== 0ccfc025dd97 2026-09-29T05:42:58Z
05:44:06 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.
05:44:15 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)
05:44:15 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)
[TKV autotune] start: kv_cache shape=(1024, 256, 1088) data_ptr=787940000000 in_capture=False (total_pages=1024) H_kv=4 num_sms=128 sw=0 buckets=(1024, 4096, 8192, 16384) batches=(1, 2, 4, 8) (capped at max_batch=4) reusable_cells=0 inherited=0 variants=['splitk'] tile_tokens=[4, 8, 16] min_blocks=[0, 3]
05:47:20 WARNING [TKV warm-autotune] swept at boot: no baked table for K4/V4, H_kv=4, head=256, page=256, sliding=0 (fingerprint c372f4eee356dd13). The image bakes tables for its served default configuration only; this one is swept once per host and then read from the tune dir.
05:47:20 INFO    [TKV warm-autotune] arbi-serve 1 fingerprint(s): 1 swept here, 0 trusted from a same-surface prior (no kernel launched), 0 cache-hit (measured against sweep-src=b5332b8dfc1c8389)
05:47:20 INFO    TKV warm-autotune complete: 1 fp × 16 layers, MTP verify block_m=none (MTP off) — loaded 1/1 combo(s) (13 kernel variants), 1 swept, 0 cache-hit, 184.7s
05:49:06 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)
05:49:06 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)
05:49:06 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)
05:49:06 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
05:49:17 INFO    engine ready, accepting requests on 0.0.0.0:8000
== a31655e2d0e5 2026-09-29T05:49:49Z
05:50:04 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)
05:50:04 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)
05:50:04 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)
05:50:04 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
05:50:15 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)
05:50:15 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)
05:50:15 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)
05:50:15 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
05:50:27 INFO    engine ready, accepting requests on 0.0.0.0:8000
== 2d44db81a37f 2026-09-29T06:08:10Z
06:08:19 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.
06:08:25 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)
06:08:25 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)
06:08:25 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)
06:08:25 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
06:08:37 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)
06:08:37 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)
06:08:37 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)
06:08:37 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
06:08:48 INFO    engine ready, accepting requests on 0.0.0.0:8000
