pub struct HcLowRank {
pub norm_w: DevicePtr,
pub down_w: DevicePtr,
pub up_w: DevicePtr,
pub inject_w: DevicePtr,
pub rank: usize,
}Expand description
Qwen3.8-Flash-Next’s LOW-RANK hyper-connection parameters for one site.
Present instead of — never alongside — the Sinkhorn hc_fn/hc_base/
hc_scale triple above. DeepSeek-V4 mixes the hc_mult streams with a
Sinkhorn-normalized matrix; Qwen mixes them through a rank-rank pair.
Both share the [T, hc_mult, H] highway and hc_mult, so the presence of
this struct — not the model name — is what selects the kernel.
All BF16, matching the checkpoint. See bench/qwen4_exp/ARCHITECTURE.md §1.
Fields§
§norm_w: DevicePtrhc_norm [hc_mult*hidden]. A GROUPED RMSNorm scale: the streams
normalize independently inside the vector, group_size = hidden.
down_w: DevicePtrinput_mix_weight_down [rank, hc_mult*hidden].
up_w: DevicePtrinput_mix_weight_up [hc_mult*hidden, rank].
inject_w: DevicePtrblock_inject_weight [hc_mult, hc_mult*hidden]. NULL on the
model-level mixer, which is built use_combine=False and emits no
injection vector.
rank: usizehc_lowrank (320 on this checkpoint).