Four fixes Qwen3.8-Flash-Next needs to work - #362

Merged

TheTom merged 5 commits into

TheTom:feature/turboquant-kv-cachefrom

jasstrong:pr/flash-next-fixes

Sep 19, 2026

Commits on Sep 7, 2026

common: offer the device pinned host buffer to --override-tensor

jas
and
claude
committed
a6b74b1
quantize: keep qwen4exp's structural tensors out of the quantizer

jas
and
claude
committed
518f437
qwen4exp: let a draft-only MTP export use its own output mixer

jas
and
claude
committed
82016d2
speculative-simple: drive the draft through common_speculative_process

jas
and
claude
committed
61e7d6f
qwen4exp: set an 8-bit floor on the hyper-connection injection matrices

jasstrong
and
claude
committed
5edc0f8