DeepSpeed
Fix hook count performance regression from v0.18.5
#7886

Merged

Fix hook count performance regression from v0.18.5 #7886

tohtana merged 5 commits into deepspeedai:master from tohtana:tohtana/fix-perf-regression

Add should_refresh_expected_hook_count and harden enter_backward

e1b99f70

Fix ZeRO-2 hook to cache count_used_parameters_in_backward result

91b591bc

Fix ZeRO-3 hooks to cache count_used_parameters_in_backward result

d3a6e996

delock commented on 2026-03-05

delock approved these changes on 2026-03-05

tohtana marked this pull request as ready for review 4 days ago

tohtana requested a review from

tjruwase 4 days ago

tohtana requested a review from

loadams 4 days ago

Add regression tests for hook count performance fix

e1b41ee5

Add comment clarifying refresh-before-reenter ordering

c43e4237

tohtana force pushed from fd933f3e to c43e4237 4 days ago

tohtana merged 6c59d544 into master 4 days ago

Reviewers

delock

tjruwase

loadams

Assignees

No one assigned

Labels

None yet

Milestone

No milestone