Enable Triton MXFP4 MoE on gfx950 for GPT-OSS#220

Merged

ChuanLi1101 merged 6 commits intomainfrom

gpt-oss-triton-moe-gfx950

Mar 5, 2026

Commits on Mar 5, 2026

Enable AllReduce+RMSNorm fusion for GPT-OSS model

ChuanLi1101
and
cursoragent
committed
Move MXFP4 padding from layernorm to MLPBlock, enable full AllReduce+RMSNorm fusion

ChuanLi1101
and
cursoragent
committed
Include compute_logits in HIP Graph for decode

ChuanLi1101
and
cursoragent
committed
Guard compute_logits in HIP Graph for TP=1 only

ChuanLi1101
and
cursoragent
committed
Fix Black formatting for compute_logits guard

ChuanLi1101
and
cursoragent
committed
Enable Triton MXFP4 MoE on gfx950 when ATOM_USE_TRITON_GEMM=1

ChuanLi1101
and
cursoragent
committed