Enum StepTpGateShards 

pub enum StepTpGateShards<'a> {
    F32(&'a [CudaSlice<f32>]),
    Bf16(&'a [CudaSlice<u8>]),
}

Persistent workspace of the v2 rank-local decode-attention driver.

Buffers live in their producing rank’s CUDA context, are never freed, and events are re-recorded per call — the pp.rs BoundarySlot discipline — so the per-token path has no cuMemAlloc, no cross-stream free, and no host round-trip. Every buffer is fully overwritten before its consumers run in the same call; nothing carries state between tokens. Per-rank attn_gate row shards for the fused QKV+gate kernel, in the weight class the fused kernels read (F32 mirror or raw checkpoint bf16).

Variants

F32(&'a [CudaSlice<f32>])

Bf16(&'a [CudaSlice<u8>])