feat: stream a model's weights uncached and NZ-ordered - #2843

Closed

zhangqi-chen wants to merge 13 commits into

mainfrom

feat/l2-bypass-device-offset

Commits on Sep 20, 2026

chore(runtime): pin the runtime submodule to simpler fe8b3389
zhangqi-chen
committed
4ecc680
build(toolchain): bump PTOAS to v0.64
zhangqi-chen
committed
be91aaf
feat(codegen): issue a bypassing load against the device's uncached alias
zhangqi-chen
committed
c0d3828
fix(codegen): follow PTOAS v0.64's scalar pointer load and store rename
zhangqi-chen
committed
b289cd0

Commits on Sep 21, 2026

feat(ir): reach one layer of a stacked NZ weight through a slice
zhangqi-chen
committed
3c764a1
feat(runtime): compare a device tensor against an NZ parameter's blocked shape
zhangqi-chen
committed
13c4f3f
feat(verifier): let a device dispatch hand an ND buffer to an NZ parameter
zhangqi-chen
committed
e1abea1
feat(ir): prove a split-K remainder offset non-negative
zhangqi-chen
committed
325f813
feat(ir): flatten a whole NZ tensor so it can still be prefetched
zhangqi-chen
committed
ddf0cee
fix(codegen): restate an NZ argument in blocked terms at the entry
zhangqi-chen
committed
e683459
docs(passes): record the flatten rule and the division sign proofs
zhangqi-chen
committed
d8b8ffa
fix(ir): prove a remainder from its dividend, and exempt only NZ at a dispatch
zhangqi-chen
committed
0e9d722
feat(ir): block a ragged NZ tail and prove a strided loop's facts from its start
zhangqi-chen
committed
0c1823f