strata generate: PCIe probe: 28.7 GB/s host->device -> pcie_frac 0.55 (default 0.55)
strata generate: this CPU has no AVX-512: the expert kernels run on AVX-2 (multi-token for the i-quant gate/up rows)
strata generate: native pack: /data/strata-bench/packs/q2_0 experts (largest blob 1.38 MB), token embedding type 11 in mapped host memory (260 MiB)
strata generate: 1416 MiB of weights loaded from /data/strata-bench/packs/q2_0 (303 canonical tensors skipped: served natively)
strata generate: 301 native projection matrices, 1376.20 MiB of weights
strata generate: KV streaming: 32768 of 131072 cells per QSA layer in VRAM, the K/V in 1.55 GiB of pinned RAM
strata generate: PLE on, table 320001536 rows of /data/strata-bench/models/ISTA-ed59f920/Q2_0/Qwen3.8-Flash-Next-GSQ-RCO-Q2_0-00002-of-00002.gguf
strata mtp: draft layer loaded, 835 MiB of VRAM (experts 675, dense 111), files read in 0.20 s (3905 MiB/s)
strata generate: expert arena: cudaHostRegister PORTABLE ok; MAP_HUGETLB unavailable (no hugetlb pool configured?); using 4 KB pages
strata generate: loaded 31.64 GiB at 5.98 GiB/s
strata generate: profile /data/strata-bench/src/strata-8fc40dde/data/expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: experimental native Q5_K head, 437043200 bytes
strata generate: expert cache auto: 18.19 GiB free, 700 MiB reserved -> 13595 slots
strata generate: expert cache 13595 slots, 17.50 GiB of VRAM; policy is
strata generate: the GPU computes the experts in the cache; it rounds differently from the CPU,
                 so a reply can differ slightly from a run without the cache (same quality:
                 bench/results/2026-09-27-cache-parity).
                 PROFILE, ranked by routing frequency, no eviction.
strata generate: pre-filled 13595 of 13595 slots from the profile; slot 0 verified
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: 23 expert-pool workers + the host thread
strata generate: session is up (engine 0.1.25)
strata generate: token graph hit path: 13595 resident experts, decided on the device
strata serve: prompt chunk auto: 8192 tokens
strata serve: the prompt path borrows 3171 cache slots (4.08 GiB)
strata verify: window up to 6 tokens, 55.2 MiB of device buffers
strata mtp: draft head over 40525 tokens (68.0 MiB)
strata serve: 261 MiB of VRAM free with everything loaded
strata verify: captured the 6-token window (upload no error, sync no error)
strata verify: captured the 1-token window (upload no error, sync no error)
strata verify: captured the 4-token window (upload no error, sync no error)
strata verify: captured the 5-token window (upload no error, sync no error)
strata verify: captured the 2-token window (upload no error, sync no error)
strata verify: captured the 3-token window (upload no error, sync no error)
strata serve: prompt 4038 tokens = 0 reused + 4038 read in 1966 ms (2054.2 tok/s), 256 generated in 2217 ms (115.5 tok/s), drafts accepted 159 of 210, 1 checkpoints
strata serve: decode expert cache hit rate: 95.4% (138234 hits / 144850 lookups)
strata serve: KV streaming: 99.34% of 1941012 block reads hit VRAM, 51.5 MiB read from RAM
strata serve: suffix drafts: 3 windows, 8 of 10 drafts accepted
strata serve: prompt 4038 tokens = 0 reused + 4038 read in 1715 ms (2354.1 tok/s), 256 generated in 1951 ms (131.2 tok/s), drafts accepted 158 of 205, 1 checkpoints
strata serve: decode expert cache hit rate: 98.3% (141814 hits / 144301 lookups)
strata serve: KV streaming: 99.33% of 1904052 block reads hit VRAM, 51.6 MiB read from RAM
strata serve: suffix drafts: 5 windows, 10 of 17 drafts accepted
strata serve: prompt 4038 tokens = 0 reused + 4038 read in 1708 ms (2363.6 tok/s), 256 generated in 2034 ms (125.8 tok/s), drafts accepted 154 of 226, 1 checkpoints
strata serve: decode expert cache hit rate: 98.5% (154104 hits / 156468 lookups)
strata serve: KV streaming: 99.38% of 2058084 block reads hit VRAM, 51.5 MiB read from RAM
strata serve: suffix drafts: 3 windows, 7 of 13 drafts accepted
strata serve: prompt 32038 tokens = 0 reused + 32038 read in 11419 ms (2805.7 tok/s), 256 generated in 2124 ms (120.5 tok/s), drafts accepted 156 of 222, 2 checkpoints
strata serve: decode expert cache hit rate: 98.2% (150513 hits / 153331 lookups)
strata serve: KV streaming: 97.48% of 2021148 block reads hit VRAM, 205.3 MiB read from RAM
strata serve: suffix drafts: 1 windows, 2 of 3 drafts accepted
strata serve: prompt 32038 tokens = 0 reused + 32038 read in 11367 ms (2818.4 tok/s), 256 generated in 2132 ms (120.1 tok/s), drafts accepted 154 of 212, 2 checkpoints
strata serve: decode expert cache hit rate: 98.5% (147592 hits / 149861 lookups)
strata serve: KV streaming: 97.22% of 1971840 block reads hit VRAM, 220.5 MiB read from RAM
strata serve: prompt 32038 tokens = 0 reused + 32038 read in 11349 ms (2823.0 tok/s), 256 generated in 2021 ms (126.7 tok/s), drafts accepted 162 of 211, 2 checkpoints
strata serve: decode expert cache hit rate: 98.9% (144811 hits / 146376 lookups)
strata serve: KV streaming: 97.46% of 1922544 block reads hit VRAM, 197.0 MiB read from RAM
strata serve: suffix drafts: 1 windows, 2 of 5 drafts accepted
strata serve: prompt 128037 tokens = 0 reused + 128037 read in 44834 ms (2855.8 tok/s), 256 generated in 2408 ms (106.3 tok/s), drafts accepted 153 of 231, 6 checkpoints
strata serve: decode expert cache hit rate: 98.8% (159092 hits / 161058 lookups)
strata serve: KV streaming: 94.90% of 2113560 block reads hit VRAM, 434.3 MiB read from RAM
strata serve: prompt 128037 tokens = 0 reused + 128037 read in 44949 ms (2848.5 tok/s), 256 generated in 2276 ms (112.5 tok/s), drafts accepted 150 of 211, 6 checkpoints
strata serve: decode expert cache hit rate: 99.1% (151357 hits / 152738 lookups)
strata serve: KV streaming: 94.49% of 2002644 block reads hit VRAM, 444.4 MiB read from RAM
strata serve: prompt 128037 tokens = 0 reused + 128037 read in 45120 ms (2837.7 tok/s), 256 generated in 2321 ms (110.3 tok/s), drafts accepted 148 of 214, 6 checkpoints
strata serve: decode expert cache hit rate: 99.2% (153493 hits / 154723 lookups)
strata serve: KV streaming: 94.49% of 2027352 block reads hit VRAM, 449.7 MiB read from RAM
strata serve: suffix drafts: 1 windows, 1 of 2 drafts accepted