MinT-2M Backend Integration Plan

The current repository is an integration release. It proves the shape of the training path while keeping model-specific code reviewable. A production backend submission should be split by code ownership.

Proposed patch series

1. Add the GLM resident prefix state abstraction and lifecycle tests.
2. Add global context-parallel DSA candidate exchange and selected-state materialization beside the owning sparse-attention implementation.
3. Add response-only replay and chunked output-loss integration under the backend training schedule.
4. Add Bridge LoRA target coverage and recompute/offload controls.
5. Add conventional-versus-replay distributed correctness tests.
6. Keep MinT Runtime request translation in a separate, thin adapter commit.

Each patch should state its exact Megatron-LM and Megatron-Bridge base commit. The current runtime was not a clean set of upstream worktrees, so generated patches must include the dirty source deltas or be reconstructed and revalidated against clean pinned revisions.

Required review artifacts

• a source manifest for every evidence bundle;
• the full model/checkpoint revision and parallelism configuration;
• a container digest or reproducible environment lock;
• 32K and 64K distributed gradient/optimizer parity results;
• a fresh 2M execution receipt from the same source commit;
• license and redistribution approval for newly authored integration code.

Until those artifacts exist, this repository is the review surface; it should not be described as a ready-to-merge backend pull request.