fix(build): restore deleted extract/build.rs and align stale test/exa…

crates/larql-compute/examples/compare_decode.rs

@@ -176,6 +176,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -282,6 +283,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })

crates/larql-compute/examples/compare_formats.rs

@@ -190,6 +190,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -293,6 +294,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -396,6 +398,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })

crates/larql-compute/examples/compare_ollama.rs

@@ -194,6 +194,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -279,6 +280,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -365,6 +367,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -458,6 +461,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -541,6 +545,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })

crates/larql-compute/examples/compare_pipeline.rs

@@ -192,6 +192,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })
@@ -300,6 +301,7 @@ fn main() {
                 ffn_up_bias: None,
                 ffn_down_bias: None,
                 moe: None,
+                ffn_is_remote: false,
                 moe_combined_output_norm: false,
                 moe_outer_post_norm: None,
             })

crates/larql-compute/examples/demo_architecture.rs

@@ -184,6 +184,7 @@ fn main() {
             ffn_up_bias: None,
             ffn_down_bias: None,
             moe: None,
+            ffn_is_remote: false,
             moe_combined_output_norm: false,
             moe_outer_post_norm: None,
         };

crates/larql-compute/examples/diag_decode_pipeline.rs

@@ -132,6 +132,7 @@ fn main() {
         ffn_up_bias: None,
         ffn_down_bias: None,
         moe: None,
+        ffn_is_remote: false,
         moe_combined_output_norm: false,
         moe_outer_post_norm: None,
     };
@@ -299,6 +300,7 @@ fn main() {
             ffn_up_bias: None,
             ffn_down_bias: None,
             moe: None,
+            ffn_is_remote: false,
             moe_combined_output_norm: false,
             moe_outer_post_norm: None,
         };
@@ -388,6 +390,7 @@ fn main() {
             ffn_up_bias: None,
             ffn_down_bias: None,
             moe: None,
+            ffn_is_remote: false,
             moe_combined_output_norm: false,
             moe_outer_post_norm: None,
         };

crates/larql-server/examples/bench_expert_server.rs

@@ -269,6 +269,7 @@ fn main() {
         // first half — but we set this *after* peeking at num_experts below.
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
     };
 
     let path_str = args[1].clone();

crates/larql-server/examples/openai_demo.rs

@@ -139,6 +139,7 @@ fn load_default(path: &str) -> Result<LoadedModel, Box<dyn std::error::Error + S
         release_mmap_after_request: false,
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
     };
     load_single_vindex(path, opts)
 }

crates/larql-server/tests/common/mod.rs

@@ -141,6 +141,13 @@ pub fn model_functional(id: &str) -> Arc<LoadedModel> {
         ffn_l2_cache: larql_server::ffn_l2_cache::FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }
 
@@ -171,6 +178,13 @@ pub fn model_infer_enabled(id: &str) -> Arc<LoadedModel> {
         ffn_l2_cache: larql_server::ffn_l2_cache::FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }
 
@@ -240,6 +254,13 @@ impl ModelBuilder {
             ffn_l2_cache: FfnL2Cache::new(1),
             expert_filter: None,
             unit_filter: None,
+            moe_remote: None,
+            #[cfg(feature = "metal-experts")]
+            metal_backend: std::sync::OnceLock::new(),
+            #[cfg(feature = "metal-experts")]
+            moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+            #[cfg(feature = "metal-experts")]
+            metal_ffn_layer_bufs: std::sync::OnceLock::new(),
         })
     }
 }

crates/larql-server/tests/test_expert_endpoint.rs

@@ -221,6 +221,8 @@ fn make_loaded_model(
         down_top_k: 1,
         has_model_weights: false,
         model_config: None,
+        fp4: None,
+        ffn_layout: None,
     };
 
     // Build ModelWeights with expert data in raw_bytes (no mmap needed).
@@ -238,6 +240,7 @@ fn make_loaded_model(
         vectors,
         raw_bytes,
         packed_mmaps: HashMap::new(),
+        skipped_tensors: Vec::new(),
         packed_byte_ranges: HashMap::new(),
         embed: embed.clone(),
         lm_head: embed,
@@ -253,7 +256,7 @@ fn make_loaded_model(
     };
 
     let lock = OnceLock::new();
-    lock.set(weights).ok();
+    lock.set(std::sync::RwLock::new(weights)).ok();
 
     LoadedModel {
         id: "test-moe".into(),
@@ -272,7 +275,14 @@ fn make_loaded_model(
         probe_labels: HashMap::new(),
         ffn_l2_cache: FfnL2Cache::new(1),
         expert_filter: None,
+        unit_filter: None,
         moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     }
 }
 
@@ -307,11 +317,16 @@ fn local_output(
     router_proj: &[f32],
     pre_norm: &[f32],
 ) -> Vec<f32> {
+    let gate_up_per_expert = gate_up.len() / NUM_EXPERTS;
+    let down_per_expert = down.len() / NUM_EXPERTS;
+    let experts_gate_up: Vec<&[u8]> = gate_up.chunks(gate_up_per_expert).collect();
+    let experts_down: Vec<&[u8]> = down.chunks(down_per_expert).collect();
     cpu_moe_forward(
         h,
         &MoeLayerWeights {
-            experts_gate_up: gate_up,
-            experts_down: down,
+            experts_gate_up,
+            experts_down,
+            expert_data_format: larql_compute::QuantFormat::BF16,
             router_proj,
             router_scale: &[],
             router_per_expert_scale: &[],

crates/larql-server/tests/test_http_full_routes.rs

@@ -48,6 +48,13 @@ fn model_functional_with_labels(id: &str) -> Arc<LoadedModel> {
         ffn_l2_cache: larql_server::ffn_l2_cache::FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }
 

crates/larql-server/tests/test_unit_band_utils.rs

@@ -166,6 +166,13 @@ fn make_minimal_model(layer_bands: Option<LayerBands>) -> Arc<LoadedModel> {
         ffn_l2_cache: FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }
 

crates/larql-server/tests/test_unit_state.rs

@@ -96,6 +96,13 @@ fn make_tiny_model(id: &str) -> Arc<LoadedModel> {
         ffn_l2_cache: FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }
 
@@ -172,6 +179,13 @@ fn make_loaded_model_for_warmup() -> Arc<LoadedModel> {
         ffn_l2_cache: FfnL2Cache::new(1),
         expert_filter: None,
         unit_filter: None,
+        moe_remote: None,
+        #[cfg(feature = "metal-experts")]
+        metal_backend: std::sync::OnceLock::new(),
+        #[cfg(feature = "metal-experts")]
+        moe_scratches: std::sync::Mutex::new(std::collections::HashMap::new()),
+        #[cfg(feature = "metal-experts")]
+        metal_ffn_layer_bufs: std::sync::OnceLock::new(),
     })
 }