PR #21891 CUDA: manage NCCL communicators in context

CUDA: manage NCCL communicators in context #21891

JohannesGaessler merged 5 commits into ggml-org:master from JohannesGaessler:cuda-nccl-context

JohannesGaessler requested a review 38 days ago

JohannesGaessler requested a review from

ggerganov 38 days ago

CUDA: manage NCCL communicators in context

3a396fd7

JohannesGaessler force pushed from fd1d82bd to 3a396fd7 38 days ago

JohannesGaessler changed the title ~~CUDA: manage NCCL communictors in context~~ CUDA: manage NCCL communicators in context 38 days ago

ggerganov commented on 2026-04-14

gaugarg-nv commented on 2026-04-14

add check that all backends are CUDA

d49b1c1e

remove unused vector, limit init to > 1 GPUs

2c43969c

JohannesGaessler force pushed from 20457274 to 2c43969c 38 days ago

github-actions added Nvidia GPU

github-actions added ggml

fix warnings

db53263d

gaugarg-nv commented on 2026-04-15

ggerganov approved these changes on 2026-04-15

fix cuda device, cache allreduce

8c78c5c2

ggerganov approved these changes on 2026-04-15

am17an approved these changes on 2026-04-15

JohannesGaessler merged 014dca49 into master 37 days ago

Reviewers

am17an

ggerganov

gaugarg-nv

Assignees

No one assigned

Labels

Nvidia GPU ggml

Milestone

No milestone