Gnostral — MoE inference under GPU memory pressure

Repeated bounded local experiment, no verified external reproduction. Neither production qualification nor a GPU supervisory authority.

Q-010: avoid repeated Q2_K/Q3_K CPU expert copies

Same Qwen3-30B-A3B Q2_K on an RTX 3070 8 GiB, three server launches per arm, naturally complete answers.

Median	Original Strata	Q-010 candidate
Six-check response decoding	9.51 tok/s	20.18 tok/s
Total GPU card memory peak	2,344 MiB	2,351 MiB
Process-tree RSS	22,094 MiB	22,172 MiB

Observed median throughput ratio is 2.12×, with six byte-identical paired answers. However, the first trial failed its GPU ambient reclaim gate (+196 MiB for a 128 MiB limit). A second independent run with unchanged thresholds passed. The Rust unsafe borrowed-weight projection has not undergone a complete memory-safety audit. No superiority over llama.cpp is established.

Q-011: GPU-aware placement at load time

A separate 2,048 MiB CUDA memory reservation, without sustained competing GPU computation, changed placement for the same 48-layer model:

Three starts	Control	GPU memory reserved
GPU layers	22, 21, 21	13, 12, 13
Median six-check decoding	20.88 tok/s	20.15 tok/s

Startup placement adaptation is observed, not live remapping of loaded layers. PagedAttention was disabled for both mixed CPU/GPU configurations.

Method, limitations and falsifiability

GPU memory is whole-card usage including desktop workloads; RSS is not private PSS. Three runs per condition cannot establish general performance. Q-011 text varied across starts: task structure was checked, not general answer quality. Raw logs and model weights are not distributed with this page; a private-data SHA alone does not establish public reproducibility.

Sanitized public observations · Research registry

Statuses: external replication PENDING; mixed PagedAttention NOT_QUALIFIED; production authorization DENIED.