diff --git a/CHANGES b/CHANGES
index c7480879f2c..128e6bd83b0 100644
--- a/CHANGES
+++ b/CHANGES
@@ -1,3 +1,17 @@
+===============================================================================
+                               Changes in 3.4.1
+===============================================================================
+
+# Fix bug in GPU memory hooks that caused failures with NCCL
+
+# Fix bug in GPU-aware nonblocking op collectives
+
+# Hide symbols from embedded json-c module
+
+# Removed anonymous struct/union usage in MPL
+
+# Disable Java support in embedded UCX module
+
 ===============================================================================
                                Changes in 3.4
 ===============================================================================
diff --git a/configure.ac b/configure.ac
index e2ffe9d522f..dd7b4d7c642 100644
--- a/configure.ac
+++ b/configure.ac
@@ -1191,7 +1191,7 @@ AC_SUBST([jsonlib])
 
 PAC_PUSH_ALL_FLAGS()
 PAC_RESET_ALL_FLAGS()
-PAC_CONFIG_SUBDIR_ARGS([modules/json-c],[--disable-werror],[],[AC_MSG_ERROR(json-c configure failed)])
+PAC_CONFIG_SUBDIR_ARGS([modules/json-c],[--enable-embedded --disable-werror],[],[AC_MSG_ERROR(json-c configure failed)])
 PAC_POP_ALL_FLAGS()
 jsonsrcdir="${main_top_builddir}/modules/json-c"
 jsonlib="${main_top_builddir}/modules/json-c/libjson-c.la"
diff --git a/maint/version.m4 b/maint/version.m4
index aa771df22ac..e3bc8d9e7bf 100644
--- a/maint/version.m4
+++ b/maint/version.m4
@@ -14,7 +14,7 @@
 # changing this by playing with diversions, but then we would probably be
 # playing with autotools-fire.
 
-m4_define([MPICH_VERSION_m4],[3.4])dnl
+m4_define([MPICH_VERSION_m4],[3.4.1])dnl
 m4_define([MPICH_RELEASE_DATE_m4],[unreleased development copy])dnl
 
 # For libtool ABI versioning rules see:
@@ -35,6 +35,6 @@ m4_define([MPICH_RELEASE_DATE_m4],[unreleased development copy])dnl
 # libmpi so version only includes functionality defined in the MPI
 # standard, and does not include MPIX_ functions and C++ bindings.
 
-m4_define([libmpi_so_version_m4],[13:9:1])dnl
+m4_define([libmpi_so_version_m4],[13:10:1])dnl
 
 [#] end of __file__
diff --git a/modules/json-c b/modules/json-c
index 366f1c6c0ea..c496a735ec7 160000
--- a/modules/json-c
+++ b/modules/json-c
@@ -1 +1 @@
-Subproject commit 366f1c6c0ea2ca2f1077c1296f5cb744336fac38
+Subproject commit c496a735ec7312a22c1484ac6950f12fc8712b2f
diff --git a/src/include/mpir_gpu.h b/src/include/mpir_gpu.h
index de790a8632f..a3c06d37788 100644
--- a/src/include/mpir_gpu.h
+++ b/src/include/mpir_gpu.h
@@ -33,6 +33,14 @@
 
 extern int MPIR_CVAR_ENABLE_GPU;
 
+#undef ENABLE_GPU
+
+#ifdef MPL_HAVE_GPU
+#define ENABLE_GPU MPIR_CVAR_ENABLE_GPU
+#else
+#define ENABLE_GPU FALSE
+#endif
+
 MPL_STATIC_INLINE_PREFIX int MPIR_GPU_query_pointer_attr(const void *ptr, MPL_pointer_attr_t * attr)
 {
     int mpi_errno = MPI_SUCCESS, mpl_err = MPL_SUCCESS;
@@ -40,7 +48,7 @@ MPL_STATIC_INLINE_PREFIX int MPIR_GPU_query_pointer_attr(const void *ptr, MPL_po
     /* Skip query if GPU support is disabled by CVAR. Because we assume
      * no GPU buffer is used. If the user disables GPU at configure time
      * (e.g., --without-cuda), then MPL fallback will handle the query. */
-    if (MPIR_CVAR_ENABLE_GPU) {
+    if (ENABLE_GPU) {
         mpl_err = MPL_gpu_query_pointer_attr(ptr, attr);
         MPIR_ERR_CHKANDJUMP(mpl_err != MPL_SUCCESS, mpi_errno, MPI_ERR_OTHER, "**gpu_query_ptr");
     } else {
@@ -56,7 +64,7 @@ MPL_STATIC_INLINE_PREFIX int MPIR_GPU_query_pointer_attr(const void *ptr, MPL_po
 
 MPL_STATIC_INLINE_PREFIX bool MPIR_GPU_query_pointer_is_dev(const void *ptr)
 {
-    if (MPIR_CVAR_ENABLE_GPU && ptr != NULL) {
+    if (ENABLE_GPU && ptr != NULL) {
         MPL_pointer_attr_t attr;
         MPL_gpu_query_pointer_attr(ptr, &attr);
 
diff --git a/src/mpi/coll/src/coll_impl.c b/src/mpi/coll/src/coll_impl.c
index 0af50c61749..ffa6cc1d93d 100644
--- a/src/mpi/coll/src/coll_impl.c
+++ b/src/mpi/coll/src/coll_impl.c
@@ -218,7 +218,7 @@ void MPIR_Coll_host_buffer_alloc(const void *sendbuf, const void *recvbuf, MPI_A
     MPI_Aint extent = 0;
 
     if (sendbuf != MPI_IN_PLACE) {
-        MPL_gpu_query_pointer_attr(sendbuf, &attr);
+        MPIR_GPU_query_pointer_attr(sendbuf, &attr);
         if (attr.type == MPL_GPU_POINTER_DEV) {
             MPI_Aint true_extent;
             MPI_Aint true_lb;
@@ -233,7 +233,7 @@ void MPIR_Coll_host_buffer_alloc(const void *sendbuf, const void *recvbuf, MPI_A
         }
     }
 
-    MPL_gpu_query_pointer_attr(recvbuf, &attr);
+    MPIR_GPU_query_pointer_attr(recvbuf, &attr);
     if (attr.type == MPL_GPU_POINTER_DEV) {
         if (!extent) {
             MPI_Aint true_extent;
diff --git a/src/mpi/datatype/typerep/src/typerep_yaksa_pack.c b/src/mpi/datatype/typerep/src/typerep_yaksa_pack.c
index 068ade08f70..b7c241485bb 100644
--- a/src/mpi/datatype/typerep/src/typerep_yaksa_pack.c
+++ b/src/mpi/datatype/typerep/src/typerep_yaksa_pack.c
@@ -20,8 +20,8 @@ int MPIR_Typerep_copy(void *outbuf, const void *inbuf, MPI_Aint num_bytes)
     }
 
     MPL_pointer_attr_t inattr, outattr;
-    MPL_gpu_query_pointer_attr(inbuf, &inattr);
-    MPL_gpu_query_pointer_attr(outbuf, &outattr);
+    MPIR_GPU_query_pointer_attr(inbuf, &inattr);
+    MPIR_GPU_query_pointer_attr(outbuf, &outattr);
 
     if ((inattr.type == MPL_GPU_POINTER_UNREGISTERED_HOST ||
          inattr.type == MPL_GPU_POINTER_REGISTERED_HOST) &&
@@ -63,8 +63,8 @@ static inline bool fastpath_memcpy(const void *inbuf, void *outbuf, MPI_Datatype
      * true_lb) into the MPIR_Datatype structure */
     if (HANDLE_IS_BUILTIN(type)) {
         MPL_pointer_attr_t inattr, outattr;
-        MPL_gpu_query_pointer_attr(inbuf, &inattr);
-        MPL_gpu_query_pointer_attr(outbuf, &outattr);
+        MPIR_GPU_query_pointer_attr(inbuf, &inattr);
+        MPIR_GPU_query_pointer_attr(outbuf, &outattr);
 
         if ((inattr.type == MPL_GPU_POINTER_UNREGISTERED_HOST ||
              inattr.type == MPL_GPU_POINTER_REGISTERED_HOST) &&
@@ -86,11 +86,11 @@ static inline bool fastpath_memcpy(const void *inbuf, void *outbuf, MPI_Datatype
             MPL_pointer_attr_t inattr, outattr;
 
             if (dir == MEMCPY_DIR__PACK) {
-                MPL_gpu_query_pointer_attr((const char *) inbuf + dtp->true_lb + offset, &inattr);
-                MPL_gpu_query_pointer_attr(outbuf, &outattr);
+                MPIR_GPU_query_pointer_attr((const char *) inbuf + dtp->true_lb + offset, &inattr);
+                MPIR_GPU_query_pointer_attr(outbuf, &outattr);
             } else {
-                MPL_gpu_query_pointer_attr(inbuf, &inattr);
-                MPL_gpu_query_pointer_attr((char *) outbuf + dtp->true_lb + offset, &outattr);
+                MPIR_GPU_query_pointer_attr(inbuf, &inattr);
+                MPIR_GPU_query_pointer_attr((char *) outbuf + dtp->true_lb + offset, &outattr);
             }
 
             if ((inattr.type == MPL_GPU_POINTER_UNREGISTERED_HOST ||
diff --git a/src/mpi/misc/utils.c b/src/mpi/misc/utils.c
index 20059b35a0e..d64f1f8fedd 100644
--- a/src/mpi/misc/utils.c
+++ b/src/mpi/misc/utils.c
@@ -66,8 +66,8 @@ int MPIR_Localcopy(const void *sendbuf, MPI_Aint sendcount, MPI_Datatype sendtyp
         intptr_t sfirst;
         intptr_t rfirst;
 
-        MPL_gpu_query_pointer_attr(sendbuf, &send_attr);
-        MPL_gpu_query_pointer_attr(recvbuf, &recv_attr);
+        MPIR_GPU_query_pointer_attr(sendbuf, &send_attr);
+        MPIR_GPU_query_pointer_attr(recvbuf, &recv_attr);
 
         if (send_attr.type == MPL_GPU_POINTER_DEV && recv_attr.type == MPL_GPU_POINTER_DEV) {
             MPL_gpu_malloc((void **) &buf, COPY_BUFFER_SZ, recv_attr.device);
diff --git a/src/mpid/ch4/src/ch4_impl.h b/src/mpid/ch4/src/ch4_impl.h
index 83aee8a4c97..df7060e2b74 100644
--- a/src/mpid/ch4/src/ch4_impl.h
+++ b/src/mpid/ch4/src/ch4_impl.h
@@ -1025,7 +1025,7 @@ MPL_STATIC_INLINE_PREFIX int MPIDIG_compute_acc_op(void *source_buf, int source_
     void *in_targetbuf = target_buf;
     void *host_targetbuf = NULL;
     MPL_pointer_attr_t attr;
-    MPL_gpu_query_pointer_attr(target_buf, &attr);
+    MPIR_GPU_query_pointer_attr(target_buf, &attr);
     /* FIXME: use typerep/yaksa GPU-aware accumulate when available */
     if (attr.type == MPL_GPU_POINTER_DEV) {
         MPI_Aint extent, true_extent;
diff --git a/src/mpl/src/gpu/mpl_gpu_cuda.c b/src/mpl/src/gpu/mpl_gpu_cuda.c
index 97328fa07b5..62b84ff0a59 100644
--- a/src/mpl/src/gpu/mpl_gpu_cuda.c
+++ b/src/mpl/src/gpu/mpl_gpu_cuda.c
@@ -284,9 +284,12 @@ int MPL_gpu_get_buffer_bounds(const void *ptr, void **pbase, uintptr_t * len)
 static void gpu_free_hooks_cb(void *dptr)
 {
     gpu_free_hook_s *current = free_hook_chain;
-    while (current) {
-        current->free_hook(dptr);
-        current = current->next;
+    if (dptr != NULL) {
+        /* we call gpu hook only when dptr != NULL */
+        while (current) {
+            current->free_hook(dptr);
+            current = current->next;
+        }
     }
     return;
 }