feat: Qwen3.5 / Qwen3.5-MoE MTP speculative decoding#1330

Closed

sufubao wants to merge 34 commits into

ModelTC:mainfrom

sufubao:qw35_mtp

Commits on Jun 5, 2026

feat(qwen3_5_mtp): split linear-attn cache state for spec-decode verify
sufubao
committed
feat(qwen3_5_mtp): qwen3next GDN spec-decode verify path
sufubao
committed
feat(qwen3_5_mtp): basemodel MTP decode CUDA graphs + verify dispatch
sufubao
committed
feat(qwen3_5_mtp): scheduler MTP verify backend + accept-len transport
sufubao
committed
feat(qwen3_5_mtp): Qwen3.5 / Qwen3.5-MoE MTP draft models
sufubao
committed
fix(qwen3next): persist mtp full-attn cpu cache slots
sufubao
committed
refactor(qwen3_5_mtp): drop unused _draft_kv_slot attribute
sufubao
committed
style: fix black formatting and drop unused var for pre-commit
sufubao
committed

Commits on Jun 7, 2026

Commits on Jun 8, 2026