refactor(moe): extract hybrid decode orchestration - #491

Draft

zihaomu wants to merge 11 commits into

FlashML-org:mainfrom

zihaomu:proposal/hybrid-decode-executor-v1

Commits on Sep 21, 2026

feat: add simple ROCm GPU support for RDNA3 (gfx1100-1103)

bouclem
authored and
zihaomu
committed
7061350
fix(rocm): fall back for single-bank fast index copy

nekomario28
authored and
zihaomu
committed
66b370c
test(rocm): cover fast index copy fallback

nekomario28
authored and
zihaomu
committed
e278c5d
test(rocm): use explicit fast copy module import

nekomario28
authored and
zihaomu
committed
61931ef
fix(rocm): complete RDNA3 runtime path

nekomario28
authored and
zihaomu
committed
f95b4a2
fix(rocm): preserve CPU MoE implementation

nekomario28
authored and
zihaomu
committed
63ca97c
build(rocm): support RDNA4 discovery and modular SDKs
zihaomu
committed
8886b9e
fix(rocm): complete modular SDK runtime support
zihaomu
committed
6c7ec92
test(rocm): harden foundation platform contracts
zihaomu
committed
a060f8f
fix(rocm): make CPU MoE graph replay safe
zihaomu
committed
b395144
refactor(moe): extract hybrid decode orchestration
zihaomu
committed
00e4bb1