vulkan: add tensor-parallel allreduce function #25051
ggerganov
merged
21 commits into
ggml-org:master
from
pwilkin:vulkan-tp-p2p
netrunnereve
commented
on 2026-06-27
0cc4m
changed the title
pwilkin
changed the title
pwilkin
changed the title
TheBlueMatt
commented
on 2026-06-28
netrunnereve
commented
on 2026-06-28
JohannesGaessler
commented
on 2026-06-29
0cc4m
changed the title
vulkan: add tensor-parallel all-reduce communication
a23aac8f
vulkan: require shared semaphores for tensor communication
c42f2a3d
vulkan: reduce directly from shared host buffers
a52c1ee6
vulkan: fix ring all-reduce races with 3+ devices
b700ea79
vulkan: fold the tensor-parallel all-reduce into the compute submissions
bed9ef92
- Two-phase all-reduce over shared host memory: write own slot + sign…
3f71382d
add minimal submit split size
853cf28d
replace opaque_fd with sync_fd for spec compliance
d87d9a3c
add precision check
f3681626
add semaphore pool
5a4b2934
rename env var, add logging if requirements not met
883fe50b
reduce comments
c28174c3
code improvements
f7563967
vulkan: use posix_memalign for the tensor-parallel host buffers
468d073b
vulkan: synchronize the tensor-parallel all-reduce through the host
37f83e6d
use sync_fd on mesa drivers
7d8c3496
refactor for readability
128a849c
move sync_fd semaphore handling to thread, trigger mesa submit threads
d8446f0f
move allreduce functions to new cpp file
b5145b27
clean comments
ecfee2ec
smaller fixes, code improvements
e5793a11
0cc4m
approved these changes
on 2026-10-30
Reviewers
Assignees
No one assigned
Labels
Vulkan
ggml
merge ready
Milestone
No milestone
Login to write a write a comment.
Login via GitHub