flashinfer.comm._ulysses_lowp.finalize_stats

flashinfer.comm._ulysses_lowp.finalize_stats(gathered: Tensor, ctx: StatsContext, k: Tensor, *, enable_pdl: bool | None = None) → V2GStats

Turn the gathered statistics into final quantization inputs.

gathered is the AllGather of local_stats sends, rank-major ([world_size * N] or [world_size, N] fp32); k is the same local K shard passed to local_stats (protocol 2 computes its grouped amax here, after the global mean exists). The K mean divides by the LIVE row count (used_sequence): tail padding contributes exactly 0 to the gathered sum, so dividing by the padded length would bias the mean toward zero; with no padding the two are equal. Reductions run in fixed rank-major layout (sum(dim=0) / amax(dim=0)), so results are bit-identical on every rank.

Parameters:

  • gathered – AllGather output in group-rank order, FP32 [world_size, N] or flattened, where N is the local stats send size.

  • ctx – Context returned by local_stats on this rank.

  • k – The same local BF16/FP16 K shard used for local_stats.

  • enable_pdl – Whether to enable Programmatic Dependent Launch. None selects according to device support.

Returns:

V2GStats with global mean, scales, finalized Q/K maxima and shard geometry.