Qwen4exp correctness fixes - #27879

Closed

tarruda wants to merge 10 commits into

ggml-org:masterfrom

tarruda:qwen4exp-correctness-fixes

Commits on Aug 28, 2026

qwen4exp: fix sparse-attention block selection
tarruda
committed
c07204d
qwen4exp: support independent PLE embedding widths
tarruda
committed
fce561d
qwen4exp: validate model metadata
tarruda
committed
14b71e5
qwen4exp: update indexer cache after sequence copies
tarruda
committed
6ef2436
qwen4exp: enable recurrent state rollback
tarruda
committed
edb6dec
qwen4exp: disable tensor split
tarruda
committed
8427625
qwen4exp: fix GDN QK normalization
tarruda
committed
a7fc7e4

Commits on Aug 29, 2026

qwen4exp: avoid CUDA RMSNorm grid limit
tarruda
committed
dfc7932

Commits on Aug 30, 2026

Merge branch 'master' into qwen4exp-correctness-fixes
tarruda
committed
01ac8a2
qwen4exp: share QSA block tables across queries
tarruda
committed
50d6b3d