arch=deepseek4 experts=256 -> keep 93, used 6 wants census for 40 layers from /home/ayourtch/merged-wants.tsv ranking: measured-counts | full_router: False layers with routed experts: 43 (hash-routed: ['0', '1', '2']) bytes in 145.3 GiB -> out 57.9 GiB; writing... Writing: 0%| | 0.00/62.2G [00:00