== 5d97bb088dda 2026-09-28T15:57:08Z dflash-draft-path=/models/Qwen3.8-27B-DFlash2-exl3-6.0bpw
    runtime flags: prefix_cache=on; non-default: assert_warm_cache=True, ext_share_dir=/ext-cache, serve_supervised=True, serve_calibration_dir=/cal, tkv_bits=4, tkv_calibration_file=/cal/v59/qwen3.8-27b-exl3-4.0bpw-servedcal_v59.json, TKV_BAKED_KERNEL_DIR=/opt/cache-baked/torch_extensions, TKV_CACHE_DIR=/cache/tkv, TKV_MTP_MAX_BLOCK_M=12
15:58:54 INFO    TKV recent ring: sinks=4 w=32 slack=2 for 4 sequences (288 rows); 16 attention layer(s) read it through 16 ring layer(s); 9.16 MiB in scratch.attn_codec, 2.290 MiB per sequence.
15:59:06 INFO    TKV recent ring: sinks=4 w=32 slack=2 for 4 sequences (288 rows); 16 attention layer(s) read it through 16 ring layer(s); 9.16 MiB in scratch.attn_codec, 2.290 MiB per sequence.
15:59:35 INFO    context arenas: 568.2 MiB for a 262144-token key walk, fitted beside the KV pages into the measured 4.89 GiB (context bound 262144).
15:59:35 INFO    pool: 64 model layers (16 paged), 1112 pages × 256 tok = 4.62 GiB paged KV (17408 B/tok summed over the paged layers); 0.66 GiB PHYSICAL across ALL state pools (7.00 GiB reserved VA — paged KV + recurrent/conv/MLA state)
15:59:43 INFO    TKV recent ring: sinks=4 w=32 slack=2 for 4 sequences (288 rows); 16 attention layer(s) read it through 16 ring layer(s); 9.16 MiB in scratch.attn_codec, 2.290 MiB per sequence.
16:00:02 INFO    kv grow inputs: free 4997644288 B (4766 MiB) - unmappable 0 MiB [UNMEASURED] - floor 337259639 B (322 MiB) = growable 4444 MiB / per-page 4.25 MiB = +1046 pages costing 4416 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1069 pages (273664 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:00:02 INFO    kv grow inputs: free 4997644288 B (4766 MiB) - unmappable 0 MiB [UNMEASURED] - floor 396122624 B (378 MiB) = growable 4388 MiB / per-page 4.25 MiB = +1038 pages costing 4384 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1061 pages (271616 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:00:06 INFO    kv grow inputs: free 4997644288 B (4766 MiB) - unmappable 0 MiB [UNMEASURED] - floor 356276736 B (340 MiB) = growable 4426 MiB / per-page 4.25 MiB = +1046 pages costing 4416 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1069 pages (273664 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:00:10 INFO    kv grow inputs: free 5159124992 B (4920 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 4576 MiB / per-page 4.25 MiB = +1083 pages costing 4576 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as 
16:00:11 INFO    kv grow inputs: free 360841216 B (344 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 0 MiB / per-page 4.25 MiB = +0 pages costing 0 MiB of physical [MEASURED per-layer map cost]; cur_mapped 1106, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as a delta i
16:00:12 INFO    kv grow inputs: free 360841216 B (344 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 0 MiB / per-page 4.25 MiB = +0 pages costing 0 MiB of physical [MEASURED per-layer map cost]; cur_mapped 1106, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as a delta i
16:00:17 INFO    engine ready, accepting requests on 0.0.0.0:8000
== 682a74be792d 2026-09-28T16:00:39Z dflash-draft-path=/models/Qwen3.8-27B-DFlash2-exl3-6.0bpw
    runtime flags: prefix_cache=on; non-default: assert_warm_cache=True, ext_share_dir=/ext-cache, serve_supervised=True, serve_calibration_dir=/cal, tkv_bits=4, tkv_calibration_file=/cal/v59/qwen3.8-27b-exl3-4.0bpw-servedcal_v59.json, TKV_BAKED_KERNEL_DIR=/opt/cache-baked/torch_extensions, TKV_CACHE_DIR=/cache/tkv, TKV_MTP_MAX_BLOCK_M=12
16:03:55 INFO    context arenas: 568.2 MiB for a 262144-token key walk, fitted beside the KV pages into the measured 4.90 GiB (context bound 262144).
16:03:55 INFO    pool: 64 model layers (16 paged), 1114 pages × 256 tok = 4.62 GiB paged KV (17408 B/tok summed over the paged layers); 0.66 GiB PHYSICAL across ALL state pools (7.00 GiB reserved VA — paged KV + recurrent/conv/MLA state)
16:04:24 INFO    kv grow inputs: free 5016518656 B (4784 MiB) - unmappable 0 MiB [UNMEASURED] - floor 337259639 B (322 MiB) = growable 4462 MiB / per-page 4.25 MiB = +1053 pages costing 4448 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1076 pages (275456 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:04:24 INFO    kv grow inputs: free 5016518656 B (4784 MiB) - unmappable 0 MiB [UNMEASURED] - floor 394025472 B (376 MiB) = growable 4408 MiB / per-page 4.25 MiB = +1038 pages costing 4384 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1061 pages (271616 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:04:29 INFO    kv grow inputs: free 5016518656 B (4784 MiB) - unmappable 0 MiB [UNMEASURED] - floor 356276736 B (340 MiB) = growable 4444 MiB / per-page 4.25 MiB = +1046 pages costing 4416 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1536 -> target 1069 pages (273664 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up a
16:04:34 INFO    kv grow inputs: free 5177999360 B (4938 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 4594 MiB / per-page 4.25 MiB = +1083 pages costing 4576 MiB of physical [MEASURED per-layer map cost]; cur_mapped 23, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as 
16:04:34 INFO    kv grow inputs: free 379715584 B (362 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 18 MiB / per-page 4.25 MiB = +0 pages costing 0 MiB of physical [MEASURED per-layer map cost]; cur_mapped 1106, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as a delta 
16:04:36 INFO    kv grow inputs: free 379715584 B (362 MiB) - unmappable 4 MiB [MEASURED] - floor 356276736 B (340 MiB) = growable 18 MiB / per-page 4.25 MiB = +0 pages costing 0 MiB of physical [MEASURED per-layer map cost]; cur_mapped 1106, va_ceiling 1535 -> target 1106 pages (283136 tokens). Every term here is an input to the KV size; a boot-to-boot delta in the result must show up as a delta 
16:04:39 INFO    engine ready, accepting requests on 0.0.0.0:8000
