=== boot 1 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 81s
r0 16:34:09 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 16:34:11 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 16:34:11 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 16:34:11 INFO    post-freeze floor: this rank consumed 8 MiB of the 384 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.serving_consumption 81 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 
r0 16:34:11 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 16:34:12 INFO    Boot: ready in 74.2s — 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).
=== boot 2 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 71s
r0 16:35:37 INFO    Boot: KV cache ready at 10.34 GiB (44% of 23.54 GiB) — 4707 pages × 256 tok/page = 1204992 servable tokens, free 0.26 GiB (1% of 23.54 GiB)
r0 16:35:39 INFO    Boot: KV cache ready at 10.34 GiB (44% of 23.54 GiB) — 4707 pages × 256 tok/page = 1204992 servable tokens, free 0.26 GiB (1% of 23.54 GiB)
r1 16:35:40 INFO    post-freeze floor: this rank consumed 10 MiB of the 384 MiB held out (3356 MiB free at freeze, 3346 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 16:35:40 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (262 MiB free at freeze, 254 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 16:35:40 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 16:35:40 INFO    Boot: ready in 61.9s — serving now. GPU memory: KV cache 10.34 GiB (44% of 23.54 GiB), free 0.25 GiB (1% of 23.54 GiB) (memory locked — won't grow while serving).
=== boot 3 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 70s
r0 16:37:08 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 16:37:10 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 16:37:11 INFO    post-freeze floor: this rank consumed 8 MiB of the 384 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.serving_consumption 81 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 
r0 16:37:11 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.
r1 16:37:11 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 16:37:11 INFO    Boot: ready in 62.2s — 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).
=== boot 4 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 70s
r0 16:38:35 INFO    Boot: KV cache ready at 10.34 GiB (44% of 23.54 GiB) — 4707 pages × 256 tok/page = 1204992 servable tokens, free 0.26 GiB (1% of 23.54 GiB)
r0 16:38:37 INFO    Boot: KV cache ready at 10.34 GiB (44% of 23.54 GiB) — 4707 pages × 256 tok/page = 1204992 servable tokens, free 0.26 GiB (1% of 23.54 GiB)
r1 16:38:37 INFO    post-freeze floor: this rank consumed 10 MiB of the 384 MiB held out (3356 MiB free at freeze, 3346 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 16:38:37 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (262 MiB free at freeze, 254 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 16:38:37 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 16:38:38 INFO    Boot: ready in 62.0s — serving now. GPU memory: KV cache 10.34 GiB (44% of 23.54 GiB), free 0.25 GiB (1% of 23.54 GiB) (memory locked — won't grow while serving).
=== boot 5 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 70s
r0 16:40:06 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 16:40:08 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 16:40:09 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 16:40:09 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 16:40:09 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 16:40:09 INFO    Boot: ready in 62.8s — 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 6 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 81s
r0 16:41:51 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 16:41:53 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 16:41:53 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 16:41:53 INFO    post-freeze floor: this rank consumed 8 MiB of the 412 MiB held out (292 MiB free at freeze, 284 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 16:41:53 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 16:41:54 INFO    Boot: ready in 76.2s — 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).
=== boot 7 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 71s
r0 16:43:13 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 16:43:15 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 16:43:15 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 16:43:15 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 16:43:15 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 16:43:16 INFO    Boot: ready in 63.4s — 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 8 ===
 Container pfskew206-engine Starting 
 Container pfskew206-engine Started 
[pfskew] READY in 70s
r0 16:44:40 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 16:44:42 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 16:44:43 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 16:44:43 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 16:44:43 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 16:44:43 INFO    Boot: ready in 62.4s — 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
