[FEAT][kernels] Fused linear log-prob without materializing logits — SM90 TMA + tensor-core forward, chunked backward#122

Open

KJLdefeated wants to merge 7 commits into

RL-Align:mainfrom

KJLdefeated:feat/fused-linear-logp

Commits on Jun 14, 2026

[FEAT][kernels]: add fused linear log-prob operator (Triton + chunked backward + SM90 TMA/WGMMA)

KJLdefeated
and
claude
committed

Commits on Jun 16, 2026

(bug) add target/bias validation
KJLdefeated
committed

Commits on Jun 17, 2026

check lm_head_weight and hidden device
KJLdefeated
committed