diff --git a/projects/rccl/src/clique/AllReduceCliqueKernel.h b/projects/rccl/src/clique/AllReduceCliqueKernel.h
index 33e023ad17..5fc4752efb 100644
--- a/projects/rccl/src/clique/AllReduceCliqueKernel.h
+++ b/projects/rccl/src/clique/AllReduceCliqueKernel.h
@@ -66,7 +66,7 @@ __device__ void AllReduceCliqueSplitKernel(struct ncclWorkElem* args)
     // Perform the reduction
     #define ALL_REDUCE_CLIQUE_UNROLL 1
     ReduceOrCopyMulti<ALL_REDUCE_CLIQUE_UNROLL, FUNC, T, NUM_RANKS, NUM_RANKS, NUM_RANKS, NUM_RANKS>(
-      threadIdx.x, blockDim.x, redOp, false, false, NUM_RANKS, srcs, NUM_RANKS, dsts, blockN);
+      threadIdx.x, blockDim.x, redOp, true, true, NUM_RANKS, srcs, NUM_RANKS, dsts, blockN);
   }
 
   // Even if there was nothing for this GPU to do, it must participate in a barrier
diff --git a/projects/rccl/src/collectives/device/all_gather.h b/projects/rccl/src/collectives/device/all_gather.h
index 2ec0f21759..f030c8fbfd 100644
--- a/projects/rccl/src/collectives/device/all_gather.h
+++ b/projects/rccl/src/collectives/device/all_gather.h
@@ -79,7 +79,7 @@ namespace {
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllGather, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     using Proto = ProtoSimple<ALLGATHER_CHUNKSTEPS/ALLGATHER_SLICESTEPS, ALLGATHER_SLICESTEPS>;
     runRing<T, RedOp, Proto>(args);
   }
@@ -87,14 +87,14 @@ struct RunWorkElement<ncclFuncAllGather, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SI
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllGather, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllGather, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL128>(args);
   }
 };
diff --git a/projects/rccl/src/collectives/device/all_reduce.h b/projects/rccl/src/collectives/device/all_reduce.h
index 80254c1589..fff6e5d404 100644
--- a/projects/rccl/src/collectives/device/all_reduce.h
+++ b/projects/rccl/src/collectives/device/all_reduce.h
@@ -276,7 +276,7 @@ namespace {
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     using Proto = ProtoSimple<ALLREDUCE_CHUNKSTEPS/ALLREDUCE_SLICESTEPS, ALLREDUCE_SLICESTEPS>;
     runRing<T, RedOp, Proto>(args);
   }
@@ -284,14 +284,14 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SI
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_TREE, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runTreeUpDown<T, RedOp, ProtoSimple<1, 1>>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     static constexpr int COLLNET_COPY_THREADS = 64;
     const int tid = threadIdx.x;
     const int bid = args->coll.bid;
@@ -377,28 +377,28 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET, NCCL_PROTO
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_TREE, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runTreeUpDown<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     LAUNCH_CLIQUE_KERNEL(AllReduceCliqueSplitKernel, RedOp, T, args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_TREE, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     LAUNCH_CLIQUE_KERNEL(AllReduceCliqueSplitKernel, RedOp, T, args);
   }
 };
diff --git a/projects/rccl/src/collectives/device/broadcast.h b/projects/rccl/src/collectives/device/broadcast.h
index 82d11d6ad5..388ea8631d 100644
--- a/projects/rccl/src/collectives/device/broadcast.h
+++ b/projects/rccl/src/collectives/device/broadcast.h
@@ -77,7 +77,7 @@ namespace {
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncBroadcast, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     using Proto = ProtoSimple<BROADCAST_CHUNKSTEPS/BROADCAST_SLICESTEPS, BROADCAST_SLICESTEPS>;
     runRing<T, RedOp, Proto>(args);
   }
@@ -85,14 +85,14 @@ struct RunWorkElement<ncclFuncBroadcast, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SI
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncBroadcast, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncBroadcast, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL128>(args);
   }
 };
diff --git a/projects/rccl/src/collectives/device/common.h b/projects/rccl/src/collectives/device/common.h
index 59584e1aed..8204a06cd7 100644
--- a/projects/rccl/src/collectives/device/common.h
+++ b/projects/rccl/src/collectives/device/common.h
@@ -173,7 +173,7 @@ void NCCL_CALL_FUNCTIONS(struct ncclWorkElem* const c) noexcept {
 template <ncclFunc_t FUNCTION, int ALGO, int PROTO, class REDOP, typename T, int UNROLL>
 class ncclFunction {
   public:
-  __device__ void run(struct ncclWorkElem* args) {}
+  __device__ __attribute__((noinline)) void run(struct ncclWorkElem* args) {}
 };
 
 #ifdef ENABLE_COLLTRACE
@@ -262,14 +262,14 @@ __device__ int copyToShmem(T *dst, T const *src, int turn=0) {
 
 template<ncclFunc_t Fn, typename T, typename RedOp, int Algo, int Proto>
 struct RunWorkElement {
-  __device__ void run(ncclWorkElem*) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem*) {
     // Put NOT IMPLEMENTED behavior here.
   }
 };
 
 template<ncclFunc_t Fn, typename T, typename RedOp, int Algo, int Proto>
 struct RunWork {
-  __device__ void run(ncclWork *w) {
+  __device__ __attribute__((noinline)) void run(ncclWork *w) {
   }
 };
 
diff --git a/projects/rccl/src/collectives/device/reduce.h b/projects/rccl/src/collectives/device/reduce.h
index 9b3863653a..89b94f6637 100644
--- a/projects/rccl/src/collectives/device/reduce.h
+++ b/projects/rccl/src/collectives/device/reduce.h
@@ -71,7 +71,7 @@ namespace {
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     using Proto = ProtoSimple<REDUCE_CHUNKSTEPS/REDUCE_SLICESTEPS, REDUCE_SLICESTEPS>;
     runRing<T, RedOp, Proto>(args);
   }
@@ -79,14 +79,14 @@ struct RunWorkElement<ncclFuncReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPL
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduce, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL128>(args);
   }
 };
diff --git a/projects/rccl/src/collectives/device/reduce_scatter.h b/projects/rccl/src/collectives/device/reduce_scatter.h
index c8dc1a48fa..c7172f7a90 100644
--- a/projects/rccl/src/collectives/device/reduce_scatter.h
+++ b/projects/rccl/src/collectives/device/reduce_scatter.h
@@ -69,7 +69,7 @@ namespace {
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduceScatter, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     using Proto = ProtoSimple<REDUCESCATTER_CHUNKSTEPS/REDUCESCATTER_SLICESTEPS, REDUCESCATTER_SLICESTEPS>;
     runRing<T, RedOp, Proto>(args);
   }
@@ -77,14 +77,14 @@ struct RunWorkElement<ncclFuncReduceScatter, T, RedOp, NCCL_ALGO_RING, NCCL_PROT
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduceScatter, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL>(args);
   }
 };
 
 template<typename T, typename RedOp>
 struct RunWorkElement<ncclFuncReduceScatter, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_LL128> {
-  __device__ void run(ncclWorkElem *args) {
+  __device__ __attribute__((noinline)) void run(ncclWorkElem *args) {
     runRing<T, RedOp, ProtoLL128>(args);
   }
 };
diff --git a/projects/rccl/src/collectives/device/sendrecv.h b/projects/rccl/src/collectives/device/sendrecv.h
index aaa8a1de15..7e8b6eb5a1 100644
--- a/projects/rccl/src/collectives/device/sendrecv.h
+++ b/projects/rccl/src/collectives/device/sendrecv.h
@@ -11,7 +11,7 @@
 
 template<typename T, typename RedOp>
 struct RunWork<ncclFuncSendRecv, T, RedOp, NCCL_ALGO_RING, NCCL_PROTO_SIMPLE> {
-  __device__ void run(ncclWork *work) {
+  __device__ __attribute__((noinline)) void run(ncclWork *work) {
     int tid = threadIdx.x;
     int group = 0;
     const int rank = ncclShmem->comm.rank;