Model compilation defaults

Model-executing CLI commands use torch.compile(model, dynamic=True). The first run can take longer while kernels compile; use the opt-out below to diagnose compilation failures.

CLI	Default	Opt-out / notes
pairwise-comparison	Compiled	--disable-compile
prototype extract-kmer-prototypes	Compiled	--disable_compile; union and model-free dry runs execute no model
pretrain	Compiled	--disable_compile
finetune-modification	Compiled	--disable_compile
detect-modification	Compiled	--disable_compile
zero-shot train, train-scale, fit-whitening, calibrate	Compiled	--disable_compile
zero-shot pipeline	Compiled	Passes --disable_compile to its model stages
zero-shot detect	Compiled	--disable_compile; optional --deterministic fixes numerically sensitive kernel choices
prototype classify	Compiled backbone	--disable_compile

Commands that process stored embeddings, plots, BAM tags without a model, or configuration files have no model to compile.

zero-shot detect --deterministic retains compilation and the command precision settings, but opts into deterministic PyTorch kernels. It can change runtime speed and numerical rounding relative to the default execution policy. The option applies to detection. See the detection guide for reproducibility limits and calibration considerations.