Runtime options

Two runtime options follow one pattern: read the state via .value, set it by calling, scope it as a context manager. Both initialize from their environment variable.

specux.deterministic.value            # False
specux.deterministic(True)            # set globally

x = np.random.randn(8, 32768).astype(np.float32)
Z = specux.stft(x, n_fft=1024)
with specux.deterministic(True):      # or scoped: restores on exit
    y = specux.istft(Z, n_fft=1024, length=x.shape[-1])

• specux.deterministic: bitwise-reproducible overlap-add for istft and the training backward: ordered gather instead of atomics, at the cost of a larger intermediate buffer. torch’s use_deterministic_algorithms(True) enables it automatically.
• specux.benchmark: autotune-and-cache every configuration on first use: the first call for a new (GPU, plan configuration) problem measures the candidate kernels and persists the winner. The tuning results survive across processes in the cache directory.

specux.set_deterministic(flag) and specux.set_benchmark(on=True) are plain-function synonyms for calling the options (a set with no scoping).

Threads and caches

• specux.set_num_threads(n) / get_num_threads(): CPU backend thread count; 0 means auto (the physical cores).
• specux.clear_cache(): drop cached plans, filterbanks, twiddle tables, and idle memory-pool blocks. Compiled kernels stay cached.

Environment variables

variable	effect
SPECUX_DETERMINISTIC	initial value of specux.deterministic
SPECUX_BENCHMARK	initial value of specux.benchmark
SPECUX_CACHE_DIR	tuning + kernel cache location (default {tmp}/specux_{user})
SPECUX_CUDA_HOME	CUDA toolkit root (build and runtime discovery)
SPECUX_CPU_ONLY	1 builds CPU-only; 0 makes a missing toolkit a build error