Add native Qwen MTP speculation

tools/qwen38-artifacts.rs

@@ -657,7 +657,7 @@ fn precision_assignments(
     reject_unassigned_quantized(&core, &assignments)?;
 
     let mtp = tensor_headers(manifest, source, Some("mtp"), complete)?;
-    add_inferred_sidecar(&mtp, 64, &mut assignments)?;
+    add_inferred_sidecar(&mtp, &mut assignments)?;
     let ple = tensor_headers(manifest, source, Some("ple"), complete)?;
     let ple_group = config
         .pointer("/mtplx_recipe/ngram/group_size")
@@ -711,7 +711,6 @@ fn validate_unquantized(base: &str, tensors: &BTreeMap<String, Tensor>) -> Resul
 
 fn add_inferred_sidecar(
     tensors: &BTreeMap<String, Tensor>,
-    group_size: u64,
     assignments: &mut BTreeMap<String, Quantization>,
 ) -> Result<(), String> {
     for (name, tensor) in tensors {
@@ -732,6 +731,11 @@ fn add_inferred_sidecar(
             .shape
             .last()
             .ok_or_else(|| format!("{base}.scales has no dimensions"))?;
+        let group_size = if base.contains(".switch_mlp.") {
+            32
+        } else {
+            64
+        };
         let pack = groups
             .checked_mul(group_size)
             .and_then(|columns| columns.checked_div(packed))