Performance
Faster first calls: the compilation cache
virgil is built on JAX, which compiles each program the first time it runs. A first fit in a fresh Python process is dominated by that compile time, not by the fit itself; later calls in the same process are fast. To keep the compiled programs between processes, turn on JAX's persistent compilation cache once, before the first fit:
import virgil
virgil.enable_compilation_cache() # ~/.cache/virgil/jax, or pass a path
The first process fills the cache; later ones read from it, which cut a cold binary fit by a factor of 2 to 4 in our benchmarks. Results are unchanged.
JAX's default thresholds skip programs that compile quickly or are small, which
covers most of virgil's, so the helper sets
jax_persistent_cache_min_compile_time_secs and
jax_persistent_cache_min_entry_size_bytes to zero. The directory is
$VIRGIL_COMPILATION_CACHE if set. The cache is keyed on the program, JAX
version and hardware, so it is safe to share and to delete. It is off by default.
On a cluster, put it on node-local or per-user scratch storage rather than a shared path that many jobs write at once.