pub fn preflight(
store: &WeightStore,
config: &ModelConfig,
use_speculative: bool,
kv_cache_dtype: Option<&str>,
) -> Result<()>Expand description
Run all model-agnostic + model-type-specific pre-flight checks.
Called by spark-server/src/main.rs immediately after the
WeightStore is populated and before spark_comm::NcclBackend::new
runs, so a bad checkpoint aborts rank 0 before rank 1 even connects.
use_speculative is the final resolved flag (user --speculative
OR model default). It gates MiniMax-specific MTP-presence diagnostics:
if the user didn’t ask for speculative decoding, MTP tensors in the
checkpoint are harmless dead weight and do not warrant a bail.