diff --git a/src/collectives/device/Makefile b/src/collectives/device/Makefile
index 04bce8e..a2498a0 100644
--- a/src/collectives/device/Makefile
+++ b/src/collectives/device/Makefile
@@ -29,7 +29,7 @@ all: $(STATICLIB)
 all_deps: $(DEPENDFILES)
 
 # Auto-generating the rules per op/reduction/datatype/algorithm
-$(RULESFILE) :
+$(RULESFILE) : gen_rules.sh
 	@printf "Generating %-35s > %s\n" rules $@
 	@mkdir -p $(OBJDIR)
 	@CUDA_MAJOR=${CUDA_MAJOR} CUDA_MINOR=${CUDA_MINOR} ./gen_rules.sh $(OBJDIR) > $@
diff --git a/src/collectives/device/gen_rules.sh b/src/collectives/device/gen_rules.sh
index aaf3685..a43f005 100755
--- a/src/collectives/device/gen_rules.sh
+++ b/src/collectives/device/gen_rules.sh
@@ -13,6 +13,9 @@ then
     datatypes+=" bf16"
 fi
 
+echo "CURDIR := \$(dir \$(realpath \$(word \$(words \$(math\$(MAKEFILE_LIST))-1), \$(MAKEFILE_LIST))))"
++echo ""
+
 targets="GENOBJS := \\\\\n"
 
 for base in sendrecv all_reduce all_gather broadcast reduce reduce_scatter; do
@@ -21,10 +24,17 @@ for base in sendrecv all_reduce all_gather broadcast reduce reduce_scatter; do
     dtn=0
     # Order must match that of the ncclDataType_t enum
     for dt in ${datatypes}; do
-      echo "${dir}/${base}_${op}_${dt}.o : ${base}.cu ${dir}/${base}.dep"
+      # Generate a unique filename for each compilation unit,
+      # otherwise the __nv_module_id may conflict at link time
+      echo "${dir}/${base}_${opn}_${dtn}.cu :"
+      echo "	echo \"#include \\\"\$(CURDIR)${base}.cu\\\"\" > \$@"
+      echo ""
+      # Compile the file
+      echo "${dir}/${base}_${op}_${dt}.o : ${dir}/${base}_${opn}_${dtn}.cu ${base}.cu ${dir}/${base}.dep"
+
       echo "	@printf \"Compiling  %-35s > %s\\\\n\" ${base}.cu ${dir}/${base}_${op}_${dt}.o"
       echo "	mkdir -p ${dir}"
-      echo "	\${NVCC} -DNCCL_OP=${opn} -DNCCL_TYPE=${dtn} \${NVCUFLAGS} -dc ${base}.cu -o ${dir}/${base}_${op}_${dt}.o"
+      echo "	\${NVCC} -DNCCL_OP=${opn} -DNCCL_TYPE=${dtn} \${NVCUFLAGS} -dc \$< -o \$@"
       echo ""
       targets="$targets\t${dir}/${base}_${op}_${dt}.o \\\\\n"
       dtn=$(($dtn + 1))