llama.cpp
metal : per-op source split + parallel compile
#24021
Open

metal : per-op source split + parallel compile #24021

forforever73 wants to merge 11 commits into ggml-org:master from forforever73:metallib-split
forforever73
preliminary extract common header
4d316c4d
op source split
d2187402
split metallib into 8 libs && load in parallel
f8ba384a
derive kernel->library routing from functionNames
15a9ada2
x-macro lib list + underscore filenames, dedup QK_NL, MRC fixes
b975ce5e
op source split 8 to 20
ee8f20c0
improve robustness of source fallback
24200a77
clean up
61480fde
change bool -> atomic_bool
e519b641
only prepend headers that source actually includes
42a2495e
no semaphore, use GCD global queue
0d78e869
forforever73 forforever73 requested a review 1 day ago
ggerganov ggerganov assigned ggerganov ggerganov 1 day ago
github-actions github-actions added ggml
github-actions github-actions added Apple Metal

Login to write a write a comment.

Login via GitHub

Reviewers
No reviews
Assignees
Labels
Milestone