PR #24340 Support Step3.5/3.7 flash mtp3

Support Step3.5/3.7 flash mtp3 #24340

ggerganov merged 16 commits into ggml-org:master from stepfun-ai:step35-multilayer-mtp-rebase

github-actions added model

forforever73 marked this pull request as ready for review 12 days ago

forforever73 requested a review from

CISC 12 days ago

forforever73 requested a review from

ggerganov 12 days ago

forforever73 requested a review 12 days ago

am17an commented on 2026-06-13

add mtp_layer_offset + include nextn flags in graph reuse

e98e0b3f

add llama_set_mtp_layer_offset + llama_model_n_nextn_layer API

1208d99d

offset head select + require all MTP blocks

0b8aa51e

speculative multi-head process()

e0fb9ffa

speculative multi-head draft()

34f68f5a

gather outputs via inp_out_ids

ae013e3c

cleanup

1885d8f3

fix core

f4a2c12e

minor cleanup

9a0ff266

merged draft_multi_head into draft()

2952d834

forforever73 force pushed from cffdd9a5 to 2952d834 9 days ago

am17an commented on 2026-06-14

mtp rename nextn

48a7484a

Apply suggestions from code review

2ce24fb7

clean-up comments

9b8f3b66

am17an approved these changes on 2026-06-14

ggerganov requested changes on 2026-06-15

fix for multi seq

7a0a2475

ggerganov commented on 2026-06-15

ggerganov assigned

ggerganov 6 days ago

ggerganov commented on 2026-06-21

apply suggestions && chain-heads comment

9858fd2e

add a reference for chain_heads discussion

60b04d3f

ggerganov approved these changes on 2026-06-21

ggerganov merged d7895274 into master 2 days ago

Reviewers

ggerganov

am17an

CISC

Assignees

ggerganov

Labels

model

Milestone

No milestone

llama.cpp Support Step3.5/3.7 flash mtp3 #24340 Merged

Support Step3.5/3.7 flash mtp3 #24340

llama.cpp
Support Step3.5/3.7 flash mtp3
#24340

Merged