Repository navigation
P7: decide who chooses a tensor's device placement #1133
Description
Activity
- added sub-issues
on Aug 26, 2026 Resolved: a resolver owns placement —
AllocationResolver.resolve(weight, profile, platform), a pure function of what will be held (the resolvedWeightForm) × what the profile says (domainFor) × what the platform can do (StorageCapabilities), mirroring the patternWeightFormResolverproved. Consumers carry and obey: the plan (PlanTensor.allocation), the loader (ResolvedGguf). No graph needed — weights are per-tensor decisions, activations are a lifetime question (#1135), and with only CPU backends a device-arbitration policy would be untestable speculation.The user always wins: per-tensor
weightFormFor> uniformweightForm> resolver;explain()renders every decision with its reason before a byte of payload is read.Landed via #1142 (PR #1152 — the dead machinery deleted), #1143 + #1144 (PRs #1153/#1154, re-landed as #1155). Written up in SKEEP-003a (#1149, PR #1158).
Split out of #1131.
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.Research question
Where should placement decisions come from: a per-model policy, the
PlannerProfile, or something derived from the graph? Does this need a graph to reason about at all, or can it work over the eager construction order?Depends on
Nothing blocking — the storage model (#932) landed the types this needs. Related to #1131's open question of whether P7 is wanted for the eager path at all, or only for the recorded path where IREE already does allocation planning.