Mhc expand - #5014

Draft

muskansh-google wants to merge 6 commits into

AI-Hypercomputer:mainfrom

muskansh-google:mhc_expand

Commits on Aug 25, 2026

Support Qwen3-Next 80B with GDNv3, fine-grained remat, MoE routing, Muon sharding, and mHC-lite Pallas/Mosaic kernels on TPU v6e-256.
darisoy
committed
5ca753f
Wire sharding-aware Muon optimizer and NamedSharding in MaxText.
darisoy
committed
e0b495d
Add run_custom_qwen3_next_on_xpk.sh runner script matching qn80b-v6e256-borg.sh flags
darisoy
committed
36769e9

Commits on Aug 26, 2026

Call mhc_expand & mhc_reduce once for all layers
muskansh-google
committed
c4ad690

Commits on Aug 28, 2026

comment out adding random bias
muskansh-google
committed
2c1fb48
pdbs=6 10% MFU
muskansh-google
committed
5a66d4b