snnlab.sim.accelerator_forward

Complete declared API of the accelerator_forward module, with signatures, data fields, validation and source.

Bounded forward-only accelerator validation. This check compares graph and legacy forward outputs on the same accelerator under an explicit tolerance policy; it does not cover training, checkpoint trajectories or cross-device equality.

The signatures, defaults, fields, docstrings and implementation excerpts below are generated from the Python source. Annotations are shown as declared; unannotated means the source supplies no type annotation. These pages document callable surfaces, including legacy support utilities, without promising backend support for every declaration.

Symbol	Kind
ForwardAcceleratorResult	class
run_forward_accelerator_check	function
main	function

ForwardAcceleratorResult

View source

Result record for the bounded forward accelerator fixture and its conformance report.

Class decorators: dataclass(frozen=True).

Dataclass constructor parameters. Factory defaults are shown as field declarations; omit these arguments to create fresh values per instance:

ForwardAcceleratorResult(device: str, snnlang_reproducibility: ConformanceReport, legacy_graph_parity: ConformanceReport)

Declared fields, including fields inherited from local data classes:

Field	Annotation	Default	Meaning
`device`	`str`	required	Requested or resolved tensor execution device.
`snnlang_reproducibility`	`ConformanceReport`	required	Stored member of this data contract; see the class docstring and serialization methods.
`legacy_graph_parity`	`ConformanceReport`	required	Stored member of this data contract; see the class docstring and serialization methods.

ForwardAcceleratorResult.passed

View source

Decorators: property.

def ForwardAcceleratorResult.passed(self) -> bool

Return annotation: bool.

Return expressions (branch-dependent; names refer to the linked implementation):

self.snnlang_reproducibility.passed and self.legacy_graph_parity.passed

Implementation

def passed(self) -> bool:
        return self.snnlang_reproducibility.passed and self.legacy_graph_parity.passed

ForwardAcceleratorResult.to_dict

View source

def ForwardAcceleratorResult.to_dict(self) -> dict[str, Any]

Return annotation: dict[str, Any].

Return expressions (branch-dependent; names refer to the linked implementation):

{'device': self.device, 'passed': self.passed, 'tolerances': {'atol': ATOL, 'rtol': RTOL}, 'snnlang_reproducibility': self.snnlang_reproducibility.to_dict(), 'legacy_graph_parity': self.legacy_graph_parity.to_dict()}

Implementation

def to_dict(self) -> dict[str, Any]:
        return {
            "device": self.device,
            "passed": self.passed,
            "tolerances": {"atol": ATOL, "rtol": RTOL},
            "snnlang_reproducibility": self.snnlang_reproducibility.to_dict(),
            "legacy_graph_parity": self.legacy_graph_parity.to_dict(),
        }

ForwardAcceleratorResult.require_passed

View source

def ForwardAcceleratorResult.require_passed(self) -> None

Return annotation: None.

Implementation

def require_passed(self) -> None:
        self.snnlang_reproducibility.require_passed()
        self.legacy_graph_parity.require_passed()

Complete class implementation

class ForwardAcceleratorResult:
    device: str
    snnlang_reproducibility: ConformanceReport
    legacy_graph_parity: ConformanceReport

    @property
    def passed(self) -> bool:
        return self.snnlang_reproducibility.passed and self.legacy_graph_parity.passed

    def to_dict(self) -> dict[str, Any]:
        return {
            "device": self.device,
            "passed": self.passed,
            "tolerances": {"atol": ATOL, "rtol": RTOL},
            "snnlang_reproducibility": self.snnlang_reproducibility.to_dict(),
            "legacy_graph_parity": self.legacy_graph_parity.to_dict(),
        }

    def require_passed(self) -> None:
        self.snnlang_reproducibility.require_passed()
        self.legacy_graph_parity.require_passed()

run_forward_accelerator_check

View source

def run_forward_accelerator_check(device: str) -> ForwardAcceleratorResult

Run the bounded graph/legacy forward conformance fixture on the selected accelerator and return its named comparison report.

Parameter	Annotation	Default	Meaning
`device`	`str`	required	Requested or resolved tensor execution device.

Return annotation: ForwardAcceleratorResult.

Return expressions (branch-dependent; names refer to the linked implementation):

ForwardAcceleratorResult(device=str(resolved), snnlang_reproducibility=reproducibility, legacy_graph_parity=parity)

Explicit exceptions in this implementation; called helpers may raise additional errors:

Explicit exception expression
`ValueError('MPS was requested but is unavailable')`
`ValueError('CUDA was requested but is unavailable')`
`ValueError(f'accelerator device expected mps or cuda, got {device!r}')`

Implementation

def run_forward_accelerator_check(device: str) -> ForwardAcceleratorResult:
    resolved = torch.device(device)
    if resolved.type == "mps" and not torch.backends.mps.is_available():
        raise ValueError("MPS was requested but is unavailable")
    if resolved.type == "cuda" and not torch.cuda.is_available():
        raise ValueError("CUDA was requested but is unavailable")
    if resolved.type not in {"mps", "cuda"}:
        raise ValueError(f"accelerator device expected mps or cuda, got {device!r}")

    with tempfile.TemporaryDirectory(prefix="snnlang-forward-") as scratch:
        root = Path(scratch)
        first = _subprocess_forward("graph", device, root / "graph-first.pt")
        second = _subprocess_forward("graph", device, root / "graph-second.pt")
        legacy = _subprocess_forward("legacy", device, root / "legacy.pt")
        reproducibility = compare_conformance_layers(
            "snnlang-same-accelerator-forward",
            {"forward": first},
            {"forward": second},
            policies=_policies(first),
        )
        parity = compare_conformance_layers(
            "legacy-snnlang-same-accelerator-forward",
            {"forward": legacy},
            {"forward": first},
            policies=_policies(legacy),
        )
    return ForwardAcceleratorResult(
        device=str(resolved),
        snnlang_reproducibility=reproducibility,
        legacy_graph_parity=parity,
    )

main

View source

def main() -> None

Command entry point for the bounded accelerator validation utility.

Return annotation: None.

Return expressions (branch-dependent; names refer to the linked implementation):

None

Implementation

def main() -> None:
    parser = argparse.ArgumentParser(
        description="Check graph repeatability and legacy parity for one forward PING case."
    )
    parser.add_argument("--device", required=True, choices=("mps", "cuda"))
    parser.add_argument("--worker", choices=("graph", "legacy"), help=argparse.SUPPRESS)
    parser.add_argument("--output", type=Path, help=argparse.SUPPRESS)
    args = parser.parse_args()
    if args.worker:
        if args.output is None:
            parser.error("--worker requires --output")
        _run_worker(args.worker, args.device, args.output)
        return
    if args.output is not None:
        parser.error("--output requires --worker")
    result = run_forward_accelerator_check(args.device)
    print(json.dumps(result.to_dict(), indent=2, sort_keys=True))
    result.require_passed()

Constants and type aliases

Initial source expressions are shown, not evaluated runtime values. Legacy configuration may mutate module defaults.

Name	Annotation	Initial expression	Source
`ATOL`	`unannotated`	`1e-06`	Source
`RTOL`	`unannotated`	`1e-06`	Source

snnlab.sim.accelerator_forward

ForwardAcceleratorResultCopy Anchor Link

ForwardAcceleratorResult.passedCopy Anchor Link

ForwardAcceleratorResult.to_dictCopy Anchor Link

ForwardAcceleratorResult.require_passedCopy Anchor Link

run_forward_accelerator_checkCopy Anchor Link

mainCopy Anchor Link

Constants and type aliasesCopy Anchor Link

On this page

ForwardAcceleratorResult

ForwardAcceleratorResult.passed

ForwardAcceleratorResult.to_dict

ForwardAcceleratorResult.require_passed

run_forward_accelerator_check

main

Constants and type aliases