Repository navigation

Initialize multimodal models from language checkpoints - #7316

Merged

tylerpoon merged 17 commits into

NVIDIA:mainfrom

tylerpoon:tpoon/github-vlm-sft-04-llm-only-checkpoint-init

Sep 22, 2026

Merged

Initialize multimodal models from language checkpoints#7316
tylerpoon merged 17 commits into
NVIDIA:mainfrom
tylerpoon:tpoon/github-vlm-sft-04-llm-only-checkpoint-init

Commits on Sep 11, 2026
- Add multimodal vision token utilities
  tylerpoon
  committed
  40d85dc
- Test multimodal vision token utilities
  tylerpoon
  committed
  c806aa7
- Support multimodal metadata and vision context parallelism
  tylerpoon
  committed
  4bf5ae1
- Provide an image placeholder in the temporal video forward test
  tylerpoon
  committed
  ea7eee9
- Add copyright header to tensor-parallel data tests
  tylerpoon
  committed
  c4ca837
- Restore legacy checkpoint deserialization compatibility
  tylerpoon
  committed
  be3f8b4
- Initialize multimodal models from language checkpoints
  tylerpoon
  committed
  b487695
Commits on Sep 14, 2026
Commits on Sep 16, 2026
- Merge branch 'main' into tpoon/github-vlm-sft-04-llm-only-checkpoint-init
  tylerpoon
  authored
  1f0ce9b
Commits on Sep 17, 2026
- Merge branch 'main' into tpoon/github-vlm-sft-04-llm-only-checkpoint-init
  tylerpoon
  authored
  58af71f
Commits on Sep 18, 2026
- Merge main and preserve multimodal checkpoint initialization and saving
  tylerpoon
  committed
  5869a36
Commits on Sep 19, 2026
- Merge main and retain multimodal checkpoint initialization
  tylerpoon
  committed
  c540df7
Commits on Sep 21, 2026