Skip to content

Allow subselecting the appropriate config for llama4 #1815

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 20 commits into from
May 5, 2025

Conversation

dakinggg
Copy link
Collaborator

@dakinggg dakinggg commented May 3, 2025

Multimodal models may return a config with subconfigs from AutoConfig. This PR allows subclasses to automatically select a subconfig (and sets up the causal LM training class to select text config). With this, llama4 scout finetuning works (llama4-scout-hf-real-7-a3JTUJ). llama4 also works with flex attention (llama4-scout-hf-flex-1-EUH7iy)

Along for the ride, install the new HF extra that supposedly speeds up download hf_xet, and bumps the minimum hf hub version that supports it.

@dakinggg dakinggg marked this pull request as ready for review May 3, 2025 04:14
@dakinggg dakinggg requested review from a team as code owners May 3, 2025 04:14
@dakinggg dakinggg changed the title Allow subselecting the appropriate config Allow subselecting the appropriate config for llama4 May 3, 2025
Copy link
Contributor

@bowenyang008 bowenyang008 left a comment

Choose a reason for hiding this comment

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

LGTM

Copy link
Contributor

@irenedea irenedea left a comment

Choose a reason for hiding this comment

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

couple of nits! lgtm

@dakinggg dakinggg enabled auto-merge (squash) May 5, 2025 20:35
@dakinggg dakinggg merged commit 5c47350 into mosaicml:main May 5, 2025
11 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.

4 participants