diff --git a/makefiles/common.mk b/makefiles/common.mk
index 37e81be204..2e448268a6 100644
--- a/makefiles/common.mk
+++ b/makefiles/common.mk
@@ -42,9 +42,10 @@ else
 endif
 #$(info NVCC_GENCODE is ${NVCC_GENCODE})
 
-CXXFLAGS   := -DCUDA_MAJOR=$(CUDA_MAJOR) -DCUDA_MINOR=$(CUDA_MINOR) -fPIC -fvisibility=hidden
-CXXFLAGS   += -Wall -Wno-unused-function -Wno-sign-compare -std=c++11 -Wvla
-CXXFLAGS   += -I $(CUDA_INC)
+CXXFLAGS   := -DCUDA_MAJOR=$(CUDA_MAJOR) -DCUDA_MINOR=$(CUDA_MINOR) -fPIC -fvisibility=hidden \
+              -Wall -Wno-unused-function -Wno-sign-compare -std=c++11 -Wvla \
+              -I $(CUDA_INC) \
+              $(CXXFLAGS)
 # Maxrregcount needs to be set accordingly to NCCL_MAX_NTHREADS (otherwise it will cause kernel launch errors)
 # 512 : 120, 640 : 96, 768 : 80, 1024 : 60
 # We would not have to set this if we used __launch_bounds__, but this only works on kernels, not on functions.
diff --git a/makefiles/version.mk b/makefiles/version.mk
index efbdee7258..05abbc75c5 100644
--- a/makefiles/version.mk
+++ b/makefiles/version.mk
@@ -1,6 +1,6 @@
 ##### version
 NCCL_MAJOR   := 2
 NCCL_MINOR   := 5
-NCCL_PATCH   := 6
+NCCL_PATCH   := 7
 NCCL_SUFFIX  :=
 PKG_REVISION := 1
diff --git a/pkg/debian/Makefile b/pkg/debian/Makefile
index 2ce4390365..0494f3e032 100644
--- a/pkg/debian/Makefile
+++ b/pkg/debian/Makefile
@@ -16,13 +16,8 @@ DEBFILES   := compat copyright libnccl-dev.install rules $(DEBGEN)
 DEBTARGETS := $(patsubst %, $(DEBPREPDIR)/%, $(DEBFILES))
 
 PKG_TIMESTAMP  := $(shell date -R)
-ARCH           := $(shell uname -m)
-PKG_ARCH       ?= $(shell uname -m | sed -e "s/x86_64/amd64/g" | sed -e "s/ppc64le/ppc64el/g"| sed -e "s/aarch64/arm64/g")
-PKG_MULTIARCH  ?= $(shell $(CXX) -print-multiarch)
-ifeq ($(PKG_MULTIARCH),)
-# Hardwire the PKG_MULTIARCH directory as the RHEL6 distribution agnostic compiler (gcc 4.8.3) doesn't set it
-PKG_MULTIARCH  := $(ARCH)-linux-gnu
-endif
+PKG_ARCH       ?= $(shell dpkg-architecture -qDEB_HOST_ARCH)
+PKG_MULTIARCH  ?= $(shell dpkg-architecture -qDEB_HOST_MULTIARCH)
 
 prep : $(DEBTARGETS)
 	$(MAKE) -C ../.. lic BUILDDIR=$(BUILDDIR)
diff --git a/src/collectives/device/primitives.h b/src/collectives/device/primitives.h
index f2a175229f..8b2ac5a3ba 100644
--- a/src/collectives/device/primitives.h
+++ b/src/collectives/device/primitives.h
@@ -182,7 +182,7 @@ class ncclPrimitives {
   }
 
   template <int DIRECTRECV>
-  inline __device__ const T* directRecvPtr(int i, int directOffset) {
+  inline __device__ const T* directRecvPtr(int i, ssize_t directOffset) {
 #if defined(RCCL_USE_DIRECT_BUFFER)
     return DIRECTRECV && r.recvDirectBuff[i] ? r.recvDirectBuff[i]+directOffset : recvPtr(i);
 #else
@@ -191,7 +191,7 @@ class ncclPrimitives {
   }
 
   template <int DIRECTSEND>
-  inline __device__ T* directSendPtr(int i, int directOffset) {
+  inline __device__ T* directSendPtr(int i, ssize_t directOffset) {
 #if defined(RCCL_USE_DIRECT_BUFFER)
     return DIRECTSEND && s.sendDirectBuff[i] ? s.sendDirectBuff[i]+directOffset : sendPtr(i);
 #else
@@ -219,7 +219,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
 
   template <int DIRECTRECV, int DIRECTSEND, int RECV, int SEND, int SRC, int DST>
   inline __device__ void
-  GenericOp(const T* srcPtr, T* dstPtr, int nelem, int directOffset) {
+  GenericOp(const T* srcPtr, T* dstPtr, int nelem, ssize_t directOffset) {
     int offset = 0;
     int sliceSize = stepSize*SLICESTEPS;
     int dataSize = max(DIVUP(nelem, 16*SLICESPERCHUNK)*16, sliceSize/32);
@@ -386,7 +386,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
     GenericOp<0, 0, 0, 1, 1, 0>(src, NULL, nelem, 0);
   }
   __device__ __forceinline__ void
-  directSend(const T* src, int directOffset, int nelem) {
+  directSend(const T* src, ssize_t directOffset, int nelem) {
     GenericOp<0, 1, 0, 1, 1, 0>(src, NULL, nelem, directOffset);
   }
 
@@ -395,7 +395,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
     GenericOp<0, 0, 1, 0, 0, 1>(NULL, dst, nelem, 0);
   }
   __device__ __forceinline__ void
-  directRecv(T* dst, int directOffset, int nelem) {
+  directRecv(T* dst, ssize_t directOffset, int nelem) {
     GenericOp<1, 0, 1, 0, 0, 1>(NULL, dst, nelem, directOffset);
   }
 
@@ -404,7 +404,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
     GenericOp<0, 0, 0, 1, 1, 1>(src, dst, nelem, 0);
   }
   __device__ __forceinline__ void
-  directCopySend(const T* src, T* dst, int directOffset, int nelem) {
+  directCopySend(const T* src, T* dst, ssize_t directOffset, int nelem) {
     GenericOp<0, 1, 0, 1, 1, 1>(src, dst, nelem, directOffset);
   }
 
@@ -413,7 +413,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
     GenericOp<0, 0, 1, 1, 0, 1>(NULL, dst, nelem, 0);
   }
   __device__ __forceinline__ void
-  directRecvCopySend(T* dst, int directOffset, int nelem) {
+  directRecvCopySend(T* dst, ssize_t directOffset, int nelem) {
     GenericOp<1, 1, 1, 1, 0, 1>(NULL, dst, nelem, directOffset);
   }
 
@@ -432,7 +432,7 @@ inline __device__ int directSendInc(int i, int directInc, int sliceInc) {
     GenericOp<0, 0, 1, 1, 1, 1>(src, dst, nelem, 0);
   }
   __device__ __forceinline__ void
-  directRecvReduceCopySend(const T* src, T* dst, int directOffset, int nelem) {
+  directRecvReduceCopySend(const T* src, T* dst, ssize_t directOffset, int nelem) {
     // Direct is only for the send part
     GenericOp<0, 1, 1, 1, 1, 1>(src, dst, nelem, directOffset);
   }
diff --git a/src/graph/paths.cc b/src/graph/paths.cc
index ce1772c39a..eba1964f44 100644
--- a/src/graph/paths.cc
+++ b/src/graph/paths.cc
@@ -179,11 +179,18 @@ static ncclResult_t addCpuStep(struct ncclTopoSystem* system, int c, int t1, int
 // Remove/free paths for a given type
 static void ncclTopoRemovePathType(struct ncclTopoSystem* system, int nodeType) {
   for (int t=0; t<NCCL_TOPO_NODE_TYPES; t++) {
+    // Remove links _to_ the given type
     for (int n=0; n<system->nodes[t].count; n++) {
       struct ncclTopoNode* node = system->nodes[t].nodes+n;
       free(node->paths[nodeType]);
       node->paths[nodeType] = NULL;
     }
+    // Remove links _from_ the given type
+    for (int n=0; n<system->nodes[nodeType].count; n++) {
+      struct ncclTopoNode* node = system->nodes[nodeType].nodes+n;
+      free(node->paths[t]);
+      node->paths[t] = NULL;
+    }
   }
 }
 
@@ -309,6 +316,22 @@ ncclResult_t ncclTopoTrimSystem(struct ncclTopoSystem* system, struct ncclComm*
     // Trim network
     ncclTopoRemovePathType(system, NET);
     system->nodes[NET].count = 0;
+    for (int t=0; t<NCCL_TOPO_NODE_TYPES; t++) {
+      for (int n=0; n<system->nodes[t].count; n++) {
+        struct ncclTopoNode* node = system->nodes[t].nodes+n;
+        for (int l=0; l<node->nlinks; l++) {
+          struct ncclTopoLink* link = &(node->links[l]);
+          if (link->remNode->type == NET) {
+            // Remove the link
+            for (int i=l; i<(node->nlinks-1); i++) {
+              memcpy(&(node->links[i]), &(node->links[i+1]), sizeof(ncclTopoLink));
+            }
+            node->nlinks--;
+            l--;  // revisit the same value of "l" for the next iteration, since we edited the list in the middle of the loop
+          }
+        }
+      }
+    }
   }
   free(domains);
   free(ids);
diff --git a/src/graph/tuning.cc b/src/graph/tuning.cc
index df2a901feb..957c63e973 100644
--- a/src/graph/tuning.cc
+++ b/src/graph/tuning.cc
@@ -193,7 +193,7 @@ ncclResult_t ncclSetThresholds(struct ncclComm* comm, int minCompCap, int maxCom
   // Override defaults with user env
   char* str = getenv("NCCL_THREAD_THRESHOLDS");
   if (str) {
-    ssize_t t[NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = { -2 };
+    ssize_t t[NCCL_NUM_ALGORITHMS][NCCL_NUM_PROTOCOLS] = {{ -2, -2, -2 }, { -2, -2, -2}};
     sscanf(str, "%ld %ld %ld %ld %ld %ld", t[0], t[0]+1, t[0]+2, t[1], t[1]+1, t[1]+2);
     for (int a=0; a<NCCL_NUM_ALGORITHMS; a++) {
       for (int p=0; p<NCCL_NUM_PROTOCOLS; p++) {
diff --git a/src/init.cc b/src/init.cc
index 998d0a13f1..f22b867c3f 100644
--- a/src/init.cc
+++ b/src/init.cc
@@ -127,7 +127,13 @@ ncclResult_t ncclGetUniqueId(ncclUniqueId* out) {
 }
 
 // Prevent compiler from optimizing out these operations
-void __attribute__((optimize("O0"))) commPoison(ncclComm_t comm) {
+#ifdef __clang__
+#define NCCL_NO_OPTIMIZE __attribute__((optnone))
+#else
+#define NCCL_NO_OPTIMIZE __attribute__((optimize("O0")))
+#endif
+
+void NCCL_NO_OPTIMIZE commPoison(ncclComm_t comm) {
   comm->rank = comm->cudaDev = comm->busId = comm->nRanks = -1;
 }
 
@@ -179,6 +185,8 @@ void *ncclCommThreadMain(void *arg) {
 }
 #endif
 
+#undef NCCL_NO_OPTIMIZE
+
 static ncclResult_t commFree(ncclComm_t comm) {
   if (comm == NULL)
     return ncclSuccess;