[yashaswikarnati]

What changed

• expose EncoderPrefetchLoader.save_state() -> None so checkpointing can treat encoder lookahead as derived state while the canonical language-loader cursor remains authoritative.
• allow text-only MIMO microbatches to pass through encoder prefetch without invoking the image encoder.
• guard CUDA stream and debug-timing bookkeeping when a batch produces no encoder features.

Why

Encoder prefetch from #5833 can wrap a checkpointable data iterator and can receive image-sparse batches from heterogeneous multimodal datasets. The outer wrapper must participate in the checkpoint interface without persisting its speculative queue, and batches without images must preserve schedule progress without manufacturing encoder work.

The model still validates missing inputs when matching modality tokens are present; this only permits genuinely text-only batches.

Validation

• python3 -m py_compile examples/mimo/training/encoder_prefetch.py tests/unit_tests/models/mimo/test_mimo_encoder_prefetch.py
• OCI-HSG 5-node Mistral VLM smoke job 5590459: 200 iterations completed with both patches in the staging branch.
• focused regression coverage added for non-authoritative prefetch state and text-only batch pass-through.

Local pytest was not rerun on the Codex host because that host does not provide uv or Torch; CI should run the focused tests in the supported container.

[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.

[yashaswikarnati]

/ok to test d16427a

[svcnvidia-nemo-ci]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/31058063807