fhelium.runtime.memory

Cross-platform snapshots of host and CUDA memory availability.

MemorySnapshot

class View source

MemorySnapshot(device: torch.device, observed_at_ns: int, capacity_bytes: int, available_bytes: int, torch_allocated_bytes: int | None = None, torch_reserved_bytes: int | None = None)

Record memory capacity and availability at one point in time.

CPU snapshots use psutil.virtual_memory, whose available value estimates memory that the operating system can provide without swapping. CUDA snapshots use torch.cuda.mem_get_info, whose free value covers the whole selected device, including use by other processes and non-PyTorch allocators. CUDA snapshots additionally report the current process's PyTorch caching-allocator allocation and reservation counters. observed_at_ns is a Unix timestamp in nanoseconds recorded after the counters are read.

Attributes

Name	Type	Default/value
device	torch.device
observed_at_ns	int
capacity_bytes	int
available_bytes	int
torch_allocated_bytes	int | None	None
torch_reserved_bytes	int | None	None

read

method

def read(device: str | torch.device='cpu') -> MemorySnapshot: ...

Read current host or CUDA memory counters for device.

as_dict

method

def as_dict() -> dict[str, object]: ...

Return a serializable representation of the snapshot.