LAB: cuda-unit section, aprender-serve --lib test binary killed by host global OOM during qwen35_session gpu test #4988

LAB (C324): the fat-job cuda-unit section on the sm_89 runner is not a required check, so this red blocks no merge or release. This is its one ticket.

What was measured

• Run 37973784963, job 113967873127, PR ci: remove a fat job's section containers when its run is cancelled #4962 head 1f9a7c7, section cuda-unit, step 4 failed after 712.9 s.
• The aprender-serve --lib test binary (realizar) exited with status 137 (SIGKILL) during gguf::qwen35_session::tests::gpu::gpu_a_turn_that_extends_the_last_one_reuses_the_state_and_matches_one_shot, run with --test-threads 1 under the host GPU lock.
• Runner kernel log, 18:48:54Z: global_oom, constraint CONSTRAINT_NONE. The kernel killed the test binary at total-vm 22.6 GB, anon-rss 12.45 GB, oom_score_adj 500. The OOM was triggered from a different docker cpuset than the killed task's memcg, so other work on the host was also allocating.
• PR ci: remove a fat job's section containers when its run is cancelled #4962 changes only the fat-job teardown step, which runs after this section. The killed binary is unrelated to the diff.
• Same class, earlier: a global OOM on the same runner at 23:52:01Z on 2026-10-08 killed a rustc of aprender-contracts in this section.

Not measured

• Whether 12.45 GB anon RSS is normal for this binary at this test, or a regression. No per-test RSS baseline exists.
• What else held host memory at 18:48:54Z. The OOM dump lists the tasks, but they were not tallied.

Related

• FLAKE: intermittent SIGSEGV in libcuda during gdn device tests (aprender-gpu --lib) #4956: SIGSEGV in libcuda during gdn device tests. Different signal, same section.
• yoga cuda-unit: cta64_vs_cta32_vs_cublas_fp16 CUDA_ERROR_OUT_OF_MEMORY under the host GPU lock (flaky, 1 of 2 on yoga-eph) #4593: CUDA_ERROR_OUT_OF_MEMORY on the device under the host GPU lock. That is device memory; this ticket is host memory.

Done when

The cause is measured: either a per-test RSS baseline for the realizar binary shows a regression to fix, or the host OOM traces to concurrent jobs and the runner's memory caps are fixed. In both cases one green night of this section is required.

🤖 Generated with Claude Code