[serge] Fix 8 integration tests for model mamba2 failing with OOM (other (8))

[sergereview]

Original CI failure

• Failure group: 1/1: 8 integration tests for model mamba2failing withOOM (other (8))
• tests/models/mamba2/test_modeling_mamba2.py::Mamba2IntegrationTest::test_batched_equivalence_with_cache [multi-gpu] (OOM, seen 7/7)

Relates to #47053

The mamba2 integration tests were OOMing on 22 GB CI GPUs. Root cause: the tests loaded the 7B Mamba-Codestral-7B-v0.1 checkpoint with dtype=torch.bfloat16, which is not the from_pretrained argument for weight dtype; the kwarg is ignored, so weights load in float32 and roughly double memory usage. The patch switches those calls to torch_dtype=torch.bfloat16, adds a torch.cuda.empty_cache() in setUp to reduce fragmentation between tests, and lowers the mixer equivalence test sequence length from 512 to 128 to further reduce activation memory.

---

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 · 48 LLM turns · 49 tool calls · 205.6s · 2131666 in / 15914 out tokens

[github-actions]

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

run-slow: mamba2

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 28699495193:1
Result: success | Jobs: 5 | Tests: 335 | Failures: 0 | Duration: 4m 38s

[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.