Language

unilab.ipc.replay_pipelines.transfer.cuda_like¶

CUDA/ROCm replay transfer backend.

Classes

CudaLikeReplayTransferBackend

Pinned host to CUDA-like device transfer backend.

class unilab.ipc.replay_pipelines.transfer.cuda_like.CudaLikeReplayTransferBackend[source]¶

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)

host_memory_kind = 'registered_pinned_shared'¶

supports_async_submit = True¶

supports_timing_events = True¶

__init__(*, device, ring_depth)[source]¶

Parameters:

device (device)
ring_depth (int)

register_host_slots(slots)[source]¶

Parameters:: slots (list[Tensor])
Return type:: None

allocate_device_slots(*, count, shape, dtype)[source]¶

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)[source]¶

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)[source]¶

Parameters:: slot (int)
Return type:: None

ready_query(slot)[source]¶

Parameters:: slot (int)
Return type:: bool

synchronize_ready(slot)[source]¶

Parameters:: slot (int)
Return type:: None

wait_current_stream_for_ready(slot)[source]¶

Parameters:: slot (int)
Return type:: None

close()[source]¶

Return type:: None