Current guides live at the top level of docs/; dated validation runs, audits and one-off experiments are kept under
archive/ unchanged (paths inside them were rewritten, content was not).
| Where | What |
|---|---|
| repositories.md | repository map, engine pin history, validation notes |
| pipeline.md | training phases and configs |
| known-issues.md | validation limits and open problems |
| kernels/ | kernel development guides and status reports |
| operations/b200-environment.md | B200 nodes: environment, GPUs, lock file, tests |
| operations/ | cluster and launcher procedures |
| paper/ | technical report draft |
| graphical-abstract/ | figures |
| archive/ | dated checks and records: graph-training-20260923, trimul-fusion, repository-audit-20260927, history, ... |
Related checks and commands: bash scripts/check.sh (lint, format, type check) and
bash scripts/b200/run_tests.sh (the test suite on a B200 node).
team-gm no longer exists as a dependency. MiniWorld and miniworld-engine are the only two repositories.
Was (team_gm.*) |
Is now |
|---|---|
core.client, core.batch |
miniworld.training.client, miniworld.training.batch |
core.callbacks (incl. ModelEMA) |
miniworld.training.callbacks |
diffusion (base, edm, decoupled_xpred, augmentation) |
miniworld_engine.architecture.diffusion |
modules.blocks, modules.layers, modules.primitives, modules.exceptions, modules.tensor_layout |
miniworld_engine.architecture.{blocks,layers,primitives,exceptions,tensor_layout} |
utils.se3, utils.align, utils.transform |
miniworld_engine.architecture.utils.* |
utils.script_utils |
miniworld.utils.script_utils |
typing.typecheck |
miniworld_engine.architecture.typing |
Parameter names are unchanged, so checkpoints written by team-gm-based versions load as before. Documents below that
link into libs/team-gm/... describe the code as it was when they were written.
The kernel and integration switches are fields of miniworld_engine.settings.Settings; a run sets them with
train.engine_settings={field: value} (or settings.configure(...)). The run options of the training scripts are train.*
fields (configs/run_options.py). Variables that are still read from the environment are the ones an external payload or
tool reads (CUDA_*, XDG_CACHE_HOME, OPT_CORE_DIR, TRIMUL_NATIVE_BUILD_DIR, ...) and the ones that have to reach child
processes or be read when the kernel modules import (MINIWORLD_COMPILE_WRAP, MINIWORLD_BUILD_*, ...). A removed variable that is
still exported is reported once when the engine imports its settings.
The raw experiment directories that docs/kernels/* and docs/graphical-abstract/* link to (runs/triattn_20260921,
runs/trimul_*, runs/anthropic_*, runs/norm_cuda_20260926, ...) are no longer tracked. They are kept, unchanged, under the git tag
archive/h100-kernel-experiments-20260927 (commit df0d94b5). bash scripts/restore_kernel_experiments.sh extracts them into runs/, after which the links and
docs/graphical-abstract/collect_evidence.py resolve again.