diff --git a/ggml/src/ggml-cuda/allreduce.cu b/ggml/src/ggml-cuda/allreduce.cu
index c1df8d8d2c..270a3d0e7f 100644
--- a/ggml/src/ggml-cuda/allreduce.cu
+++ b/ggml/src/ggml-cuda/allreduce.cu
@@ -322,8 +322,11 @@ static void ggml_cuda_ar_wdog_thread(ggml_cuda_ar_pipeline * p) {
 // ---------------------------------------------------------------------------
 
 ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(
-        const int * devices, int n_devices, size_t max_bytes) {
-    GGML_ASSERT(n_devices >= 2 && n_devices <= GGML_CUDA_MAX_DEVICES);
+        const int * devices, size_t n_devices, size_t max_bytes) {
+
+    if ((n_devices != 2) || (n_devices > GGML_CUDA_MAX_DEVICES)) {
+        return nullptr;
+    }
 
     auto * p = new ggml_cuda_ar_pipeline{};
     p->n_devices  = n_devices;
diff --git a/ggml/src/ggml-cuda/allreduce.cuh b/ggml/src/ggml-cuda/allreduce.cuh
index 4c2be43962..a30a4f8c28 100644
--- a/ggml/src/ggml-cuda/allreduce.cuh
+++ b/ggml/src/ggml-cuda/allreduce.cuh
@@ -18,7 +18,7 @@ struct ggml_cuda_ar_pipeline;
 // as the largest tensor that will be reduced.
 // Returns nullptr on allocation failure.
 ggml_cuda_ar_pipeline * ggml_cuda_ar_pipeline_init(
-    const int * devices, int n_devices, size_t max_bytes);
+    const int * devices, size_t n_devices, size_t max_bytes);
 
 // Release all resources owned by the pipeline.
 void ggml_cuda_ar_pipeline_free(ggml_cuda_ar_pipeline * pipeline);
diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 9a4dd062aa..498679daa6 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -1234,7 +1234,7 @@ static void * ggml_backend_cuda_comm_init(ggml_backend_t * backends, size_t n_ba
 
     if (provider == GGML_CUDA_ALLREDUCE_INTERNAL) {
         ret->ar_pipeline = ggml_cuda_ar_pipeline_init(
-            dev_ids.data(), static_cast<int>(n_backends), GGML_CUDA_AR_MAX_BYTES);
+            dev_ids.data(), n_backends, GGML_CUDA_AR_MAX_BYTES);
         if (ret->ar_pipeline != nullptr) {
             return ret;
         }
@@ -1388,13 +1388,15 @@ static ggml_cuda_comm_allreduce_result ggml_backend_cuda_comm_try_allreduce_inte
             return GGML_CUDA_COMM_ALLREDUCE_FAILED;
         }
         if (!ggml_is_contiguously_allocated(tensors[i])) {
-            GGML_LOG_WARN("%s: internal tensor[%zu] is not contiguously allocated: ne=%" PRId64 " nbytes=%zu packed=%zu type=%d\n",
+            GGML_LOG_WARN("%s: internal unsupported: tensor[%zu] is not contiguously allocated: ne=%" PRId64 " nbytes=%zu packed=%zu type=%d\n",
                           __func__, i, ne, ggml_nbytes(tensors[i]),
                           (size_t) ne * ggml_type_size(type) / ggml_blck_size(type), (int) type);
+            return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
         }
         if (((uintptr_t) tensors[i]->data & 0xF) != 0) {
-            GGML_LOG_WARN("%s: internal tensor[%zu] data pointer is not 16-byte aligned: %p type=%d ne=%" PRId64 "\n",
+            GGML_LOG_WARN("%s: internal unsupported: tensor[%zu] data pointer is not 16-byte aligned: %p type=%d ne=%" PRId64 "\n",
                           __func__, i, tensors[i]->data, (int) type, ne);
+            return GGML_CUDA_COMM_ALLREDUCE_UNSUPPORTED;
         }
     }