Repository navigation
P7/P8: device placement and graph-level memory planning #1131
Description
Activity
- added sub-issues
on Aug 26, 2026 All three research questions answered and the eager-scope implementation landed — see SKEEP-003a (docs/modules/skeep/pages/003a-placement-and-planning-resolution.adoc, #1149) for the full record.
P7 — placement (#1133): resolver-owned.
AllocationResolverdecides from (resolved form × profile × platform); theMemoryPlan.kthardcode is gone; the loader consults the resolution (ResolvedGguf); the user outranks the resolver through a documented precedence chain;explain()makes every decision printable pre-load. The annotations and the dead planner that pretended to own this are deleted.Eager lifetime (#1135): the Scope split is the mechanism;
memoryScopeis wired into creation, opt-in, with steady-state zero-allocation decode pinned by test. Adoption of op outputs: #1146.P8 — graph planning (#1134): would duplicate IREE; core decides and carries; carriage (#1147) gated on the
.vmfbparity harness (#1148).Follow-ups #1146/#1147/#1148 carry the remaining engineering; none of them blocks this umbrella. Legacy loader axes removal (#1159) is in flight as the first slice of façade removal.
Split out of #932's closing summary, where this was recorded as post-M2 but never given an issue.
The SKEEP-003 proposal's phases 7 and 8 were always scoped after M2, and the storage model was built so they could land without another rewrite:
PlacementandMemoryDomainexist andAllocationSpeccarries them, but nothing chooses. A tensor's home is decided by whoever constructs it, not by a policy that can see the whole model.MemoryPlanis arithmetic over a tensor list; it does not know the shape of the computation, so it cannot reuse a buffer whose live range has ended or size the forward slab from actual liveness rather than a worst-case formula.Why this is a research issue, not a coding one
Both need a decision about where the information comes from. The eager path has no graph to analyse — by construction, it executes as it is called. The recorded path (
RecordingExecution→HloGenerator→ StableHLO → IREE) does, and IREE already performs its own allocation planning, so P8 may be less "build a planner" than "stop duplicating one".Worth establishing before designing: what the recorded path already gives us, and whether P7/P8 are wanted for the eager path at all.
Depends on
Nothing blocking. The storage model (#932) landed the types both phases need.