feat: native MTP speculative decoding for Qwen3.5#990

Open

AirRunner wants to merge 10 commits intoml-explore:mainfrom

AirRunner:feat/mtp-native

Commits on Apr 1, 2026

feat: native MTP speculative decoding for Qwen3.5
AirRunner
committed
fix(mtp): eliminate SSM state contamination on draft rejection
AirRunner
committed
fix(mtp): server integration (yield types, cache fallback, batching)
AirRunner
committed
fix(mtp): address @janhilgard code review feedback (double-norm, quant_predicate)
AirRunner
committed
feat(mtp): add --mtp CLI flag for generate and server
AirRunner
committed
test(mtp): add unit tests for MTP speculative decoding
AirRunner
committed
fix(mtp): warn when --mtp flag is used with a model without MTP head
AirRunner
committed
style: apply black and isort formatting
AirRunner
committed
fix(mtp): stack per-expert MTP weights for MoE models in sanitize()

AirRunner
and
Thump604
committed
fix(mtp): raise clear error when config has MTP but weights do not
AirRunner
committed