experiment: keep-93 = the PURE UNION of the coder-grown and general-grown keep-64 expert sets
requested:  ayourtch 2026-08-16 21:24 CEST via Webex ("build the 87 expert one", coding smoke + opencode)
built:      2026-08-16T19:25:26+00:00
host:       [build host — RTX PRO 6000 workstation, hostname redacted]
source_model: /home/ayourtch/llm/ds4/gguf/DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf
wants:      /home/ayourtch/merged-wants.tsv (same file as merge-128: normalised sum of the two censuses, +1e7 bonus on every union member)
why 93 not 87/88: GGUF needs one expert count for all layers; union size per layer is 78..93 (mean 87.3). keep=93 keeps EVERY union member in EVERY layer; layers whose union is <93 get 0-15 filler by merged rank. keep=88 would drop union members in the larger layers.
keep:       93, used 6
prune_cmd:  PRUNE_WANTS=~/merged-wants.tsv ~/hf-venv/bin/python ~/prune_experts.py SRC ~/grow-union/deepseek4-keep93.gguf 93 6
serve_cmd:  CUDA_VISIBLE_DEVICES=0 ~/llama.cpp/build-mtp/bin/llama-server -m ~/grow-union/deepseek4-keep93.gguf -ngl 99 -c 262144 -fa on