[serge] Fix 2 integration tests regressed by commit bcb9849a4e66 (PR #48714)

[sergereview]

Original CI failure

• Failure group: 2 integration tests regressed by commit bcb9849a4e66 (PR #48714)
• tests/models/qwen3_omni_moe/test_modeling_qwen3_omni_moe.py::Qwen3OmniModelIntegrationTest::test_small_model_integration_test_batch_audio_matches_single [multi-gpu] (other, seen 7/7)

CI traceback — tests/models/qwen3_omni_moe/test_modeling_qwen3_omni_moe.py::Qwen3OmniModelIntegrationTest::test_small_model_integration_test_batch_audio_matches_single

(line 311)  RuntimeError: We encountered some issues during automatic conversion of the weights. For details look at the `CONVERSION` entries of the above report!

Where to watch it:

• test_small_model_integration_test_batch_audio_matches_single — Test dashboard

Relates to #48954

The Qwen3-Omni-MoE integration tests OOM while MergeModulelist stacks MoE expert gate_up_proj weights during from_pretrained. The 30B checkpoint in bfloat16 is ~60 GiB; device_map="auto" fills each ~22 GiB GPU to capacity, leaving no room for the 384 MiB torch.stack temporary buffer.

Reserve 85% of each accelerator for auto placement and keep a CPU offload budget, leaving headroom for weight-conversion temporaries. This keeps PR #48714's kernels version bump intact; it only changes how the test asks for memory.

---

✅ Verified on GPU

serge ran the targeted @slow test(s) on a GPU runner and opened this PR only after they passed with this patch. Each targeted test was run 5× on both the pre-patch and patched trees to rule out flakiness — the result held on every run.

• Passes with this patch: https://github.com/huggingface/transformers/actions/runs/35474398928
• Failed before the fix (base commit): https://github.com/huggingface/transformers/actions/runs/35472698566

Possibly related

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

• #48903 — Fix tests due to dropping attn mask (PR, closed, updated 2026-09-18)
• #48750 — Switch daily CI to torch 2.14 — update expected outputs (PR, closed, updated 2026-09-13)
• #47186 — Add support for batched Qwen2.5/3-Omni audio generation (PR, closed, updated 2026-08-03)

---

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 · 51 LLM turns · 50 tool calls · 68.4s · 2076648 in / 11583 out tokens

[github-actions]

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

run-slow: qwen3_omni_moe

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 35476724874
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.

[vasqu]

Makes no sense tbh and is some OOM handling which is not related to bumping kernels

[ydshieh]

well, the PR description is not in a good shape, but the fix itself make some sense, except the 300G one.

I will open a new one.