diff --git a/projects/rccl/makefiles/common.mk b/projects/rccl/makefiles/common.mk
index 1a1c2b66f8..0c0d04a86f 100644
--- a/projects/rccl/makefiles/common.mk
+++ b/projects/rccl/makefiles/common.mk
@@ -31,13 +31,17 @@ CUDA8_GENCODE = -gencode=arch=compute_35,code=sm_35 \
                 -gencode=arch=compute_61,code=sm_61
 CUDA9_GENCODE = -gencode=arch=compute_70,code=sm_70
 CUDA11_GENCODE = -gencode=arch=compute_80,code=sm_80
+CUDA11_8_GENCODE = -gencode=arch=compute_90,code=sm_90
 
 CUDA8_PTX     = -gencode=arch=compute_61,code=compute_61
 CUDA9_PTX     = -gencode=arch=compute_70,code=compute_70
 CUDA11_PTX    = -gencode=arch=compute_80,code=compute_80
+CUDA11_8_PTX  = -gencode=arch=compute_90,code=compute_90
 
-# Include Ampere support if we're using CUDA11 or above
-ifeq ($(shell test "0$(CUDA_MAJOR)" -ge 11; echo $$?),0)
+ifeq ($(shell test "0$(CUDA_MAJOR)" -eq 11 -a "0$(CUDA_MINOR)" -ge 8 -o "0$(CUDA_MAJOR)" -gt 11; echo $$?),0)
+# Include Hopper support if we're using CUDA11.8 or above
+  NVCC_GENCODE ?= $(CUDA8_GENCODE) $(CUDA9_GENCODE) $(CUDA11_GENCODE) $(CUDA11_8_GENCODE) $(CUDA11_8_PTX)
+else ifeq ($(shell test "0$(CUDA_MAJOR)" -ge 11; echo $$?),0)
   NVCC_GENCODE ?= $(CUDA8_GENCODE) $(CUDA9_GENCODE) $(CUDA11_GENCODE) $(CUDA11_PTX)
 # Include Volta support if we're using CUDA9 or above
 else ifeq ($(shell test "0$(CUDA_MAJOR)" -ge 9; echo $$?),0)
@@ -45,7 +49,7 @@ else ifeq ($(shell test "0$(CUDA_MAJOR)" -ge 9; echo $$?),0)
 else
   NVCC_GENCODE ?= $(CUDA8_GENCODE) $(CUDA8_PTX)
 endif
-#$(info NVCC_GENCODE is ${NVCC_GENCODE})
+$(info NVCC_GENCODE is ${NVCC_GENCODE})
 
 CXXFLAGS   := -DCUDA_MAJOR=$(CUDA_MAJOR) -DCUDA_MINOR=$(CUDA_MINOR) -fPIC -fvisibility=hidden \
               -Wall -Wno-unused-function -Wno-sign-compare -std=c++11 -Wvla \
diff --git a/projects/rccl/makefiles/version.mk b/projects/rccl/makefiles/version.mk
index 55fa6cc273..be64e9a90c 100644
--- a/projects/rccl/makefiles/version.mk
+++ b/projects/rccl/makefiles/version.mk
@@ -1,6 +1,6 @@
 ##### version
 NCCL_MAJOR   := 2
-NCCL_MINOR   := 14
-NCCL_PATCH   := 3
+NCCL_MINOR   := 15
+NCCL_PATCH   := 5
 NCCL_SUFFIX  :=
 PKG_REVISION := 1
diff --git a/projects/rccl/src/channel.cc b/projects/rccl/src/channel.cc
index 95870086ec..05140766b1 100644
--- a/projects/rccl/src/channel.cc
+++ b/projects/rccl/src/channel.cc
@@ -20,14 +20,14 @@ ncclResult_t initChannel(struct ncclComm* comm, int channelId) {
 
   // The extra on nRanks+1 is for collnet root (i.e. network)
   channel->peers = ncclMemoryStackAlloc<struct ncclChannelPeer>(&comm->memPermanent, nRanks+1);
-  NCCLCHECK(ncclCudaCallocAsync(&channel->devPeers, nRanks+1, comm->deviceStream.stream));
+  NCCLCHECK(ncclCudaCallocAsync(&channel->devPeers, nRanks+1, comm->deviceStream.cudaStream));
   ncclCommPushCudaFree(comm, channel->devPeers);
 
   channel->ring.userRanks = ncclMemoryStackAlloc<int>(&comm->memPermanent, nRanks);
-  NCCLCHECK(ncclCudaCallocAsync(&channel->devRingUserRanks, nRanks, comm->deviceStream.stream));
+  NCCLCHECK(ncclCudaCallocAsync(&channel->devRingUserRanks, nRanks, comm->deviceStream.cudaStream));
   ncclCommPushCudaFree(comm, channel->devRingUserRanks);
 
-  NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNull(), &comm->deviceStream));
+  NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNone(), &comm->deviceStream));
 
   for (int r=0; r < nRanks+1; ++r) {
     for (int b=0; b < NCCL_MAX_CONNS; b++) {
diff --git a/projects/rccl/src/enqueue.cc b/projects/rccl/src/enqueue.cc
index 4fc77a3175..bf42e446f9 100644
--- a/projects/rccl/src/enqueue.cc
+++ b/projects/rccl/src/enqueue.cc
@@ -366,7 +366,7 @@ static void finishPlan(struct ncclKernelPlan* plan) {
   plan->channelCount = channelCount;
   plan->channelMask = channelMask;
   plan->hasProxyOps = hasProxyOps;
-  plan->threadPerBlock = std::max(plan->threadPerBlock, 4*WARP_SIZE);
+  plan->threadPerBlock = std::max(plan->threadPerBlock, 4*plan->comm->WarpSize);
 }
 
 static ncclResult_t registerIntraNodeBuffers(
@@ -918,19 +918,38 @@ ncclResult_t ncclLaunchPrepare(struct ncclComm* comm) {
     struct ncclKernelPlan* planHead = ncclIntruQueueHead(&comm->planQueue);
     comm->unlaunchedPlansHead = planHead;
 
+    // Semantically we want these dependencies for the kernels launched:
+    //   1. Launch host task on hostStream.
+    //   2. Launch kernel, depends on all of {deviceStream, hostStream, userStream[i]...}
+    //   3. {deviceStream, userStream[i]...} depend on kernel.
+    // We achieve this by:
+    //   1. userStream[0] waits on deviceStream
+    //   2. deviceStream waits on each of userStream[1...]
+    //   3. host task launch on hostStream
+    //   4. userStream[0] waits on hostStream
+    //   5. kernel launch on userStream[0]
+    //   6. deviceStream waits on userStream[0]
+    //   7. userStream[1...] each waits on deviceStream
+    // The two-level fan-in fan-out is because ncclStrongStreamWaitStream() requires
+    // at least one of the two streams to be strong-stream.
+    hipStream_t launchStream = tasks->streams->stream;
     NCCLCHECKGOTO(ncclStrongStreamAcquire(tasks->capturingGraph, &comm->deviceStream), result, failure);
 
-    // Create dependency for nccl device work on user streams.
-    for (struct ncclCudaStreamList* l=tasks->streams; l != nullptr && tasks->numStreams != 1; l = l->next) {
-      NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, &comm->deviceStream, l->stream), result, failure);
-    }
-
-    if (tasks->numStreams == 1 && tasks->streams->stream != comm->lastStream) {
+    if (tasks->numStreams != 1) {
+      // Create dependency for device stream on user streams. First from extra user
+      // streams to deviceStream. Then deviceStream to first user stream.
+      for (struct ncclCudaStreamList* l=tasks->streams->next; l != nullptr; l = l->next) {
+        NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, &comm->deviceStream, l->stream), result, failure);
+      }
+      NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, launchStream, &comm->deviceStream), result, failure);
+    } else if (tasks->streams->stream != comm->lastStream) {
       // Stream changed from last call, create dependency against last NCCL kernel launch
       CUDACHECK(hipStreamWaitEvent(tasks->streams->stream, comm->doneEvent, 0));
     }
 
     if (persistent || comm->persistentRefs != 0) {
+      // We have to launch host tasks to push proxy args. We are careful to only
+      // do this if necessary since host tasks impose a high performance cost in CUDA.
       bool acquired = false;
       for (struct ncclKernelPlan* plan=planHead; plan != nullptr; plan = plan->next) {
         if (plan->hasProxyOps) {
@@ -942,6 +961,8 @@ ncclResult_t ncclLaunchPrepare(struct ncclComm* comm) {
         }
       }
       if (acquired) {
+        // Make to-be-launched kernels dependent on just-launched host stream tasks.
+        if (tasks->numStreams != 1) NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, launchStream, &comm->hostStream), result, failure);
         NCCLCHECKGOTO(ncclStrongStreamRelease(tasks->capturingGraph, &comm->hostStream), result, failure);
       }
     }
@@ -967,8 +988,15 @@ ncclResult_t ncclLaunchKernelBefore_NoUncapturedCuda(struct ncclComm* comm, stru
   return ncclSuccess;
 }
 
+#if CUDART_VERSION >= 11080
+#define NCCL_MAX_CGA_CLUSTER_SIZE 8
+NCCL_PARAM(CGAClusterSize, "CGA_CLUSTER_SIZE", 0);
+#endif
+
 ncclResult_t ncclLaunchKernel(struct ncclComm* comm, struct ncclKernelPlan* plan) {
   struct ncclTasks* tasks = &comm->tasks;
+  void *fn = plan->kernelFn;
+  hipStream_t launchStream = tasks->streams->stream;
   dim3 grid = {(unsigned)plan->channelCount, 1, 1};
   dim3 block = {(unsigned)plan->threadPerBlock, 1, 1};
   void *args[3] = {&comm->devComm, &plan->channelMask, &plan->workHead};
@@ -976,9 +1004,54 @@ ncclResult_t ncclLaunchKernel(struct ncclComm* comm, struct ncclKernelPlan* plan
     CUDACHECK(hipExtLaunchKernel(plan->kernelFn, grid, block, args, 0, tasks->streams->stream, NULL, comm->doneEvent, 0));
     comm->lastStream = tasks->streams->stream;
   } else {
-    NCCLCHECK(ncclStrongStreamLaunchKernel(
-      tasks->capturingGraph, &comm->deviceStream, plan->kernelFn, grid, block, args, 0
-    ));
+    #if CUDART_VERSION >= 11080
+    int driverVersion;
+    NCCLCHECK(ncclCudaDriverVersion(&driverVersion));
+
+    unsigned int clusterSize = 0;
+    clusterSize = ncclParamCGAClusterSize();
+    if (clusterSize > NCCL_MAX_CGA_CLUSTER_SIZE) {
+      static bool warned = false;
+      if (warned == false) {
+        WARN("NCCL_CGA_CLUSTER_SIZE value %d is too big. Limiting value to %d.",
+             clusterSize, NCCL_MAX_CGA_CLUSTER_SIZE);
+        warned = true;
+      }
+      clusterSize = NCCL_MAX_CGA_CLUSTER_SIZE;
+    }
+
+    if (clusterSize && driverVersion >= 11080) {
+      cudaLaunchConfig_t launchConfig = {0};
+      cudaLaunchAttribute launchAttrs[2];
+      /* Cooperative Group Array (CGA)
+       * On sm90 and later we have an extra level of hierarchy where we
+       * can group together several blocks within the Grid, called
+       * Thread Block Clusters.
+       * Clusters enable multiple thread blocks running concurrently
+       * across multiple SMs to synchronize and collaboratively fetch
+       * and exchange data. A cluster of blocks are guaranteed to be
+       * concurrently scheduled onto a group of SMs.
+       * The maximum value is 8 and it must be divisible into the grid dimensions
+       */
+      // Grid dimension must be divisible by clusterSize
+      if (grid.x % clusterSize) clusterSize = 1;
+      launchAttrs[0].id = cudaLaunchAttributeClusterDimension;
+      launchAttrs[0].val.clusterDim = {clusterSize, 1, 1};
+      launchAttrs[1].id = cudaLaunchAttributeClusterSchedulingPolicyPreference;
+      launchAttrs[1].val.clusterSchedulingPolicyPreference = cudaClusterSchedulingPolicySpread;
+
+      launchConfig.gridDim = grid;
+      launchConfig.blockDim = block;
+      launchConfig.attrs = launchAttrs;
+      launchConfig.numAttrs = sizeof(launchAttrs)/sizeof(launchAttrs[0]);
+      launchConfig.stream = launchStream;
+
+      CUDACHECK(cudaLaunchKernelExC(&launchConfig, fn, args));
+      return ncclSuccess;
+    }
+    #endif
+    // Standard kernel launch
+    CUDACHECK(hipLaunchKernel(fn, grid, block, args, 0, launchStream));
   }
   return ncclSuccess;
 }
@@ -1005,18 +1078,22 @@ ncclResult_t ncclLaunchFinish(struct ncclComm* comm) {
     // Reset queue to empty without destroying plans since those will be sent
     // back to us for reclaiming via callbackQueue.
     ncclIntruQueueConstruct(&comm->planQueue);
-    // Close strong stream "transaction" encompassing cuda launches
-    NCCLCHECKGOTO(ncclStrongStreamRelease(tasks->capturingGraph, &comm->deviceStream), result, resume1);
+    hipStream_t launchStream = tasks->streams->stream; // First user stream gets launch
+    // Create dependency for deviceStream on launchStream.
+    if (tasks->numStreams != 1) NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, &comm->deviceStream, launchStream), result, resume1);
   resume1:
-    // Create dependency for user streams on nccl device work.
-    struct ncclCudaStreamList* sl = tasks->streams;
-    tasks->streams = nullptr; // reset streams to empty
+    // Create dependency for other user streams (skip launch stream).
+    struct ncclCudaStreamList* sl = tasks->streams->next;
+    tasks->streams = nullptr; // Reset comm->tasks.streams to empty.
     while (sl != nullptr && tasks->numStreams != 1) {
       NCCLCHECKGOTO(ncclStrongStreamWaitStream(tasks->capturingGraph, sl->stream, &comm->deviceStream), result, resume2);
     resume2:
       sl = sl->next;
     }
     tasks->numStreams = 0;
+    // Release device stream as acquired in ncclLaunchPrepare()
+    NCCLCHECKGOTO(ncclStrongStreamRelease(tasks->capturingGraph, &comm->deviceStream), result, resume3);
+  resume3:;
   }
   return result;
 }
@@ -1412,20 +1489,20 @@ static ncclResult_t taskAppend(struct ncclComm* comm, struct ncclInfo const* inf
           NCCLCHECK(ncclChannelComputeFromBase(comm, channelBaseId, c, &channelId));
           if (isSendNotRecv) {
             if (comm->channels[channelId].peers[peer].send[1].connected == 0) { // P2P uses only 1 connector
-              comm->connectSend[peer] |= (1<<channelId);
+              comm->connectSend[peer] |= (1UL<<channelId);
               ncclGroupCommPreconnect(comm);
             }
             if (comm->p2pNet && comm->channels[channelId].peers[peer].send[NCCL_CONN_IDX_P2P_NET].connected == 0) {
-              comm->connectSend[peer+comm->nRanks*NCCL_CONN_IDX_P2P_NET] |= (1<<channelId);
+              comm->connectSend[peer+comm->nRanks*NCCL_CONN_IDX_P2P_NET] |= (1UL<<channelId);
               ncclGroupCommPreconnect(comm);
             }
           } else {
             if (comm->channels[channelId].peers[peer].recv[1].connected == 0) { // P2P uses only 1 connector
-              comm->connectRecv[peer] |= (1<<channelId);
+              comm->connectRecv[peer] |= (1UL<<channelId);
               ncclGroupCommPreconnect(comm);
             }
             if (comm->p2pNet && comm->channels[channelId].peers[peer].recv[NCCL_CONN_IDX_P2P_NET].connected == 0) {
-              comm->connectRecv[peer+comm->nRanks*NCCL_CONN_IDX_P2P_NET] |= (1<<channelId);
+              comm->connectRecv[peer+comm->nRanks*NCCL_CONN_IDX_P2P_NET] |= (1UL<<channelId);
               ncclGroupCommPreconnect(comm);
             }
           }
@@ -1486,6 +1563,7 @@ static ncclResult_t taskAppend(struct ncclComm* comm, struct ncclInfo const* inf
       }
       if (l->stream == info->stream)
         break; // Already seen stream.
+      l = l->next;
     }
   }
   return ncclSuccess;
diff --git a/projects/rccl/src/graph/connect.cc b/projects/rccl/src/graph/connect.cc
index db1cd4495f..3ba3f8806d 100644
--- a/projects/rccl/src/graph/connect.cc
+++ b/projects/rccl/src/graph/connect.cc
@@ -35,7 +35,7 @@
 /******************************************************************/
 
 ncclResult_t ncclTopoPreset(struct ncclComm* comm,
-    struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph,
+    struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph,
     struct ncclTopoRanks* topoRanks) {
   int rank = comm->rank;
   int nChannels = comm->nChannels;
@@ -60,6 +60,7 @@ ncclResult_t ncclTopoPreset(struct ncclComm* comm,
 
     int* ringIntra = ringGraph->intra+c*localRanks;
     int* treeIntra = treeGraph->intra+c*localRanks;
+    int* collNetIntra = collNetGraph->intra+c*localRanks;
 
     for (int i=0; i<localRanks; i++) {
       if (ringIntra[i] == rank) {
@@ -78,8 +79,10 @@ ncclResult_t ncclTopoPreset(struct ncclComm* comm,
         topoRanks->treeToChild1[c] = treeIntra[child1Index];
         channel->tree.up         = i == 0 ? -1 : treeIntra[i-1];
         channel->tree.down[0]    = i == localRanks-1 ? -1 : treeIntra[i+1];
-        channel->collnetChain.up         = i == 0 ? comm->nRanks : treeIntra[i-1];
-        channel->collnetChain.down[0]    = i == localRanks-1 ? -1 : treeIntra[i+1];
+      }
+      if (collNetIntra[i] == rank) {
+        channel->collnetChain.up      = i == 0 ? comm->nRanks : collNetIntra[i-1];
+        channel->collnetChain.down[0] = i == localRanks-1 ? -1 : collNetIntra[i+1];
       }
     }
     topoRanks->ringPrev[c] = channel->ring.prev;
diff --git a/projects/rccl/src/graph/paths.cc b/projects/rccl/src/graph/paths.cc
index 9519a21d9a..bf95319b22 100644
--- a/projects/rccl/src/graph/paths.cc
+++ b/projects/rccl/src/graph/paths.cc
@@ -424,6 +424,19 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* system, int64_t busId, int
   return ncclSuccess;
 }
 
+// Set to 0 to disable the flush on Hopper when using GDR
+NCCL_PARAM(NetForceFlush, "NET_FORCE_FLUSH", 1);
+
+// Determine whether we need to flush the GDR recv buffers
+ncclResult_t ncclTopoNeedFlush(struct ncclTopoSystem* system, int64_t busId, int* flush) {
+  int g;
+  NCCLCHECK(ncclTopoIdToIndex(system, GPU, busId, &g));
+  struct ncclTopoNode* gpu = system->nodes[GPU].nodes+g;
+  // Flush is required on Ampere and earlier
+  *flush = gpu->gpu.cudaCompCap < 90 ? 1 : ncclParamNetForceFlush();
+  return ncclSuccess;
+}
+
 NCCL_PARAM(NetDisableIntra, "NET_DISABLE_INTRA", 1);
 
 // Check whether going through the network would be faster than going through P2P/SHM.
diff --git a/projects/rccl/src/graph/search.cc b/projects/rccl/src/graph/search.cc
index 8a8c551153..f680619ab5 100644
--- a/projects/rccl/src/graph/search.cc
+++ b/projects/rccl/src/graph/search.cc
@@ -797,7 +797,7 @@ float speedArrayIntra[] = { 24.0, 20.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5
 float speedArrayInter[] = { 24.0, 20.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.4, 1.2, 0.24, 0.12 };
 #else
 float speedArrayIntra[] = { 44.0, 30.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0 };
-float speedArrayInter[] = { 48.0, 30.0, 24.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.4, 1.2, 0.24, 0.12 };
+float speedArrayInter[] = { 48.0, 30.0, 28.0, 24.0, 22.0, 18.0, 15.0, 12.0, 10.0, 9.0, 7.0, 6.0, 5.0, 4.0, 3.0, 2.4, 1.2, 0.24, 0.12 };
 #endif
 #define NSPEEDSINTRA (sizeof(speedArrayIntra)/sizeof(float))
 #define NSPEEDSINTER (sizeof(speedArrayInter)/sizeof(float))
@@ -918,6 +918,7 @@ ncclResult_t ncclTopoCompute(ncclTopoSystem* system, struct ncclTopoGraph* graph
   while (speedArray[speedIndex] > system->maxBw && speedIndex < nspeeds-1) speedIndex++;
   tmpGraph.bwIntra = tmpGraph.bwInter = speedArray[speedIndex];
   int64_t globalTimeout = NCCL_SEARCH_GLOBAL_TIMEOUT;
+
 search:
   int time = tmpGraph.sameChannels ? NCCL_SEARCH_TIMEOUT_SAMECHANNELS :
     tmpGraph.pattern == NCCL_TOPO_PATTERN_TREE ? NCCL_SEARCH_TIMEOUT_TREE : NCCL_SEARCH_TIMEOUT;
@@ -953,6 +954,7 @@ search:
     if (time != -1) globalTimeout += time;
     else globalTimeout = NCCL_SEARCH_GLOBAL_TIMEOUT;
     if (globalTimeout < 0 && graph->nChannels) goto done;
+
     int maxTypeIntra = system->nodes[NET].count > 0 ? tmpGraph.typeInter : PATH_SYS;
     if (tmpGraph.typeIntra < maxTypeIntra && (graph->nChannels == 0 || tmpGraph.typeIntra < graph->typeIntra)) {
       tmpGraph.typeIntra += 1;
diff --git a/projects/rccl/src/graph/tuning.cc b/projects/rccl/src/graph/tuning.cc
index 1527cfa848..f45b881cc5 100644
--- a/projects/rccl/src/graph/tuning.cc
+++ b/projects/rccl/src/graph/tuning.cc
@@ -13,11 +13,11 @@
 NCCL_PARAM(Nthreads, "NTHREADS", -2);
 NCCL_PARAM(Ll128Nthreads, "LL128_NTHREADS", -2);
 
-static int getNthreads(const char* name, int env, int min, int max, int def) {
+static int getNthreads(const char* name, int env, int min, int max, int def, int WarpSize) {
   int nt = env;
   if (nt > 0) {
-    if (nt % WARP_SIZE != 0) {
-      WARN("Invalid %s %d (must be a multiple of %d)", name, nt, WARP_SIZE);
+    if (nt % WarpSize != 0) {
+      WARN("Invalid %s %d (must be a multiple of %d)", name, nt, WarpSize);
       nt = max;
     } else if (nt > max) {
       WARN("Invalid %s %d (maximum %d).", name, nt, max);
@@ -226,22 +226,36 @@ static struct tuningModel rcclTuningModel[] = {
   tuning_model_4,
 };
 
+/* Array indexes used below */
+#define VOLTA_COMPCAP_IDX 0
+#define AMPERE_COMPCAP_IDX 1
+#define HOPPER_COMPCAP_IDX 2
+
 // LL128 max BW per channel
 static const double ll128MaxBwPerCh = 20.0;
-static const double llMaxBws[2][3] = { /* Volta-N1/Intel-N2/Intel-N4) */ {39.0, 39.0, 20.4}, /* Ampere-N1/AMD-N2/AMD-N4) */ {87.7, 22.5 /*avg of ring & tree*/, 19.0} };
-static const double perChMaxTreeBws[2][3] = { /* Volta (N1/N2/N4) */ {26.5, 18.5, 10.0}, /* Ampere (N1/N2/N4) */ {24.0, 23.6, 17.8} };
+static const double llMaxBws[3][3] = {
+  /* Volta-N1/Intel-N2/Intel-N4) */ {39.0, 39.0, 20.4},
+  /* Ampere-N1/AMD-N2/AMD-N4) */ {87.7, 22.5 /*avg of ring & tree*/, 19.0},
+  /* Hopper-N1/AMD-N2/AMD-N4) */ {87.7, 22.5 /*avg of ring & tree*/, 19.0}
+};
+
+static const double perChMaxTreeBws[3][3] = {
+  /* Volta (N1/N2/N4) */ {26.5, 18.5, 10.0},
+  /* Ampere (N1/N2/N4) */ {24.0, 23.6, 17.8},
+  /* Hopper (N1/N2/N4) */ {24.0, 23.6, 17.8},
+};
 
 ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCompCap, struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph) {
   int simpleDefaultThreads = (ringGraph->bwIntra*ringGraph->nChannels <= PCI_BW) ? 256 : NCCL_SIMPLE_MAX_NTHREADS;
   comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_SIMPLE] =
 #if defined(__HIP_PLATFORM_HCC__) || defined(__HCC__) || defined(__HIPCC__)
-    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, simpleDefaultThreads);
+    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, simpleDefaultThreads, comm->WarpSize);
   comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = comm->maxThreads[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_SIMPLE] =
-    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
+    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS, comm->WarpSize);
   comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL] = comm->maxThreads[NCCL_ALGO_COLLNET_DIRECT][NCCL_PROTO_LL] =
-    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS);
+    getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 4*comm->WarpSize, NCCL_MAX_NTHREADS, NCCL_MAX_NTHREADS, comm->WarpSize);
   comm->maxThreads[NCCL_ALGO_RING][NCCL_PROTO_LL128] = comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_LL128] =
-    getNthreads("NCCL_LL128_NTHREADS", ncclParamLl128Nthreads(), 4*comm->WarpSize, NCCL_LL128_MAX_NTHREADS, NCCL_LL128_MAX_NTHREADS);
+    getNthreads("NCCL_LL128_NTHREADS", ncclParamLl128Nthreads(), 4*comm->WarpSize, NCCL_LL128_MAX_NTHREADS, NCCL_LL128_MAX_NTHREADS, comm->WarpSize);
 #else
     getNthreads("NCCL_NTHREADS", ncclParamNthreads(), 2*WARP_SIZE, NCCL_SIMPLE_MAX_NTHREADS, simpleDefaultThreads);
   comm->maxThreads[NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] =
@@ -258,14 +272,14 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
   int nRanks = comm->nRanks;
   if (nRanks <= 1) return ncclSuccess;
 
-  int compCap80 = minCompCap == 80 && maxCompCap == 80 ? 1 : 0;
+  int compCapIndex = (minCompCap == 80 && maxCompCap == 80) ? AMPERE_COMPCAP_IDX : ((minCompCap == 90 && maxCompCap == 90) ? HOPPER_COMPCAP_IDX : VOLTA_COMPCAP_IDX);
   int cpuArch, cpuVendor, cpuModel;
   NCCLCHECK(ncclTopoCpuType(comm->topo, &cpuArch, &cpuVendor, &cpuModel));
   int index2 = nNodes <= 2 ? nNodes-1 : 2;
   // LL: for single node, we look at GPU type; for multi-node, we look at CPU type
-  int index1 = nNodes == 1 ? compCap80 : cpuVendor == NCCL_TOPO_CPU_VENDOR_AMD ? 1 : 0;
+  int index1 = nNodes == 1 ? compCapIndex : cpuVendor == NCCL_TOPO_CPU_VENDOR_AMD ? 1 : 0;
   double llMaxBw = llMaxBws[index1][index2];
-  double perChMaxTreeBw = perChMaxTreeBws[compCap80][index2];
+  double perChMaxTreeBw = perChMaxTreeBws[compCapIndex][index2];
   // De-penalize Tree/Simple latency on Power systems to favor Tree than Ring
   //if (cpuArch == NCCL_TOPO_CPU_ARCH_POWER) hwLat[NCCL_HW_PCI][NCCL_ALGO_TREE][NCCL_PROTO_SIMPLE] = hwLat[NCCL_HW_PCI][NCCL_ALGO_RING][NCCL_PROTO_SIMPLE];
   float ppn = (float)nRanks / nNodes; // if ppn < 2, then we are sending/receiving at the same GPU through the NIC, apply some bw discount
@@ -298,7 +312,7 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
         else
           busBw *= rcclTuningModel[comm->topo->tuning].bwRatio[1][a][p];
 #else
-        if (compCap80) busBw = std::min(busBw, 235.0f);
+        if (compCapIndex == AMPERE_COMPCAP_IDX) busBw = std::min(busBw, 235.0f);
         if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL) { busBw = std::min(llMaxBw, busBw * ((nNodes > 1 || coll == ncclFuncAllReduce || coll == ncclFuncReduce) ? 1.0/4.0 : 1.0/3.0)); }
         if (a == NCCL_ALGO_RING && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (ppn < 2 ? 0.7 : 0.92 /*120.0/128.0*/), ll128MaxBwPerCh*graphs[a]->nChannels);
         if (a == NCCL_ALGO_TREE) busBw = std::min(busBw*.92, graphs[a]->nChannels*perChMaxTreeBw);
@@ -306,14 +320,14 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
         if (a == NCCL_ALGO_TREE && p == NCCL_PROTO_LL128) busBw = std::min(busBw * (nNodes == 1 ? 7.0/9.0 : 120.0/128.0), ll128MaxBwPerCh*graphs[a]->nChannels);
         if (a == NCCL_ALGO_COLLNET_DIRECT && p != NCCL_PROTO_SIMPLE) busBw = 0;  // Not used
         if (a == NCCL_ALGO_COLLNET_CHAIN && p != NCCL_PROTO_SIMPLE) busBw = 0;  // Not used
-      	if (a == NCCL_ALGO_COLLNET_DIRECT && p == NCCL_PROTO_SIMPLE) {
-                // Collnet+Direct requires all GPUs to have a local NIC to work at full speed
-                float factor = ppn / (1.0*graphs[a]->nChannels); // GPU/NIC ratio
-      	  factor -= (factor-1)/2;
-                busBw /= factor;
-      	}
-      	if (a == NCCL_ALGO_COLLNET_CHAIN && p == NCCL_PROTO_SIMPLE) busBw *= .75;
+        if (a == NCCL_ALGO_COLLNET_DIRECT && p == NCCL_PROTO_SIMPLE) {
+          // Collnet+Direct requires all GPUs to have a local NIC to work at full speed
+          float factor = ppn / (1.0*graphs[a]->nChannels); // GPU/NIC ratio
+          factor -= (factor-1)/2;
+          busBw /= factor;
+        }
 #endif
+        if (a == NCCL_ALGO_COLLNET_CHAIN && p == NCCL_PROTO_SIMPLE) busBw *= .75;
 
         // Convert bus BW to algorithm BW
         float ratio = (a != NCCL_ALGO_RING) ? .5 : (1.0 * nRanks) / nsteps;
@@ -387,11 +401,18 @@ ncclResult_t ncclTopoTuneModel(struct ncclComm* comm, int minCompCap, int maxCom
       pEnable = (graphs[a]->typeInter <= PATH_PXB) && graphs[a]->typeIntra <= PATH_NVL &&
         (comm->topo->nodes[GPU].nodes[0].gpu.gcn == 910 && comm->topo->ll128Enabled) ? 1 : 0;
 #else
-      // Enable LL128 by default only on Volta/Ampere+NVLink. Other cases are not tested and may cause silent data corruption.
-      pEnable = (graphs[a]->typeInter <= PATH_PXB) && graphs[a]->typeIntra <= PATH_NVL &&
-        ((minCompCap == 70 && maxCompCap == 70) || (minCompCap == 80 && maxCompCap == 80)) ? 1 : 0;
+      // Enable LL128 by default only on Volta/Ampere/Hopper+NVLink. Other cases are not tested and may cause silent data corruption.
+      pEnable = 1;
+      pEnable &= (graphs[a]->typeInter <= PATH_PXB);
+      pEnable &= (graphs[a]->typeIntra <= PATH_NVL);
+      pEnable &= (minCompCap == maxCompCap);
+      switch (minCompCap) {
+      case 70: pEnable &= 1; break;
+      case 80: pEnable &= 1; break;
+      case 90: pEnable &= !(CUDART_VERSION == 11080 && c == ncclFuncAllReduce && a == NCCL_ALGO_RING && comm->nRanks == 2); break;
+      default: pEnable &= 0; break;
+      }
 #endif
-      if (comm->rank == 0 && c == 0 && a == 0) INFO(NCCL_INIT, "Using tuning table %d with LL128 %s", comm->topo->tuning, pEnable ? "enabled" : "disabled");
     }
     if (pEnable == 0) comm->bandwidths[c][a][p] = 0;
     // Only disable algo for Allreduce since others only have one
diff --git a/projects/rccl/src/graph/xml.cc b/projects/rccl/src/graph/xml.cc
index 99dd5ab9f7..3b8691c6d8 100644
--- a/projects/rccl/src/graph/xml.cc
+++ b/projects/rccl/src/graph/xml.cc
@@ -688,7 +688,7 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, uint32_t rocmDev
     }
 #else
     // NVML NVLink detection
-    int maxNvLinks = (sm < 60) ? 0 : (sm < 70) ? 4 : (sm < 80) ? 6 : 12;
+    int maxNvLinks = (sm < 60) ? 0 : (sm < 70) ? 4 : (sm < 80) ? 6 : (sm < 90) ? 12 : 18;
 
     if (maxNvLinks > 0 && nvmlDev == NULL) {
       WARN("No NVML device handle. Skipping nvlink detection.");
@@ -701,8 +701,21 @@ ncclResult_t ncclTopoGetXmlFromGpu(struct ncclXmlNode* pciNode, uint32_t rocmDev
       if ((ncclNvmlDeviceGetNvLinkCapability(nvmlDev, l, NVML_NVLINK_CAP_P2P_SUPPORTED, &canP2P) != ncclSuccess) || !canP2P) continue;
 
       // Make sure the Nvlink is up. The previous call should have trained the link.
-      nvmlEnableState_t isActive;
-      if ((ncclNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive) != ncclSuccess) || (isActive != NVML_FEATURE_ENABLED)) continue;
+      nvmlEnableState_t isActive = NVML_FEATURE_DISABLED;
+#if CUDART_VERSION >= 11080
+      if (sm >= 90) {
+        nvmlFieldValue_t fv;
+        fv.fieldId = NVML_FI_DEV_NVLINK_GET_STATE;
+        fv.scopeId = l;
+        // fv.value will contain NV_FEATURE_ENABLED or NV_FEATURE_DISABLED
+        if ((ncclNvmlDeviceGetFieldValues(nvmlDev, 1, &fv) == ncclSuccess) && (fv.nvmlReturn == NVML_SUCCESS))
+          isActive = (nvmlEnableState_t) fv.value.uiVal;
+      } else /* FALLTHRU to GetNvLinkState if before SM90 */
+#endif
+      {
+        (void) ncclNvmlDeviceGetNvLinkState(nvmlDev, l, &isActive);
+      }
+      if (isActive != NVML_FEATURE_ENABLED) continue;
 
       // Try to figure out what's on the other side of the NVLink
       nvmlPciInfo_t remoteProc;
diff --git a/projects/rccl/src/group.cc b/projects/rccl/src/group.cc
index af03c1fd20..bc0e2a54a1 100644
--- a/projects/rccl/src/group.cc
+++ b/projects/rccl/src/group.cc
@@ -213,8 +213,8 @@ static void groupCleanup(struct ncclComm** groupCommHeadPtr, struct ncclComm** g
     for (int i = 0; i < comm->nRanks; i++) {
       comm->tasks.peers[i].sendSeen = false;
       comm->tasks.peers[i].recvSeen = false;
-      comm->connectSend[i] = 0;
-      comm->connectRecv[i] = 0;
+      comm->connectSend[i] = 0UL;
+      comm->connectRecv[i] = 0UL;
     }
     comm->unlaunchedPlansHead = nullptr;
     // Reclaim abandoned kernel plan memory. Note ncclWork structs were already
@@ -333,6 +333,8 @@ static ncclResult_t groupLaunch(struct ncclAsyncJob *job_) {
 
         job = job->next;
       } while (job != nullptr);
+      // Let preconnect threads progress.
+      if (jobsDone == false) usleep(1);
     } while (jobsDone == false);
 
     if (ret != ncclSuccess) goto fail;
diff --git a/projects/rccl/src/include/comm.h b/projects/rccl/src/include/comm.h
index 41997b35d6..ef7bc5c94f 100644
--- a/projects/rccl/src/include/comm.h
+++ b/projects/rccl/src/include/comm.h
@@ -174,8 +174,8 @@ struct ncclComm {
   ncclCollNet_t* ncclCollNet;
   void* bootstrap;
   // Bitmasks for ncclTransportP2pSetup
-  uint32_t* connectSend;
-  uint32_t* connectRecv;
+  uint64_t* connectSend;
+  uint64_t* connectRecv;
 
   int rank;    // my rank in the communicator
   int nRanks;  // number of GPUs in communicator
diff --git a/projects/rccl/src/include/cudawrap.h b/projects/rccl/src/include/cudawrap.h
index 2bd3b4dab0..0fd594582a 100644
--- a/projects/rccl/src/include/cudawrap.h
+++ b/projects/rccl/src/include/cudawrap.h
@@ -8,6 +8,8 @@
 #define NCCL_CUDAWRAP_H_
 
 #include <cuda.h>
+#include <cuda_runtime.h>
+#include "checks.h"
 
 #if CUDART_VERSION >= 11030
 #include <cudaTypedefs.h>
@@ -83,6 +85,18 @@ DECLARE_CUDA_PFN_EXTERN(cuDriverGetVersion, 2020);
 DECLARE_CUDA_PFN_EXTERN(cuGetProcAddress, 11030);
 
 
-ncclResult_t cudaLibraryInit(void);
+ncclResult_t ncclCudaLibraryInit(void);
+
+extern int ncclCudaDriverVersionCache;
+
+inline ncclResult_t ncclCudaDriverVersion(int* driver) {
+  int version = __atomic_load_n(&ncclCudaDriverVersionCache, __ATOMIC_RELAXED);
+  if (version == -1) {
+    CUDACHECK(cudaDriverGetVersion(&version));
+    __atomic_store_n(&ncclCudaDriverVersionCache, version, __ATOMIC_RELAXED);
+  }
+  *driver = version;
+  return ncclSuccess;
+}
 
 #endif
diff --git a/projects/rccl/src/include/graph.h b/projects/rccl/src/include/graph.h
index 2e4e6769b4..063c1b1289 100644
--- a/projects/rccl/src/include/graph.h
+++ b/projects/rccl/src/include/graph.h
@@ -37,6 +37,7 @@ ncclResult_t ncclTopoCheckGdr(struct ncclTopoSystem* topo, int64_t busId, int ne
 #define MAX_XGMI_INTER_GPUS 4
 ncclResult_t ncclTopoGetIntraNetDev(struct ncclTopoSystem* system, int rank, struct ncclTopoGraph* graph, int channelId, int type, int* dev);
 ncclResult_t ncclTopoGetLinkType(struct ncclTopoSystem* system, int cudaDev1, int cudaDev2, bool* isXGMI, int maxInter=MAX_XGMI_INTER_GPUS, int nInter=0, int *inter=nullptr);
+ncclResult_t ncclTopoNeedFlush(struct ncclTopoSystem* system, int64_t busId, int* flush);
 ncclResult_t ncclTopoCheckNet(struct ncclTopoSystem* system, int64_t id1, int64_t id2, int* net);
 int ncclPxnDisable(struct ncclComm* comm);
 ncclResult_t ncclTopoGetPxnRanks(struct ncclComm* comm, int** intermediateRanks, int* nranks);
@@ -109,7 +110,7 @@ struct ncclTopoRanks {
 };
 
 ncclResult_t ncclTopoPreset(struct ncclComm* comm,
-    struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph,
+    struct ncclTopoGraph* treeGraph, struct ncclTopoGraph* ringGraph, struct ncclTopoGraph* collNetGraph,
     struct ncclTopoRanks* topoRanks);
 
 ncclResult_t ncclTopoPostset(struct ncclComm* comm, int* firstRanks, int* treePatterns,
diff --git a/projects/rccl/src/include/nvmlwrap.h b/projects/rccl/src/include/nvmlwrap.h
index 29731dd835..fa1f5cf835 100644
--- a/projects/rccl/src/include/nvmlwrap.h
+++ b/projects/rccl/src/include/nvmlwrap.h
@@ -107,6 +107,75 @@ typedef enum nvmlGpuP2PCapsIndex_enum
     NVML_P2P_CAPS_INDEX_UNKNOWN
 } nvmlGpuP2PCapsIndex_t;
 
+/**
+ * Represents the type for sample value returned
+ */
+typedef enum nvmlValueType_enum
+{
+    NVML_VALUE_TYPE_DOUBLE = 0,
+    NVML_VALUE_TYPE_UNSIGNED_INT = 1,
+    NVML_VALUE_TYPE_UNSIGNED_LONG = 2,
+    NVML_VALUE_TYPE_UNSIGNED_LONG_LONG = 3,
+    NVML_VALUE_TYPE_SIGNED_LONG_LONG = 4,
+
+    // Keep this last
+    NVML_VALUE_TYPE_COUNT
+}nvmlValueType_t;
+
+
+/**
+ * Union to represent different types of Value
+ */
+typedef union nvmlValue_st
+{
+    double dVal;                    //!< If the value is double
+    unsigned int uiVal;             //!< If the value is unsigned int
+    unsigned long ulVal;            //!< If the value is unsigned long
+    unsigned long long ullVal;      //!< If the value is unsigned long long
+    signed long long sllVal;        //!< If the value is signed long long
+}nvmlValue_t;
+
+/**
+ * Field Identifiers.
+ *
+ * All Identifiers pertain to a device. Each ID is only used once and is guaranteed never to change.
+ */
+
+/* NVLink Speed */
+#define NVML_FI_DEV_NVLINK_SPEED_MBPS_COMMON 90  //!< Common NVLink Speed in MBps for active links
+#define NVML_FI_DEV_NVLINK_LINK_COUNT        91  //!< Number of NVLinks present on the device
+
+/**
+ * Remote device NVLink ID
+ *
+ * Link ID needs to be specified in the scopeId field in nvmlFieldValue_t.
+ */
+#define NVML_FI_DEV_NVLINK_REMOTE_NVLINK_ID     146 //!< Remote device NVLink ID
+
+/**
+ * NVSwitch: connected NVLink count
+ */
+#define NVML_FI_DEV_NVSWITCH_CONNECTED_LINK_COUNT   147  //!< Number of NVLinks connected to NVSwitch
+
+#define NVML_FI_DEV_NVLINK_GET_SPEED                  164
+#define NVML_FI_DEV_NVLINK_GET_STATE                  165
+#define NVML_FI_DEV_NVLINK_GET_VERSION                166
+#define NVML_FI_MAX 167 //!< One greater than the largest field ID defined above
+
+/**
+ * Information for a Field Value Sample
+ */
+typedef struct nvmlFieldValue_st
+{
+    unsigned int fieldId;       //!< ID of the NVML field to retrieve. This must be set before any call that uses this struct. See the constants starting with NVML_FI_ above.
+    unsigned int scopeId;       //!< Scope ID can represent data used by NVML depending on fieldId's context. For example, for NVLink throughput counter data, scopeId can represent linkId.
+    long long timestamp;        //!< CPU Timestamp of this value in microseconds since 1970
+    long long latencyUsec;      //!< How long this field value took to update (in usec) within NVML. This may be averaged across several fields that are serviced by the same driver call.
+    nvmlValueType_t valueType;  //!< Type of the value stored in value
+    nvmlReturn_t nvmlReturn;    //!< Return code for retrieving this value. This must be checked before looking at value, as value is undefined if nvmlReturn != NVML_SUCCESS
+    nvmlValue_t value;          //!< Value for this field. This is only valid if nvmlReturn == NVML_SUCCESS
+} nvmlFieldValue_t;
+
 /* End of nvml.h */
 #endif // NCCL_NVML_DIRECT
 
@@ -135,4 +204,6 @@ ncclResult_t ncclNvmlDeviceGetNvLinkRemotePciInfo(nvmlDevice_t device, unsigned
 ncclResult_t ncclNvmlDeviceGetNvLinkCapability(nvmlDevice_t device, unsigned int link, nvmlNvLinkCapability_t capability, unsigned int *capResult);
 ncclResult_t ncclNvmlDeviceGetCudaComputeCapability(nvmlDevice_t device, int* major, int* minor);
 ncclResult_t ncclNvmlDeviceGetP2PStatus(nvmlDevice_t device1, nvmlDevice_t device2, nvmlGpuP2PCapsIndex_t p2pIndex, nvmlGpuP2PStatus_t* p2pStatus);
+ncclResult_t ncclNvmlDeviceGetFieldValues(nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values);
+
 #endif // End include guard
diff --git a/projects/rccl/src/include/strongstream.h b/projects/rccl/src/include/strongstream.h
index 2882e55aa5..b57649e491 100644
--- a/projects/rccl/src/include/strongstream.h
+++ b/projects/rccl/src/include/strongstream.h
@@ -18,11 +18,11 @@
 struct ncclCudaGraph {
 #if CUDART_VERSION >= 11030
   cudaGraph_t graph;
-  uint64_t graphId;
+  unsigned long long graphId;
 #endif
 };
 
-inline struct ncclCudaGraph ncclCudaGraphNull() {
+inline struct ncclCudaGraph ncclCudaGraphNone() {
   struct ncclCudaGraph tmp;
   #if CUDART_VERSION >= 11030
     tmp.graph = nullptr;
@@ -50,7 +50,6 @@ inline bool ncclCudaGraphSame(struct ncclCudaGraph a, struct ncclCudaGraph b) {
 ncclResult_t ncclCudaGetCapturingGraph(struct ncclCudaGraph* graph, hipStream_t stream);
 ncclResult_t ncclCudaGraphAddDestructor(struct ncclCudaGraph graph, hipHostFn_t fn, void* arg);
 
-
 /* ncclStrongStream: An abstraction over CUDA streams that do not lose their
  * identity while being captured. Regular streams have the deficiency that the
  * captured form of a stream in one graph launch has no relation to the
@@ -58,37 +57,30 @@ ncclResult_t ncclCudaGraphAddDestructor(struct ncclCudaGraph graph, hipHostFn_t
  * streams unfit for the use of serializing access to a persistent resource.
  * Strong streams have been introduced to address this need.
  *
- * Constraints of using strong streams:
+ * - All updates to a strong stream must be enclosed by a Acquire/Release pair.
  *
- * - Operations that enqueue work to the strong stream need to be enclosed by
- *   ncclStrongStream[Acquire/Release] pairs. Acquire/release act like fences,
- *   the strong stream is not stateful so there is no harm in redundant acquire
- *   or releases.
+ * - The Acquire, Release, and all updates take a ncclCudaGraph parameter
+ *   indicating the currently capturing graph (or none). This parameter must be
+ *   the same for the entire sequence of {Acquire; ...; Release}.
  *
  * - An {Acquire; ...; Release} sequence must not be concurrent with any
  *   other operations against the strong stream including graph launches which
  *   reference this stream.
- *
- * - All strong stream functions take a "graph" parameter which must reference
- *   the currently capturing graph, or null if none.
  */
 struct ncclStrongStream;
 
 ncclResult_t ncclStrongStreamConstruct(struct ncclStrongStream* ss);
 ncclResult_t ncclStrongStreamDestruct(struct ncclStrongStream* ss);
 
-// Has this strong stream ever been captured in a graph.
-bool ncclStrongStreamEverCaptured(struct ncclStrongStream* ss);
-
 // Acquire-fence the strong stream.
 ncclResult_t ncclStrongStreamAcquire(
   struct ncclCudaGraph graph, struct ncclStrongStream* ss
 );
 
 // Acquire-fence the strong stream assuming no graph is capturing. This permits
-// the caller to enqueue directly to the `ss->stream` member using native CUDA
-// calls. Strong stream must be released via:
-//   ncclStrongStreamRelease(ncclCudaGraphNull(), graphRefs, ss);
+// the caller to enqueue directly to the `ss->cudaStream` member using native CUDA
+// calls. Strong stream still must be released via:
+//   ncclStrongStreamRelease(ncclCudaGraphNone(), ss);
 ncclResult_t ncclStrongStreamAcquireUncaptured(struct ncclStrongStream* ss);
 
 // Release-fence of the strong stream.
@@ -104,6 +96,7 @@ ncclResult_t ncclStrongStreamLaunchKernel(
   struct ncclCudaGraph graph, struct ncclStrongStream* ss,
   void* fn, dim3 grid, dim3 block, void** args, size_t sharedMemBytes
 );
+
 // Cause `a` to wait for the current state `b`. Both `a` and `b` must be acquired.
 ncclResult_t ncclStrongStreamWaitStream(
   struct ncclCudaGraph graph, struct ncclStrongStream* a, struct ncclStrongStream* b
@@ -122,21 +115,23 @@ ncclResult_t ncclStrongStreamSynchronize(struct ncclStrongStream* ss);
 
 ////////////////////////////////////////////////////////////////////////////////
 
+struct ncclStrongStreamGraph; // internal to ncclStrongStream
+
 struct ncclStrongStream {
-  hipStream_t stream;
-  hipEvent_t event;
-  #if CUDART_VERSION >= 11030
-  cudaGraphNode_t node; // null if never captured, otherwise never null again
-  uint64_t graphId:63, eventIsLagging:1;
-  #endif
+  // Used when not graph capturing.
+  hipStream_t cudaStream;
+#if CUDART_VERSION >= 11030
+  // The event used to establish order between graphs and streams. During acquire
+  // this event is waited on, during release it is recorded to.
+  cudaEvent_t serialEvent;
+  // This stream ever appeared in a graph capture.
+  bool everCaptured;
+  // Tracks whether serialEvent needs to be recorded to upon Release().
+  bool serialEventNeedsRecord;
+  struct ncclStrongStreamGraph* graphHead;
+#else
+  hipEvent_t scratchEvent;
+#endif
 };
 
-inline bool ncclStrongStreamEverCaptured(struct ncclStrongStream* ss) {
-  #if CUDART_VERSION >= 11030
-    return ss->node != nullptr;
-  #else
-    return false;
-  #endif
-}
-
 #endif
diff --git a/projects/rccl/src/init.cc b/projects/rccl/src/init.cc
index 7e3ccef8f6..8c6be42fa7 100644
--- a/projects/rccl/src/init.cc
+++ b/projects/rccl/src/init.cc
@@ -540,7 +540,7 @@ static ncclResult_t devCommSetup(ncclComm_t comm) {
 
   int nRanks = comm->nRanks;
   struct ncclDevCommAndChannels *devCommAndChans, tmpCommAndChans;
-  NCCLCHECK(ncclCudaCallocAsync(&devCommAndChans, 1, comm->deviceStream.stream));
+  NCCLCHECK(ncclCudaCallocAsync(&devCommAndChans, 1, comm->deviceStream.cudaStream));
   ncclCommPushCudaFree(comm, devCommAndChans);
   comm->devComm = &devCommAndChans->comm;
   tmpCommAndChans.comm.rank = comm->rank;
@@ -587,7 +587,7 @@ static ncclResult_t devCommSetup(ncclComm_t comm) {
     tmpCommAndChans.channels[c].workFifoDone = &comm->workFifoDone[c];
 
     if (comm->channels[c].ring.userRanks != nullptr) {
-      NCCLCHECK(ncclCudaMemcpyAsync(tmpCommAndChans.channels[c].ring.userRanks, comm->channels[c].ring.userRanks, nRanks, comm->deviceStream.stream));
+      NCCLCHECK(ncclCudaMemcpyAsync(tmpCommAndChans.channels[c].ring.userRanks, comm->channels[c].ring.userRanks, nRanks, comm->deviceStream.cudaStream));
     }
   }
 
@@ -608,10 +608,9 @@ static ncclResult_t devCommSetup(ncclComm_t comm) {
   NCCLCHECK(ncclCudaCalloc(&tmpCommAndChans.comm.devProf, MAXCHANNELS*PROFILE_NUM_LAUNCHES), comm->sideStream);
 #endif
 
-  NCCLCHECK(ncclCudaMemcpyAsync(devCommAndChans, &tmpCommAndChans, 1, comm->deviceStream.stream));
-  CUDACHECK(hipStreamSynchronize(comm->deviceStream.stream));
-  NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNull(), &comm->deviceStream));
-
+  NCCLCHECK(ncclCudaMemcpyAsync(devCommAndChans, &tmpCommAndChans, 1, comm->deviceStream.cudaStream));
+  CUDACHECK(hipStreamSynchronize(comm->deviceStream.cudaStream));
+  NCCLCHECK(ncclStrongStreamRelease(ncclCudaGraphNone(), &comm->deviceStream));
   return ncclSuccess;
 }
 
@@ -942,7 +941,7 @@ static ncclResult_t initTransportsRank(struct ncclComm* comm, ncclUniqueId* comm
 
   comm->nChannels = (comm->topo->nodes[GPU].count != comm->topo->nRanks && comm->topo->nodes[NET].count)
     ? std::min(treeGraph.nChannels, ringGraph.nChannels) : ringGraph.nChannels;
-  NCCLCHECK(ncclTopoPreset(comm, &treeGraph, &ringGraph, &allGather3Data[rank].topoRanks));
+  NCCLCHECK(ncclTopoPreset(comm, &treeGraph, &ringGraph, &collNetGraph, &allGather3Data[rank].topoRanks));
 
   NCCLCHECK(bootstrapAllGather(comm->bootstrap, allGather3Data, sizeof(*allGather3Data)));
 
@@ -1288,13 +1287,13 @@ collnet_cleanup:
       for (int c=0; c<comm->p2pnChannelsPerPeer; c++) {
         NCCLCHECK(ncclChannelCompute(comm, peer, c, ncclFuncSend, &channelId));
         if (comm->channels[channelId].peers[peer].send[1].connected == 0) {
-          comm->connectSend[peer] |= (1<<channelId);
+          comm->connectSend[peer] |= (1UL<<channelId);
         }
       }
       for (int c=0; c<comm->p2pnChannelsPerPeer; c++) {
         NCCLCHECK(ncclChannelCompute(comm, peer, c, ncclFuncRecv, &channelId));
         if (comm->channels[channelId].peers[peer].recv[1].connected == 0) {
-          comm->connectRecv[peer] |= (1<<channelId);
+          comm->connectRecv[peer] |= (1UL<<channelId);
         }
       }
     }
@@ -1562,6 +1561,7 @@ ncclResult_t ncclCommInitAll(ncclComm_t* comms, int ndev, const int* devlist) {
       gpuFlags[devlist[i]] = 1;
     }
     free(gpuFlags);
+    gpuFlags = nullptr;
   }
 
   ncclUniqueId uniqueId;
@@ -1573,11 +1573,9 @@ ncclResult_t ncclCommInitAll(ncclComm_t* comms, int ndev, const int* devlist) {
   }
   NCCLCHECKGOTO(ncclGroupEnd(), ret, fail);
 
-exit:
-  return ret;
 fail:
-  if (gpuFlags) free(gpuFlags);
-  goto exit;
+  free(gpuFlags);
+  return ret;
 }
 
 ncclResult_t ncclCommSetAsyncError(ncclComm_t comm, ncclResult_t nextState) {
diff --git a/projects/rccl/src/misc/cudawrap.cc b/projects/rccl/src/misc/cudawrap.cc
index 52663b5bf7..b1786f4160 100644
--- a/projects/rccl/src/misc/cudawrap.cc
+++ b/projects/rccl/src/misc/cudawrap.cc
@@ -38,7 +38,7 @@ DECLARE_CUDA_PFN(cuGetProcAddress, 11030);
 #define CUDA_DRIVER_MIN_VERSION 11030
 
 static void *cudaLib;
-static int cudaDriverVersion;
+int ncclCudaDriverVersionCache = -1;
 
 #if CUDART_VERSION >= 11030
 /*
@@ -107,16 +107,17 @@ static void initOnceFunc() {
     goto error;
   }
 
-  res = pfn_cuDriverGetVersion(&cudaDriverVersion);
+  int driverVersion;
+  res = pfn_cuDriverGetVersion(&driverVersion);
   if (res != 0) {
     WARN("cuDriverGetVersion failed with %d", res);
     goto error;
   }
 
-  INFO(NCCL_INIT, "cudaDriverVersion %d", cudaDriverVersion);
+  INFO(NCCL_INIT, "cudaDriverVersion %d", driverVersion);
 
-  if (cudaDriverVersion < CUDA_DRIVER_MIN_VERSION) {
-    // WARN("CUDA Driver version found is %d. Minimum requirement is %d", cudaDriverVersion, CUDA_DRIVER_MIN_VERSION);
+  if (driverVersion < CUDA_DRIVER_MIN_VERSION) {
+    // WARN("CUDA Driver version found is %d. Minimum requirement is %d", driverVersion, CUDA_DRIVER_MIN_VERSION);
     // Silently ignore version check mismatch for backwards compatibility
     goto error;
   }
@@ -148,7 +149,7 @@ error:
   return;
 }
 
-ncclResult_t cudaLibraryInit() {
+ncclResult_t ncclCudaLibraryInit() {
   pthread_once(&initOnceControl, initOnceFunc);
   return initResult;
 }
diff --git a/projects/rccl/src/misc/nvmlwrap.cc b/projects/rccl/src/misc/nvmlwrap.cc
index 5db7c6be5c..2de993a6e5 100644
--- a/projects/rccl/src/misc/nvmlwrap.cc
+++ b/projects/rccl/src/misc/nvmlwrap.cc
@@ -38,6 +38,7 @@ namespace {
   NCCL_NVML_FN(nvmlDeviceGetNvLinkCapability, nvmlReturn_t, (nvmlDevice_t device, unsigned int link, nvmlNvLinkCapability_t capability, unsigned int *capResult))
   NCCL_NVML_FN(nvmlDeviceGetCudaComputeCapability, nvmlReturn_t, (nvmlDevice_t device, int* major, int* minor))
   NCCL_NVML_FN(nvmlDeviceGetP2PStatus, nvmlReturn_t, (nvmlDevice_t device1, nvmlDevice_t device2, nvmlGpuP2PCapsIndex_t p2pIndex, nvmlGpuP2PStatus_t* p2pStatus))
+  NCCL_NVML_FN(nvmlDeviceGetFieldValues, nvmlReturn_t, (nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values))
 
   std::mutex lock; // NVML has had some thread safety bugs
   bool initialized = false;
@@ -80,7 +81,8 @@ ncclResult_t ncclNvmlEnsureInitialized() {
       {(void**)&pfn_nvmlDeviceGetNvLinkRemotePciInfo, "nvmlDeviceGetNvLinkRemotePciInfo"},
       {(void**)&pfn_nvmlDeviceGetNvLinkCapability, "nvmlDeviceGetNvLinkCapability"},
       {(void**)&pfn_nvmlDeviceGetCudaComputeCapability, "nvmlDeviceGetCudaComputeCapability"},
-      {(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"}
+      {(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"},
+      {(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"}
     };
     for(Symbol sym: symbols) {
       *sym.ppfn = dlsym(libhandle, sym.name);
@@ -260,3 +262,10 @@ ncclResult_t ncclNvmlDeviceGetP2PStatus(
   }
   return ncclSuccess;
 }
+
+ncclResult_t ncclNvmlDeviceGetFieldValues(nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values) {
+  NCCLCHECK(ncclNvmlEnsureInitialized());
+  std::lock_guard<std::mutex> locked(lock);
+  NVMLTRY(nvmlDeviceGetFieldValues, device, valuesCount, values);
+  return ncclSuccess;
+}
diff --git a/projects/rccl/src/misc/strongstream.cc b/projects/rccl/src/misc/strongstream.cc
index 936d720a85..9e0dcc4666 100644
--- a/projects/rccl/src/misc/strongstream.cc
+++ b/projects/rccl/src/misc/strongstream.cc
@@ -5,37 +5,65 @@
  ************************************************************************/
 
 #include "strongstream.h"
+#include "rocmwrap.h"
 #include "checks.h"
 #include "param.h"
 
+// Tracks the chain of graph nodes for a given graph captured identified by
+// its graph id. This state has to live for as long as captured work is being
+// submitted. CUDA doesn't have mechanism to inform us when the user ends capture
+// so the best we can do is get notified when the graph is destroyed.
+struct ncclStrongStreamGraph {
+  struct ncclStrongStreamGraph* next;
+  // Atomically exchanged to false by both the main thread or the graph destructor
+  // callback. The last to arrive deletes the node.
+  bool alive;
+  unsigned long long graphId;
+  // For each graph we track the "tip" of the chain of graph nodes. A linear
+  // chain would always have just one node at its tip, but since we have to merge
+  // in chains from other streams (via ncclStrongStreamWaitStream) some spots
+  // in the chain can be wider than a single node and thus need a list, so we
+  // maintain a dynamically sized array of tip nodes.
+  int tipCount, tipCapacity;
+  hipGraphNode_t* tipNodes;
+};
+
+static void ncclStrongStreamGraphDelete(struct ncclStrongStreamGraph* g) {
+  free(g->tipNodes);
+  free(g);
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 ncclResult_t ncclCudaGetCapturingGraph(
     struct ncclCudaGraph* graph, hipStream_t stream
   ) {
-  #if CUDART_VERSION >= 11030
-    thread_local int driver = -1;
-    if (driver == -1) {
-      CUDACHECK(cudaDriverGetVersion(&driver));
-    }
-    if (driver < 11030) {
+  #if CUDART_VERSION >= 10000 // cudaStreamGetCaptureInfo
+    int driver;
+    NCCLCHECK(ncclCudaDriverVersion(&driver));
+    if (CUDART_VERSION < 11030 || driver < 11030) {
       cudaStreamCaptureStatus status;
       unsigned long long gid;
-      graph->graph = nullptr;
       CUDACHECK(cudaStreamGetCaptureInfo(stream, &status, &gid));
+      #if CUDART_VERSION >= 11030
+        graph->graph = nullptr;
+        graph->graphId = ULLONG_MAX;
+      #endif
       if (status != cudaStreamCaptureStatusNone) {
-        WARN("The installed CUDA driver is older than the minimum version (R465) required for NCCL's CUDA Graphs support");
+        WARN("NCCL cannot be captured in a graph if either it wasn't built with CUDA runtime >= 11.3 or if the installed CUDA driver < R465.");
         return ncclInvalidUsage;
       }
     } else {
-      cudaStreamCaptureStatus status;
-      unsigned long long gid;
-      CUDACHECK(cudaStreamGetCaptureInfo_v2(stream, &status, &gid, &graph->graph, nullptr, nullptr));
-      if (status != cudaStreamCaptureStatusActive) {
-        graph->graph = nullptr;
-        gid = ULLONG_MAX;
-      }
-      graph->graphId = gid;
+      #if CUDART_VERSION >= 11030
+        cudaStreamCaptureStatus status;
+        unsigned long long gid;
+        CUDACHECK(cudaStreamGetCaptureInfo_v2(stream, &status, &gid, &graph->graph, nullptr, nullptr));
+        if (status != cudaStreamCaptureStatusActive) {
+          graph->graph = nullptr;
+          gid = ULLONG_MAX;
+        }
+        graph->graphId = gid;
+      #endif
     }
   #endif
   return ncclSuccess;
@@ -58,52 +86,114 @@ ncclResult_t ncclCudaGraphAddDestructor(struct ncclCudaGraph graph, hipHostFn_t
 ////////////////////////////////////////////////////////////////////////////////
 
 ncclResult_t ncclStrongStreamConstruct(struct ncclStrongStream* ss) {
-  CUDACHECK(hipStreamCreateWithFlags(&ss->stream, hipStreamNonBlocking));
-  CUDACHECK(hipEventCreateWithFlags(&ss->event, hipEventDisableTiming));
+  CUDACHECK(hipStreamCreateWithFlags(&ss->cudaStream, hipStreamNonBlocking));
   #if CUDART_VERSION >= 11030
-    ss->node = nullptr;
-    ss->graphId = (1ull<<(8*sizeof(long long)-1))-1;
-    ss->eventIsLagging = 0;
+    CUDACHECK(cudaEventCreateWithFlags(&ss->serialEvent, cudaEventDisableTiming));
+    ss->everCaptured = false;
+    ss->serialEventNeedsRecord = false;
+    ss->graphHead = nullptr;
+  #else
+    CUDACHECK(hipEventCreateWithFlags(&ss->scratchEvent, hipEventDisableTiming));
   #endif
   return ncclSuccess;
 }
 
+static void graphDestructor(void* arg) {
+  struct ncclStrongStreamGraph* g = (struct ncclStrongStreamGraph*)arg;
+  if (false == __atomic_exchange_n(&g->alive, false, __ATOMIC_ACQ_REL)) {
+    // Last to arrive deletes list node.
+    ncclStrongStreamGraphDelete(g);
+  }
+}
+
 ncclResult_t ncclStrongStreamDestruct(struct ncclStrongStream* ss) {
+  CUDACHECK(hipStreamDestroy(ss->cudaStream));
   #if CUDART_VERSION >= 11030
-    CUDACHECK(cudaEventDestroy(ss->event));
+    CUDACHECK(hipEventDestroy(ss->serialEvent));
+    // Delete list of per-graph chains.
+    struct ncclStrongStreamGraph* g = ss->graphHead;
+    while (g != nullptr) {
+      struct ncclStrongStreamGraph* next = g->next;
+      if (false == __atomic_exchange_n(&g->alive, false, __ATOMIC_ACQ_REL)) {
+        // Last to arrive deletes list node.
+        ncclStrongStreamGraphDelete(g);
+      }
+      g = next;
+    }
+  #else
+    CUDACHECK(hipEventDestroy(ss->scratchEvent));
   #endif
-  CUDACHECK(hipStreamDestroy(ss->stream));
   return ncclSuccess;
 }
 
 NCCL_PARAM(GraphMixingSupport, "GRAPH_MIXING_SUPPORT", 1)
 
+static void ensureTips(struct ncclStrongStreamGraph* g, int n) {
+  if (g->tipCapacity < n) {
+    g->tipNodes = (hipGraphNode_t*)realloc(g->tipNodes, n*sizeof(hipGraphNode_t));
+    g->tipCapacity = n;
+  }
+}
+
 ncclResult_t ncclStrongStreamAcquire(
     struct ncclCudaGraph graph, struct ncclStrongStream* ss
   ) {
   #if CUDART_VERSION >= 11030
     bool mixing = ncclParamGraphMixingSupport();
     if (graph.graph == nullptr) {
-      if (mixing && ncclStrongStreamEverCaptured(ss)) {
-        CUDACHECK(cudaStreamWaitEvent(ss->stream, ss->event, 0));
-        ss->eventIsLagging = 0;
+      if (mixing && ss->everCaptured) {
+        CUDACHECK(cudaStreamWaitEvent(ss->cudaStream, ss->serialEvent, 0));
+        ss->serialEventNeedsRecord = false;
       }
     } else {
-      if (ss->graphId != graph.graphId) {
-        if (mixing && ss->eventIsLagging) {
-          // Can only be here if previous release was for uncaptured work that
-          // elided updating the event because no capture had yet occurred.
-          CUDACHECK(cudaStreamWaitEvent(ss->stream, ss->event, 0));
-          CUDACHECK(cudaEventRecord(ss->event, ss->stream));
-        }
-        ss->graphId = graph.graphId;
-        ss->eventIsLagging = 0;
-        if (mixing) {
-          CUDACHECK(cudaGraphAddEventWaitNode(&ss->node, graph.graph, nullptr, 0, ss->event));
+      ss->everCaptured = true;
+      // Find the current graph in our list of graphs if it exists.
+      struct ncclStrongStreamGraph** pg = &ss->graphHead;
+      struct ncclStrongStreamGraph* g;
+      while (*pg != nullptr) {
+        g = *pg;
+        if (g->graphId == graph.graphId) {
+          // Move to front of list so that operations after acquire don't have to search the list.
+          *pg = g->next;
+          g->next = ss->graphHead;
+          ss->graphHead = g;
+          return ncclSuccess;
+        } else if (false == __atomic_load_n(&g->alive, __ATOMIC_ACQUIRE)) {
+          // Unrelated graph that has been destroyed. Remove and delete.
+          *pg = g->next;
+          ncclStrongStreamGraphDelete(g);
         } else {
-          CUDACHECK(cudaGraphAddEmptyNode(&ss->node, graph.graph, nullptr, 0));
+          pg = &g->next;
         }
       }
+
+      // This is a new graph so add to the list.
+      g = (struct ncclStrongStreamGraph*)malloc(sizeof(struct ncclStrongStreamGraph));
+      g->graphId = graph.graphId;
+      g->tipNodes = nullptr;
+      g->tipCapacity = 0;
+      g->tipCount = 0;
+      g->next = ss->graphHead;
+      ss->graphHead = g;
+      g->alive = true;
+      NCCLCHECK(ncclCudaGraphAddDestructor(graph, graphDestructor, (void*)g));
+
+      if (mixing && ss->serialEventNeedsRecord) {
+        // Can only be here if previous release was for uncaptured work that
+        // elided updating the event because no capture had yet occurred.
+        CUDACHECK(cudaStreamWaitEvent(ss->cudaStream, ss->serialEvent, 0));
+        CUDACHECK(cudaEventRecord(ss->serialEvent, ss->cudaStream));
+      }
+      ss->serialEventNeedsRecord = false;
+
+      // First node in the chain must be a wait on the serialEvent.
+      if (mixing) {
+        ensureTips(g, 1);
+        CUDACHECK(cudaGraphAddEventWaitNode(&g->tipNodes[0], graph.graph, nullptr, 0, ss->serialEvent));
+        g->tipCount = 1;
+      } else {
+        g->tipCount = 0;
+      }
     }
   #endif
   return ncclSuccess;
@@ -112,26 +202,38 @@ ncclResult_t ncclStrongStreamAcquire(
 ncclResult_t ncclStrongStreamAcquireUncaptured(struct ncclStrongStream* ss) {
   #if CUDART_VERSION >= 11030
     bool mixing = ncclParamGraphMixingSupport();
-    if (mixing && ncclStrongStreamEverCaptured(ss)) {
-      CUDACHECK(cudaStreamWaitEvent(ss->stream, ss->event, 0));
+    if (mixing && ss->everCaptured) {
+      CUDACHECK(cudaStreamWaitEvent(ss->cudaStream, ss->serialEvent, 0));
     }
-    ss->eventIsLagging = 1; // Assume the caller is going to add work to stream.
+    ss->serialEventNeedsRecord = true; // Assume the caller is going to add work to stream.
   #endif
   return ncclSuccess;
 }
 
+static ncclResult_t checkGraphId(struct ncclStrongStreamGraph* g, unsigned long long id) {
+  if (g == nullptr || g->graphId != id) {
+    WARN("Expected graph id=%llu was not at head of strong stream's internal list.", id);
+    return ncclInternalError;
+  }
+  return ncclSuccess;
+}
+
 ncclResult_t ncclStrongStreamRelease(struct ncclCudaGraph graph, struct ncclStrongStream* ss) {
   #if CUDART_VERSION >= 11030
     bool mixing = ncclParamGraphMixingSupport();
-    if (mixing && ss->eventIsLagging) {
+    if (mixing && ss->serialEventNeedsRecord) {
       if (graph.graph == nullptr) {
-        if (ncclStrongStreamEverCaptured(ss)) {
-          CUDACHECK(cudaEventRecord(ss->event, ss->stream));
-          ss->eventIsLagging = 0;
+        if (ss->everCaptured) {
+          CUDACHECK(cudaEventRecord(ss->serialEvent, ss->cudaStream));
+          ss->serialEventNeedsRecord = false;
         }
       } else {
-        CUDACHECK(cudaGraphAddEventRecordNode(&ss->node, graph.graph, &ss->node, 1, ss->event));
-        ss->eventIsLagging = 0;
+        struct ncclStrongStreamGraph* g = ss->graphHead;
+        NCCLCHECK(checkGraphId(g, graph.graphId));
+        ensureTips(g, 1);
+        CUDACHECK(cudaGraphAddEventRecordNode(&g->tipNodes[0], graph.graph, g->tipNodes, g->tipCount, ss->serialEvent));
+        g->tipCount = 1;
+        ss->serialEventNeedsRecord = false;
       }
     }
   #endif
@@ -143,17 +245,20 @@ ncclResult_t ncclStrongStreamLaunchHost(
   ) {
   #if CUDART_VERSION >= 11030
     if (graph.graph == nullptr) {
-      CUDACHECK(cudaLaunchHostFunc(ss->stream, fn, arg));
+      CUDACHECK(cudaLaunchHostFunc(ss->cudaStream, fn, arg));
     } else {
       cudaHostNodeParams p;
       p.fn = fn;
       p.userData = arg;
-      CUDACHECK(cudaGraphAddHostNode(&ss->node, graph.graph, &ss->node, 1, &p));
+      struct ncclStrongStreamGraph* g = ss->graphHead;
+      NCCLCHECK(checkGraphId(g, graph.graphId));
+      ensureTips(g, 1);
+      CUDACHECK(cudaGraphAddHostNode(&g->tipNodes[0], graph.graph, g->tipNodes, g->tipCount, &p));
+      g->tipCount = 1;
     }
-    ss->eventIsLagging = 1;
+    ss->serialEventNeedsRecord = true;
   #else
-    //CUDACHECK(hipLaunchHostFunc(ss->stream, fn, arg));
-    CUDACHECK(hipStreamAddCallback(ss->stream, (hipStreamCallback_t)fn, arg, 0));
+    CUDACHECK(hipStreamAddCallback(ss->cudaStream, (hipStreamCallback_t)fn, arg, 0));
   #endif
   return ncclSuccess;
 }
@@ -164,9 +269,8 @@ ncclResult_t ncclStrongStreamLaunchKernel(
   ) {
   #if CUDART_VERSION >= 11030
     if (graph.graph == nullptr) {
-      CUDACHECK(cudaLaunchKernel(fn, grid, block, args, sharedMemBytes, ss->stream));
+      CUDACHECK(cudaLaunchKernel(fn, grid, block, args, sharedMemBytes, ss->cudaStream));
     } else {
-      cudaGraphNode_t tip = ss->node;
       cudaKernelNodeParams p;
       p.func = fn;
       p.gridDim = grid;
@@ -174,33 +278,53 @@ ncclResult_t ncclStrongStreamLaunchKernel(
       p.kernelParams = args;
       p.sharedMemBytes = sharedMemBytes;
       p.extra = nullptr;
-      CUDACHECK(cudaGraphAddKernelNode(&ss->node, graph.graph, &tip, 1, &p));
+      struct ncclStrongStreamGraph* g = ss->graphHead;
+      NCCLCHECK(checkGraphId(g, graph.graphId));
+      ensureTips(g, 1);
+      CUDACHECK(cudaGraphAddKernelNode(&g->tipNodes[0], graph.graph, g->tipNodes, g->tipCount, &p));
+      g->tipCount = 1;
     }
-    ss->eventIsLagging = 1;
+    ss->serialEventNeedsRecord = true;
   #else
-    CUDACHECK(hipLaunchKernel(fn, grid, block, args, sharedMemBytes, ss->stream));
+    CUDACHECK(hipLaunchKernel(fn, grid, block, args, sharedMemBytes, ss->cudaStream));
   #endif
   return ncclSuccess;
 }
 
+// Merge node list `b` into list `a` but don't add duplicates.
+static void mergeTips(struct ncclStrongStreamGraph* a, hipGraphNode_t const* bNodes, int bn) {
+  int an = a->tipCount;
+  ensureTips(a, an + bn);
+  for (int bi=0; bi < bn; bi++) {
+    for (int ai=0; ai < an; ai++) {
+      if (a->tipNodes[ai] == bNodes[bi]) goto next_b;
+    }
+    a->tipNodes[a->tipCount++] = bNodes[bi];
+  next_b:;
+  }
+}
+
 ncclResult_t ncclStrongStreamWaitStream(
     struct ncclCudaGraph graph, struct ncclStrongStream* a, struct ncclStrongStream* b
   ) {
   #if CUDART_VERSION >= 11030
     if (graph.graph == nullptr) {
-      if (b->eventIsLagging) {
-        b->eventIsLagging = 0;
-        CUDACHECK(cudaEventRecord(b->event, b->stream));
+      if (b->serialEventNeedsRecord) {
+        b->serialEventNeedsRecord = false;
+        CUDACHECK(cudaEventRecord(b->serialEvent, b->cudaStream));
       }
-      CUDACHECK(cudaStreamWaitEvent(a->stream, b->event, 0));
-      a->eventIsLagging = 1;
+      CUDACHECK(cudaStreamWaitEvent(a->cudaStream, b->serialEvent, 0));
     } else {
-      cudaGraphNode_t pair[2] = {a->node, b->node};
-      CUDACHECK(cudaGraphAddEmptyNode(&a->node, graph.graph, pair, 2));
+      struct ncclStrongStreamGraph* ag = a->graphHead;
+      NCCLCHECK(checkGraphId(ag, graph.graphId));
+      struct ncclStrongStreamGraph* bg = b->graphHead;
+      NCCLCHECK(checkGraphId(bg, graph.graphId));
+      mergeTips(ag, bg->tipNodes, bg->tipCount);
     }
+    a->serialEventNeedsRecord = true;
   #else
-    CUDACHECK(hipEventRecord(b->event, b->stream));
-    CUDACHECK(hipStreamWaitEvent(a->stream, b->event, 0));
+    CUDACHECK(hipEventRecord(b->scratchEvent, b->cudaStream));
+    CUDACHECK(hipStreamWaitEvent(a->cudaStream, b->scratchEvent, 0));
   #endif
   return ncclSuccess;
 }
@@ -210,36 +334,29 @@ ncclResult_t ncclStrongStreamWaitStream(
   ) {
   #if CUDART_VERSION >= 11030
     if (graph.graph == nullptr) {
-      CUDACHECK(cudaEventRecord(a->event, b));
-      CUDACHECK(cudaStreamWaitEvent(a->stream, a->event, 0));
-      // We used a->event to record b so it no longer reflects anything about a.
-      a->eventIsLagging = 1;
+      // It is ok to use a->serialEvent to record b since we'll be setting
+      // a->serialEventNeedsRecord so the event won't be considered accurate
+      // until re-recorded.
+      CUDACHECK(cudaEventRecord(a->serialEvent, b));
+      CUDACHECK(cudaStreamWaitEvent(a->cudaStream, a->serialEvent, 0));
     } else {
       cudaStreamCaptureStatus status;
-      unsigned long long gid1;
-      cudaGraphNode_t const* deps;
-      size_t depN = 0;
-      CUDACHECK(cudaStreamGetCaptureInfo_v2(b, &status, &gid1, nullptr, &deps, &depN));
-      if (status != cudaStreamCaptureStatusActive || graph.graphId != gid1) {
+      unsigned long long bGraphId;
+      cudaGraphNode_t const* bNodes;
+      size_t bCount = 0;
+      CUDACHECK(cudaStreamGetCaptureInfo_v2(b, &status, &bGraphId, nullptr, &bNodes, &bCount));
+      if (status != cudaStreamCaptureStatusActive || graph.graphId != bGraphId) {
         WARN("Stream is not being captured by the expected graph.");
         return ncclInvalidUsage;
       }
-      if (depN > 0 && (depN > 1 || deps[0] != a->node)) {
-        cudaGraphNode_t tie;
-        if (depN == 1) {
-          tie = deps[0];
-        } else {
-          CUDACHECK(cudaGraphAddEmptyNode(&tie, graph.graph, deps, depN));
-        }
-        cudaGraphNode_t pair[2] = {a->node, tie};
-        CUDACHECK(cudaGraphAddEmptyNode(&a->node, graph.graph, pair, 2));
-      }
-      // a->eventIsLagging doesn't change since we are just updating the
-      // dependencies of a->node.
+      struct ncclStrongStreamGraph* ag = a->graphHead;
+      NCCLCHECK(checkGraphId(ag, graph.graphId));
+      mergeTips(ag, bNodes, bCount);
     }
+    a->serialEventNeedsRecord = true;
   #else
-    CUDACHECK(hipEventRecord(a->event, b));
-    CUDACHECK(hipStreamWaitEvent(a->stream, a->event, 0));
+    CUDACHECK(hipEventRecord(a->scratchEvent, b));
+    CUDACHECK(hipStreamWaitEvent(a->cudaStream, a->scratchEvent, 0));
   #endif
   return ncclSuccess;
 }
@@ -249,25 +366,28 @@ ncclResult_t ncclStrongStreamWaitStream(
   ) {
   #if CUDART_VERSION >= 11030
     if (graph.graph == nullptr) {
-      if (b->eventIsLagging) {
-        b->eventIsLagging = 0;
-        CUDACHECK(cudaEventRecord(b->event, b->stream));
+      if (b->serialEventNeedsRecord) {
+        b->serialEventNeedsRecord = false;
+        CUDACHECK(cudaEventRecord(b->serialEvent, b->cudaStream));
       }
-      CUDACHECK(cudaStreamWaitEvent(a, b->event, 0));
+      CUDACHECK(cudaStreamWaitEvent(a, b->serialEvent, 0));
     } else {
-      CUDACHECK(cudaStreamUpdateCaptureDependencies(a, &b->node, 1, cudaStreamAddCaptureDependencies));
+      struct ncclStrongStreamGraph* bg = b->graphHead;
+      NCCLCHECK(checkGraphId(bg, graph.graphId));
+      CUDACHECK(cudaStreamUpdateCaptureDependencies(a, bg->tipNodes, bg->tipCount, cudaStreamAddCaptureDependencies));
     }
   #else
-    CUDACHECK(hipEventRecord(b->event, b->stream));
-    CUDACHECK(hipStreamWaitEvent(a, b->event, 0));
+    CUDACHECK(hipEventRecord(b->scratchEvent, b->cudaStream));
+    CUDACHECK(hipStreamWaitEvent(a, b->scratchEvent, 0));
   #endif
   return ncclSuccess;
 }
 
 ncclResult_t ncclStrongStreamSynchronize(struct ncclStrongStream* ss) {
   #if CUDART_VERSION >= 11030
-    CUDACHECK(cudaStreamWaitEvent(ss->stream, ss->event, 0));
+    CUDACHECK(cudaStreamWaitEvent(ss->cudaStream, ss->serialEvent, 0));
+    ss->serialEventNeedsRecord = false;
   #endif
-  CUDACHECK(hipStreamSynchronize(ss->stream));
+  CUDACHECK(hipStreamSynchronize(ss->cudaStream));
   return ncclSuccess;
 }
diff --git a/projects/rccl/src/proxy.cc b/projects/rccl/src/proxy.cc
index 6638bb8607..26ca0453e7 100644
--- a/projects/rccl/src/proxy.cc
+++ b/projects/rccl/src/proxy.cc
@@ -1056,8 +1056,8 @@ void* ncclProxyService(void* _args) {
   int asyncOpCount = 0;
   while ((stop == 0 || (stop == 1 && npeers > 0)) && *comm->abortFlag == 0) {
     /* never let proxy service thread blocks in poll, or it cannot receive abortFlag. */
-    if (int error = poll(pollfds, NCCL_MAX_LOCAL_RANKS+1, asyncOpCount ? 0 : 500) < 0) {
-      WARN("[Proxy Service] Poll failed with error %d", error);
+    if (poll(pollfds, NCCL_MAX_LOCAL_RANKS+1, asyncOpCount ? 0 : 500) < 0) {
+      WARN("[Proxy Service] Poll failed: %s\n", strerror(errno));
       return NULL;
     }
     if (pollfds[NCCL_MAX_LOCAL_RANKS].revents) {
diff --git a/projects/rccl/src/transport.cc b/projects/rccl/src/transport.cc
index 82c0c86a26..943c9da75f 100644
--- a/projects/rccl/src/transport.cc
+++ b/projects/rccl/src/transport.cc
@@ -54,7 +54,7 @@ static ncclResult_t selectTransport(struct ncclComm* comm, struct ncclTopoGraph*
 ncclResult_t ncclTransportP2pConnect(struct ncclComm* comm, int channelId, int nrecv, int* peerRecv, int nsend, int* peerSend, int connIndex) {
   TRACE(NCCL_INIT, "nsend %d nrecv %d", nsend, nrecv);
   struct ncclChannel* channel = &comm->channels[channelId];
-  uint32_t mask = 1 << channelId;
+  uint64_t mask = 1UL << channel->id;
   for (int i=0; i<nrecv; i++) {
     int peer = peerRecv[i];
     if (peer == -1 || peer >= comm->nRanks || peer == comm->rank || channel->peers[peer].recv[connIndex].connected) continue;
@@ -85,15 +85,15 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
     int bootstrapTag = (i<<8) + (graph ? graph->id+1 : 0);
     int recvPeer = (comm->rank - i + comm->nRanks) % comm->nRanks;
     int sendPeer = (comm->rank + i) % comm->nRanks;
-    uint32_t recvMask = comm->connectRecv[recvPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)];
-    uint32_t sendMask = comm->connectSend[sendPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)];
+    uint64_t recvMask = comm->connectRecv[recvPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)];
+    uint64_t sendMask = comm->connectSend[sendPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)];
 
     struct ncclConnect* recvData = data;
     int sendChannels = 0, recvChannels = 0;
     int type;
     TIME_START(0);
     for (int c=0; c<MAXCHANNELS; c++) {
-      if (recvMask & (1<<c)) {
+      if (recvMask & (1UL<<c)) {
         NCCLCHECK(selectTransport<0>(comm, graph, recvData+recvChannels++, c, recvPeer, connIndex, &type));
         if (type > highestType) highestType = type;
       }
@@ -102,7 +102,7 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
     TIME_START(1);
     struct ncclConnect* sendData = recvData+recvChannels;
     for (int c=0; c<MAXCHANNELS; c++) {
-      if (sendMask & (1<<c)) {
+      if (sendMask & (1UL<<c)) {
         NCCLCHECK(selectTransport<1>(comm, graph, sendData+sendChannels++, c, sendPeer, connIndex, &type));
         if (type > highestType) highestType = type;
       }
@@ -127,7 +127,7 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
 
     TIME_START(3);
     for (int c=0; c<MAXCHANNELS; c++) {
-      if (sendMask & (1<<c)) {
+      if (sendMask & (1UL<<c)) {
         struct ncclConnector* conn = comm->channels[c].peers[sendPeer].send + connIndex;
         NCCLCHECK(conn->transportComm->connect(comm, sendData++, 1, comm->rank, conn));
         conn->connected = 1;
@@ -139,7 +139,7 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
     TIME_STOP(3);
     TIME_START(4);
     for (int c=0; c<MAXCHANNELS; c++) {
-      if (recvMask & (1<<c)) {
+      if (recvMask & (1UL<<c)) {
         struct ncclConnector* conn = comm->channels[c].peers[recvPeer].recv + connIndex;
         NCCLCHECK(conn->transportComm->connect(comm, recvData++, 1, comm->rank, conn));
         conn->connected = 1;
@@ -148,7 +148,7 @@ ncclResult_t ncclTransportP2pSetup(struct ncclComm* comm, struct ncclTopoGraph*
       }
     }
     TIME_STOP(4);
-    comm->connectRecv[recvPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)] = comm->connectSend[sendPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)] = 0;
+    comm->connectRecv[recvPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)] = comm->connectSend[sendPeer+comm->nRanks*(connIndex == NCCL_CONN_IDX_P2P_NET ? NCCL_CONN_IDX_P2P_NET : 0)] = 0UL;
   }
   CUDACHECK(hipStreamSynchronize(comm->sideStream));
   if (highestTransportType != NULL) *highestTransportType = highestType;
diff --git a/projects/rccl/src/transport/coll_net.cc b/projects/rccl/src/transport/coll_net.cc
index da158d9d21..28aff7a53f 100644
--- a/projects/rccl/src/transport/coll_net.cc
+++ b/projects/rccl/src/transport/coll_net.cc
@@ -123,6 +123,7 @@ struct recvResources {
   int netDev;
   int useGdr;
   int useDmaBuf;
+  int needFlush;
   uint64_t* gdcSync;
   uint64_t* gdcFlush;
   void* gdrDesc;
@@ -142,6 +143,7 @@ static ncclResult_t canConnect(int* ret, struct ncclTopoSystem* topo, struct ncc
 struct setupReq {
   int netDev;
   int useGdr;
+  int needFlush;
 };
 
 
@@ -154,6 +156,8 @@ static ncclResult_t sendSetup(struct ncclComm* comm, struct ncclTopoGraph* graph
   NCCLCHECK(ncclTopoGetNetDev(comm, myInfo->rank, graph, channelId, -1, &req.netDev, &proxyRank));
   NCCLCHECK(ncclTopoCheckGdr(comm->topo, myInfo->busId, req.netDev, 1, &req.useGdr));
   send->conn.direct |= req.useGdr ? NCCL_DIRECT_NIC : 0;
+  // Determine whether we need to flush the GDR buffer on recv or not
+  if (req.useGdr) NCCLCHECK(ncclTopoNeedFlush(comm->topo, myInfo->busId, &req.needFlush));
 
   NCCLCHECK(ncclTopoGetLocalRank(comm->topo, myInfo->rank, &send->proxyConn.localRank));
   NCCLCHECK(ncclProxyConnect(comm, TRANSPORT_COLLNET, 1, myInfo->rank, &send->proxyConn));
@@ -395,6 +399,7 @@ static ncclResult_t recvProxySetup(struct ncclProxyConnection* connection, struc
 
   resources->netDev = req->netDev;
   resources->useGdr = req->useGdr;
+  resources->needFlush = req->needFlush;
   ncclNetProperties_t props;
   NCCLCHECK(collNetGetProperties(comm, req->netDev, &props));
   /* DMA-BUF support */
@@ -763,7 +768,7 @@ static ncclResult_t recvProxyProgress(struct ncclComm* comm, struct ncclProxyArg
           TRACE(NCCL_NET, "recvProxy [%lu/%d/%d] received, size %d", sub->received, group, buffSlot, totalSize);
           sub->received += args->sliceSteps;
           sub->requests[buffSlot] = NULL;
-          if (1 && reqFifo[group][buffSlot].size > 0 && resources->useGdr) {
+          if (reqFifo[group][buffSlot].size > 0 && resources->useGdr && resources->needFlush) {
             // GDRCOPY support
             if (resources->gdcFlush) {
 #if defined (__x86_64__)
diff --git a/projects/rccl/src/transport/net.cc b/projects/rccl/src/transport/net.cc
index 8253449541..b8f11b3edc 100644
--- a/projects/rccl/src/transport/net.cc
+++ b/projects/rccl/src/transport/net.cc
@@ -125,6 +125,7 @@ struct recvResources {
   int netDev;
   int useGdr;
   int useDmaBuf;
+  int needFlush;
   int maxRecvs;
   uint64_t* gdcSync;
   uint64_t* gdcFlush;
@@ -163,6 +164,7 @@ struct setupReq {
   int shared;
   int netDev;
   int useGdr;
+  int needFlush;
   int channelId;
   int connIndex;
   uint32_t* curr_hdp_reg;
@@ -226,6 +228,9 @@ static ncclResult_t recvSetup(struct ncclComm* comm, struct ncclTopoGraph* graph
   if (req.netDev < 0) NCCLCHECK(ncclTopoGetNetDev(comm, myInfo->rank, graph, channelId, myInfo->rank, &req.netDev, &proxyRank));
   NCCLCHECK(ncclTopoCheckGdr(comm->topo, myInfo->busId, req.netDev, 0, &req.useGdr));
 
+  // Determine whether we need to flush the GDR buffer on recv or not
+  if (req.useGdr) NCCLCHECK(ncclTopoNeedFlush(comm->topo, myInfo->busId, &req.needFlush));
+
   // We don't support PXN on receive yet
   NCCLCHECK(ncclProxyConnect(comm, TRANSPORT_NET, 0, myInfo->rank, &recv->proxyConn));
 
@@ -492,6 +497,7 @@ static ncclResult_t recvProxySetup(struct ncclProxyConnection* connection, struc
   resources->netDev = req->netDev;
   resources->shared = connection->shared = req->shared;
   resources->useGdr = req->useGdr;
+  resources->needFlush = req->needFlush;
   resources->channelId = req->channelId;
   resources->connIndex = req->connIndex;
   ncclNetProperties_t props;
@@ -1150,7 +1156,7 @@ static ncclResult_t recvProxyProgress(struct ncclComm* comm, struct ncclProxyArg
         for (int i=0; i<NCCL_PROXY_MAX_SUBS; i++) sizes[i] = 0;
         NCCLCHECK(ncclNetTest(comm, subGroup->requests[step%NCCL_STEPS], &done, sizes));
         if (done) {
-          int useGdr = 0;
+          int needFlush = 0;
           int totalSize = 0;
           for (int i=0; i<NCCL_PROXY_MAX_SUBS; i++) totalSize += sizes[i];
           for (int i=0; i<subGroup->groupSize; i++) {
@@ -1175,11 +1181,11 @@ static ncclResult_t recvProxyProgress(struct ncclComm* comm, struct ncclProxyArg
             for (uint64_t step=sub->received-args->sliceSteps; step<sub->received; step++) ncclProfilingRecord(args, s+i, step, ncclProxyProfileRecvFlushWait);
             if (step < sub->nsteps) {
               struct recvResources* resources = (struct recvResources*) (sub->connection->transportResources);
-              if (resources->useGdr) useGdr = 1;
+              if (resources->useGdr) needFlush |= resources->needFlush;
             }
           }
           subGroup->requests[step%NCCL_STEPS] = NULL;
-          if (totalSize > 0 && p == NCCL_PROTO_SIMPLE && useGdr) {
+          if (totalSize > 0 && p == NCCL_PROTO_SIMPLE && needFlush) {
             // GDRCOPY support
             struct recvResources* resources = (struct recvResources*) (subGroup->connection->transportResources);
             if (resources->gdcFlush) {