Skip to content
TIDE

Module: tide.storage

Snapshot storage helpers for backward propagation.

Functions

Section titled “Functions”
  • storage_mode_to_int

Storage Modes

Section titled “Storage Modes”
  • device: snapshots stay on compute device, fastest and most memory-heavy
  • cpu: host-backed storage, lower device memory use
  • disk: file-backed storage, smallest memory footprint and highest latency
  • none: disable snapshot storage

The user-facing APIs also accept auto, which chooses mode based on configured byte limits.

storage_mode_to_int

Section titled “storage_mode_to_int”

Maps mode strings to backend integer constants used by native kernels.

Compression

Section titled “Compression”
  • storage_compression supports none and bf16 on default compute path
  • bf16 storage is valid for float32 workflows

See Also

Section titled “See Also”

Structured storage policy

Section titled “Structured storage policy”

Application code should construct StorageOptions rather than call backend integer helpers:

storage = tide.StorageOptions(
    mode=tide.StorageMode.CPU,
    compression=False,
    bytes_limit_device=None,
    bytes_limit_host=None,
    chunk_steps=0,
)

Internally, resolve_snapshot_storage converts this policy into an immutable SnapshotStorageSpec containing normalized mode, representation, sample count, and shot shape. SnapshotAllocator then owns the concrete device, host, or disk buffers. Disk allocations use tempfile.TemporaryDirectory; the allocator retains its lifetime and cleanup responsibility.

Capacity planning

Section titled “Capacity planning”

auto placement considers configured byte limits, not every allocation in the process. Reserve headroom for live fields, model gradients, optimizer state, receiver tensors, and runtime libraries. Validate the resolved mode during a small run before scaling the shot batch.