[Ascend] support qwen35 mtp on Ascend-A3#337

Open

wanfengcxz wants to merge 31 commits into

DeepLink-org:mainfrom

wanfengcxz:wq/support_qwen35_mtp

Commits on Jul 3, 2026

[Ascend] MTP speculative decoding support (qwen3_5_mtp_final_2)

authored and
wanfengcxz
committed
[Ascend] Patch MTP graph and sampling runtime

tangzhiyi11
authored and
wanfengcxz
committed
[Ascend] Reduce MTP replay state snapshot peak memory

tangzhiyi11
authored and
wanfengcxz
committed
[Ascend] fix gdn kernel in speculative decoding
wanfengcxz
committed
fix graph

Super User
authored and
wanfengcxz
committed
fix: ensure state is contiguous
wanfengcxz
committed
refactor fill buffers
wanfengcxz
committed
impl ring buffer for gdn state
wanfengcxz
committed
impl ring buffer for conv1d state
wanfengcxz
committed
Refactor GDN and conv1d computation flow
wanfengcxz
committed
refactor gdn buffer
wanfengcxz
committed
fix ascend graph
wanfengcxz
committed
fix graph
wanfengcxz
committed
remove unused patch
wanfengcxz
committed
impl reject_sample triton function on ascend
wanfengcxz
committed
fix mismatch shape error in chunked prefill
wanfengcxz
committed
refactor: lazy load mtp ops
wanfengcxz
committed
update qwen35 patch func
wanfengcxz
committed
fix no mtp
wanfengcxz
committed
refactor graph buffer and kernel
wanfengcxz
committed
fix DPTP comm buffer size and padding_batch_size
wanfengcxz
committed
update qwen35 patch
wanfengcxz
committed
fix aclnnGather error 161002 by sanitizing negative indices
wanfengcxz
committed
remove comma
wanfengcxz
committed
fix padding

Super User
authored and
wanfengcxz
committed
fix global decoding

Super User
authored and
wanfengcxz
committed
fallback reject_sample to torch

WangQing
authored and
wanfengcxz
committed
refactor reject sample
wanfengcxz
committed
refactor: remove redundant if branches
wanfengcxz
committed
format code
wanfengcxz
committed
format code
wanfengcxz
committed