PASS bash -n run_A_gpu_multi.sh
PASS py_compile owned files
INFO remaining single-device literals (defaults only expected): 3 lines -> logs/static_device_grep.txt
PASS other importers: from ac_common import * + load_model(tiny, cpu)
PASS test_helpers.py (SUMMARY 29/29 passed)
PASS measure bs=2: new run_measure_A (unsharded) == original code, byte-identical
PASS measure bs=2: 4 shards (device cpu) merged == unsharded, byte-identical
PASS measure bs=2: 4 shards (device auto2) merged == unsharded, byte-identical
PASS measure bs=7: new run_measure_A (unsharded) == original code, byte-identical
PASS measure bs=7: 3 shards (device cpu) merged == unsharded, byte-identical
PASS measure bs=7: 3 shards (device auto2) merged == unsharded, byte-identical
INFO measure bs=2 rows=449 shard sizes: shard 0/4: 57 of 225;shard 1/4: 56 of 225;shard 2/4: 56 of 225;shard 3/4: 56 of 225;
PASS measure shard crossed a 50-batch checkpoint boundary
PASS merge refuses a missing shard
PASS merge refuses shards from different runs (bs/nshards)
PASS no output written on refused merge
PASS ablate single path: new == original (parts + ablation.json)
PASS ablate split: 4 replicas ended with PARTS_OK
PASS ablate split: replicas wrote no ablation.json
PASS ablate assemble_only: parts untouched, nothing recomputed, 3s
PASS ablate split+assemble == original (config differs only in device/out)
PASS ablate race: 3 overlapping writers computed each of 16 parts exactly once (got 16)
PASS ablate race+assemble == original
SUMMARY 21 PASS, 0 FAIL
