== 06acd60c097d 2026-09-29T08:27:16Z
08:27:20 INFO    TKV recent ring: w=128 slack=2 for 4 sequences (1024 rows); 16 attention layer(s) read it through 16 ring layer(s); 32.53 MiB in scratch.attn_codec, 8.132 MiB per sequence.
08:27:26 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:27:26 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:27:26 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:27:26 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.2s
08:27:44 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:27:44 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:27:44 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:27:44 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.0s
08:27:59 INFO    engine ready, accepting requests on 0.0.0.0:8000
== 4af06b5ae3b4 2026-09-29T08:28:36Z
08:28:50 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:28:50 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:28:50 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:28:50 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (7 kernel variants), 0 swept, 1 cache-hit, 0.1s
08:29:02 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:29:02 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:29:02 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:29:02 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (7 kernel variants), 0 swept, 1 cache-hit, 0.0s
08:29:14 INFO    engine ready, accepting requests on 0.0.0.0:8000
== 1b317a62e8b5 2026-09-29T08:40:37Z
08:40:47 INFO    TKV recent ring: w=128 slack=2 for 4 sequences (1024 rows); 16 attention layer(s) read it through 16 ring layer(s); 32.53 MiB in scratch.attn_codec, 8.132 MiB per sequence.
08:40:57 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:40:57 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:40:57 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:40:57 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.2s
08:41:14 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
08:41:14 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
08:41:14 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
08:41:14 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.0s
08:41:28 INFO    engine ready, accepting requests on 0.0.0.0:8000
== 2ecfbcaa30e0 2026-09-29T09:03:39Z
09:03:52 INFO    TKV recent ring: w=128 slack=2 for 4 sequences (1024 rows); 16 attention layer(s) read it through 16 ring layer(s); 32.53 MiB in scratch.attn_codec, 8.132 MiB per sequence.
09:03:58 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
09:03:58 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
09:03:58 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
09:03:58 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.2s
09:04:10 INFO    TKV warm-autotune: 1 fp × 16 layers, MTP verify block_m=none (MTP off) (miss → 1-time nvcc JIT, GPU idle, NOT a hang → /tkv-jit-cache)
09:04:10 INFO    TKV warm-autotune fingerprint 1/1: K4/V4 bits, codec=Klloyd/Vlloyd, head_dim=256, kv_heads=4, q_heads=24 (gqa=6), 16 layer(s)
09:04:10 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 0 swept here, 0 trusted from a same-surface prior (no kernel launched), 1 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
09:04:10 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 0 swept, 1 cache-hit, 0.0s
09:04:21 INFO    engine ready, accepting requests on 0.0.0.0:8000
