diff --git a/projects/rccl/src/collectives/device/all_reduce.h b/projects/rccl/src/collectives/device/all_reduce.h
index f319b4333e..9810990cc8 100644
--- a/projects/rccl/src/collectives/device/all_reduce.h
+++ b/projects/rccl/src/collectives/device/all_reduce.h
@@ -102,8 +102,7 @@ __device__ void ncclAllReduceRingKernel(struct CollectiveArgs* args) {
 #endif
 }
 
-template<int UNROLL, class FUNC, typename T>
-__attribute__((noinline))
+template<int UNUSED, class FUNC, typename T>
 __device__ void ncclAllReduceTreeKernel(struct CollectiveArgs* args) {
   const int tid = threadIdx.x;
   const int nthreads = blockDim.x;
@@ -122,7 +121,7 @@ __device__ void ncclAllReduceTreeKernel(struct CollectiveArgs* args) {
 
   do {
     // Reduce : max number of recv is 3, max number of send is 1 (binary tree + local)
-    ncclPrimitives<UNROLL, 1, 1, T, NCCL_MAX_TREE_ARITY, 1, FUNC> prims(tid, nthreads, tree->down, &tree->up, NULL, stepSize, channel, comm, args->opCount);
+    ncclPrimitives<1, 1, 1, T, NCCL_MAX_TREE_ARITY, 1, FUNC> prims(tid, nthreads, tree->down, &tree->up, NULL, stepSize, channel, comm, args->opCount);
     for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
       // Up
       ssize_t offset = gridOffset + bid*chunkSize;
@@ -139,7 +138,7 @@ __device__ void ncclAllReduceTreeKernel(struct CollectiveArgs* args) {
 
   do {
     // Broadcast : max number of recv is 1, max number of send is 3 (binary tree + local)
-    ncclPrimitives<UNROLL, 1, 1, T, 1, NCCL_MAX_TREE_ARITY, FUNC> prims(tid, nthreads, &tree->up, tree->down, NULL, stepSize, channel, comm, args->opCount);
+    ncclPrimitives<1, 1, 1, T, 1, NCCL_MAX_TREE_ARITY, FUNC> prims(tid, nthreads, &tree->up, tree->down, NULL, stepSize, channel, comm, args->opCount);
     for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
       // Down
       ssize_t offset = gridOffset + bid*chunkSize;