[CI] Restore MiniCPMV transformers cap, scoped to HF runner only

[hmellor]

Purpose

Restores the max_transformers_version cap on the MiniCPMV test registry entry, scoped to the HF runner only (transformers_version_reason={"hf": ...}).

This PR previously flipped an existing "vllm" reason to "hf". Since then #48413 removed the cap outright, so it has been rewritten: the cap is re-added, with the correct reason.

Why the cap is needed again

Nightly build 83094 (Multi-Modal Models (Extended Generation 3), an optional: true step that only runs in the scheduled full CI run) fails 9 tests, all with:

AttributeError: 'MiniCPMV' object has no attribute 'all_tied_weights_keys'

• test_single_image_models[minicpmv_25-test_case100..102]
• test_single_image_models[minicpmv_26-test_case103..105]
• test_multi_image_models[minicpmv_26-test_case77..79]

The failure is entirely on the HF reference side. The traceback ends in transformers/modeling_utils.py:

for key in missing_keys - self.all_tied_weights_keys.keys()

all_tied_weights_keys is set by PreTrainedModel.post_init(), and MiniCPMV's remote code never calls it: post_init appears zero times in modeling_minicpmv.py for MiniCPM-Llama3-V-2_5, MiniCPM-V-4 and MiniCPM-V-4_5 (2_6 is gated on the Hub now, but fails identically). Same failure mode as SkyworkR1VChatModel (#42104) and the sibling MiniCPMO entry, both already capped.

Why it surfaced now

The MiniCPMV entry used to carry max_transformers_version="4.57" with a "vllm" reason. Because vlm_utils/core.py::run_test calls check_transformers_version(on_fail="skip") with the default check_max_version=True, that cap gated the HF-runner tests too. #48413 fixed the vLLM-side MiniCPMVBatchFeature problem and removed the cap entirely, which correctly re-enabled vLLM-only coverage but also un-gated the HF comparisons and exposed this separate, still-unfixed HF-side break.

Using an "hf"-scoped reason keeps both halves right: HF-runner comparisons skip, vLLM-only tests (including the ones #48413 added in tests/models/multimodal/processing/test_minicpmv.py) keep running, because those callers pass check_max_version=False, check_version_reason="vllm".

Tracks under #38379.

Not a duplicate

This is the pre-existing PR for this fix, rewritten rather than superseded by a new one. Duplicate checks run:

gh pr list --repo vllm-project/vllm --state open --search "minicpmv"
gh pr list --repo vllm-project/vllm --state open --search "all_tied_weights_keys in:body"
gh pr list --repo vllm-project/vllm --state open --search "48413 in:body"

No other open PR touches the MiniCPMV registry entry. #43760 and #42785 are runtime/CUDA-graph changes to the model itself.

Testing

pytest tests/models/test_registry.py -q
# 368 passed, 2 failed, 16 skipped

The two failures (test_registry_imports[HCXVisionForCausalLM], test_registry_imports[KananaVForConditionalGeneration]) are pre-existing and unrelated: confirmed by re-running them on a stashed tree.

Behaviour of the gate itself, against transformers==5.16.0.dev0:

for m in ["openbmb/MiniCPM-Llama3-V-2_5", "openbmb/MiniCPM-V-2_6",
          "openbmb/MiniCPM-V-4", "openbmb/MiniCPM-V-4_5"]:
    info = HF_EXAMPLE_MODELS.find_hf_info(m)
    info.check_transformers_version(on_fail="return")                    # -> skip message
    info.check_transformers_version(on_fail="return",
                                   check_max_version=False,
                                   check_version_reason="vllm")          # -> None

All four variants skip HF-runner tests and none skip vLLM-only tests, which is the intended split.

No model evaluation results: this is a test-gating change only, with no effect on model output, accuracy or serving.

AI assistance

AI assistance was used for this change (investigating the Buildkite failure, bisecting the cause to #48413, and drafting the patch and this description). I have reviewed every changed line and run the tests reported above.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @hmellor.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[hmellor]

/ci run

[github-actions]

✅ Triggered Buildkite CI #83373 for commit 7f92af551114.

[AndreasKaratzas]

LGTM