pub fn build_overlay(
gpu: &dyn GpuBackend,
kernels: &OverlayKernels,
slot: &OverlayRawSlot,
served_embed: DevicePtr,
served_lmhead: DevicePtr,
vocab: usize,
h: usize,
tied: bool,
stream: u64,
) -> Result<Option<EmbedOverlay>>Expand description
Stage 2 (set_lora_weights): row-diff against the served tables, compact the
override rows, build the slot_map, and free the Stage-1 raw scratch.
Ok(None) when the slot overrides nothing (silently inert = correct).
tied means the lm_head aliases the embed table (buffer aliasing OR a
quantized head derived from embed) — the caller then reuses the embed rows
for the logit recompute (overlay_tables), so no distinct lm_head build.