row bsw-r-s139-iq3-24g (R-S139-IQ3-24G, ROW_KIND=strata) · started 2026-10-09T09:31:50Z · uid 1000 · python 3.14.4
CUDA_DEVICE_ORDER=PCI_BUS_ID CUDA_VISIBLE_DEVICES=1 TZ=UTC
pins: 52 files read their sha256 again in 2.6 s (the 2 shards were read at the window's open)
the build target: /workshop/card24-strata/build/strata-6f32ec07-sm86/CMakeCache.txt reads CMAKE_CUDA_ARCHITECTURES=86
band of record: predicted 39-149 tok/s (decode at 4K, median of 3; 17.43-55.42 rounds/s x 2.25-2.69 tokens a round; C_round 124.53-149.78; outside 15-39, or 149 up to that run's tpr x 241.35; under the floor below 15; VOID above that run's tpr x 241.35 and above 1,448): AMENDMENT section 3.3
identity before the launch: the client's nine prompt digests equal L-IQ3-t8's (9/9)
port withheld-2026-10-13 reads free on the box (ss -ltn)
the Strata tree: HEAD 6f32ec070f23ced9f50e704d854d775da52591ab and `git status --porcelain --ignored` empty
the engine's build identity: the build tree names the source tree, whose HEAD is the arm's commit (/workshop/card24-strata/build/strata-6f32ec07-sm86/CMakeCache.txt)
the engine line (the laptop's, pinned): --pack /workshop/bench-flashnext/packs-6f32ec07/iq3_s --native /workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf --ple-gguf /workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf --expert-profile /workshop/bench-flashnext/strata-6f32ec07/data/expert-profile.bin --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp /workshop/bench-flashnext/mtp/rt-6f32ec07 --max-context 131072 --kv int8
setup's rule for this box's MemTotal (123.39 GiB) would write: --pack /workshop/bench-flashnext/packs-6f32ec07/iq3_s --native /workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf --ple-gguf /workshop/bench-flashnext/models/IQ3_S/Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf --expert-profile /workshop/bench-flashnext/strata-6f32ec07/data/expert-profile.bin --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp /workshop/bench-flashnext/mtp/rt-6f32ec07 --max-context 131072 --kv int8 --kv-resident 32768  [NOT USED]
engine env {'STRATA_NO_LARGEPAGES': '1'} (the client's allowlist passed); THP 'enabled=madvise defrag=madvise' as staged
server argv: /usr/bin/python3 /workshop/bench-flashnext/strata-6f32ec07/serve/server.py --engine strata --config /workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/strata-config.json --port withheld-2026-10-13
the second fence (the unit's IP rules), inner.py's ipfilter probe: NOT ENFORCED (the connect to the box's own non-loopback address was not dropped: failed (ECONNREFUSED) (a refusal means the packet reached the host's stack))
sandbox: /usr/bin/bwrap --unshare-net --unshare-pid --die-with-parent --ro-bind / / --dev-bind /dev /dev --proc /proc --tmpfs /run --tmpfs /tmp --bind /workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run /workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run --ro-bind /run/nvidia-persistenced /run/nvidia-persistenced -- /usr/bin/python3 /workshop/card24-strata/kit/client/inner.py --spec /workshop/card24-strata/results/R-S139-IQ3-24G/R-S139-IQ3-24G.run/spec.json
sandbox note: /var/run is a link to /run (or absent): the /run tmpfs covers it
sandbox note: /run/nvidia-persistenced bound read-only (the driver's persistence socket)
cards before:
index, pci.bus_id, name, memory.used [MiB], memory.total [MiB], enforced.power.limit [W], temperature.gpu
0, withheld-2026-10-13, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 85276 MiB, 97887 MiB, 420.00 W, 32
1, withheld-2026-10-13, NVIDIA GeForce RTX 3090 Ti, 1 MiB, 24564 MiB, 300.00 W, 38
the pinned card's memory.used before the launch: 1.0 MiB
residency at ready: partial residency (8144 of 24,576 slots; hit rate unread) (failed: expert cache 24576 slots; pre-filled 24576 of 24576 slots; token graph hit path: 24576 resident experts; the verifier's zero-doorbell line (100% VRAM resident); no hit-rate line on any finished request (zero-doorbell: lookups 0); no PCIe share on any finished request (absent or 0); the desktop's smaller card's VRAM growth >= 49,431 MiB (the arena 46.84 GiB + dense 1,538,035,200 B))
server maps: /workshop/bench-flashnext/cuda-13.4/targets/x86_64-linux/lib/libcublas.so.13.8.0.4, /workshop/bench-flashnext/cuda-13.4/targets/x86_64-linux/lib/libcublasLt.so.13.8.0.4, /workshop/bench-flashnext/cuda-13.4/targets/x86_64-linux/lib/libcudart.so.13.4.92, /workshop/card24-strata/build/strata-6f32ec07-sm86/strata, /usr/lib/x86_64-linux-gnu/libcuda.so.595.84
cards after:
index, pci.bus_id, name, memory.used [MiB], memory.total [MiB], enforced.power.limit [W], temperature.gpu
0, withheld-2026-10-13, NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 85276 MiB, 97887 MiB, 420.00 W, 32
1, withheld-2026-10-13, NVIDIA GeForce RTX 3090 Ti, 1 MiB, 24564 MiB, 300.00 W, 56
residency at the end: partial residency (8144 of 24,576 slots; hit rate 0.922) (failed: expert cache 24576 slots; pre-filled 24576 of 24576 slots; token graph hit path: 24576 resident experts; the verifier's zero-doorbell line (100% VRAM resident); no hit-rate line on any finished request (zero-doorbell: lookups 0); no PCIe share on any finished request (absent or 0); the desktop's smaller card's VRAM growth >= 49,431 MiB (the arena 46.84 GiB + dense 1,538,035,200 B))
ended 2026-10-09T09:36:35Z · inner rc 0 · last phase done · identity 9/9 · residency partial residency (8144 of 24,576 slots; hit rate 0.922) · fences held ['netns']