]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
spec : Support Step3.5/3.7 flash mtp3 (#24340)
authorYiChen Lv <redacted>
Sun, 21 Jun 2026 08:33:18 +0000 (16:33 +0800)
committerGitHub <redacted>
Sun, 21 Jun 2026 08:33:18 +0000 (11:33 +0300)
commitd789527482d925156d7c4adfecebf5fb8481e0ee
treed3988d2856295b6e869c7b78fc757fc1961097ef
parent063d9c156e816ae3cf62db01f429a07a099afe97
spec : Support Step3.5/3.7 flash mtp3 (#24340)

* add mtp_layer_offset + include nextn flags in graph reuse

* add llama_set_mtp_layer_offset + llama_model_n_nextn_layer API

* offset head select + require all MTP blocks

* speculative multi-head process()

* speculative multi-head draft()

* gather outputs via inp_out_ids

* cleanup

* fix core

* minor cleanup

* merged draft_multi_head into draft()

* mtp rename nextn

* Apply suggestions from code review

Co-authored-by: Aman Gupta <redacted>
* clean-up comments

* fix for multi seq

* apply suggestions && chain-heads comment

* add a reference for chain_heads discussion

---------

Co-authored-by: Aman Gupta <redacted>
common/speculative.cpp
include/llama.h
src/llama-context.cpp
src/llama-context.h
src/llama-cparams.h
src/llama-ext.h
src/llama-graph.h
src/llama-model.cpp
src/models/step35.cpp