axelera.runtime.device_config

Device settings a model asks a board for before it runs.

Keyed by board MODEL (110m), not form factor (m2): a 110m and a 113m are both m2, so a form factor cannot say which of them a card's value was measured on. axdevice resolves the model and owns that mapping.

Applied per AI core, to the cores running that model, as the YAML/GStreamer path does with its static core ranges. Unlike that path, which binds a core to a task for the whole run and so writes each setting once, this scheduler shares cores, so a setting is rewritten whenever a core changes to a model that wants a different value.

Three sources, first that answers wins: AXELERA_CONFIGURE_BOARD (see env_override), then scheduler.create(device_config=...), then the model's own op.load(device_config=...). A model's config travels with the artifact; the scheduler argument belongs to the deployment; the variable belongs to whoever runs the binary.

Summary​

board_models​

board_models() -> frozenset[str]

Every board model a DeviceConfig may name, from axdevice's own table.

Form factors are excluded: m2 names both the 110m and the 113m. Empty without axelera-runtime, which leaves names unchecked on a machine that could not act on them anyway.

BoardConfig​

Settings a model wants on one board model; one field per SETTINGS entry.

Args:

• mvm_limitation: see SETTINGS. Hardware quantises to ~1.5% steps.

Constructor:

__init__(mvm_limitation: int | None = None)

DeviceConfig​

Settings a model wants, per board model. A board not named is left alone.

Args:

• boards: board model (one of board_models()) to its settings. An unknown name raises, so a typo fails where it is written.

Example:

limits = DeviceConfig({'110m': BoardConfig(mvm_limitation=57)})
op.load('yolov8l-pose-coco.axm', device_config=limits)

Constructor:

__init__(boards: dict[str, BoardConfig] = field(default_factory=dict))

Methods:

as_spec​

as_spec() -> dict[str, dict[str, int]]

Plain data for a manifest. TOML has no null, so an unset value is absent.

classmethod from_spec​

from_spec(spec: dict[str, dict[str, int]]) -> DeviceConfig

Rebuild a DeviceConfig from as_spec output.

explain​

explain() -> list[str]

Return the reason recorded for each setting in use, for a generated recipe.

for_model​

for_model(model: str | None) -> BoardConfig | None

Return the entry for a board model, or None if this config names none.

Setting​

One setting's meaning, written once and rendered into generated recipes.

Args:

• summary: what the setting is and what values it takes.
• boards: board models a packaged model declares it for. Advisory, not a support claim.
• reason: optional, why. Printed above a recipe's DEVICE_CONFIG.

Constructor:

__init__(summary: str, boards: tuple[str, ...] = (), reason: str = '')

board_model​

board_model(device) -> str | None

Return the board model axdevice reports, or a form factor it cannot place.

env_override​

env_override() -> tuple[bool, int | None, int | None]

Read AXELERA_CONFIGURE_BOARD, returning (configure, clock, mvm).

The grammar is the one the YAML/GStreamer path already documents, so one variable governs both paths:

• 0: do not configure the board at all. configure comes back False and the model's own request is dropped.
• 1 (the default, and what an unset variable means): configure the board from the model's request. Both overrides come back None.
• [clock],[mvm]: the user's own values, either part may be empty. A value given here replaces the model's, so ,90 caps MVM at 90% whatever the model asked for and 400 sets the clock and leaves MVM to the model.

Raises:

• ValueError: if the variable is set to something this grammar does not read.