Skip to content

Update batch size for mbridge qwen3-32B on MI300X#902

Merged
GeneDer merged 1 commit into
release/v26.5from
cp-pr897
Jul 21, 2026
Merged

Update batch size for mbridge qwen3-32B on MI300X#902
GeneDer merged 1 commit into
release/v26.5from
cp-pr897

Conversation

@GeneDer

@GeneDer GeneDer commented Jul 21, 2026

Copy link
Copy Markdown
Member

Cherry-pick #897 onto the v26.5 release branch.

Copilot AI review requested due to automatic review settings July 21, 2026 15:23
@GeneDer
GeneDer changed the base branch from main to release/v26.5 July 21, 2026 15:23
@GeneDer

GeneDer commented Jul 21, 2026

Copy link
Copy Markdown
Member Author

Not going to wait for #897 to be merged into main as we want to release this asap

@GeneDer
GeneDer merged commit 9d59e81 into release/v26.5 Jul 21, 2026
5 checks passed

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

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

Pull request overview

This PR cherry-picks PR #897 onto the v26.5 release branch to adjust the effective batch sizing for MI300X runs of the Megatron Bridge Qwen3-32B post-training configs, likely to improve memory fit/stability at the current sequence length.

Changes:

  • Reduce micro_batch_size from 2 to 1 for Qwen3-32B SFT post-training on MI300X.
  • Reduce micro_batch_size from 2 to 1 for Qwen3-32B LoRA post-training on MI300X.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File Description
examples/megatron_bridge/configs/MI300X/qwen3_32b_sft_posttrain.yaml Lowers micro_batch_size to 1 for the SFT post-train MI300X configuration.
examples/megatron_bridge/configs/MI300X/qwen3_32b_lora_posttrain.yaml Lowers micro_batch_size to 1 for the LoRA post-train MI300X configuration.

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.

3 participants