MTP for Qwen3.8-Flash-Next - #144

Open

danielhanchen wants to merge 8 commits into

base/upstream-de8656bd9from

mtp/qwen4exp-nextn

Commits on Sep 2, 2026

gguf-py : register the qwen4exp NextN tensors

ryanmonsurate
authored and
danielhanchen
committed
518efa2
model : add the qwen4exp NextN/MTP draft head

ryanmonsurate
authored and
danielhanchen
committed
e52b4b0
convert : export the qwen4exp NextN/MTP draft head

ryanmonsurate
authored and
danielhanchen
committed
2b984b9
llama: let an MTP draft borrow the target's embeddings and lm head
danielhanchen
committed
78ede59
qwen4exp: allow loading a draft-only MTP export
danielhanchen
committed
b0653f3
ggml-cuda: key the CUDA graph cache by shape
danielhanchen
committed
5a08a71
qwen4exp mtp: trim comments
danielhanchen
committed
b761996

Commits on Sep 3, 2026

Merge b10786: n_ff_exp is per-layer now
danielhanchen
committed
a9e9c3c