test: keep Hunyuan LoRA CPU path on native linear

[andyluo7]

Summary

• keep the Hunyuan Image 3 PEFT QKV numerical contract on native CPU linear math
• leave real QKV construction, tensor-parallel sharding, adapter loading, and LoRA application on production code
• avoid dispatching CPU tensors into the ROCm-only unquantized GEMM custom operator

Root cause

The test is intentionally CPU-only, but vLLM selects unquantized GEMM from the process-wide platform. In the ROCm image that selects vllm::rocm_unquantized_gemm even though the test tensors are on CPU, producing NotImplementedError for the CPU backend for TP sizes 1, 2, and 4.

Example failure: https://buildkite.com/vllm/vllm-omni-amd-ci/builds/12820/table?jid=01a0dab5-3687-400c-b435-24f5786b8c98&tab=output

Validation

• python3 -m py_compile tests/diffusion/lora/test_lora_manager.py
• Ruff format check with the repository 120-column setting
• git diff --check
• exact-head AMD runtime validation requested after opening

Signed-off-by: andyluo7 andy.luo@amd.com

[vllm-omni-review-bot]

This PR matches CODEOWNERS paths: /tests/.

Code owners: @NickCao @yenuo26

Routing: @NickCao via CODEOWNERS; @yenuo26 via CODEOWNERS

@andyluo7, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

[vllm-omni-review-bot]

Omni ReviewBot: superseded

The CI failure noted on 1214805ebc44 refers to an earlier head; the pull request now points at 8f2553ed2dfb.

[vllm-omni-review-bot]

Omni ReviewBot routing record

Assigned Direct under experiment vllm-omni-strict-5050-20260829.

[vllm-omni-review-bot]

Omni ReviewBot review

PR description

This is a test-only change to the existing Hunyuan Image 3 PEFT QKV forward check. The test still builds a real QKVParallelLinear, loads a disk PEFT adapter through DiffusionLoRAManager, and compares wrapped.apply(x) to an explicit row-index oracle for TP sizes 1, 2, and 4. The new piece stubs dispatch_unquantized_gemm with torch.nn.functional.linear so CPU tensors are not sent into the process-wide ROCm unquantized GEMM custom operator. User-visible serving behavior is unchanged; the intended effect is that this CPU core_model test can keep its LoRA/QKV numerical contract on AMD images.

Change flow

flowchart TD
  A["[EXISTING] CPU PEFT QKV test TP 1/2/4"]:::existing
  B["[CHANGED] test_hunyuan_image3_peft_qkv_forward"]:::changed
  C["[NEW] cpu_unquantized_gemm via F.linear"]:::new
  D["[EXISTING] QKVParallelLinear and LoRA manager"]:::existing
  E["[EXISTING] Oracle x@A.T@B[rows].T * 0.5"]:::existing
  A --> B
  B --> C
  C --> D
  D --> E
  classDef existing fill:#e5e7eb,stroke:#6b7280,color:#111827
  classDef changed fill:#fef3c7,stroke:#d97706,color:#451a03,stroke-width:2px
  classDef new fill:#dcfce7,stroke:#16a34a,color:#052e16,stroke-width:2px
  classDef removed fill:#fee2e2,stroke:#dc2626,color:#450a0a,stroke-width:2px

Loading

CI at 1214805ebc44 (2026-09-26T20:31:16.574327+00:00): required check(s) blocking: buildkite/vllm-omni (failed), and -5 more. Observed Buildkite: buildkite/vllm-omni-amd-ci (failed), buildkite/vllm-omni-npu-ci (failed), buildkite/vllm-omni (failed), and 1 more.

See inline comments below.

[vllm-omni-review-bot]

[P1] Changed PEFT QKV CPU test has no runtime result

The only new runtime behavior is the dispatch_unquantized_gemm stub at this line, added so wrapped.apply(x) stays on native CPU linear math. The PR Validation section ran py_compile, Ruff format, and git diff --check, and deferred exact-head AMD runtime until after open. Those checks cannot tell whether the monkeypatch name is the one QKVParallelLinear.apply actually looks up, or whether an import-time GEMM cache still selects vllm::rocm_unquantized_gemm. Frozen CI at 1214805ebc4466b8cf04b0713240d2b3464cd9b9 (2026-09-26T20:31:16Z) has required buildkite/vllm-omni failed and operator-watched AMD/NPU/Intel lanes failed; .buildkite step commands were not readable in this sparse checkout, so I am not naming a guard selector. Run tests/diffusion/lora/test_lora_manager.py::test_hunyuan_image3_peft_qkv_forward (parametrized TP 1/2/4) and record that result against the CPU core_model lane before treating the ROCm CPU crash as fixed.

[vllm-omni-review-bot]

Omni ReviewBot: finding feedback

[p1] Changed PEFT QKV CPU test has no runtime result — tests/diffusion/lora/test_lora_manager.py:1167

The only new runtime behavior is the dispatch_unquantized_gemm stub at this line, added so wrapped.apply(x) stays on native CPU linear math. The PR Validation section ran py_compile, Ruff format, and git diff --check, and deferred exact-head AMD runtime until after open. Those checks cannot tell whether the monkeypatch name is the one QKVParallelLinear.apply actually looks up, or whether an import-time GEMM cache still selects vllm::rocm_unquantized_gemm. Frozen CI at 1214805ebc4466b8cf04b0713240d2b3464cd9b9 (2026-09-26T20:31:16Z) has required buildkite/vllm-omni failed and operator-watched AMD/NPU/Intel lanes failed; .buildkite step commands were not readable in this sparse checkout, so I am not naming a guard selector. Run tests/diffusion/lora/test_lora_manager.py::test_hunyuan_image3_peft_qkv_forward (parametrized TP 1/2/4) and record that result against the CPU core_model lane before treating the ROCm CPU crash as fixed.

This finding appears in the bot's COMMENT review, but GitHub could not place it as an inline diff comment. If you are the PR author and disagree, react 👎 to this comment. The disagreement will be shown to the maintainer; it does not approve or merge the PR.

[andyluo7]

Self-review at 8f2553ed: the branch is rebased onto current main, including the merged Krea2 collection fix. I checked vLLM v0.30.0: UnquantizedLinearMethod calls dispatch_unquantized_gemm(linear_backend) during construction and caches the returned callable. The test stub now accepts dispatcher arguments and is installed before each real QKVParallelLinear is constructed, so the cached implementation is native torch.nn.functional.linear for this CPU-only numerical contract. py_compile and git diff --check pass; exact-head AMD runtime validation is being retriggered.

[vllm-omni-review-bot]

Omni ReviewBot: superseded

The CI failure noted on 8f2553ed2dfb refers to an earlier head; the pull request now points at e9b2c0f1deee.

[yenuo26]

LGTM