Add persistent program cache for Program.compile#1912

Merged

leofang merged 34 commits intoNVIDIA:mainfrom

cpcloud:persistent-program-cache-178

May 6, 2026

Commits on May 6, 2026

feat(core.utils): persistent program cache
cpcloud
committed
feat(core): Program.compile(cache=...) convenience wrapper
cpcloud
committed
refactor(core.utils): import program cache eagerly, fix ruff findings
cpcloud
committed
docs(core): nest Program caches under CUDA compilation toolchain
cpcloud
committed
docs(core.utils): nudge cache callers toward close()
cpcloud
committed
docs(core.utils): teach cache examples with public ObjectCode API
cpcloud
committed
feat(core): require Program.compile(cache=...) to be a ProgramCacheResource
cpcloud
committed
test(core): mock driver_version to keep _can_load_generated_ptx as cpdef
cpcloud
committed
docs(core): note cache hits do not write to logs
cpcloud
committed
fix(core.utils): cache _LinkerBackend driver decision per key call
cpcloud
committed
fix(core): lowercase target_type in Program.compile and cache key
cpcloud
committed
refactor(core.utils): collapse _LINKER_RELEVANT_FIELDS into _LINKER_FIELD_GATES
cpcloud
committed
fix(core.utils): contextual error when path-backed ObjectCode vanishes
cpcloud
committed
docs(core): clarify cache= vs ProgramOptions.no_cache independence
cpcloud
committed
test(core.utils): tighten multiprocess cache assertions
cpcloud
committed
test(core.utils): add InMemoryProgramCache thread test and FileStream overwrite test
cpcloud
committed
refactor(core.utils): drop dead empty-key branch in _path_for_key
cpcloud
committed
perf(core.utils): drop redundant stat in FileStreamProgramCache.__len__
cpcloud
committed
feat(core.utils): reject max_size_bytes=0 in cache backends
cpcloud
committed
docs(core.utils): note ptxas_options order is intentionally significant
cpcloud
committed
docs(core.utils): document the _keys.py <-> _program.pyx import contract
cpcloud
committed
docs(core.utils): note options.name double-hash on PTX is intentional
cpcloud
committed
docs(core.utils): document burst-write over-eviction trade-off
cpcloud
committed
docs(core): explain why Program.compile cache decode is provably safe
cpcloud
committed
refactor(core.utils): extract _stat_key for the four stat-guarded paths
cpcloud
committed
refactor(core.utils): collapse Windows sharing-retry loops
cpcloud
committed
fix(test): walk sharded entries dir in FileStream overwrite assertion
cpcloud
committed
fix(test): collapse nested with into a single contextmanager group (SIM117)
cpcloud
committed
fix(core.utils): keep path-backed ObjectCode error message Windows-friendly
cpcloud
committed
perf(core.utils): track FileStream cache size incrementally
cpcloud
committed
test(core.utils): drop SUPPORTED_TARGETS cross-check
cpcloud
committed
refactor(core.utils): idiomatic with-as on os.scandir, yield from inner
cpcloud
committed
refactor(core.utils): extract _iter_tmp_entries and _sum_tmp_sizes
cpcloud
committed
fix(core.utils): clamp tracked size at zero on __delitem__
cpcloud
committed