llama.cpp
ggml : update CUDA/Metal radix Top-K to make tied selections deterministic
#31084
Merged
Merged
ggml : update CUDA/Metal radix Top-K to make tied selections deterministic #31084
ggerganov
merged
19 commits into
ggml-org:master
from
sanmai:ggml/top-k-deterministic-ties
ggml : fix for the radix TOP_K thread timing bug
09fdff62
make the key an ordered 64-bit pair
e3c72088
a rewrite without shift-skip logic
61cce0fa
fix casts
722116f8
warp-contiguous layout
87ff2828
reduce the scope removing vulkan+metal changes
e8fb7cf5
turn it off for Metal/ROCm, note about cub::DeviceTopK
19a17316
fix the hip build
e61e87e1
skip Vulkan as well
5f703e3c
sanmai
changed the title
ORippler
commented
on 2026-11-05
popc lambda from fattn
108e254e
do block_cols math in int64_t
0a374bd4
sanmai
commented
on 2026-11-05
make it allowlist bc SYCL looks non deterministic too
dddb7951
ggerganov
commented
on 2026-11-06
ggml : make Metal radix Top-K tied selections deterministic
81ea5f5b
tests : enable deterministic top-k check on Metal, add radix-path tie…
a4d0db49
ServeurpersoCom
approved these changes
on 2026-11-06
ggml : use coalesced accesses in the Metal radix Top-K compaction
c2889ab6
tests : remove the extra top-k tie and perf cases
efcf2de2
ggerganov
approved these changes
on 2026-11-06
metal: scan the radix bins and warp totals with a simdgroup
599da5be
tests: enable the deterministic top-k check on ROCm
6f4b7498
cuda: remove the redundant radix top-k counter reset
ce8aeaec
sanmai
changed the title Reviewers
Assignees
No one assigned
Labels
testing
ggml
merge ready
Apple Metal
CUDA
Milestone
No milestone
Login to write a write a comment.
Login via GitHub