particula.execution.gpu_coupled_restart
Exact boundary restart for fixed GPU events and adaptive condensation.
Classes
| Name | Description |
|---|---|
GPUCoupledCheckpoint | Save event/RNG bytes and the complete deterministic boundary policy. |
Functions
| Name | Description |
|---|---|
_validate | Check the complete host envelope without allocating device state. |
checkpoint_coupled | Download exact continuation only after a successfully completed cycle. |
load_coupled_checkpoint | Load validated host records without physics or CUDA allocation. |
resume_coupled | Restore exact event streams and controls on the saved explicit device. |
save_coupled_checkpoint | Atomically save canonical host bytes without pickle or uploads. |
GPUCoupledCheckpointclassdataclass#
class GPUCoupledCheckpoint(schema_version: int, runtime: tuple, device: tuple, events: SimulationCheckpoint, configuration: object, controls: GPUCondensationControls, order: str, statistics: StepStatistics, counters: tuple[int, int, int, int], resident: bool)Save event/RNG bytes and the complete deterministic boundary policy.
No in-progress adaptive controller is carried across a fixed event boundary: every cycle deliberately resets it from the new accepted physical state. Future forcing and source schedules are supplied again by the caller.
configurationattributeinstance attribute#
configuration: objectcontrolsattributeinstance attribute#
controls: GPUCondensationControlscountersattributeinstance attribute#
counters: tuple[int, int, int, int]deviceattributeinstance attribute#
device: tupleeventsattributeinstance attribute#
events: SimulationCheckpointorderattributeinstance attribute#
order: strresidentattributeinstance attribute#
resident: boolruntimeattributeinstance attribute#
runtime: tupleschema_versionattributeinstance attribute#
schema_version: intstatisticsattributeinstance attribute#
statistics: StepStatistics__init__method#
def __init__(schema_version: int, runtime: tuple, device: tuple, events: SimulationCheckpoint, configuration: object, controls: GPUCondensationControls, order: str, statistics: StepStatistics, counters: tuple[int, int, int, int], resident: bool) -> None_validatefunction#
def _validate(saved)Check the complete host envelope without allocating device state.
checkpoint_coupledfunction#
def checkpoint_coupled(owner)Download exact continuation only after a successfully completed cycle.
load_coupled_checkpointfunction#
def load_coupled_checkpoint(path, *, max_bytes=512 * 1024 * 1024)Load validated host records without physics or CUDA allocation.
resume_coupledfunction#
def resume_coupled(saved, events, strategy, *, device)Restore exact event streams and controls on the saved explicit device.
save_coupled_checkpointfunction#
def save_coupled_checkpoint(saved, path, *, compressed=True)Atomically save canonical host bytes without pickle or uploads.