mirror of
https://github.com/Helldez/BigMoeOnEdge.git
synced 2026-10-03 03:25:42 +00:00
* Mark the two doc-invoked host scripts executable README, AGENTS, CONTRIBUTING and docs/seam.md run scripts/build-host.sh directly, and docs/benchmarks.md runs scripts/bench-run.sh the same way, but both are committed 100644, so a fresh POSIX clone answers permission denied at the first documented step. Windows trees never see the mode bit, which is likely how it went unnoticed. * Mark the remaining host scripts executable too Per the discussion in #163: leaving only the two doc-invoked scripts executable hides the next instance of the same failure behind an inconsistent set, so all five scripts/*.sh get the bit.
24 lines
1.1 KiB
Bash
Executable file
24 lines
1.1 KiB
Bash
Executable file
#!/system/bin/sh
|
|
# Pure-mmap baseline for gpt-oss: NO expert streaming (no --moe-stream, no cache/lanes/
|
|
# O_DIRECT/overlap/prefetch) -- llama.cpp mmaps the whole 58 GB model and the OS pages it
|
|
# in/out of the page cache on demand. This is the >RAM baseline the streaming modes beat.
|
|
# top-k override still applies (it is a load-time kv_override, valid without streaming).
|
|
# Appends to the same log as the streaming matrix. k=4 then k=2, per request.
|
|
#
|
|
# Overridable from the environment, and M/OUT must match the streaming matrix run this is the
|
|
# baseline for: OUT is appended to, not replaced.
|
|
M=${M:-/data/local/tmp/bmoe/gpt-oss-120b-Q4_K_M.gguf}
|
|
P="What is 17 times 23? Then name the capital of Australia."
|
|
OUT=${OUT:-/data/local/tmp/gptoss-matrix.out}
|
|
COOLDOWN=${COOLDOWN:-20}
|
|
cd /data/local/tmp || exit 1
|
|
for K in 4 2; do
|
|
TAG="mmap_k${K}"
|
|
echo "==================== $TAG ====================" >> "$OUT"
|
|
sleep "$COOLDOWN"
|
|
LD_LIBRARY_PATH=/data/local/tmp ./bmoe-cli -m "$M" --chatml --no-think -c 2048 -n 24 \
|
|
--n-expert-used "$K" --csv /data/local/tmp/gptoss_${TAG}.csv \
|
|
-p "$P" >> "$OUT" 2>/dev/null
|
|
echo "" >> "$OUT"
|
|
done
|
|
echo "MMAP DONE" >> "$OUT"
|