pub fn plan_tensor_bytes(
name: &str,
shape: &[usize],
element_bytes: usize,
mixer: MixerKind,
config: &ModelConfig,
) -> Result<TpBytePlan>Expand description
Plan checkpoint storage BEFORE allocating a device buffer.
BF16/F32 keep their dtype. Packed experts require [N, K/2] bytes plus
[N, K/32] E8M0 scales; a column split cannot bisect a 32-value group.
Unknown non-packed tensors remain replicated. Runtime must separately
enforce the model’s tensor inventory and exclude unsupported modalities.