2.14.3-1

Add support for improved fault tolerance: non-blocking mode, new
init function with config, and ncclCommFinalize function.
Reintroduce collnet+chain algorithm, alongside collnet+direct.
Add LL protocol for intra-node P2P (on by default) and network
communication (off by default).
Use network instead of shared memory when performance is better.
Fix: wait for CUDA graph destroy before destroying comm with linked
graph resources.
Remove aggressive polling during enqueue.
Fix DMABUF fallback on MOFED 5.4 and earlier.

src/include/collectives.h

@@ -38,8 +38,9 @@ struct ncclDevRedOpFull {
   extern __device__ void NCCL_FUNC_NAME(func, algo, proto, devredop, type)(); \
   extern __global__ void NCCL_KERN_NAME(func, algo, proto, devredop, type)(struct ncclDevComm* comm, uint64_t channelMask, struct ncclWork* workHead); \
 
+#define SINGLE_ARG(...) __VA_ARGS__
 #define CONCAT(a,b) a##b
-#define MACRO_IF(cond, t, f) CONCAT(MACRO_IF_, cond)(t, f)
+#define MACRO_IF(cond, t, f) CONCAT(MACRO_IF_, cond)(SINGLE_ARG(t), SINGLE_ARG(f))
 #define MACRO_IF_0(t, f) f
 #define MACRO_IF_1(t, f) t
 
@@ -51,7 +52,8 @@ struct ncclDevRedOpFull {
 #define DECL3(func, devredop, type, undef) \
   DECL4(func, RING,    devredop, type, undef) \
   DECL4(func, TREE,    devredop, type, undef) \
-  DECL4(func, COLLNET, devredop, type, undef)
+  DECL4(func, COLLNET_DIRECT, devredop, type, undef) \
+  DECL4(func, COLLNET_CHAIN, devredop, type, undef)
 
 #if defined(__CUDA_BF16_TYPES_EXIST__)
 #define DECL2(func, devredop, undefForFloat) \
@@ -117,7 +119,6 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)();
 #define BROADCAST_CHUNKSTEPS 1
 #define REDUCE_SLICESTEPS 1
 #define REDUCE_CHUNKSTEPS 1
-#define SENDRECV_SLICEFACTOR 4
 #define NCCL_MAX_SLICE_PER_CHUNK 2  // max value for CHUNKSTEPS/SLICESTEPS, must accord with above
 
 #endif