qwen4exp : port NextN/MTP speculative decoding to master (+50% decode at 70k) - #28104

Closed

JayToltTech wants to merge 6 commits into

ggml-org:masterfrom

JayToltTech:feat/qwen4exp-mtp-upstream

Commits on Aug 31, 2026

qwen4exp: load NextN/MTP draft tensors (sidecar and in-file)

authored and
Jay Beavers
committed
4e6ba97
qwen4exp: NextN/MTP draft graph and context - spec decode works end to end

authored and
Jay Beavers
committed
af89cc2
Reapply "common: use full checkpoints for MTP rollback"

Nathanw1014
authored and
Jay Beavers
committed
a477054
server : do not re-verify replayed draft tokens after a checkpoint restore

Nathanw1014
authored and
Jay Beavers
committed
4b1f6f8
qwen4exp : defer the output-row gather past the h_nextn export
Jay Beavers
committed
77476ce
server : keep speculative checkpoints on device
Jay Beavers
committed
175b66c