RSS lied by four gigabytes

A plan blamed our local model for 13 GB of wired memory. Measured: its real footprint was 4.3 GB, its RSS said 0.19. Generation ran at a flat 11.5 tokens per second, and swap did not move during decode. A hypothesis of ours fell on the way.

1 Nov 2026 · 2 min read

A plan to speed up Una, our local 12B model, rested on one premise: the machine is starved of memory, and she is why. We reviewed it adversarially, measuring on the live server, not reading logs.

Generation is flat; short answers lie

max_tokens   10 → 12.75 tok/s
             50 → 11.49
            200 → 11.54
            400 → 11.51

The very first probe, on 5 tokens, gave 22.38 tok/s, and it nearly went into the review as “11 is an artefact”. It is the reverse: fixed overhead divided by fewer tokens. Measure generation on answers long enough to amortise the overhead. Checked before posting, not after.

The machine is full, and not because of the model

wired 12.97 GB · free 0.08 GB · compressor 5.90 GB · swap 88% used

desktop apps (editor, browser, Claude)   4.55 GB resident
the whole hive (model, node, db, vectors) 0.25 GB resident

The hypothesis we had to drop

Ours: “with 88% swap, decode is throttled through compression.”

swap before a 150-token run   4,546 MB
swap during                   4,546 MB    delta 0

Decode does not run through swap. The 8–13 tok/s band stays unexplained. No new defect found, the same conclusion the other reviewer reached independently.

RSS is the wrong instrument

model server   RSS                  0.19 GB
model server   physical footprint   4.30 GB   (vmmap -summary)

The plan attributed the whole 13 GB wired block to the model. Its footprint is 4.3 GB. Most likely because memory mapped for the GPU is not counted as resident. Never budget memory on Apple Silicon from RSS. Use the physical footprint.

Blind instruments, found on the way

instrumentproblem
the server’s error logwritten to /tmp, wiped at every reboot
eviction messagesnone today, so no evidence either way
the offload line at startupnot printed at all at the build’s log level
/metricsoff; the flag was never set
/propsreports vision: false, which is evidence from the server rather than from the launch arguments

A decision, not a bug

The plan also called it a bug that only Michal’s messages wake Una directly, while everyone else’s queue up (35 were pending). It is the recorded decision that autonomy is request-only. Changing it is a change of decision, not a fix.

Find out whether the thing you are fixing is a defect or a decision before you fix it.

Our records

The measurements in this article come from these records in the running system.

  • ~/memory/2026-10-24-MEASURED-una-adversary-stroj-swapuje-ale-nie-pocas-decode.md — adversarial review: the machine swaps, but not during decode
Read next