easydel.layers.attention_operator._attention_impl

class easydel.layers.attention_operator._attention_impl.AttentionImpl(metadata: AttentionMetadata)

Bases: ABC

create_stable_sharding(state_ps: Optional[PartitionSpec] = None, preserved_indices: List[int] = None, clone_ps: Optional[PartitionSpec] = None, dep: Optional[Union[PartitionSpec, bool]] = True) → Optional[PartitionSpec]

current_backend() → Literal['tpu', 'gpu', 'cpu']

abstract forward_cpu(*args, **kwargs) → AttentionOutput

abstract forward_cuda(*args, **kwargs) → AttentionOutput

abstract forward_gpu(*args, **kwargs) → AttentionOutput

abstract forward_native(*args, **kwargs) → AttentionOutput

abstract forward_rocm(*args, **kwargs) → AttentionOutput

abstract forward_tpu(*args, **kwargs) → AttentionOutput

abstract get_impl_metadata() → AttentionMetadata

abstract classmethod get_impl_name() → Union[str, Tuple[str]]

get_runtime_type(q: Array, BTHD: bool = True) → RuntimeType

static repeat_kv_heads(k: Array, v: Array, num_reps: int) → Tuple[Array, Array]

Repeats k and v heads to match q heads.

class easydel.layers.attention_operator._attention_impl.AttentionMetadata(runtime_dtype: Union[str, type[Any], numpy.dtype, jax._src.typing.SupportsDType], runtime_softmax_dtype: Union[str, type[Any], numpy.dtype, jax._src.typing.SupportsDType, NoneType] = None, sequence_axis_name: str = Ellipsis, mesh: Optional[jax._src.mesh.Mesh] = Ellipsis, platform: easydel.infra.etils.EasyDeLPlatforms = Ellipsis, backend: easydel.infra.etils.EasyDeLBackends = Ellipsis, partition_axis: eformer.escale.partition.constraints.PartitionAxis = Ellipsis, base_config: Optional[easydel.infra.base_config.EasyDeLBaseConfig] = None, scan_ring_attention: bool = Ellipsis, softmax_scale: float = Ellipsis, dropout_prob: float = Ellipsis, blocksize_q: int = Ellipsis, blocksize_k: int = Ellipsis, blocksize_b: int = Ellipsis)

Bases: object

backend: EasyDeLBackends = Ellipsis

base_config: Optional[EasyDeLBaseConfig] = None

blocksize_b: int = Ellipsis

blocksize_k: int = Ellipsis

blocksize_q: int = Ellipsis

dropout_prob: float = Ellipsis

classmethod from_config(config: EasyDeLBaseConfig, softmax_scale: float, dropout_prob: float = 0.0)

get_partition_specs(mode: RuntimeType, BTHD: bool = True)

mesh: Optional[Mesh] = Ellipsis

partition_axis: PartitionAxis = Ellipsis

platform: EasyDeLPlatforms = Ellipsis

runtime_dtype: Union[str, type[Any], dtype, SupportsDType]

runtime_softmax_dtype: Optional[Union[str, type[Any], dtype, SupportsDType]] = None

scan_ring_attention: bool = Ellipsis

sequence_axis_name: str = Ellipsis

set_attrs_carefully(attr_name: str, default: Optional[Any], pickup_name: Optional[str] = None, use_base_config: bool = True)

softmax_scale: float = Ellipsis

class easydel.layers.attention_operator._attention_impl.AttentionOutput(attention_weights: Optional[jax.Array] = None, attention_outputs: Optional[jax.Array] = None)

Bases: object

attention_outputs: Optional[Array] = None

attention_weights: Optional[Array] = None

class easydel.layers.attention_operator._attention_impl.AttentionRegistry

Bases: object

Registry for attention implementations.

classmethod create(impl_name: str, metadata: AttentionMetadata) → AttentionImpl

Create an instance of an attention implementation by name.

classmethod get(impl_name: str) → Type[AttentionImpl]

Get an attention implementation by name.

classmethod list_implementations() → List[str]

List all registered attention implementations.

classmethod register(impl_cls: Type[AttentionImpl]) → Type[AttentionImpl]

Decorator to register an attention implementation.

Example usage:

@AttentionRegistry.register class CustomAttention(AttentionImpl):

…

class easydel.layers.attention_operator._attention_impl.RuntimeType(value, names=None, *, module=None, qualname=None, type=None, start=1, boundary=None)

Bases: Enum

generation = 'generation'

normal = 'normal'