test_cpu_threadpool: the 100x dispatch ratio is calibrated on 20 cores and flakes on 4-vCPU CI runners

[VikashLoomba]

Row: QUANT-GGUF-CPU-THREADPOOL

tests/vt/test_cpu_threadpool.cpp, case oversubscribed dispatch does not cost a scheduler timeslice, asserts ratio < 100.0 on ratio = median_dispatch_us(cores + 1) / median_dispatch_us(cores / 2).

Observed on CI (4-vCPU hosted runner)

build-test-cpu job 102683560829 (PR #3095, run 34416865779), tests/vt/test_cpu_threadpool.cpp:610:

MESSAGE: empty-op dispatch: 2 threads 0.421 us, 5 threads 42.75 us, ratio 101.544
ERROR: CHECK( ratio < 100.0 ) is NOT correct!

The arms are 2 and 5 threads, which proves the runner reported hardware_concurrency() == 4.

Why this reads as a calibration gap rather than a detected defect

• The recorded defect signature in this test's own comment is an over arm of 2999-5996 us at 21 threads against 18-20 us fixed. Here the over arm is 42.75 us, roughly 70x below that signature, so the waiter was not spinning through a scheduler timeslice.
• The comment's calibration ("Verified both ways on 20 cores") and its margin ("~30x below the defect and ~28x above the fixed behaviour") come from a 20-core box with arms 10 and 21. On a 4-vCPU runner the arms are 2 and 5 and the denominator measured 0.421 us, about 4x below the 1.68 us the same comment calibrates against, so the ratio is dominated by denominator noise.
• std::thread::hardware_concurrency() is not reduced by CPU affinity on this glibc: taskset -c 0-3 <binary> still selects 24/49-thread arms, so the 4-vCPU geometry cannot be reproduced locally by pinning cores.
• main's own build-test-cpu job 103074166342 passes the same test on the same runner class, so the check is marginal there rather than deterministically red.

Suggested fix (owner's call; changing an assertion needs its own spec)

Floor the denominator, for example over_us < std::max(100.0 * fits_us, <absolute floor>), or scale the threshold with cores, or skip the case below 8 cores. Any of these must keep the case able to fail for the defect it was written for.

Found while triaging an unrelated red check on PR #3095: that diff changes attention and embedding gating in src/vt/ops.cpp and never calls the measured path, which is Threadpool::Run with an empty body.

[VikashLoomba]

Verified against the latest main (57220ba34), not against a branch.

git show origin/main:tests/vt/test_cpu_threadpool.cpp carries, at that revision:

• line 574 const int cores = static_cast<int>(std::thread::hardware_concurrency());
• line 575 if (cores < 4) { ... return; } — a 4-core runner is deliberately not skipped
• line 586 const int fits = cores / 2;
• line 587 const int over = cores + 1;
• line 613 CHECK_MESSAGE(ratio < 100.0, ...) — no denominator floor, no core-count scaling

So on a 4-vCPU runner the arms are 2 and 5 and the threshold is 100, which is the failing CI geometry.

git diff --stat 6db4bef906859e864c82523c01107473f7dcca29..origin/main -- tests/vt/test_cpu_threadpool.cpp src/vt/cpu/cpu_threadpool.cpp is empty: the test and the threadpool are unchanged between this report's base and current main, and the pull request that showed the failure does not touch either file.

Pull-request CI also builds the merge of the head into main's tip (the pr-size step log in the same run reads Merge 1f9107a27f2b0caf244ed0e3e74bb9ec307a55df into 2207bd4c5057a3f35778e247b532cf4fcdd314c6), so the failing tree already contained main's code plus that head. Nothing about syncing a branch changes this.