Module preflight 

Pre-flight weight-store / config consistency checks.

Runs before NCCL init and model construction so obvious checkpoint mismatches (wrong expert count, missing lm_head, MiniMax checkpoint shipped with MTP tensors that the loader can’t consume, etc.) fail fast with a readable error instead of surfacing later as:

• an ncclCommInitRank hang (when only one rank bails before reaching collective init),
• a cryptic build_model error ~10 minutes into startup,
• an opaque “Received NCCL unique ID from master” log trail with no explanation — the failure mode Discord users have been posting.

The checks are intentionally cheap: they only consult tensor NAMES already in the WeightStore (loaded lazily by the safetensors index pass), never touch GPU memory, and never issue collectives. Safe to call on every rank.

Functions

preflight

Run all model-agnostic + model-type-specific pre-flight checks.