[sergereview]

Original CI failure

• Failure group: 2 integration tests regressed by commit 83d46aa2a2c4 (PR #47625)
• tests/models/minicpmv4_6/test_modeling_minicpmv4_6.py::MiniCPMV4_6IntegrationTest::test_small_model_vision_generation_batch [multi-gpu] (output_mismatch, seen 7/7)

CI traceback — tests/models/minicpmv4_6/test_modeling_minicpmv4_6.py::MiniCPMV4_6IntegrationTest::test_small_model_vision_generation_batch

(line 505)  AssertionError: Lists differ: ['The animal in the image is a Pystylus, a[145 chars] It'] != ["The animal in the image is a Pystylus, a[161 chars]s a"]

Where to watch it:

• test_small_model_vision_generation_batch — Test dashboard

PR #47625 fixed Gated Delta Net correctness in the Qwen3.5 text backbone used by MiniCPMV4_6, so generated text legitimately changed. The failing batch test's (cuda, (10, 0)) expectation still used the pre-fix string ending in "It's a"; the CI trace shows the new output ends in " It", matching the (8, 6) expectation. Update the (10, 0) expected text to the post-PR value so the integration test keeps the correctness fix intact.

Relates to #48603

Possibly related

Existing issues/PRs mentioning test_small_model_vision_generation_batch (keyword match — not verified to share a root cause):

• #48599 — [WIP] [MiniCPMV4_6] Fix vit_merger crash when video frames are sliced (non-uniform target_sizes) (PR, open, updated 2026-09-07)
• #48454 — [Fix] Fix A10 expectations for a test (PR, closed, updated 2026-09-07)
• #48406 — [MiniCPMV4_6] Update test_small_model_vision_generation_batch expected output (value drift) (PR, closed, updated 2026-09-01)

---

This change was produced automatically by serge from a CI failure report. The patch was generated by an LLM and applied by serge; review before merging.

serge v0.1.0 · model: moonshotai/Kimi-K2.7-Code · 63 LLM turns · 66 tool calls · 383.3s · 2060827 in / 13052 out tokens

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: minicpmv4_6

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 34167590344
Result: success | Grafana metrics are not available yet.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[tarekziade]

⚠️ Post-hoc GPU verification: already_passing — closing

This PR was opened without the GPU verify gate running. A serge deploy on 2026-09-04 shipped a stale values file that silently unset VERIFY_ON_GPU, so for three nightlies serge opened fix PRs without running the targeted @slow tests — and without the footer that would have said so either way. The configuration is restored; this is the retroactive verification.

Run: https://github.com/huggingface/transformers/actions/runs/34200524970
(serge-verify-caller.yml, baseline 8eaf75f8 vs this PR bb058d7f, 5 runs per tree, plus the full tests/models/minicpmv4_6 slow suite on both trees.)

	baseline 8eaf75f8	this PR bb058d7f
MiniCPMV4_6IntegrationTest::test_small_model_vision_generation_batch	green	green
new failures in tests/models/minicpmv4_6	—	0

The targeted test was already passing before this patch. serge's own gate calls that already_passing — "a targeted test was NOT failing on the pre-patch baseline (self-healed or flaky) — nothing to fix" — and it is one of the verdicts that stops a PR being opened at all. With VERIFY_ON_GPU set, this branch would have been torn down and you would never have seen it.

Two further reasons not to keep it:

1. It is an expectation-only patch. reviewbot/expectation_guard.py classifies it as one: every changed line is a string literal in a test, so re-running the test after the change proves nothing — it passes by construction. The published PR should have carried ⚠️ Not verified — this patch changed the expectation instead of nothing at all.
2. It is the second PR for the same failure group. This branch is serge/fix/itf-62b451006663-8f1d4cd4; [serge] Fix 2 integration tests regressed by commit 83d46aa2a2c4 (PR #47625) #48580 was serge/fix/itf-62b451006663-f800400a — same fingerprint, dispatched again the following night, and [serge] Fix 2 integration tests regressed by commit 83d46aa2a2c4 (PR #47625) #48580 (also an expectation rewrite) was merged in between. So this rewrites a device-keyed expectation on a test whose expectations were rewritten yesterday.

Closing. If the minicpmv4_6 expectations are actually wrong, that deserves one deliberate change checked against real model output on a known device — not two automated rewrites on consecutive nights, neither of which was verified.