Repository navigation
tools/test-run.sh takes the fleet's run-time correctness slot itself #1004
Description
From lukstafi/ludics#160 (PR lukstafi/ludics#166). The fleet's run-time correctness cap is now fleet-worker.sh execution slot -- <command>, an N-holder flock a worker wraps around each test batch on its box. Today every wave brief has to remember the wrapper and every worker has to honour it. If tools/test-run.sh run takes the slot itself when FLEET_SLOT_STATE (or the fleet roster) is configured on the box, no brief has to say it and no worker can forget it; a box without the fleet configuration runs as before. One function: locate fleet-worker.sh through the deployed skill link, exec through execution slot when present, otherwise run directly. The ludics-lite side is done; this is the runner's half.
Activity
lukstafi commentedon Sep 23, 2026
New motivation from #1033 (staging#769). The native per-slot widths (hip
BOX_JOBS_SDMA_CAP / BOX_JOBS_NATIVE_GPU_SLOTS= 8 / 2, cuda 8) hard-code the fleet's slot count 2 (lukstafi/ludics#316). Iftest-run.shtook the fleet slot itself, the per-slot width could follow the fleet's count instead of restating it in a second repository.lukstafi commentedon Sep 26, 2026
In review as lukstafi#803, with lukstafi/ludics#435 as its companion. On a fleet box
tools/test-run.sh run/startnow takes the slot itself. It declares--cpuonly whenocannl_read_configresolves a CPU backend from every test configuration, so an unsetOCANNL_BACKENDis fine; otherwise it declares--gpu. A worker's ownexecution slotwrapper runs the batch inside that one slot through ludics-lite's newFLEET_SLOT_HELDmarker. The runner only takes a slot when the deployed fleet-worker answersexecution slot --probe, so the two PRs can land in either order.lukstafi commentedon Sep 26, 2026
Landed in lukstafi#803, with lukstafi/ludics#435 as its companion.
tools/test-run.sh run/starttake the fleet's run-time correctness slot themselves, throughtools/fleet-slot-run.sh. A worker's ownexecution slotwrapper is harmless: the ludics-lite side exportsFLEET_SLOT_HELD, and a nested slot runs inside the enclosing one.--cpu/--gpu) is resolved, and fails closed:--gpuif the argv can pick its own backend (dune exec, or a command-line--ocannl_backendnaming a non-CPU backend).--gpuif the targets can reach a stanza that names a GPU backend.Test_utils.Slot_kindworks this out from the markersenv_var_depsenforces, reading only a closed set of argv shapes; any other shape counts as GPU.--cpuonly whenocannl_read_configresolves a CPU backend from bothtest/configandarrayjit/test. An empty backend counts as GPU, becauseContext.autotries GPUs first.SLOT REFUSEDwith exit 75, and the slot's wait takes at most half of--cap.repeattakes no slot, andOCANNL_TOOL_FLEET_WORKER=noneturns it off.execution slot --probe. Until then the runner runs without a slot, as before.Worked examples: on rog,
runtest test/operationsis a GPU batch because it reachestest_cuda_pool_offset. A single CPU test's alias,@test/operations/scans,runtest test/einsumand@slowtake a CPU slot.Merge pull request #1004 from lukstafi/claude/gh-1217-strided-resnet-…