Configuration Reference

This page is the single index of every entry-point keyword argument for a solve: the explicit parameters of solve_ivp() and Solver (__init__() and solve()), plus how the remaining loose keyword arguments are routed to the six underlying settings groups. Deep-dive pages for each group are linked at the end of each section.

Entry-point signatures

solve_ivp()

solve_ivp() builds a Solver and runs a single batch solve.

Argument	Default	Effect
method	"euler"	Integration algorithm name (see Choosing an Algorithm).
duration	1.0	Total integration time.
settling_time	0.0	Warm-up period before outputs are recorded.
t0	0.0	Initial integration time.
save_variables	None	State/observable names to save; None saves all, [] saves none.
summarise_variables	None	State/observable names to summarise; None mirrors save_variables.
grid_type	"combinatorial"	"combinatorial" takes every combination of the supplied inputs; "verbatim" pairs them run-for-run. This default differs from Solver.solve below.
time_logging_level	None	Timing verbosity: 'default', 'verbose', 'debug', 'silent' (record GPU events, print nothing), or None/'None' to disable.
nan_error_trajectories	True	Trajectories with a nonzero status code are NaN-masked in the output.

Any other keyword argument is forwarded to Solver.

Solver.__init__()

Argument	Default	Effect
algorithm	"euler"	Integration algorithm name or a supplied ButcherTableau.
lineinfo	None	Compiles every kernel and device function with source-line correlation data for profilers such as Nsight Compute. None defers to the CUBIE_LINEINFO environment variable (default off); an explicit argument always wins. Updating it later triggers a full rebuild.
cache	True	Compiled-kernel disk caching; accepts bool, a cache-mode string, or a Path. See Caching and Recompilation.
auto_performance	True	Whether to use in-built heuristics for buffer locations in memory, loop unrolling, block size and resident blocks per SM. Does not override manual unroll_*, *_location or blocksize arguments.
time_logging_level	None	Same options as above.

step_control_settings, algorithm_settings, output_settings, memory_settings, and loop_settings are explicit dictionaries that override the per-group defaults; any of their keys may also be supplied as loose keyword arguments (see “Kwarg routing” below).

Solver.solve()

Argument	Default	Effect
duration	1.0	Total integration time.
settling_time	0.0	Warm-up period before outputs are recorded.
t0	0.0	Initial integration time.
blocksize	None	CUDA threads per block for this launch; None uses the solver’s blocksize setting, or lets the solver pick when that setting was not given.
grid_type	"verbatim"	Differs from solve_ivp’s default of "combinatorial" above — only relevant when dict inputs trigger grid construction.
on_device	False	Return a DeviceSolveResult of device-array handles with no device-to-host copy. See Working with Results.
nan_error_trajectories	True	Same behaviour as above; ignored when on_device=True.

Any other keyword argument is forwarded to update(), routing through the same six settings groups described next.

Kwarg routing

Beyond the explicit parameters above, Solver accepts loose keyword arguments and sorts them into settings groups. Any parameter below may be passed directly to solve_ivp(), Solver, or update():

Group	Parameters	Deep-dive page
Step control	step_controller, dt, dt_min, dt_max, atol, rtol, safety, min_step_shrink, max_step_growth, integral_gain, proportional_gain, derivative_gain, filter_coefficients, deadband_min, deadband_max, newton_target_iters	Solver Options Reference
Algorithm	newton_atol, newton_rtol, newton_max_iters, krylov_atol, krylov_rtol, krylov_max_iters, krylov_residual_reduction, krylov_residual_floor, linear_correction_type, preconditioner_type, preconditioner_order, use_smoothed_error	Solver Options Reference, Choosing an Algorithm
Output	output_types, saved_state_indices, saved_observable_indices, summarised_state_indices, summarised_observable_indices	Working with Results
Timing / loop	save_every, summarise_every, sample_summaries_every, save_last, plus the per-buffer *_location placement overrides	Timing Parameters, Making it Faster (basic)
Memory	mem_proportion, stream_group	GPU Memory Management
Cache	cache (True/False, a cache-mode string, or a Path)	Caching and Recompilation
Kernel	max_registers, blocksize	Making it Faster (basic)

A keyword argument that is neither a setting nor a constant of the system raises KeyError. Constants are given by name, at construction or through update.

Notes on selected parameters

step_controller

Selects the step-size controller by name: "fixed", "i", "pi", "pid", or "gustafsson". When omitted, the chosen algorithm’s own default controller is used — fixed for schemes without an embedded error estimate, otherwise PI for explicit Runge–Kutta and Gustafsson for the implicit families. See Choosing an Algorithm and the algorithm defaults table in the API reference.

max_registers

Per-thread register cap forwarded to cuda.jit. Default None leaves register allocation to ptxas; capping trades spill traffic for more resident warps. See Making it Faster (basic).

blocksize

Threads per block for every launch; solve(blocksize=...) overrides one launch. Unset, the solver picks a block size itself.

mem_proportion

Proportion of VRAM (0.0–1.0) reserved for this solver’s allocations. Default None places the solver in the automatic pool, which shares the VRAM remaining after manually-proportioned solvers are accounted for equally between its members. See GPU Memory Management.