pub fn load_fp8_weight(
store: &WeightStore,
name: &str,
gpu: &dyn GpuBackend,
) -> Result<Fp8Weight>Expand description
2026-09-25: Load an FP8 E4M3 checkpoint weight with per-row scales as an
Fp8Weight tagged Fp8PerRow.
Expects {name}.weight FP8E4M3 [N, K] and {name}.weight_scale [N].
The weight is the store’s pointer. An FP32 scale is the store’s pointer
too; a BF16 scale is widened on the host into a new buffer; any other dtype
is an error.