pub fn load_checkpoint_kv_scales(
store: &WeightStore,
attn_prefix: &str,
gpu: &dyn GpuBackend,
) -> Result<Option<(f32, f32)>>Expand description
2026-09-28: The checkpoint’s FP8 KV-cache scales (k_scale, v_scale) for the
attention layer at attn_prefix, or None when it ships none. The keys come
from WeightStore::kv_scale_keys, the resolver serve’s census also uses.
Each scale is the dequant multiplier (amax / 448): the KV write divides by
it and the FP8 attention kernels multiply by it. A scale that is not a
finite positive number is an error, since the write computes 1 / scale.