Summary

• Add a minimal long-context Dynamic CP packing demo; it does not train a model or launch distributed workers.
• Use Megatron-Core dev DefaultDynamicCPScheduler to schedule toy variable-length samples, then print the packed THD metadata per DPxCP rank: tokens.shape, cu_seqlens, max_seqlen, and local_cp_size.
• Add examples/training_features/long_context/README.md explaining how to run the demo, how to read the output, and how the printed metadata maps to the real DCP forward-step path.
• Show the Bridge config knobs users set in a real run: dynamic_context_parallel=True, sequence_packing_scheduler="default_dynamic_cp", max_seqlen_per_dp_cp_rank, min_dynamic_context_parallel_size, and micro_batch_size=1.
• Pass Dynamic CP initialization kwargs through Bridge distributed setup when the installed MCore supports them, so dev Dynamic CP groups are created correctly after the MCore bump.

Testing

• python -m py_compile examples/training_features/long_context/dynamic_context_parallel.py
• git diff --check
• tmux window 1, Docker container 7366a763896a after ./scripts/switch_mcore.sh dev + uv sync: uv run python examples/training_features/long_context/dynamic_context_parallel.py (prints per-sample gpus_needed and scheduled packed microbatches with local_cp_size)
• tmux window 1: uv run ruff check examples/training_features/long_context/dynamic_context_parallel.py src/megatron/bridge/training/initialize.py
• tmux window 1: uv run ruff format --check examples/training_features/long_context/dynamic_context_parallel.py src/megatron/bridge/training/initialize.py
• tmux window 1: uv run pre-commit run --all-files