test(ci): GPU-tests speedups (OPS-8033, OPS-8113, OPS-8117)

[dmitry-tokarev-nv]

Summary

Cherry-pick 3 merged CI test-scheduling changes from main onto release/1.4.0 as one PR. All three shrink the pre_merge GPU stage by moving CPU-only work off the VRAM-budgeted lane or by correcting an over-stated VRAM reservation; none of them change product code.

Cherry-pick	Origin	Effect on the GPU stage
test(ci): reclassify CPU-only sglang unit tests from gpu_1 to gpu_0 (OPS-8033)	#12718 (8d4a014873)	8 CPU-only sglang unit files leave the VRAM stage (~8,400s cumulative); loadscope runs them in ~21s
test(ci): reclassify CPU-only vllm unit tests from gpu_1 to gpu_0 (OPS-8113)	#12826 (a274300453)	322 vLLM tests / ~12,024s of 19,562s cumulative leave the VRAM stage
test(ci): correct the mm_router VRAM profile from 18.7 to the measured 7.6 GiB (OPS-8117)	#12842 (3405438c75)	18.7 GiB exceeded the multi-process budget and made every test in those two files exclusive; 7.6 GiB is the measured solo peak

Each commit is cherry-picked individually with -x, so the origin commit is recorded in each message and per-PR attribution is preserved.

Linear: OPS-8033, OPS-8113, OPS-8117

Conflict resolution

Only #12718 conflicted, in two files, and both conflicts were main-vs-release/1.4.0 drift rather than anything in the backported change. Resolved to apply the gpu_1 → gpu_0 intent and nothing else:

• test_sglang_multimodal_embedding_cache.py — took gpu_0, kept this branch's timeout(60) block, and did not pull in main's pytest.mark.multimodal (not present on release/1.4.0).
• test_sglang_unit.py — took the _cpu_engine_when_no_accelerator autouse fixture (the actual mechanism behind the reclassification) and dropped test_configured_engine_route_cannot_replace_built_in_route, which is pre-existing main context unrelated to this change.

The other two cherry-picks applied cleanly.

Validation

• Combined diff is 19 files (8 + 9 + 2), matching the sum of the three source PRs — no extra files rode along.
• Both conflict-resolved files parse; no conflict markers remain anywhere in the tree.
• pre-commit run --all-files passes on this branch with no autofixes.
• Verified the mm_router comment's cross-reference still holds here: tests/mm_router/test_vllm_mm_router_e2e.py is also at profiled_vram_gib(7.6) on release/1.4.0.
• The one remaining pytest.mark.gpu_1 under the touched vLLM test dirs (omni/test_omni_nixl_connector.py) is intentional — test(ci): reclassify CPU-only vllm unit tests from gpu_1 to gpu_0 (OPS-8113) #12826 left it at gpu_1 on main too.
• All three commits are GPG-signed and DCO signed-off.

CI on the release branch still needs a maintainer /ok to test b2ba74f959.

🤖 Generated with Claude Code

---

[devin-ai-integration]

Devin Review found 1 potential issue.

[datadog-official]

🎯 Code Coverage (details)
• Patch Coverage: 89.66%
• Overall Coverage: 44.16% (+3.98%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: b2ba74f | Docs | Datadog PR Page | Give us feedback!}