flashinfer.comm._ulysses_lowp.boundary_descriptors

flashinfer.comm._ulysses_lowp.boundary_descriptors(grouped_amax: Tensor, *, rank: int, local_sequence: int, group: int, world_size: int) → Tensor

Extract the [B, H, 2] boundary descriptor for one stats collective.

slot0 is the first touched group’s partial amax and slot1 the last’s. When the rank touches a single group both slots carry the same value; the downstream max merge is idempotent, so no dedup special case is needed.

Parameters:

• grouped_amax – FP32 local maxima [B, H, slots(local_sequence, group)].

• rank – Sequence-shard rank in [0, world_size).

• local_sequence – Number of tokens per source shard, including zero padding.

• group – Positive token count per quantization group.

• world_size – Ulysses group size, one of 2, 4, 6 or 8; 6 is supported structurally but has not been validated.

Returns:

FP32 CUDA boundary maxima of shape [B, H, 2].