feat(inkling): TOPP — adaptive routed-expert trimming, and report what it trims

[JustVugg]

colibri.c has had TOPP and kimi_k3.c has K3_TOPP (#676). inkling.c was the only engine without it — and it is the engine where it matters most.

Why here

Inkling on a small host is disk-bound, not compute-bound: topk=6 experts × ~28 MB × 66 layers is ~11 GB read per token, and a measured run spent 92% of its time in expert I/O (fill 1533 s vs expert-mm 59 s). Every expert TOPP drops is an expert not read.

What it does

Same semantics as the other two engines: sort the routed (weight, id) pairs descending, keep them until the cumulative weight crosses p, drop the tail, and do not renormalise — the dropped weight simply does not contribute.

One detail worth noting: si[] is ordered by sigmoid(logit) + bias while the weight is sigmoid(logit) without the bias, so the weights are not guaranteed descending. The pair is re-sorted before the cut, exactly as colibri.c does.

It reports its own effect

This is a quality lever, not a free speed-up — it changes the computation. So it is opt-in, off by default, announced at startup, and (new here, not in the other engines) it prints what it actually trimmed:

[topp] 0.30: 36/72 routed used (50.0% trimmed, 1.00 experts/token avg)

Measured on the tiny fixture at TOPP=0.30: 50% of routed experts trimmed, expert loads 47 → 27 (−43%). That's the mechanism the disk-bound host needs — and a number you can A/B rather than a claim you have to believe.

Testing

Check	Result
Oracle, TOPP off, cap=4	36/36 teacher-forced, 24/24 tokens
Oracle, TOPP off, cap=1	24/24 tokens
TOPP=0.30 trimming	50% trimmed, loads −43%
TOPP=1.5 (invalid)	refused with a message, not silently clamped
Builds	inkling, glm, olmoe, kimi_k3 all clean

TOPP=0 (default) leaves the computation bit-identical. Nobody's existing runs change.

Honest note

I have not yet A/B'd output quality on the real 975B — the tiny fixture is random-init, so its argmax happens not to flip when the tail is dropped, which proves the plumbing but says nothing about quality on a trained model. The docs label TOPP accordingly, and the per-run report is there so the trade can be measured on a real host instead of assumed. Same caution the GLM side takes with the TOPP=0.75 numbers in #693.