Skip to content

Upgrade/megatron v0.18.2 - #1284

Merged
Darryl233 merged 9 commits into
flagos-ai:mainfrom
zhaoyinglia:upgrade/megatron-v0.18.2
Sep 7, 2026
Merged

Upgrade/megatron v0.18.2#1284
Darryl233 merged 9 commits into
flagos-ai:mainfrom
zhaoyinglia:upgrade/megatron-v0.18.2

Conversation

@zhaoyinglia

Copy link
Copy Markdown
Collaborator

PR Category

PR Types

PR Description

zhaoyinglia and others added 6 commits September 7, 2026 00:26
Adapt model construction to Megatron v0.18.2 and restore MTP with mRoPE.
Preserve checkpoint vocabulary padding, correct conversion metadata and
vision layout, and support the Energon 6 ChatML path decoder.

Remove per-microbatch cache clearing and provide smoke and eight-GPU
throughput configurations. Document data preparation, conversion, resume,
export and measured throughput, including the current 67 GiB memory limit.
TP overlap requires TransformerEngine-FL 12595d06 or an equivalent fix.

Validation: 17 regression tests and repository checks passed; prior full
weight roundtrip, strict TP2 loading, training/resume/export and eight-GPU
15-step training completed. No machine paths or runtime artifacts included.

@Darryl233 Darryl233 left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTMM

@Darryl233
Darryl233 merged commit b5741d0 into flagos-ai:main Sep 7, 2026
38 of 49 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants