[cuda backend] share AOTI weights by FQN across methods - #21823

Merged

Gasoonjia merged 12 commits into

mainfrom

codex/cuda-fqn-weight-sharing

Aug 26, 2026

Commits on Aug 25, 2026

cuda: share AOTI weights by FQN across methods
Gasoonjia
committed
1d6744e
Address FQN weight sharing review feedback
Gasoonjia
committed
6cf890d
cuda: keep leaked AOTI libraries loaded
Gasoonjia
committed
02304be
cuda: preserve FQN storage semantics
Gasoonjia
committed
fb5f24b
cuda: apply Python formatting
Gasoonjia
committed
89de71f
cuda: centralize multimethod FQN weights
Gasoonjia
committed
38b0c99
cuda: fold weight metadata into cache
Gasoonjia
committed
27ae1e8
cuda: simplify weight collector finalization
Gasoonjia
committed
29c08b5
cuda: format backend source list
Gasoonjia
committed
29939a0
cuda: run weight cache test in CI
Gasoonjia
committed
905edad
cuda: allow per-method views of shared weights
Gasoonjia
committed
915db81
cuda: scope AOTI tensor constants per library
Gasoonjia
committed
fa7c98a