=== boot 1 ===
=== boot 1 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] CONTAINER EXITED after 1864s
=== boot 2 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 81s
r0 16:17:21 WARNING budget cache: DISCARDED the cached plan at this configuration — 079619f01f3b3224.json, 079619f01f3b3224.overhang.json, 079619f01f3b3224.residual.json. The boot that wrote it took the plan at 2026-09-13T15:45:26.577072+00:00 and never reached ready, so nothing has shown this sizing can serve and it must not be read back as a measurement. This boot sizes from the cold path instead, which over-reserves where it cannot measure; the entries are kept beside the cache as '*.discarded' for diagnosis. Recovery is automatic and costs this one cold boot.
r0 16:18:08 INFO    Boot: KV cache ready at 10.07 GiB (43% of 23.54 GiB) — 4585 pages × 256 tok/page = 1173760 servable tokens, free 0.51 GiB (2% of 23.54 GiB)
r0 16:18:10 INFO    Boot: KV cache ready at 10.07 GiB (43% of 23.54 GiB) — 4585 pages × 256 tok/page = 1173760 servable tokens, free 0.51 GiB (2% of 23.54 GiB)
r1 16:18:11 INFO    post-freeze floor: this rank consumed 10 MiB of the 384 MiB held out (3612 MiB free at freeze, 3602 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: unclaimed.serving_consumption 228 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regions (KV slab, state arenas) +0 MiB, driver.local_memory (kernel-stack pool
r0 16:18:11 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (518 MiB free at freeze, 510 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: transient.serving_step.draft 141 MiB, unclaimed.residency_gate 109 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, transient.serving_step.driver_growth 6 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regi
r0 16:18:11 WARNING activation admission: this rank holds the group's LARGEST non-step reserve, 268357120 B against the smallest rank's 114609664 B, and the gate subtracts the largest from every rank's free VRAM because every rank executes the slate rank 0 builds. The 153747456 B of spread is KV no rank was given. This rank's rows: draft 147456000 B, verify_logits 64225280 B, verify_tail 50384384 B, driver_growth 6291456 B.
r0 16:18:11 INFO    Boot: ready in 71.2s — serving now. GPU memory: KV cache 10.09 GiB (43% of 23.54 GiB), free 0.50 GiB (2% of 23.54 GiB) (memory locked — won't grow while serving).
=== boot 3 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 81s
r0 16:21:18 INFO    Boot: KV cache ready at 10.31 GiB (44% of 23.54 GiB) — 4693 pages × 256 tok/page = 1201408 servable tokens, free 0.29 GiB (1% of 23.54 GiB)
r0 16:21:20 INFO    Boot: KV cache ready at 10.31 GiB (44% of 23.54 GiB) — 4693 pages × 256 tok/page = 1201408 servable tokens, free 0.29 GiB (1% of 23.54 GiB)
r0 16:21:21 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (294 MiB free at freeze, 286 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: transient.serving_step.draft 141 MiB, unclaimed.residency_gate 109 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, transient.serving_step.driver_growth 6 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regi
r1 16:21:21 INFO    post-freeze floor: this rank consumed 10 MiB of the 384 MiB held out (3388 MiB free at freeze, 3378 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: unclaimed.serving_consumption 228 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regions (KV slab, state arenas) +0 MiB, driver.local_memory (kernel-stack pool
r0 16:21:21 WARNING activation admission: this rank holds the group's LARGEST non-step reserve, 268357120 B against the smallest rank's 114609664 B, and the gate subtracts the largest from every rank's free VRAM because every rank executes the slate rank 0 builds. The 153747456 B of spread is KV no rank was given. This rank's rows: draft 147456000 B, verify_logits 64225280 B, verify_tail 50384384 B, driver_growth 6291456 B.
r0 16:21:21 INFO    Boot: ready in 73.1s — serving now. GPU memory: KV cache 10.31 GiB (44% of 23.54 GiB), free 0.28 GiB (1% of 23.54 GiB) (memory locked — won't grow while serving).
=== boot 4 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 80s
r0 16:24:21 INFO    Boot: KV cache ready at 10.31 GiB (44% of 23.54 GiB) — 4693 pages × 256 tok/page = 1201408 servable tokens, free 0.29 GiB (1% of 23.54 GiB)
r0 16:24:23 INFO    Boot: KV cache ready at 10.31 GiB (44% of 23.54 GiB) — 4693 pages × 256 tok/page = 1201408 servable tokens, free 0.29 GiB (1% of 23.54 GiB)
r1 16:24:23 INFO    post-freeze floor: this rank consumed 10 MiB of the 384 MiB held out (3388 MiB free at freeze, 3378 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: unclaimed.serving_consumption 228 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regions (KV slab, state arenas) +0 MiB, driver.local_memory (kernel-stack pool
r0 16:24:23 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (294 MiB free at freeze, 286 MiB now). The serve-kernel warmup and the readiness forward are the floor's first consumers and run past the last driver-residency bracket, so this is what those classes cost. Held out by: transient.serving_step.draft 141 MiB, unclaimed.residency_gate 109 MiB, transient.serving_step.verify_logits 61 MiB, transient.serving_step.verify_tail 48 MiB, transient.serving_step.arena_regrow 43 MiB, transient.serving_step.driver_growth 6 MiB, unclaimed.allocator_slack 4 MiB, transient.serving_step.base 0 MiB. Taken by: named pools +0 MiB, default allocator pool +2 MiB, cuMem-direct regi
r0 16:24:23 WARNING activation admission: this rank holds the group's LARGEST non-step reserve, 268357120 B against the smallest rank's 114609664 B, and the gate subtracts the largest from every rank's free VRAM because every rank executes the slate rank 0 builds. The 153747456 B of spread is KV no rank was given. This rank's rows: draft 147456000 B, verify_logits 64225280 B, verify_tail 50384384 B, driver_growth 6291456 B.
r0 16:24:24 INFO    Boot: ready in 71.2s — serving now. GPU memory: KV cache 10.31 GiB (44% of 23.54 GiB), free 0.28 GiB (1% of 23.54 GiB) (memory locked — won't grow while serving).
SERIES DONE
