strata generate: CUDA1 context ready, 7.51 GiB free before expert arena registration
strata generate: profile <path>/expert-profile.bin: 24576 ranked pairs, built for 24576 slots
strata generate: expert arena read through the file cache (buffered (not Windows))
strata generate: expert arena: cudaHostRegister PORTABLE ok; MAP_HUGETLB unavailable (needed 23983 2 MiB pages, vm.nr_hugepages=0); transparent huge pages requested (MADV_HUGEPAGE)
strata generate: expert cache auto: 9.48 GiB free, 713 MiB reserved (+86 MiB for the draft head) -> 3508 slots
strata generate: expert cache 4560 slots, 8.70 GiB of VRAM; policy is
strata generate: pre-filled 4560 of 4560 slots from the profile in 0.2 s (42597 MB/s); slot 0 verified
strata generate: CUDA1: 3300 additional experts, 6.21 GiB; results return through pinned host rows
strata generate: R4 hit path ON - resident experts are computed on the GPU
strata generate: token graph hit path: 4560 resident experts, decided on the device
prefill: the CPU share is ON by default for prompt chunks below 1024 tokens (the idle CPU takes some of the experts the GPU would stream; answers can differ slightly from 0.1.40.3, mean KL ~0.004). STRATA_PREFILL_CPU_SHARE=0 turns it off.
strata serve: the prompt path borrows 3232 CUDA0 cache slots (6.13 GiB)
strata serve: 541 MiB of VRAM free with everything loaded
strata verify: capturing the 4-token window (517 MiB of VRAM free)
strata verify: capturing the 1-token window (511 MiB of VRAM free)
strata verify: capturing the 6-token window (503 MiB of VRAM free)
strata verify: capturing the 2-token window (497 MiB of VRAM free)
strata verify: capturing the 3-token window (491 MiB of VRAM free)
