model: the host union's ungated relu² form, for Bolt's routed experts (rebase of #355)

crates/model/src/arch/nemotron_h_moe/host.rs

@@ -0,0 +1,45 @@
+//! The host tier's view of a Bolt layer's routed experts: relu² — the up and
+//! down stacks, the layer's own expert count and width, read in the
+//! `moe_latent_size` latent the experts run in. The one place that wires them
+//! into a [`Relu2Spec`] (`HostLayer::build_relu2`); this architecture's files
+//! hold no gate stack.
+
+use super::hparams::{Experts, Hparams};
+use crate::ModelError;
+use crate::arch::names::routed_stacks;
+use crate::moe::{HostLayer, Relu2Spec};
+use crate::r8file::R8Source;
+
+/// Layer `l`'s routed experts as the host tier serves them: ungated relu²,
+/// two stacks at the layer's own expert count and width in the latent width —
+/// the MoE layers' shapes differ from layer to layer, so a caller serves each
+/// layer through the layer its own shapes make. `None` for a layer that does
+/// not route (a Mamba-2 or attention layer), an error for a layer past the
+/// model. Load-time only: the stacks are checked here.
+pub fn layer(src: R8Source<'_>, hp: &Hparams, l: usize) -> Result<Option<HostLayer>, ModelError> {
+    let Some(experts) = hp.experts.get(l) else {
+        return Err(ModelError::Shape {
+            what: "nemotron_h_moe host layer: the highest layer index, a layer inside the model's",
+            want_ne0: hp.experts.len().saturating_sub(1),
+            want_ne1: 1,
+            got_ne0: l,
+            got_ne1: 1,
+        });
+    };
+    let Some(Experts { n_expert, ff, .. }) = *experts else {
+        return Ok(None);
+    };
+    // The common spelling of a routed stack; this architecture's files hold no gate.
+    let [_, up, down] = routed_stacks(l);
+    HostLayer::build_relu2(
+        src,
+        &Relu2Spec {
+            up: &up,
+            down: &down,
+            n_expert,
+            embd: hp.latent,
+            ff,
+        },
+    )
+    .map(Some)
+}

crates/model/src/arch/nemotron_h_moe/mod.rs

@@ -10,6 +10,7 @@
 //! llama.cpp's `src/models/nemotron-h.cpp`, which loads both architectures.
 
 pub mod fixture;
+pub mod host;
 pub mod hparams;
 pub mod names;
 pub mod place;