core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order

Weak, phase-tagged execution sequences with a sequential cursor.

Module Contents

Classes

ExecutionEvent	A module and its compute phase.
ExecutionOrder	Store an execution trace with one sequential cursor across all phases.

API

class core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent

Bases: typing.NamedTuple

A module and its compute phase.

module: torch.nn.Module

None

phase: core.distributed.fsdp.src.megatron_fsdp.experimental.phase.Phase

None

class core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionOrder(

items: collections.abc.Iterable[core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent] = (),

)

Store an execution trace with one sequential cursor across all phases.

Initialization

Create a weak sequence, preserving every repeated occurrence.

property position: int

Index of the next event to consume.

append(

item: core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent,

) → None

Append an event without consuming it.

__iter__() → collections.abc.Iterator[core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent]

Iterate over live occurrences without retaining their modules.

peek(

offset: int = 0,

) → core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent | None

Read an event relative to the cursor without wrapping or advancing.

successors() → collections.abc.Iterator[core.distributed.fsdp.src.megatron_fsdp.experimental.execution_order.ExecutionEvent]

Look ahead from the current occurrence without wrapping or advancing.

advance() → None

Consume one occurrence and automatically wrap at the end of the trace.