| name | ml-training |
| description | PyTorch training correctness for optimizers and schedules, distributed/DDP, variable-size masked losses, checkpoint save/load, CUDA runtime settings, and equivariant model checks. Use when writing or reviewing training loops, optimizer/scheduler setup, multi-GPU code, loss reduction, or checkpoints.
|
ML Training
Treat optimizer, schedule, DDP, and checkpoint choices as project contracts,
not universal defaults.
Universal Invariants
- Establish a simple AdamW or project baseline before adopting specialized
optimizers or schedules.
- Declare the loss unit and denominator. Exclude padded and missing labels
before reduction.
- In DDP, every rank must execute the same collectives in the same order.
- Save portable unwrapped model weights; rank 0 writes, every rank loads.
- Resume must restore every state that affects the next update: model,
optimizer, scheduler, scaler, step/epoch, sampler position when applicable,
and RNG state required by the reproducibility contract.
- Do not enable
static_graph or find_unused_parameters without evidence
that the model graph requires and supports it.
- Load untrusted checkpoints with
weights_only=True; never deserialize
arbitrary Python objects from an untrusted source.
Route to the Relevant References
Start with the closest reference. Load additional references only when the
request crosses contracts such as distributed masking plus checkpointing.
- Parameter grouping, AdamW baseline, optional Muon, short-run-safe schedules:
optimizer-schedule.md
- DDP initialization, sampler, collectives, validation, cleanup:
distributed.md
- Variable atoms/residues/pairs and sparse-label loss normalization:
loss-masking.md
- Portable save/load and exact-resume state:
checkpoint.md
- SE(3)/equivariance implementation and transformation tests:
equivariance.md
For experiment design use research-method; for molecule/protein data splits,
labels, and metrics use chem-bio-ml; for cluster allocation use slurm-hpc.
Verification Ladder
- One CPU batch: forward, loss, backward, optimizer step.
- One device batch with the production dtype/masking path.
- Checkpoint save/load round trip and next-step equivalence when resume matters.
- Two-rank smoke test for DDP code, including unequal valid counts.
- Only then run a long or multi-node job.
