Language Language English 简体中文

unilab.ipc.replay_pipelines.transfer.cuda_like.CudaLikeReplayTransferBackend

class unilab.ipc.replay_pipelines.transfer.cuda_like.CudaLikeReplayTransferBackend

Bases: object

Pinned host to CUDA-like device transfer backend.

PyTorch ROCm exposes the same torch.cuda surface for runtime streams and events, so this backend intentionally keys on the PyTorch device type instead of NVIDIA-specific platform names.

Parameters:

• device (device)

• ring_depth (int)

Methods

__init__(*, device, ring_depth)
allocate_device_slots(*, count, shape, dtype)
clear_ready(slot)
close()
ready_query(slot)
register_host_slots(slots)
submit_h2d(*, slot, dst, src, metadata, ...)
synchronize_ready(slot)
wait_current_stream_for_ready(slot)

Attributes

host_memory_kind
supports_async_submit
supports_timing_events

host_memory_kind = 'registered_pinned_shared'

supports_async_submit = True

supports_timing_events = True

__init__(*, device, ring_depth)

Parameters:

• device (device)

• ring_depth (int)

register_host_slots(slots)

Parameters:

slots (list[Tensor])

Return type:

None

allocate_device_slots(*, count, shape, dtype)

Parameters:

• count (int)

• shape (tuple[int, int])

• dtype (dtype)

Return type:

list[Tensor]

submit_h2d(*, slot, dst, src, metadata, trace_recorder, trace_cuda_events, h2d_bytes, pack_layout, pack_executor)

Parameters:

• slot (int)

• dst (Tensor)

• src (Tensor)

• metadata (ReplayTickMetadata | None)

• trace_cuda_events (bool)

• h2d_bytes (int)

• pack_layout (str)

• pack_executor (str)

Return type:

float

clear_ready(slot)

Parameters:

slot (int)

Return type:

None

ready_query(slot)

Parameters:

slot (int)

Return type:

bool

synchronize_ready(slot)

Parameters:

slot (int)

Return type:

None

wait_current_stream_for_ready(slot)

Parameters:

slot (int)

Return type:

None

close()

Return type:

None