Pass the config container to pretrain in the multimodal example

[huthvincent]

• I, the PR author, have personally reviewed every line of this PR.

What does this PR do?

Makes examples/multimodal/train.py start again. It currently cannot.

Two halves of one incomplete migration, both in the __main__ block:

1. Line 391 calls pretrain_cfg_container_from_args(args), but the module's only import from that package is from megatron.training.arguments import parse_and_validate_args. The symbol lives in megatron.training.argument_utils, which this file never imports, and there is no star-import — so the line raises NameError.
2. The full_config it builds is never passed. pretrain() now begins (cfg_container, train_valid_test_dataset_provider, model_type, forward_step_func, model_provider=None, ...), and the call here begins with the dataloader provider, so every positional argument binds one slot early: cfg_container gets the dataloader provider, train_valid_test_dataset_provider gets ModelType.encoder_or_decoder, model_type gets forward_step. pretrain dereferences its first argument immediately as cfg_container.logger.log_progress, so there is no legacy order to fall back on.

The ModelBuilder migration in #5516 applied both halves correctly to pretrain_vlm.py, examples/mimo/train.py and examples/bert/pretrain_bert.py; this brings the multimodal entry point in line with those three. pretrain_vlm.py:500-502 is the exact shape used here.

A reviewer raised this on #5516 — #5516 (comment) — and the thread was closed as "resolved in af7f1b6". That sha is not in the repository (the PR was squash-merged), and the merged result contains only the args_defaults / extra_args_provider half of the fix; the import and the first positional argument are still missing on main and on dev.

Issue tracking

For PRs from open-source community contributors:

• New features: a linked issue is required.
• Small updates (bug fixes, minor improvements): a linked issue is recommended.

Linked issue: none — 2-line bug fix. Related to #5516 (comment linked above).

Contribution process

Pre-checks

• I have added relevant unit tests — nothing in tests/ imports examples/multimodal/train.py, which is why the migration could land without breaking anything visible. Happy to add a smoke test if you want one.
• I have added relevant functional tests
• I have added proper typing to my code — the touched lines carry no annotations.
• I have added relevant documentation
• I have run the autoformatter.sh on my PR — the changeset is empty under its scope (megatron/core and tests/), and the new import is placed where isort would put it (argument_utils sorts before arguments).

How to reproduce, and how we checked it

examples/multimodal/pretrain_mistral_clip.sh:133 runs torchrun --nproc_per_node 8 examples/multimodal/train.py ${OPTIONS}, and examples/multimodal/README.md:89 documents running that script. The failure is at module __main__, before any GPU work, so no GPU or data is needed to see it.

Checked on 8xL4 / torch 2.12.0+cu130: an AST pass collected every name bound at module scope (no star-imports present), confirmed pretrain_cfg_container_from_args is not among them, read inspect.signature(pretrain) as beginning ['cfg_container', 'train_valid_test_dataset_provider', 'model_type', 'forward_step_func', 'model_provider'], and executed the file's real __main__ block with every bound name stubbed — which raises NameError: name 'pretrain_cfg_container_from_args' is not defined.

What we did not do: we did not run the LLaVA example end to end, so this is verified to remove the NameError and the argument shift, not to demonstrate that multimodal pretraining converges after the ModelBuilder migration. Nothing was benchmarked and no performance claim is made.

The strongest objection to this change

This is an examples/ script that no CI job executes, and it has been broken identically on dev for hundreds of commits — so the two-line patch fixes the entry point without showing that the documented example works after it. If examples/multimodal/train.py is superseded by pretrain_vlm.py, the right change is deleting the file or the README pointer rather than fixing it, and we would rather be told that than guess.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[huthvincent]

@NVIDIA/mcore-oncall — same ask as on #6859: Ready, but copy-pr-bot is still waiting on validation so this has not been built. Could someone run /ok to test fffd9d9 when convenient?

Two added lines in examples/multimodal/train.py so the documented LLaVA pretraining entry point can start. Happy to rebase first if you would prefer it on a newer main.