DeepSpeed
feat(comm): one-shot P2P custom all-reduce for TP=2 inference
#8997
Merged

feat(comm): one-shot P2P custom all-reduce for TP=2 inference #8997

delock merged 16 commits into master from gma/custom_ar_standalone
delock
delock feat(comm): one-shot P2P custom all-reduce for TP=2 inference
d9054634
delock delock requested a review from hwchen2017 hwchen2017 15 days ago
delock delock requested a review from loadams loadams 15 days ago
delock delock requested a review from tjruwase tjruwase 15 days ago
delock delock requested a review from tohtana tohtana 15 days ago
chatgpt-codex-connector
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-11-03
delock Remove reduce-scatter+all-gather fallback from RowParallel inference …
5b67adb1
delock Guard RowParallel all-reduce against non-contiguous inputs
735e6a56
delock Let inference_all_reduce own backend selection in RowParallel
817612e2
delock delock marked this pull request as draft 15 days ago
delock Restore VocabParallelLinear fused_loss_capture override
e1f96180
delock delock force pushed from ae7a5636 to 8f756c95 15 days ago
delock Restructure custom collective as an accelerator backend contract
9b2d0ecc
delock delock force pushed from 8f756c95 to 9b2d0ecc 15 days ago
delock delock marked this pull request as ready for review 15 days ago
chatgpt-codex-connector
chatgpt-codex-connector commented on 2026-11-03
delock Drop top-level torch import from the custom_collective builder
9ec7d953
delock delock force pushed from 4d02ca7a to 9ec7d953 15 days ago
delock
delock Address review findings on the custom collective fast path
fe8a15f3
delock Add two-device parity test for the custom collective fast path
6413ab6c
hwchen2017
delock Keep the custom collective route off the CPU accelerator for now
a2fb1038
delock
jinyouzhi
delock Make the custom collective link setup safe and correct on real fabrics
aa6e8f15
delock clang-format the custom collective kernel
1cfdbed3
delock
jinyouzhi
delock
tohtana tohtana added needs-gpu-test
delock
delock fix(comm): restrict the CPU shm inference all-reduce to world-wide gr…
4738e29a
delock Merge branch 'master' into gma/custom_ar_standalone
a1ed4d5d
pengdurice Merge branch 'master' into gma/custom_ar_standalone
b24f6757
pengdurice
pengdurice approved these changes on 2026-11-13
pengdurice pengdurice enabled auto-merge 5 days ago
hwchen2017
tohtana tohtana removed needs-gpu-test
delock delock force pushed from f86be4e2 to 50219aa3 4 days ago
delock delock force pushed from 50219aa3 to c80239cc 4 days ago
delock delock force pushed from c80239cc to d5f950c1 4 days ago
delock delock force pushed from d5f950c1 to 2b0a1782 4 days ago
delock Address the review findings on link setup, routing, and tests
27f6d793
delock delock force pushed from 2b0a1782 to 27f6d793 4 days ago
delock
delock delock merged 7d30f920 into master 4 days ago
delock delock deleted the gma/custom_ar_standalone branch 4 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone