Struct Glm5NextDsaWorkspace 

pub struct Glm5NextDsaWorkspace { /* private fields */ }

Scratch reused across decode steps. Allocated once per layer.

Implementations

impl Glm5NextDsaWorkspace

pub fn new( gpu: &dyn GpuBackend, cfg: &Glm5NextDsaConfig, max_rows: usize, ) -> Result<Self>

max_rows is the widest speculative verify this workspace serves. The projection buffers, attn_out and the selector’s OUTPUT row scale with it; the indexer staging rows, the block table and the selector’s within-pass temporaries stay per-row, because Glm5NextDsaLayer::decode_k still SELECTS one token at a time.