eager`,
`qwen3_5_mtp`, fully-sharded LoRA, and two `--lora-modules` names, then
checks `/v1/models` and `/v1/chat/completions` for `mix-lora-a` vs
`mix-lora-b`.

- vLLM version: v0.27.1
- vLLM main:
https://github.com/vllm-project/vllm/commit/ba07e4a48fc951300d97eb506217dd530583dea3

Signed-off-by: Agoni-02 <114491222+Agoni-02@users.noreply.github.com>
A
Agoni-02 committed
9c7ef37c5079ab81365944a85339301bc8306406
Parent: 80c833f
Committed by GitHub <noreply@github.com> on 8/26/2026, 7:30:58 AM