Repository navigation

feat(offload): memory.offload_exec=passback — scheduled co-inference of the spilled layers - #809

Open

aldrouil wants to merge 37 commits into

warpfront:masterfrom

aldrouil:feature/cpu-offload-pcie-passback

Commits on Oct 1, 2026

feat(offload): CPU-exec GPU-idle accounting, and the pass-back headroom record
Avery Drouillard
committed
eb62684
fix(docs)
Avery Drouillard
committed
c696b30
feat(offload): memory.offload_exec=passback — split a spilled step across both engines
Avery Drouillard
committed
965c9a4
docs(offload): pass-back mode — design § 6.2.2, config/env tables, TUI row, AGENTS, changelog
Avery Drouillard
committed
f907988
fix(offload): time the pass-back probe with a real sync, and read the join against its own floor
Avery Drouillard
committed
80bb649
docs(perf): pass-back split checkpoint — gfx1201, 9B MQ4, budget 24 (8 of 32 spilled)
Avery Drouillard
committed
02a324a
feat(offload): put the scheduler's no-wait floor on the split trace line
Avery Drouillard
committed
619116a
test(cli): cover offload-bench's refusal predicate and share recommendation
Avery Drouillard
committed
42c5430
docs(perf): offload-amount sweep chart, 27B comparison point, in-place record amendment
Avery Drouillard
committed
e728392
docs(perf): chart labels that cannot collide — values in the legend, gains on their own axis
Avery Drouillard
committed
ac01393
docs(perf): fix the amended record's stale ranges, its dangling section, and sharpen the 27B claim
Avery Drouillard
committed
62e172a
docs(perf): drop the long-horizon arc from the record, keep what the attempt taught
Avery Drouillard
committed
534aff7

Commits on Oct 2, 2026

fix(dispatch): censored, probe-anchored estimator for the passback share
Avery Drouillard
committed
8ee62f5
perf(dispatch): re-estimate the passback share every 8 steps, not 64
Avery Drouillard
committed
308f8a9
docs: describe pass-back as scheduled co-inference of the spilled layers
Avery Drouillard
committed
99d8ca5
revert(dispatch): the 8-step estimator cadence is a regression
Avery Drouillard
committed
7262136
perf(dispatch): co-infer the dense FFN down-projection in passback
Avery Drouillard
committed
fc5c072
revert(dispatch): the censored estimator is a regression; keep the shipped one
Avery Drouillard
committed
8f656c4
docs: correct the FFN-down gain to the shipped-estimator measurement
Avery Drouillard
committed
01f0457
perf(dispatch): let the passback latch re-open when the balance point moves
Avery Drouillard
committed
9c72282
docs: add the passback perf-gap and scheduling-algorithm investigations
Avery Drouillard
committed
f97f836
perf(dispatch): tune the passback latch's re-open policy
Avery Drouillard
committed
ce8f3e5
docs(perf): the passback coverage term is 4-6% of wall, not the deficit
Avery Drouillard
committed
00bb0fe
docs(perf): record the passback latch band A/B as the null it is
Avery Drouillard
committed
0cc2cb5
docs(perf): amendments to the two 2026-10-02 passback checkpoints
Avery Drouillard
committed
d85faa9
docs(perf): correct the oracle bound direction (amendments 1 -> 2)
Avery Drouillard
committed
6a082f6
docs(perf): amendment 3 — fix the amendment-2 self-contradiction, close the chain
Avery Drouillard
committed
7443401
docs(perf): current-HEAD passback mode A/B; make the changelog claim traceable
Avery Drouillard
committed
a7e3e1c
docs(perf): pass-back cross-model spread (2B/9B/27B); pass-back loses to pcie on a small model
Avery Drouillard
committed
289b38a
docs(perf): quant spread, pass-back nondeterminism, mq6 blocker, model-spread amendment
Avery Drouillard
committed
b0b4141
docs(perf): correct determinism framing, 27B point, mq6 root cause
Avery Drouillard
committed
2b1d60d
docs(perf): quant comparison figure for the pass-back quant spread
Avery Drouillard
committed
c32e41c
docs(perf): scatter figure for the pass-back spreads (PR evidence)
Avery Drouillard
committed
d76f6be
fix(offload): route the FFN down-projection through the step seam; one exec-mode snapshot
Avery Drouillard
committed
059f879
fix(rdna-compute): serialize JIT compiles of one cache key
Avery Drouillard
committed
196feda
chore(crate-maps): refresh the generated blocks this branch drifted
Avery Drouillard
committed
a39c7d4
fix(hipfire-xdna): allow clippy::mut_from_ref on the documented BO view
Avery Drouillard
committed
9599bd7