fix benches

crates/engine-gpu/benches/gpu_engine_bench.rs

@@ -6,6 +6,14 @@ use primitive_types::U512;
 use rand::RngCore;
 use std::sync::atomic::AtomicBool;
 
+/// Drop thread-local wgpu buffers before `GpuEngine` is dropped. Criterion
+/// creates a fresh engine per group; without this, TLS buffers outlive the
+/// device and the next group panics (`Buffer[…] does not exist`).
+fn teardown_gpu(engine: GpuEngine) {
+    GpuEngine::clear_worker_resources();
+    drop(engine);
+}
+
 fn bench_cpu_vs_gpu_small(c: &mut Criterion) {
     let cpu_engine = FastCpuEngine::new(10_000);
     let gpu_engine = GpuEngine::try_new(10_000_000, 0, false).expect("Failed to init GPU");
@@ -55,6 +63,7 @@ fn bench_cpu_vs_gpu_small(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 fn bench_cpu_vs_gpu_medium(c: &mut Criterion) {
@@ -106,6 +115,7 @@ fn bench_cpu_vs_gpu_medium(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 fn bench_cpu_vs_gpu_large(c: &mut Criterion) {
@@ -157,6 +167,7 @@ fn bench_cpu_vs_gpu_large(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 fn bench_solution_finding(c: &mut Criterion) {
@@ -208,6 +219,7 @@ fn bench_solution_finding(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 fn bench_throughput_per_second(c: &mut Criterion) {
@@ -259,6 +271,7 @@ fn bench_throughput_per_second(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 fn bench_gpu_batch_efficiency(c: &mut Criterion) {
@@ -335,6 +348,7 @@ fn bench_gpu_batch_efficiency(c: &mut Criterion) {
     });
 
     group.finish();
+    teardown_gpu(gpu_engine);
 }
 
 criterion_group!(

crates/engine-gpu/src/lib.rs

@@ -505,11 +505,20 @@ impl GpuEngine {
     }
 
     /// Explicitly clear thread-local GPU resources.
-    /// Call this before thread exit to avoid TLS destruction order issues with wgpu.
+    /// Call this before dropping a `GpuEngine` (or before thread exit) so
+    /// buffers are destroyed while the device is still alive — otherwise the
+    /// next engine on this thread can reuse dead buffer IDs and panic with
+    /// `Buffer[…] does not exist`.
     pub fn clear_worker_resources() {
         WORKER_RESOURCES.with(|resources| {
             *resources.borrow_mut() = None;
         });
+        ASSIGNED_GPU_DEVICE.with(|assigned| {
+            *assigned.borrow_mut() = None;
+        });
+        DEVICE_LOST.with(|lost| {
+            *lost.borrow_mut() = None;
+        });
     }
 }