perf(moe): add frozen grouped GEMM for adapter training - #3

Merged

taufeeque9 merged 11 commits into

tf-at/torch-gmm-full-recompute

Aug 29, 2026

Commits on Aug 26, 2026

perf(moe): add frozen torch grouped GEMM backend
taufeeque9
committed
5916f5c
test(moe): cover zero-token grouped expert splits
taufeeque9
committed
ca36c35
fix(moe): validate every grouped expert weight view
taufeeque9
committed
ede5504
fix(recompute): keep chunk outputs differentiable with a frozen embedding
taufeeque9
committed
4947c1a
style(moe): format grouped GEMM tests
taufeeque9
committed
b992f28
build: record Megatron-LM fork base
taufeeque9
committed
97f7d41
test(recompute): cover frozen-input adapter gradients
taufeeque9
committed
96ecb4b
test(moe): simplify no-token setup
taufeeque9
committed
dcee153
fix(moe): reject TE op fuser with torch backend
taufeeque9
committed
8447c0d

Commits on Aug 29, 2026

chore: merge farai/main into torch grouped GEMM
taufeeque9
committed
73e3b36
fix(moe): harden torch grouped expert backend
taufeeque9
committed
c03fbc4