CUDA: Factor out and re-use `block_reduce` function #18785
am17an
merged
17 commits into
ggml-org:master
from
ORippler:osimons/factor_out_two_stage_warp_reductions
CUDA: Refactor and expose two_stage_warp_reduce_* function
418fb728
Use `two_stage_warp_reduce` also in softmax kernel, move smem out of it
1ebe58d8
Update ggml/src/ggml-cuda/common.cuh
c63c148f
Use two_stage_warp_reduce in group_norm_f32
41bab53c
Use two_stage_warp_reduce in rms_norm_f32
a40b15fa
Fix smem calculation which expects bytes
dbd449e0
Make `two_stage_warp_reduce` accept all values warp_reduce accepts
7f43e648
Use two_stage_warp_reduce in l2_norm_f32
67a9c139
Use type traits for block reduction for better legibility
612874fc
Make norm tests cover all cuda paths
82a3458b
Mark columns % WARP_SIZE !=0 as supported for RMS_NORM_BACK
8bc326a5
Use `enum class` for `block_reduce_method`
bd6ffff4
Rename variables as suggested in code review by @am17an
c1a048bf
Rename two_stage_warp_reduce -> block_reduce
767eba93
ORippler
changed the title
am17an
approved these changes
on 2026-01-13
JohannesGaessler
approved these changes
on 2026-01-13
Fix trailing whitespace in common.cuh
0ed37212
Make condition of static_assert type-dependent
b275e42e
Inline definitions
38e3040e
Reviewers
Assignees
No one assigned
Labels
testing
Nvidia GPU
ggml
Milestone
No milestone
Login to write a write a comment.
Login via GitHub