perf(glm52): MLA decode arena + CUDA graph capture on top of #535 (−36%/layer) + kernel bench#533

Closed

n-WN wants to merge 8 commits into

openinfer-project:mainfrom

n-WN:feat/glm52-kernel-bench

Commits on Jul 3, 2026

WIP glm52 kernel bench
n-WN
committed
docs(lessons): megakernel decode-latency research — what transfers here
n-WN
committed
WIP glm52 scratch forward + A/B bench
n-WN
committed
feat(glm52): zero-alloc scratch MLA decode forward + A/B parity bench
n-WN
committed
style(glm52): rustfmt + review fixups (unused import, doc/attr adjacency)
n-WN
committed
test(glm52): oracle gate replays the scratch forward and asserts bitwise parity
n-WN
committed
perf(glm52): step-0 decode kernel optimizations — arena + schedule hoist + CUDA graph (-42%/layer)
n-WN
committed
fix(glm52): the decode scratch owns its FlashMLA contract (review round)
n-WN
committed