power: zram swap ahead of disk swapfile for large model loads
This commit is contained in:
parent
22746bcc25
commit
2eb1008ec0
1 changed files with 11 additions and 0 deletions
|
|
@ -15,6 +15,17 @@
|
|||
# suspend-then-hibernate so long sleeps drain to disk.
|
||||
boot.resumeDevice = "/dev/mapper/luks-71c97ce7-dd29-4b07-8d2a-8cc3985a65bd";
|
||||
|
||||
# zram ahead of the disk swapfile: when a large LLM load pins 20-30 GiB of
|
||||
# GTT, evicted app pages compress into RAM instead of contending with model
|
||||
# reads on the same NVMe (this contention caused 8-minute model loads).
|
||||
# Hibernate is unaffected: the image goes to resumeDevice/resume_offset
|
||||
# (the swapfile), which stays available as overflow swap at priority -2.
|
||||
zramSwap = {
|
||||
enable = true;
|
||||
memoryPercent = 50;
|
||||
priority = 100;
|
||||
};
|
||||
|
||||
# On battery: suspend-then-hibernate (s2idle then hibernate after
|
||||
# HibernateDelaySec). Previously this looped suspend->wake->suspend due
|
||||
# to ACPI EC interrupts, but acpi.ec_no_wakeup=1 (above) plus disabling
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue