== run_row_arbiserve CONFIG=BYPASS (arm=BYPASS) GPU=1 host=arbi-whitebox-201 ==
[image] using=registry.arbi.work/arbi-serve:test-fa8d4e2 is_pin=1 digest=<none>
-- kernel cache: /home/arbi-admin/ringab27/ladder_cache/arbiserve  state=COLD (will compile kernels this run)  (/cache/{tkv,fa4-cute,triton,arbi-serve,xgrammar} + /home/arbi-admin/ringab27/ladder_cache/arbiserve/torch_extensions_gpu1)
-- arbi-serve env forwarded: ARBI_TKV_RECENT_RING=0
-- MOUNT_ARBI_SERVE: /home/arbi-admin/ringab27/src/arbi_serve -> /opt/venv/lib/python3.12/site-packages/arbi_serve (src sha f3661ead9d96c6ea)
-- scoring BYPASS -> /mnt/k8scache/paper_kv_ablation/run_arbiserve_27b_ring_ladder_20260928/out/warmup_BYPASS.json (chunk=2048 ctx=16384 gmu=1.0)
[arbiserve_score] BYPASS -> backend paged_kv:tkv-bypass  (tkv=/opt/venv/lib/python3.12/site-packages/tkv/__init__.py src=d2edb26a05ece0b7)
activation admission: the widest slate the step limits allow needs 405 MiB but the card has 100 MiB free for one step — SHORT by 305 MiB. The padded gather is bounded at one chunk and costs nothing here, so this shortfall is elsewhere in the step. Admission narrows the prefill chunk to what fits rather than letting that slate's forward OOM after the layout freeze, so the shortfall costs chunk wi
[arbiserve_score] engine up in 1071.1s tkv_autotune_table_sha256=51072d5d0b0ffb2fbfbbaebc0007fd0c286b57cfeb88b08f801c74df7f74036e
FATAL: alignment preflight FAILED at 3 of 4 greedy positions (first: position 1, fire argmax 248046, emitted 198). At temperature 0 the emitted token IS the argmax of the distribution that produced it, so the fire index this scorer reads is not the position it thinks. Refusing to score.
FATAL[run_row_arbiserve BYPASS @ arbi-whitebox-201:gpu1]: docker/score exited 1
