pub type ExpertFn<'a> = &'a dyn Fn(usize) -> Result<Glm5NextExpertWeights>;Expand description
One routed expert, by GLOBAL id.
🔴 A closure rather than a name→bytes loader on purpose. The routed experts are the only
thing here that is NOT sharded — an expert is owned whole — so the caller can hand over the
checkpoint’s own device pointers with no copy. Routing 3.85 GiB per layer through a host
f32 round trip, as the TP-sliced halves must, would be pure waste. It also keeps the packed
e2m1 codes and e4m3 block scales from ever passing through f32.