feat(minimax): support MiniMax-Text-01 and MiniMax-M2 - #583

Open

y258dd wants to merge 8 commits into

InfiniTensor:InfiniLM-v0.2.9cfrom

y258dd:feat/minimax-model-support

Commits on Sep 19, 2026

feat(minimax): support MiniMax-Text-01 (lightning attention + MoE)
y258dd
committed
e737a25
fix(minimax): correct InfiniOps lightning attention test (disjoint state rows, dtype-aware reference)
y258dd
committed
a0d84a5
test(minimax): allow CUDA and multi-expert validation
y258dd
committed
c739a26
fix(engine): copy input offsets to host before sampling
y258dd
committed
524489d
test(minimax): use disjoint state rows on CUDA
y258dd
committed
5752601
test(minimax): support fp16 MoE validation
y258dd
committed
8aaec91
fix(minimax): map HuggingFace MoE expert weights
y258dd
committed
3ad578c
fix(minimax): use model dtype for lightning state
y258dd
committed
e23eba7