FHElium

Appearance

fhelium.distributed._limb_collectives ​

Structural scatter and gather collectives for ciphertext RNS limbs.

scatter_ciphertext_limbs ​

function View source

python
def scatter_ciphertext_limbs(value_or_none: Ciphertext | None, *, limb_ranges: Sequence[tuple[int, int]] | None=None, src: int=0, group: torch.distributed.ProcessGroup | None=None) -> Ciphertext: ...

Slice and scatter caller-selected RNS intervals of one ciphertext.

The source provides one nonempty half-open interval per group rank. The intervals refer to positions on the source's limb axis, form a consecutive range in group-rank order, and may have unequal lengths. slice_limbs selects both Tensor rows and their declared prime_ids. The caller chooses the partition; the collective performs no arithmetic or balancing. It is synchronous and accepts no async_op argument.

Parameters

  • value_or_none: Source ciphertext on src; None on every other rank. The source may itself represent an interval of the full basis.
  • limb_ranges: On src, one (start, stop) pair per process-group rank, indexing the source's stored rows rather than global prime IDs. Other ranks supply None. The intervals may select a consecutive portion of the source without covering every row.
  • src: Global rank of the source process, which must belong to group.
  • group: Participating process group. None selects the default group.

Returns

The local ciphertext shard with its selected prime IDs and unchanged arithmetic state. On src, it shares the source Tensor storage; other ranks receive newly allocated storage. World size one returns the requested source view without communication or copying.

Raises

  • ValueError: If the source rank or source/non-source arguments are invalid; the range count differs from group size; or intervals are empty, out of bounds, or not consecutive in group-rank order.
  • RuntimeError: If distributed communication is uninitialized or the caller is not a member of group.

gather_ciphertext_limbs ​

function View source

python
def gather_ciphertext_limbs(local_shard: Ciphertext, *, dst: int=0, group: torch.distributed.ProcessGroup | None=None) -> Ciphertext | None: ...

Gather and reconstruct one ciphertext from disjoint RNS limb shards.

Group-rank order defines prime-interval order. On dst, compatible, nonempty, contiguous parameter intervals are concatenated along the RNS limb dimension. This is structural reconstruction, not ciphertext addition; use reduce_ciphertext for additive partials. The operation is synchronous, accepts no async_op argument, and returns no torch.distributed.Work.

Parameters

  • local_shard: Caller-owned rank-local ciphertext shard. It is read but not mutated.
  • dst: Global rank that reconstructs the ciphertext, which must belong to group. This is not a process-group-relative rank.
  • group: Participating process group. None selects the default process group.

Returns

The reconstructed ciphertext on dst and None on every other rank. The destination allocates receive buffers and the concatenated result. For world size one, the validated local_shard object is returned unchanged without concatenation.

Raises

  • ValueError: If dst is outside group or shards differ in logical ciphertext metadata, dtype, device, or contiguous prime layout.
  • RuntimeError: If distributed communication is uninitialized or the caller is not a member of group.