DeepSpeed
[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model
#8851

Merged

[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model #8851

delock merged 9 commits into deepspeedai:master from jinyouzhi:autotp-frozen-replicated-hook

Support frozen AutoTP models and keep the AutoTP mode per model

6b6f4b84

Reject unfreezing AutoTP replicated params after initialize

53777bfa

Carry the AutoTP training mode per model in AutoTPMeta

68647c93

Keep layers without an explicit AutoTP mode following set_autotp_mode

63bb7f58

jinyouzhi marked this pull request as ready for review 20 days ago

jinyouzhi requested a review from

tjruwase 20 days ago

jinyouzhi requested a review from

tohtana 20 days ago

jinyouzhi requested a review from

hwchen2017 20 days ago

jinyouzhi requested a review from

loadams 20 days ago

chatgpt-codex-connector commented on 2026-10-21

Keep training_mode=None in AutoTP layer metadata

11a3cb34

Resolve an unspecified AutoTP mode when it is used, not at construction

d93285fb

delock requested a review from

delock 11 days ago

0z5a commented on 2026-10-31

Run the frozen AutoTP model tests on CPU

880bc503

AutoTP: use per-layer mode in AsyncColumnParallel bias; assert teache…

e79cdb84

Merge branch 'master' into autotp-frozen-replicated-hook

b61d6b00

delock approved these changes on 2026-11-04

delock merged d8241a9b into master 6 days ago

Reviewers

delock

chatgpt-codex-connector

0z5a

tjruwase

tohtana

hwchen2017

loadams

Assignees

No one assigned

Labels

None yet

Milestone

No milestone

DeepSpeed [AutoTP] Support frozen AutoTP models and manage AutoTP mode per model #8851 Merged

[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model #8851

DeepSpeed
[AutoTP] Support frozen AutoTP models and manage AutoTP mode per model
#8851

Merged