chore: reformat tree with cargo fmt and add fmt to CI gate

.github/PULL_REQUEST_TEMPLATE.md

@@ -33,7 +33,7 @@ What did you run to convince yourself this works? Be specific.
 For inference changes, real-checkpoint validation is required — synthetic-only is not enough.
 -->
 
-- [ ] `cargo fmt --check`
+- [ ] `cargo fmt --all -- --check` (enforced by CI — violations block merge)
 - [ ] `cargo clippy --all-targets -- -D warnings`
 - [ ] `cargo test --release`
 - [ ] `cargo deny check`

.github/workflows/ci.yml

@@ -1,8 +1,8 @@
-# PR / push CI for security audit. Builds and tests are covered by
+# PR / push CI for lightweight quality gates. Builds and tests are covered by
 # release.yml (signed artifacts) and pipeline-parallel-ci.yml (distributed
-# runs); this workflow runs the lightweight cargo-deny gate on every
-# touched-Rust change so license drift and new advisories are caught at
-# PR time.
+# runs); this workflow runs cargo-deny (license / advisory) and cargo-fmt
+# checks on every touched-Rust change so formatting drift and license issues
+# are caught at PR time.
 
 name: CI
 
@@ -54,3 +54,19 @@ jobs:
         with:
           command: check
           log-level: warn
+
+  fmt:
+    name: cargo-fmt
+    needs: changes
+    if: needs.changes.outputs.rust == 'true'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v6
+        with:
+          persist-credentials: false
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: rustfmt
+      - run: cargo fmt --all -- --check

CONTRIBUTING.md

@@ -41,9 +41,9 @@ Thank you for your interest in contributing to mlxcel! This document covers the
    ```
 4. Run the local quality gates:
    ```bash
-   cargo fmt --check
+   cargo fmt --all -- --check   # enforced by CI; fmt violations block merge
    cargo clippy --all-targets -- -D warnings
-   cargo deny check        # advisories + licenses + sources
+   cargo deny check             # advisories + licenses + sources
    ```
 5. For inference changes, validate against a real checkpoint — synthetic or build-only validation is not enough (see [`AGENTS.md`](AGENTS.md) for why).
 6. Commit with a conventional prefix (see below) and a clear message.

src/lib/mlxcel-core/src/cache.rs

@@ -2373,10 +2373,8 @@ impl KVCache {
 
                 let k_slice = ffi::slice(k_int8, &[0, 0, 0, 0], &[ks[0], ks[1], live_len, ks[3]]);
                 let v_slice = ffi::slice(v_int8, &[0, 0, 0, 0], &[vs[0], vs[1], live_len, vs[3]]);
-                let ks_slice =
-                    ffi::slice(k_scales, &[0, 0, 0, 0], &[kss[0], kss[1], live_len, 1]);
-                let vs_slice =
-                    ffi::slice(v_scales, &[0, 0, 0, 0], &[vss[0], vss[1], live_len, 1]);
+                let ks_slice = ffi::slice(k_scales, &[0, 0, 0, 0], &[kss[0], kss[1], live_len, 1]);
+                let vs_slice = ffi::slice(v_scales, &[0, 0, 0, 0], &[vss[0], vss[1], live_len, 1]);
 
                 (
                     dequantize(&k_slice, &ks_slice),
@@ -3785,8 +3783,7 @@ impl RotatingKVCache {
             return (new_keys, new_values);
         }
 
-        let (base_k, base_v, current_seq_len) =
-            self.visible_fp16_prefix_for_concat();
+        let (base_k, base_v, current_seq_len) = self.visible_fp16_prefix_for_concat();
 
         let concat_k = concatenate(&base_k, &new_keys, 2);
         let concat_v = concatenate(&base_v, &new_values, 2);
@@ -5681,10 +5678,7 @@ mod tests {
                 .collect::<Vec<_>>()
         };
         assert_eq!(to_f32(&visible_keys), vec![1.0, 5.0, 6.0, 7.0, 8.0]);
-        assert_eq!(
-            to_f32(&visible_values),
-            vec![10.0, 50.0, 60.0, 70.0, 80.0]
-        );
+        assert_eq!(to_f32(&visible_values), vec![10.0, 50.0, 60.0, 70.0, 80.0]);
     }
 
     #[test]
@@ -6190,10 +6184,8 @@ mod tests {
         }
         let (q_unrot, _) = unit_token(42);
         let q_ref = rotate_at(&q_unrot, M);
-        let (k_ref, v_ref) = cache_ref.update_and_fetch(
-            rotate_at(&unit_token(99).0, M),
-            unit_token(99).1,
-        );
+        let (k_ref, v_ref) =
+            cache_ref.update_and_fetch(rotate_at(&unit_token(99).0, M), unit_token(99).1);
         let out_ref = mlxcel_core::causal_attention(&q_ref, &k_ref, &v_ref, scale, 0.0, 0);
         let out_ref_f32 = to_f32(&out_ref);
 
@@ -6216,11 +6208,10 @@ mod tests {
         // position `M` to simulate the pre-fix offset decrement.
         assert_eq!(cache_broken.trim_front(N), N);
         let q_broken = rotate_at(&q_unrot, M);
-        let (k_broken, v_broken) = cache_broken.update_and_fetch(
-            rotate_at(&unit_token(99).0, M),
-            unit_token(99).1,
-        );
-        let out_broken = mlxcel_core::causal_attention(&q_broken, &k_broken, &v_broken, scale, 0.0, 0);
+        let (k_broken, v_broken) =
+            cache_broken.update_and_fetch(rotate_at(&unit_token(99).0, M), unit_token(99).1);
+        let out_broken =
+            mlxcel_core::causal_attention(&q_broken, &k_broken, &v_broken, scale, 0.0, 0);
         let out_broken_f32 = to_f32(&out_broken);
 
         let mut sq_err = 0.0_f64;

src/lib/mlxcel-core/src/cache/detach.rs

@@ -241,19 +241,14 @@ impl DetachedKVCache {
 
         // Slice axis 2 (seq-len) of each tensor to the new length. Width axes
         // (B, H, head_dim / packed_dim) come from the existing shape.
-        let trim_axis_seq = |a: &Option<UniquePtr<MlxArray>>,
-                             tail_axis: i32|
-         -> Option<UniquePtr<MlxArray>> {
-            a.as_ref().map(|arr| {
-                let shape = ffi::array_shape(arr);
-                let last = if tail_axis == 0 { shape[3] } else { tail_axis };
-                ffi::slice(
-                    arr,
-                    &[0, 0, 0, 0],
-                    &[shape[0], shape[1], new_len, last],
-                )
-            })
-        };
+        let trim_axis_seq =
+            |a: &Option<UniquePtr<MlxArray>>, tail_axis: i32| -> Option<UniquePtr<MlxArray>> {
+                a.as_ref().map(|arr| {
+                    let shape = ffi::array_shape(arr);
+                    let last = if tail_axis == 0 { shape[3] } else { tail_axis };
+                    ffi::slice(arr, &[0, 0, 0, 0], &[shape[0], shape[1], new_len, last])
+                })
+            };
 
         if self.mode == KVCacheMode::Turbo4Delegated {
             // K is unified — same shape contract as Fp16. Slice to `new_len`.

src/lib/mlxcel-core/src/cache/detach_tests.rs

@@ -677,10 +677,7 @@ fn detached_kvcache_trim_to_mid_buffer_resets_offset_and_visible_keys() {
 #[test]
 fn detached_kvcache_trim_to_zero_drops_storage_and_keeps_mode() {
     let mut live = KVCache::new_with_mode(KVCacheMode::Int8);
-    live.update(
-        fp32_tokens(&[1.0, 2.0, 3.0]),
-        fp32_tokens(&[4.0, 5.0, 6.0]),
-    );
+    live.update(fp32_tokens(&[1.0, 2.0, 3.0]), fp32_tokens(&[4.0, 5.0, 6.0]));
     let mut handle = live.clone_handle();
     assert_eq!(handle.seq_len(), 3);
     assert_eq!(handle.mode(), KVCacheMode::Int8);
@@ -698,7 +695,9 @@ fn detached_kvcache_trim_to_exact_offset_is_noop() {
     live.update(fp32_tokens(&[1.0, 2.0]), fp32_tokens(&[3.0, 4.0]));
     let mut handle = live.clone_handle();
 
-    handle.trim_to(2).expect("trim_to exact offset must succeed");
+    handle
+        .trim_to(2)
+        .expect("trim_to exact offset must succeed");
     assert_eq!(handle.seq_len(), 2);
 
     let mut restored = KVCache::new();
@@ -825,7 +824,9 @@ fn detached_cache_set_truncate_to_zero_drops_every_layer() {
     let v: Vec<f32> = (1..=4).map(|i| i as f32 * 10.0).collect();
     let mut detached = detached_set_with_fp16_layers(2, &k, &v);
 
-    detached.truncate_to(0).expect("truncate_to zero must succeed");
+    detached
+        .truncate_to(0)
+        .expect("truncate_to zero must succeed");
     assert!(detached.caches.iter().all(|c| c.is_empty()));
     assert_eq!(detached.current_offset, 0);
     assert_eq!(detached.prompt_len, 0);
@@ -940,10 +941,7 @@ fn detached_cache_set_truncate_to_int8_preserves_dequantization() {
     let seq_adopted = pool.adopt(&model, detached).unwrap();
     let (k_a, v_a) = {
         let caches = pool.get_caches_mut(seq_adopted).unwrap();
-        caches[0].update_and_fetch(
-            fp32_tokens(&full_k[5..]),
-            fp32_tokens(&full_v[5..]),
-        )
+        caches[0].update_and_fetch(fp32_tokens(&full_k[5..]), fp32_tokens(&full_v[5..]))
     };
     eval(&k_a);
     eval(&v_a);

src/lib/mlxcel-core/src/drafter/dflash/attention.rs

@@ -441,4 +441,3 @@ mod tests {
         );
     }
 }
-

src/lib/mlxcel-core/src/drafter/dflash/drafter.rs

@@ -106,12 +106,11 @@ impl DFlashDrafter {
                 path: config_path.display().to_string(),
                 source: e,
             })?;
-        let config = DFlashConfig::from_json(&config_json).map_err(|e| {
-            DrafterError::ConfigParse {
+        let config =
+            DFlashConfig::from_json(&config_json).map_err(|e| DrafterError::ConfigParse {
                 path: config_path.display().to_string(),
                 source: serde::de::Error::custom(e),
-            }
-        })?;
+            })?;
 
         let mut weights = crate::weights::load_weights_from_dir(path)
             .map_err(|msg| DrafterError::LoadFailed { reason: msg })?;
@@ -209,8 +208,9 @@ impl Drafter for DFlashDrafter {
         //    expose `embed_tokens` (e.g. a non-Qwen-3.5 model fed in by
         //    mistake).
         if self.model.needs_embed_binding() {
-            let embed = target.embed_tokens_module().ok_or_else(|| {
-                DrafterError::BindFailed {
+            let embed = target
+                .embed_tokens_module()
+                .ok_or_else(|| DrafterError::BindFailed {
                     reason: format!(
                         "DFlash drafter checkpoint omits embed_tokens.weight \
                          and the target does not expose embed_tokens_module(); \
@@ -220,8 +220,7 @@ impl Drafter for DFlashDrafter {
                          (kind = {})",
                         self.kind()
                     ),
-                }
-            })?;
+                })?;
             self.model.bind_target_embedding(embed);
         } else {
             // Legacy capability smoke-test for self-contained checkpoints:
@@ -427,10 +426,7 @@ fn sample_block_per_position_batched(
     sampler: &SamplingConfig,
 ) -> Result<Vec<Vec<i32>>, DrafterError> {
     let shape = ffi::array_shape(logits);
-    if shape.len() != 3
-        || shape[0] != batch_size as i32
-        || shape[1] != block_size as i32
-    {
+    if shape.len() != 3 || shape[0] != batch_size as i32 || shape[1] != block_size as i32 {
         return Err(DrafterError::DraftFailed {
             reason: format!(
                 "DFlash drafter (batched) expected logits shape \
@@ -468,11 +464,7 @@ fn sample_block_per_position_batched(
         for i in 0..n {
             // Row `(b, i+1)` of the [B, L, V] logits.
             let pos = (i + 1) as i32;
-            let row = ffi::slice(
-                logits,
-                &[b, pos, 0_i32],
-                &[b + 1, pos + 1, vocab],
-            );
+            let row = ffi::slice(logits, &[b, pos, 0_i32], &[b + 1, pos + 1, vocab]);
             // Drop the seq axis so we get a `[1, vocab]` 2D slice (fused_sample
             // / argmax expect `[batch, vocab]`).
             let row = ffi::reshape(&row, &[1_i32, vocab]);
@@ -536,7 +528,11 @@ fn sample_block_per_position(
     for i in 0..n {
         // Row `i + 1` of the [1, L, V] logits.
         let row_idx = (i + 1) as i32;
-        let row = ffi::slice(logits, &[0_i32, row_idx, 0_i32], &[1_i32, row_idx + 1, vocab]);
+        let row = ffi::slice(
+            logits,
+            &[0_i32, row_idx, 0_i32],
+            &[1_i32, row_idx + 1, vocab],
+        );
         // Drop the seq axis so we get a `[1, vocab]` 2D slice (fused_sample
         // / argmax expect `[batch, vocab]`).
         let row = ffi::reshape(&row, &[1_i32, vocab]);
@@ -588,10 +584,7 @@ fn sample_block_per_position_array(
     }
 
     let tokens = sample_block_per_position(logits, block_size, sampler)?;
-    Ok(ffi::from_slice_i32(
-        &tokens,
-        &[1, (block_size - 1) as i32],
-    ))
+    Ok(ffi::from_slice_i32(&tokens, &[1, (block_size - 1) as i32]))
 }
 
 #[cfg(test)]
@@ -620,10 +613,7 @@ mod tests {
             ffi::zeros(&[4, 4], dtype::BFLOAT16),
         );
         // A non-bf16 tensor: should pass through.
-        weights.insert(
-            "fc.weight".to_string(),
-            ffi::zeros(&[4, 4], dtype::FLOAT16),
-        );
+        weights.insert("fc.weight".to_string(), ffi::zeros(&[4, 4], dtype::FLOAT16));
 
         convert_bf16_to_f16_non_quantized(&mut weights);
 

src/lib/mlxcel-core/src/drafter/dflash/layer.rs

@@ -83,18 +83,13 @@ impl DFlashDecoderLayer {
         let post_attention_layernorm_w = weights
             .get(&format!("{prefix}.post_attention_layernorm.weight"))
             .map(|w| ffi::copy(w))
-            .ok_or_else(|| {
-                format!("Weight not found: {prefix}.post_attention_layernorm.weight")
-            })?;
+            .ok_or_else(|| format!("Weight not found: {prefix}.post_attention_layernorm.weight"))?;
 
         Ok(Self {
             self_attn,
             mlp,
             input_layernorm: RMSNorm::new(input_layernorm_w, config.rms_norm_eps),
-            post_attention_layernorm: RMSNorm::new(
-                post_attention_layernorm_w,
-                config.rms_norm_eps,
-            ),
+            post_attention_layernorm: RMSNorm::new(post_attention_layernorm_w, config.rms_norm_eps),
         })
     }
 }

src/lib/mlxcel-core/src/drafter/dflash/mlp.rs

@@ -67,24 +67,12 @@ impl DFlashMlp {
         group_size: i32,
         bits: i32,
     ) -> Result<Self, String> {
-        let gate_proj = UnifiedLinear::from_weights(
-            weights,
-            &format!("{prefix}.gate_proj"),
-            group_size,
-            bits,
-        )?;
-        let up_proj = UnifiedLinear::from_weights(
-            weights,
-            &format!("{prefix}.up_proj"),
-            group_size,
-            bits,
-        )?;
-        let down_proj = UnifiedLinear::from_weights(
-            weights,
-            &format!("{prefix}.down_proj"),
-            group_size,
-            bits,
-        )?;
+        let gate_proj =
+            UnifiedLinear::from_weights(weights, &format!("{prefix}.gate_proj"), group_size, bits)?;
+        let up_proj =
+            UnifiedLinear::from_weights(weights, &format!("{prefix}.up_proj"), group_size, bits)?;
+        let down_proj =
+            UnifiedLinear::from_weights(weights, &format!("{prefix}.down_proj"), group_size, bits)?;
         Ok(Self {
             gate_proj,
             up_proj,

src/lib/mlxcel-core/src/drafter/dflash/mod.rs

@@ -104,8 +104,7 @@ pub(crate) fn materialize_argmax_i32_vec(argmax: &MlxArray, expected_len: usize)
             .take(expected_len)
             .map(|chunk| {
                 i64::from_ne_bytes([
-                    chunk[0], chunk[1], chunk[2], chunk[3], chunk[4], chunk[5], chunk[6],
-                    chunk[7],
+                    chunk[0], chunk[1], chunk[2], chunk[3], chunk[4], chunk[5], chunk[6], chunk[7],
                 ]) as i32
             })
             .collect(),

src/lib/mlxcel-core/src/drafter/dflash/model.rs

@@ -464,8 +464,14 @@ mod tests {
     fn tiny_weights_without_embed() -> WeightMap {
         let mut w: WeightMap = std::collections::HashMap::new();
         // fc: [hidden, num_target_layers * hidden] = [4, 16]
-        w.insert("fc.weight".to_string(), ffi::zeros(&[4, 16], dtype::FLOAT32));
-        w.insert("hidden_norm.weight".to_string(), ffi::zeros(&[4], dtype::FLOAT32));
+        w.insert(
+            "fc.weight".to_string(),
+            ffi::zeros(&[4, 16], dtype::FLOAT32),
+        );
+        w.insert(
+            "hidden_norm.weight".to_string(),
+            ffi::zeros(&[4], dtype::FLOAT32),
+        );
         w.insert("norm.weight".to_string(), ffi::zeros(&[4], dtype::FLOAT32));
         // Layer 0 projections. q out = n_heads*head_dim = 4; k/v out =
         // n_kv_heads*head_dim = 2; o in = 4.
@@ -593,9 +599,9 @@ mod tests {
 
         // Stand-in for the target's embedding table — a regular
         // [vocab, hidden] tensor, the same shape Qwen 3.5 hands out.
-        let target_embed = crate::layers::UnifiedEmbedding::Regular(
-            crate::layers::Embedding::new(ffi::zeros(&[8, 4], dtype::FLOAT32)),
-        );
+        let target_embed = crate::layers::UnifiedEmbedding::Regular(crate::layers::Embedding::new(
+            ffi::zeros(&[8, 4], dtype::FLOAT32),
+        ));
         model.bind_target_embedding(target_embed);
 
         assert!(