diff --git a/test/layer_tests.jl b/test/layer_tests.jl
index 6cadcc5..c702051 100644
--- a/test/layer_tests.jl
+++ b/test/layer_tests.jl
@@ -20,7 +20,8 @@
                 @jet model(x, ps, st)
 
                 __f = (x, ps) -> sum(abs2, first(model(x, ps, st)))
-                test_gradients(__f, x, ps; atol=1e-3, rtol=1e-3)
+                test_gradients(
+                    __f, x, ps; atol=1e-3, rtol=1e-3, soft_fail=[AutoFiniteDiff()])
             end
         end
     end
diff --git a/test/runtests.jl b/test/runtests.jl
index c5e84f6..5ed2d20 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -19,11 +19,15 @@ end
 using Boltz
 
 const BOLTZ_TEST_GROUP = get(ENV, "BOLTZ_TEST_GROUP", "all")
+const RETESTITEMS_NWORKERS = parse(
+    Int, get(ENV, "RETESTITEMS_NWORKERS", string(min(Hwloc.num_physical_cores(), 16))))
 const RETESTITEMS_NWORKER_THREADS = parse(Int,
-    get(ENV, "RETESTITEMS_NWORKER_THREADS", string(max(Hwloc.num_virtual_cores(), 1))))
+    get(ENV, "RETESTITEMS_NWORKER_THREADS",
+        string(max(Hwloc.num_virtual_cores() ÷ RETESTITEMS_NWORKERS, 1))))
 
-@info "Running tests for group: $BOLTZ_TEST_GROUP"
+@info "Running tests for group: $BOLTZ_TEST_GROUP with $RETESTITEMS_NWORKERS workers"
 
 ReTestItems.runtests(
     Boltz; tags=(BOLTZ_TEST_GROUP == "all" ? nothing : [Symbol(BOLTZ_TEST_GROUP)]),
-    nworkers=0, nworker_threads=RETESTITEMS_NWORKER_THREADS, testitem_timeout=3600)
+    nworkers=ifelse(BACKEND_GROUP ∈ ("cuda", "amdgpu"), 0, RETESTITEMS_NWORKERS),
+    nworker_threads=RETESTITEMS_NWORKER_THREADS, testitem_timeout=3600)