Behavior banks and population runners

Module contract

Behavior-bank quantization stores component scales and W4/W8 values separately from the shared model. Quantization is a deployment transformation, while training uses floating-point parameters. The low-level GRU population runners and deployment records expose the contracts used for direct runtime construction.

Source revision: 306a51ed950b. Public export index.

QuantizedNeuronBehaviorBank

Source

Physical low-bit storage for a logical behavior-adaptation bank.

QuantizedNeuronBehaviorBank(values: torch.Tensor, scales: torch.Tensor, bits: int, logical_parameters_per_neuron: int) -> None

Fields

Parameter	Type	Default
values	torch.Tensor	required
scales	torch.Tensor	required
bits	int	required
logical_parameters_per_neuron	int	required

bits: Quantization precision, restricted to the supported bit widths.

QuantizedNeuronBehaviorBank.storage_bytes

Source

QuantizedNeuronBehaviorBank.storage_bytes: int

Read-only property. Access as instance.storage_bytes; do not call it as a function.

Returns int.

MixedQuantizedNeuronBehaviorBank

Source

Component-wise W4/W8 storage for a logical adaptation bank.

MixedQuantizedNeuronBehaviorBank(values: torch.Tensor, scales: torch.Tensor, byte_slices: dict[str, slice], component_bits: dict[str, int], logical_parameters_per_neuron: int) -> None

Fields

Parameter	Type	Default
values	torch.Tensor	required
scales	torch.Tensor	required
byte_slices	dict[str, slice]	required
component_bits	dict[str, int]	required
logical_parameters_per_neuron	int	required

MixedQuantizedNeuronBehaviorBank.storage_bytes

Source

MixedQuantizedNeuronBehaviorBank.storage_bytes: int

Read-only property. Access as instance.storage_bytes; do not call it as a function.

Returns int.

quantize_neuron_behavior_parameters

Source

quantize_neuron_behavior_parameters(parameters: torch.Tensor, *, adaptation: NeuronBehaviorAdaptation, bits: int) -> QuantizedNeuronBehaviorBank

Quantize each adaptation component with one symmetric scale.

parameters is floating-point (N,adaptation.parameter_count). bits is 4 or 8. W4 requires even logical width and packs two values per byte. Returns physical values and one scale per adaptation component.

Parameter	Type	Default
parameters	torch.Tensor	required
adaptation	NeuronBehaviorAdaptation	required
bits	int	required

bits: Quantization precision, restricted to the supported bit widths.

Returns QuantizedNeuronBehaviorBank.

quantize_mixed_neuron_behavior_parameters

Source

quantize_mixed_neuron_behavior_parameters(parameters: torch.Tensor, *, adaptation: NeuronBehaviorAdaptation, component_bits: dict[str, int]) -> MixedQuantizedNeuronBehaviorBank

Quantize each adaptation component to its declared W4/W8 tier.

parameters is floating-point (N,adaptation.parameter_count). component_bits declares W4/W8 for each named adaptation component. Returns packed byte slices and scale metadata needed for reconstruction.

Parameter	Type	Default
parameters	torch.Tensor	required
adaptation	NeuronBehaviorAdaptation	required
component_bits	dict[str, int]	required

Returns MixedQuantizedNeuronBehaviorBank.

dequantize_neuron_behavior_parameters

Source

dequantize_neuron_behavior_parameters(bank: QuantizedNeuronBehaviorBank, *, adaptation: NeuronBehaviorAdaptation) -> torch.Tensor

Materialize a low-bit behavior bank for validation or export.

Parameter	Type	Default
bank	QuantizedNeuronBehaviorBank	required
adaptation	NeuronBehaviorAdaptation	required

Returns torch.Tensor.

dequantize_mixed_neuron_behavior_parameters

Source

dequantize_mixed_neuron_behavior_parameters(bank: MixedQuantizedNeuronBehaviorBank, *, adaptation: NeuronBehaviorAdaptation) -> torch.Tensor

Materialize a component-wise W4/W8 adaptation bank.

Parameter	Type	Default
bank	MixedQuantizedNeuronBehaviorBank	required
adaptation	NeuronBehaviorAdaptation	required

Returns torch.Tensor.

pack_neuron_behavior_adapter

Source

pack_neuron_behavior_adapter(adapter: NeuronBehaviorAdapter, *, adaptation: NeuronBehaviorAdaptation) -> torch.Tensor

Pack a trained behavior adapter into the population ABI.

Parameter	Type	Default
adapter	NeuronBehaviorAdapter	required
adaptation	NeuronBehaviorAdaptation	required

Returns torch.Tensor.

GRUPopulationRunner

Source

Compiled persistent-state inference for a shared-weight GRU population.

GRUPopulationRunner.init

Source

__init__(self, model: AxoTemporalModel, *, population: int, chunk_size: int, compile_step: bool=True, compile_mode: str='reduce-overhead', population_embedding_adapters: torch.Tensor | None=None) -> None

Parameter	Type	Default
model	AxoTemporalModel	required
population	int	required
chunk_size	int	required
compile_step	bool	True
compile_mode	str	'reduce-overhead'
population_embedding_adapters	torch.Tensor | None	None

Returns None.

GRUPopulationRunner.recurrent_state

Source

GRUPopulationRunner.recurrent_state: torch.Tensor

Read-only property. Access as instance.recurrent_state; do not call it as a function.

Returns torch.Tensor.

GRUPopulationRunner.patch_sum

Source

GRUPopulationRunner.patch_sum: torch.Tensor | None

Read-only property. Access as instance.patch_sum; do not call it as a function.

Returns torch.Tensor \| None.

GRUPopulationRunner.patch_correction

Source

GRUPopulationRunner.patch_correction: torch.Tensor | None

Read-only property. Access as instance.patch_correction; do not call it as a function.

Returns torch.Tensor \| None.

GRUPopulationRunner.patch_position

Source

GRUPopulationRunner.patch_position: int

Read-only property. Access as instance.patch_position; do not call it as a function.

Returns int.

GRUPopulationRunner.step

Source

step(self, event_indices: torch.Tensor, event_values: torch.Tensor, morphology_indices: torch.Tensor) -> torch.Tensor

Parameter	Type	Default
event_indices	torch.Tensor	required
event_values	torch.Tensor	required
morphology_indices	torch.Tensor	required

Returns torch.Tensor.

GRUBranchPopulationRunner

Source

Compiled GRU inference from pre-aggregated raw branch inputs.

Bases: GRUPopulationRunner.

GRUBranchPopulationRunner.init

Source

__init__(self, model: AxoTemporalModel, *, population: int, chunk_size: int, compile_step: bool=True, compile_mode: str='reduce-overhead', population_embedding_adapters: torch.Tensor | None=None) -> None

Parameter	Type	Default
model	AxoTemporalModel	required
population	int	required
chunk_size	int	required
compile_step	bool	True
compile_mode	str	'reduce-overhead'
population_embedding_adapters	torch.Tensor | None	None

Returns None.

GRUBranchPopulationRunner.step

Source

step(self, branch_inputs: torch.Tensor, morphology_indices: torch.Tensor) -> torch.Tensor

Parameter	Type	Default
branch_inputs	torch.Tensor	required
morphology_indices	torch.Tensor	required

Returns torch.Tensor.

AdaptedGRUBranchPopulationRunner

Source

Packed per-neuron behavior adaptation around one shared GRU base.

Bases: GRUBranchPopulationRunner.

AdaptedGRUBranchPopulationRunner.init

Source

__init__(self, model: AxoTemporalModel, *, population: int, chunk_size: int, behavior_parameters: torch.Tensor | QuantizedNeuronBehaviorBank | MixedQuantizedNeuronBehaviorBank, adaptation: NeuronBehaviorAdaptation, compile_step: bool=True, compile_mode: str='max-autotune-no-cudagraphs') -> None

Parameter	Type	Default
model	AxoTemporalModel	required
population	int	required
chunk_size	int	required
behavior_parameters	torch.Tensor | QuantizedNeuronBehaviorBank | MixedQuantizedNeuronBehaviorBank	required
adaptation	NeuronBehaviorAdaptation	required
compile_step	bool	True
compile_mode	str	'max-autotune-no-cudagraphs'

Returns None.

AdaptedGRUBranchPopulationRunner.step

Source

step(self, branch_inputs: torch.Tensor, morphology_indices: torch.Tensor) -> torch.Tensor

Parameter	Type	Default
branch_inputs	torch.Tensor	required
morphology_indices	torch.Tensor	required

Returns torch.Tensor.

GroupedGRUPopulationRunner

Source

Compiled GRU inference with one recurrent parameter set per group.

GroupedGRUPopulationRunner.init

Source

__init__(self, model: AxoTemporalModel, *, population: int, groups: int, chunk_size: int, compile_step: bool=True, compile_mode: str='reduce-overhead') -> None

Parameter	Type	Default
model	AxoTemporalModel	required
population	int	required
groups	int	required
chunk_size	int	required
compile_step	bool	True
compile_mode	str	'reduce-overhead'

Returns None.

GroupedGRUPopulationRunner.resident_parameter_count

Source

GroupedGRUPopulationRunner.resident_parameter_count: int

Read-only property. Access as instance.resident_parameter_count; do not call it as a function.

Returns int.

GroupedGRUPopulationRunner.patch_position

Source

GroupedGRUPopulationRunner.patch_position: int

Read-only property. Access as instance.patch_position; do not call it as a function.

Returns int.

GroupedGRUPopulationRunner.step

Source

step(self, event_indices: torch.Tensor, event_values: torch.Tensor, morphology_indices: torch.Tensor) -> torch.Tensor

Parameter	Type	Default
event_indices	torch.Tensor	required
event_values	torch.Tensor	required
morphology_indices	torch.Tensor	required

Returns torch.Tensor.

GroupedGRUPopulationRunner.load_group_recurrent_parameters

Source

load_group_recurrent_parameters(self, group: int, source: AxoTemporalModel) -> None

Load one group’s recurrent weights from a compatible GRU model.

Parameter	Type	Default
group	int	required
source	AxoTemporalModel	required

Returns None.

NeuronBehaviorAdaptation

Source

Mechanism-level parameters that remain unique to each neuron.

NeuronBehaviorAdaptation(width: int, branches: int, outputs: int, matrix_rank: int, adapt_recurrent: bool, branch_token_offset: bool = False) -> None

Fields

Parameter	Type	Default
width	int	required
branches	int	required
outputs	int	required
matrix_rank	int	required
adapt_recurrent	bool	required
branch_token_offset	bool	False

NeuronBehaviorAdaptation.tensor_shapes

Source

NeuronBehaviorAdaptation.tensor_shapes: dict[str, tuple[int, ...]]

Read-only property. Access as instance.tensor_shapes; do not call it as a function.

Returns dict[str, tuple[int, ...]].

NeuronBehaviorAdaptation.tensor_slices

Source

NeuronBehaviorAdaptation.tensor_slices: dict[str, slice]

Read-only property. Access as instance.tensor_slices; do not call it as a function.

Returns dict[str, slice].

NeuronBehaviorAdaptation.parameter_counts

Source

NeuronBehaviorAdaptation.parameter_counts: dict[str, int]

Read-only property. Access as instance.parameter_counts; do not call it as a function.

Returns dict[str, int].

NeuronBehaviorAdaptation.parameter_count

Source

NeuronBehaviorAdaptation.parameter_count: int

Read-only property. Access as instance.parameter_count; do not call it as a function.

Returns int.

NeuronBehaviorAdaptation.adapted_components

Source

NeuronBehaviorAdaptation.adapted_components: frozenset[str]

Read-only property. Access as instance.adapted_components; do not call it as a function.

Returns frozenset[str].

NeuronBehaviorAdaptation.shared_components

Source

NeuronBehaviorAdaptation.shared_components: frozenset[str]

Read-only property. Access as instance.shared_components; do not call it as a function.

Returns frozenset[str].

PopulationDeploymentMeasurement

Source

One complete population-simulation measurement.

PopulationDeploymentMeasurement(population: int, adapted_parameters_per_neuron: int, adapted_components: tuple[str, ...], shared_components: tuple[str, ...], unique_adaptation_per_neuron: bool, adaptation_quality_validated: bool, shared_frozen_base: bool, output_materialized: bool, routing_feeds_model: bool, model_latency_ms: float, routing_latency_ms: float | None, complete_stack_latency_ms: float | None, model_peak_bytes: int, routing_peak_bytes: int | None, complete_stack_peak_bytes: int | None, local_source_identity_retained: bool = False, long_range_source_identity_retained: bool = False, morphology_assignment_independent: bool = False, activity_generation_mode: str = 'unknown', activity_dynamics_validated: bool = False, adaptation_execution_mode: str = 'direct', adaptation_reference_envelope_validated: bool = False, forecast_block_size: int = 1, complete_block_latency_ms: float | None = None) -> None

Fields

Parameter	Type	Default
population	int	required
adapted_parameters_per_neuron	int	required
adapted_components	tuple[str, ...]	required
shared_components	tuple[str, ...]	required
unique_adaptation_per_neuron	bool	required
adaptation_quality_validated	bool	required
shared_frozen_base	bool	required
output_materialized	bool	required
routing_feeds_model	bool	required
model_latency_ms	float	required
routing_latency_ms	float | None	required
complete_stack_latency_ms	float | None	required
model_peak_bytes	int	required
routing_peak_bytes	int | None	required
complete_stack_peak_bytes	int | None	required
local_source_identity_retained	bool	False
long_range_source_identity_retained	bool	False
morphology_assignment_independent	bool	False
activity_generation_mode	str	'unknown'
activity_dynamics_validated	bool	False
adaptation_execution_mode	str	'direct'
adaptation_reference_envelope_validated	bool	False
forecast_block_size	int	1
complete_block_latency_ms	float | None	None

PopulationDeploymentTarget

Source

Requirements for a defensible real-time population claim.

PopulationDeploymentTarget(headline_population: int, benchmark_population: int, biological_step_ms: float, adaptation: NeuronBehaviorAdaptation, shared_frozen_base: bool, output_materialization_required: bool, network_routing_required: bool, local_source_identity_required: bool, long_range_source_identity_required: bool, independent_morphology_assignment_required: bool, activity_dynamics_validation_required: bool, quota_activity_allowed: bool, forecast_block_size: int, firing_rate: float, fanout: int) -> None

Fields

Parameter	Type	Default
headline_population	int	required
benchmark_population	int	required
biological_step_ms	float	required
adaptation	NeuronBehaviorAdaptation	required
shared_frozen_base	bool	required
output_materialization_required	bool	required
network_routing_required	bool	required
local_source_identity_required	bool	required
long_range_source_identity_required	bool	required
independent_morphology_assignment_required	bool	required
activity_dynamics_validation_required	bool	required
quota_activity_allowed	bool	required
forecast_block_size	int	required
firing_rate	float	required
fanout	int	required

PopulationDeploymentTarget.adapted_parameters_per_neuron

Source

PopulationDeploymentTarget.adapted_parameters_per_neuron: int

Read-only property. Access as instance.adapted_parameters_per_neuron; do not call it as a function.

Returns int.

PopulationDeploymentTarget.validation_errors

Source

validation_errors(self, measurement: PopulationDeploymentMeasurement) -> tuple[str, ...]

Parameter	Type	Default
measurement	PopulationDeploymentMeasurement	required

Returns tuple[str, ...].

PopulationDeploymentTarget.evaluate

Source

evaluate(self, measurement: PopulationDeploymentMeasurement) -> dict[str, int | float | bool]

Parameter	Type	Default
measurement	PopulationDeploymentMeasurement	required

Returns dict[str, int \| float \| bool].