#6248 [Engine] Retracted (reporter config error); kept: per-request strata_tune no-op on 0.1.62, boot-time pcie_frac costs 2-3x on SM86

closed · @NoBirdWasHam · 15 comments · View on GitHub

Models & quantsDocumentation

Description

**FINAL STATE (19 NOV evening). The uptime-regression claim in the original post is retracted in full - follow-ups 1-6 preserve every correction step, nothing deleted. Root cause was on the reporter's side: my A/B harness wrote `pcie_frac: 0.15` into the run-config during an interrupted first run, a second run's backup captured the already-poisoned file (so its hash-exact restore faithfully restored the poison), and a later boot inherited it. What I read as "the engine degrades over uptime" was a healthy engine running a poisoned config (1.8x at 8.9k, 3.3x at 59-61k, instantly).**

**Two findings that do stand and are worth maintainer time:**

1. **Per-request `strata_tune` appears to be a no-op on 0.1.62.** Against a live boot whose config says `pcie_frac: 0.15`: requests carrying `strata_tune {pcie_frac: 0.0}` measured 59k 63,734 ms vs 62,598 ms un-tuned (8.9k 11,947 vs control 11,581); conversely `strata_tune` 0.25/0.50 on a 0.00 boot did not slow anything (6x 59k flat). A client-sent `prefill_pipe: 1` was likewise a no-op (original post). The docs advertise per-request calibration via this channel; a silent no-op silently invalidated a day of my experiments - anyone A/B-ing these knobs in-process is measuring nothing.
2. **Boot-time `--pcie-frac 0.15 is expensive on SM86 + UD-Q4_K_XL + int8 KV**: 8.9k 6.5 to 11.4 s, 59-61k 19 to 62.6 s, verified both directions on a fresh boot (and back to 6.8-6.9 s / 20.2 s after restoring 0.00). @mercant1982's SM80 + IQ3_S control measured config 0 vs 0.5 as identical, so either the cost or the honoring of the value is platform-dependent - worth a doc note at minimum.

Everything else on the battery (decode, draft acceptance, conversation reuse, correctness) behaved identically in all states, consistent with a purely config-level effect. Box verified back at `pcie_frac: 0.00` with fresh-speed cold prefill (follow-up 6). Closing this issue; happy to re-open or hand data to a follow-up for item 1 or 2.

Related on this site

Quick links to related pages.