llama.cpp
Make the drafter probabilistic and the target verify by rejection sampling for simple draft and MTP
#27694
Merged

Make the drafter probabilistic and the target verify by rejection sampling for simple draft and MTP #27694

praneshgo
github-actions github-actions added server
praneshgo
praneshgo praneshgo force pushed 42 days ago
ruixiang63
praneshgo
praneshgo praneshgo force pushed to 88e706e9 39 days ago
praneshgo praneshgo marked this pull request as ready for review 39 days ago
praneshgo praneshgo requested a review 39 days ago
praneshgo praneshgo requested a review 39 days ago
praneshgo
praneshgo praneshgo force pushed from 88e706e9 to b11dae78 35 days ago
ruixiang63
praneshgo
ruixiang63
gaugarg-nv
gaugarg-nv commented on 2026-09-08
gaugarg-nv
ggerganov ggerganov assigned ggerganov ggerganov 34 days ago
danbev
danbev commented on 2026-09-09
praneshgo
danbev
danbev approved these changes on 2026-09-15
praneshgo
praneshgo praneshgo requested a review from gaugarg-nv gaugarg-nv 25 days ago
ggerganov
ggerganov commented on 2026-09-21
praneshgo
praneshgo praneshgo requested a review from ggerganov ggerganov 12 days ago
ggerganov
ggerganov approved these changes on 2026-10-02
praneshgo Make the drafter probabilistic and the target verify by rejection sam…
269bb4fa
praneshgo Drop stale spec_draft_q before drafting
2f82a679
praneshgo Fallback to argmax sampling for grammar-constrained requests and addi…
94d05d5f
praneshgo Support grammar-constrained requests in rejection sampling
b1b5b01c
praneshgo Fix - renormalize distribution after masking
bba1f2e1
praneshgo copy rng on sampler copy and re-accept drafted tokens on replay
d9eb9f0d
praneshgo Fix draft sampler sharing the target's rng stream
b5c26731
Simplify the rejection sampler's inputs and move replay to the server
61de64e1
Truncate the draft candidates along with the draft
c19c18a7
praneshgo praneshgo force pushed from 75471a12 to c19c18a7 10 days ago
praneshgo
ggerganov ggerganov added merge ready
ServeurpersoCom
ServeurpersoCom approved these changes on 2026-10-02
ggerganov ggerganov merged 1fb7ef3e into master 10 days ago
lbyte
praneshgo
henfiber

Login to write a write a comment.

Login via GitHub

Assignees
Labels
Milestone