third_party/nccl/archive.patch

diff --git a/src/collectives.h b/src/collectives.h
new file mode 100644
index 0000000..7d04b16
--- /dev/null
+++ b/src/collectives.h
@@ -0,0 +1 @@
+#include "collectives/collectives.h"
diff --git a/src/collectives/device/all_gather.cu b/src/collectives/device/all_gather.cu.cc
similarity index 100%
rename from src/collectives/device/all_gather.cu
rename to src/collectives/device/all_gather.cu.cc
diff --git a/src/collectives/device/all_reduce.cu b/src/collectives/device/all_reduce.cu.cc
similarity index 100%
rename from src/collectives/device/all_reduce.cu
rename to src/collectives/device/all_reduce.cu.cc
diff --git a/src/collectives/device/broadcast.cu b/src/collectives/device/broadcast.cu.cc
similarity index 100%
rename from src/collectives/device/broadcast.cu
rename to src/collectives/device/broadcast.cu.cc
diff --git a/src/collectives/device/common.h b/src/collectives/device/common.h
index 8c336bf..2eef3ae 100644
--- a/src/collectives/device/common.h
+++ b/src/collectives/device/common.h
@@ -7,7 +7,7 @@
 #ifndef NCCL_DEVICE_COMMON_H_
 #define NCCL_DEVICE_COMMON_H_
 
-#include "../collectives.h"
+#include "collectives.h"
 #include "devcomm.h"
 #include "nccl.h"
 
diff --git a/src/collectives/device/functions.cu b/src/collectives/device/functions.cu.cc
similarity index 100%
rename from src/collectives/device/functions.cu
rename to src/collectives/device/functions.cu.cc
diff --git a/src/collectives/device/reduce.cu b/src/collectives/device/reduce.cu.cc
similarity index 100%
rename from src/collectives/device/reduce.cu
rename to src/collectives/device/reduce.cu.cc
diff --git a/src/collectives/device/reduce_scatter.cu b/src/collectives/device/reduce_scatter.cu.cc
similarity index 100%
rename from src/collectives/device/reduce_scatter.cu
rename to src/collectives/device/reduce_scatter.cu.cc
diff --git a/src/nccl.h.in b/src/nccl.h
similarity index 98%
rename from src/nccl.h.in
rename to src/nccl.h
index 985274e..7ebb1e1 100644
--- a/src/nccl.h.in
+++ b/src/nccl.h
@@ -10,12 +10,12 @@
 #include <cuda_runtime.h>
 #include <cuda_fp16.h>
 
-#define NCCL_MAJOR ${nccl:Major}
-#define NCCL_MINOR ${nccl:Minor}
-#define NCCL_PATCH ${nccl:Patch}
-#define NCCL_SUFFIX "${nccl:Suffix}"
+#define NCCL_MAJOR 2
+#define NCCL_MINOR 4
+#define NCCL_PATCH 7
+#define NCCL_SUFFIX ""
 
-#define NCCL_VERSION_CODE ${nccl:Version}
+#define NCCL_VERSION_CODE 2407
 #define NCCL_VERSION(X,Y,Z) ((X) * 1000 + (Y) * 100 + (Z))
 
 #ifdef __cplusplus