pub fn hc_post_site(
gpu: &dyn GpuBackend,
kernel: KernelHandle,
hc: &HcWeights,
block_out: DevicePtr,
residual: DevicePtr,
post: DevicePtr,
comb: DevicePtr,
out: DevicePtr,
num_tokens: u32,
hidden_size: u32,
stream: u64,
) -> Result<()>Expand description
2026-09-25: Inject the block output back into every stream. out may alias residual.
Takes the whole HcWeights rather than a site because neither variant’s hc_post reads
site weights: Sinkhorn reads the comb its hc_pre emitted, low-rank the injection vector.
The variant is taken from the layer’s attention site.