Spaces:
Running
Running
File size: 10,786 Bytes
31226fd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 | """Incremental hot-path collector for runtime persistence.
The original :class:`StorageSession` remains the conservative reference path.
This collector is used by the asynchronous operator/runtime path when
``capture_policy=dirty_tracking``. It consumes the network's dirty sets after
all hooks registered before storage have run and avoids rebuilding the complete
synapse map on every tick.
Neuron membrane state is special: an active spiking neuron legitimately changes
on almost every tick. ``neuron_state_interval_ticks`` therefore controls an
explicit persistence trade-off:
* ``1`` keeps per-tick neuron-state capture for exact restart-oriented runs.
* ``>1`` captures neuron state at a bounded cadence for interactive/operator
runs while spike, synapse and topology deltas remain event-driven.
Using an interval greater than one is an engineering performance mode, not a
claim of per-tick restart equivalence.
"""
from __future__ import annotations
from collections.abc import MutableSet
from typing import cast
from .delta_codec import (
NeuronAddDelta,
NeuronRemoveDelta,
NeuronStateDelta,
SpikeEventDelta,
SynapseAddDelta,
SynapseRemoveDelta,
SynapseWeightDelta,
encode_neuron_add,
encode_neuron_remove,
encode_neuron_state,
encode_spike_event,
encode_synapse_add,
encode_synapse_remove,
encode_synapse_weight,
)
from .delta_journal import DeltaRecord
from .optical_codec import state_from_neuron
from .runtime import (
RuntimeNetworkLike,
RuntimeSynapseLike,
StepResultLike,
StorageRuntimeConfig,
StorageSession,
)
class IncrementalStorageSession(StorageSession):
"""Storage collector whose per-tick synapse work is proportional to dirties."""
def __init__(
self,
network: RuntimeNetworkLike,
config: StorageRuntimeConfig,
*,
neuron_state_interval_ticks: int = 1,
) -> None:
super().__init__(network, config)
if neuron_state_interval_ticks <= 0:
raise ValueError("neuron_state_interval_ticks must be positive")
self.neuron_state_interval_ticks = int(neuron_state_interval_ticks)
def prime(self) -> None:
"""Prime fingerprints and discard construction-time dirty markers."""
super().prime()
self._consume_network_dirty_sets()
def _network_dirty_sets(
self, result: StepResultLike
) -> tuple[set[int], set[tuple[int, int]]]:
neuron_dirty = getattr(self.network, "_dirty_neuron_ids", None)
synapse_dirty = getattr(self.network, "_dirty_synapse_ids", None)
if isinstance(neuron_dirty, set):
neuron_ids = {
int(cast(int | float | str | bytes | bytearray, value))
for value in cast(set[object], neuron_dirty)
}
else:
neuron_ids = {int(value) for value in result.dirty_neuron_ids}
if isinstance(synapse_dirty, set):
synapse_ids = {
(
int(cast(int | float | str | bytes | bytearray, source_id)),
int(cast(int | float | str | bytes | bytearray, target_id)),
)
for source_id, target_id in cast(
set[tuple[object, object]], synapse_dirty
)
}
else:
synapse_ids = {
(int(source_id), int(target_id))
for source_id, target_id in result.dirty_synapse_ids
}
return neuron_ids, synapse_ids
def _consume_network_dirty_sets(self) -> None:
"""Clear dirty markers already consumed by this storage hook.
Hooks that execute after storage can mark the sets again; those changes
are then observed on the next storage callback.
"""
for attribute in ("_dirty_neuron_ids", "_dirty_synapse_ids"):
values = getattr(self.network, attribute, None)
if isinstance(values, MutableSet):
values.clear()
def _current_synapse(
self, source_id: int, target_id: int
) -> RuntimeSynapseLike | None:
outgoing = self.network.synapses.get(source_id, ())
for synapse in outgoing:
if int(synapse.target_id) == target_id:
return synapse
return None
def collect_deltas(self, result: StepResultLike) -> tuple[DeltaRecord, ...]:
"""Collect one tick without a full O(E) topology reconstruction."""
tick = int(result.tick)
deltas: list[DeltaRecord] = []
dirty_neurons, dirty_synapses = self._network_dirty_sets(result)
try:
for neuron_id in sorted(dirty_neurons):
current_neuron = self.network.neurons.get(neuron_id)
previous_neuron = self._neurons.get(neuron_id)
if current_neuron is None and previous_neuron is not None:
deltas.append(
encode_neuron_remove(tick, NeuronRemoveDelta(neuron_id))
)
self._neurons.pop(neuron_id, None)
self._topology_deltas += 1
elif current_neuron is not None and previous_neuron is None:
optical = state_from_neuron(current_neuron)
deltas.append(
encode_neuron_add(
tick,
NeuronAddDelta(
neuron_id=neuron_id,
tick=tick,
optical=optical,
a=float(current_neuron.a),
b=float(current_neuron.b),
c=float(current_neuron.c),
d=float(current_neuron.d),
spike_cost=float(current_neuron.spike_cost),
spike_counter=int(current_neuron.spike_counter),
last_spike_tick=int(current_neuron.last_spike_tick),
),
)
)
self._neurons[neuron_id] = self._neuron_fingerprint(current_neuron)
self._topology_deltas += 1
capture_neuron_state = (
self.neuron_state_interval_ticks == 1
or (tick + 1) % self.neuron_state_interval_ticks == 0
)
if capture_neuron_state:
# O(N) at the declared cadence. There is no honest O(changes)
# shortcut for v/u because membrane state changes continuously.
for neuron_id, neuron in self.network.neurons.items():
numeric_id = int(neuron_id)
neuron_fingerprint = self._neuron_fingerprint(neuron)
previous_neuron_fingerprint = self._neurons.get(numeric_id)
if (
previous_neuron_fingerprint is not None
and neuron_fingerprint != previous_neuron_fingerprint
):
deltas.append(
encode_neuron_state(
tick,
NeuronStateDelta(
neuron_id=numeric_id,
membrane_v=neuron_fingerprint.v,
recovery_u=neuron_fingerprint.u,
energy=neuron_fingerprint.energy,
spike_counter=neuron_fingerprint.spike_counter,
last_spike_tick=neuron_fingerprint.last_spike_tick,
),
)
)
self._neuron_deltas += 1
self._neurons[numeric_id] = neuron_fingerprint
for source_id, target_id in sorted(dirty_synapses):
key = (source_id, target_id)
current_synapse = self._current_synapse(source_id, target_id)
previous_synapse = self._synapses.get(key)
if current_synapse is None:
if previous_synapse is not None:
deltas.append(
encode_synapse_remove(
tick,
SynapseRemoveDelta(
source_id=source_id, target_id=target_id
),
)
)
self._synapses.pop(key, None)
self._topology_deltas += 1
continue
synapse_fingerprint = self._synapse_fingerprint(current_synapse)
if previous_synapse is None:
deltas.append(
encode_synapse_add(
tick,
SynapseAddDelta(
source_id=source_id,
target_id=target_id,
weight=synapse_fingerprint.weight,
eligibility=synapse_fingerprint.eligibility,
delay=synapse_fingerprint.delay,
last_pre_spike=synapse_fingerprint.last_pre_spike,
),
)
)
self._topology_deltas += 1
elif synapse_fingerprint != previous_synapse:
deltas.append(
encode_synapse_weight(
tick,
SynapseWeightDelta(
source_id=source_id,
target_id=target_id,
weight=synapse_fingerprint.weight,
eligibility=synapse_fingerprint.eligibility,
last_pre_spike=synapse_fingerprint.last_pre_spike,
),
)
)
self._synapse_deltas += 1
self._synapses[key] = synapse_fingerprint
if self.config.capture_spike_events:
for neuron_id in result.spike_ids:
deltas.append(
encode_spike_event(
tick, SpikeEventDelta(neuron_id=int(neuron_id))
)
)
self._spike_events += 1
return tuple(deltas)
finally:
self._consume_network_dirty_sets()
__all__ = ["IncrementalStorageSession"]
|