diff --git a/src/bin/orangu-server/engine/arch/qwen_hybrid.rs b/src/bin/orangu-server/engine/arch/qwen_hybrid.rs
index 9ee54981..11254a17 100644
--- a/src/bin/orangu-server/engine/arch/qwen_hybrid.rs
+++ b/src/bin/orangu-server/engine/arch/qwen_hybrid.rs
@@ -3183,7 +3183,6 @@ impl<F: HybridFfn> Trunk<F> {
     // Eight because the scratch buffers are threaded rather than held in
     // `self`; grouping them into `LayerScratch` already removed one.
     #[allow(clippy::too_many_arguments)]
-    #[allow(clippy::too_many_arguments)]
     fn forward_full_attn_layer(
         &self,
         backend: &dyn Backend,
@@ -3292,7 +3291,6 @@ impl<F: HybridFfn> Trunk<F> {
     // `self`, the same reason `forward_full_attn_layer` carries this allow:
     // `forward` takes `&self` and two layers may be in flight at once.
     #[allow(clippy::too_many_arguments)]
-    #[allow(clippy::too_many_arguments)]
     fn apply_ffn(
         &self,
         backend: &dyn Backend,
diff --git a/src/bin/orangu-server/engine/backend/vulkan.rs b/src/bin/orangu-server/engine/backend/vulkan.rs
index f38d016d..f76b48cf 100644
--- a/src/bin/orangu-server/engine/backend/vulkan.rs
+++ b/src/bin/orangu-server/engine/backend/vulkan.rs
@@ -4362,7 +4362,6 @@ than half the speed. Prefer another quantization of this model, or `backend = cp
 
     /// The norm-pair bind group: `x` (the `wo` output), `w1`, the residual,
     /// `w2`, `y1` (the FFN residual), `y2` (the FFN norm's output), meta.
-    #[allow(clippy::too_many_arguments)]
     /// `q8` is the 8-bit form of `y2` the kernel writes when `meta`'s
     /// `aux` is 1 — a layer's `FfnQ8` — or the placeholder otherwise.
     #[allow(clippy::too_many_arguments)]
@@ -15388,7 +15387,6 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
         buf
     }
 
-    #[allow(clippy::too_many_arguments)]
     #[allow(clippy::too_many_arguments)]
     fn rope_meta_buffer(
         &self,
diff --git a/src/bin/orangu-server/engine/draft_control.rs b/src/bin/orangu-server/engine/draft_control.rs
index 9c6a8103..c61327c9 100644
--- a/src/bin/orangu-server/engine/draft_control.rs
+++ b/src/bin/orangu-server/engine/draft_control.rs
@@ -745,9 +745,9 @@ mod tests {
         c.planned = Some(Planned {
             widths: vec![(1, 18.0), (2, 19.0), (4, 21.0), (8, 25.0)],
         });
-        for d in 0..DEPTHS {
-            c.accept_bin[d][3] = 0.95;
-            c.accept_bin[d][0] = 0.05;
+        for bins in &mut c.accept_bin {
+            bins[3] = 0.95;
+            bins[0] = 0.05;
         }
         assert_eq!(c.cut(&[3, 3, 3, 3, 3, 3, 3, 3]), 8);
         assert_eq!(c.cut(&[3, 3, 0, 0, 0, 0, 0, 0]), 2);
diff --git a/src/bin/orangu-server/engine/kv_cache.rs b/src/bin/orangu-server/engine/kv_cache.rs
index 34bea43c..d6281e47 100644
--- a/src/bin/orangu-server/engine/kv_cache.rs
+++ b/src/bin/orangu-server/engine/kv_cache.rs
@@ -6196,17 +6196,17 @@ mod tests {
         }
 
         // And it reads what the first sequence computed, in every layer.
-        for layer in 0..2 {
+        for (layer, cache) in second.layers.iter().enumerate().take(2) {
             for r in 0..12 {
                 let expect: Vec<f32> = (0..KV).map(|d| row(r, d)).collect();
                 assert_eq!(
-                    second.layers[layer].key_at(r, 0, KV),
+                    cache.key_at(r, 0, KV),
                     &expect[..],
                     "layer {layer} row {r} keys"
                 );
                 let expect_v: Vec<f32> = expect.iter().map(|x| -x).collect();
                 assert_eq!(
-                    second.layers[layer].value_at(r, 0, KV),
+                    cache.value_at(r, 0, KV),
                     &expect_v[..],
                     "layer {layer} row {r} values"
                 );