2.17.1-1

Add new NVLS algorithm for allreduce using NVLink SHARP (intra-node only).
Add new config options: cgaClusterSize, minCTAs, maxCTAs, netName.
Enable LL128 when we use PXN to close rings.
NVTX3 includes update.
Fix crash when one CollNet (SHARP) rail fails to initialize.

src/include/collectives.h

@@ -53,7 +53,8 @@ struct ncclDevRedOpFull {
   DECL4(func, RING,    devredop, type, undef) \
   DECL4(func, TREE,    devredop, type, undef) \
   DECL4(func, COLLNET_DIRECT, devredop, type, undef) \
-  DECL4(func, COLLNET_CHAIN, devredop, type, undef)
+  DECL4(func, COLLNET_CHAIN, devredop, type, undef) \
+  DECL4(func, NVLS,    devredop, type, undef)
 
 #if defined(__CUDA_BF16_TYPES_EXIST__)
 #define DECL2(func, devredop, undefForFloat) \
@@ -121,4 +122,13 @@ extern __device__ void NCCL_ONERANK_REDUCE_NAME(PreMulSum, double)();
 #define REDUCE_CHUNKSTEPS 1
 #define NCCL_MAX_SLICE_PER_CHUNK 2  // max value for CHUNKSTEPS/SLICESTEPS, must accord with above
 
+// We can't use the enum identifiers like ncclSum, ncclFloat, etc since this
+// macro will be used in preprocessor conditionals where enums have no meaning.
+#define NCCL_NVLS_SUPPORTS(/*ncclDataType_t*/ type, /*ncclDevRedOp_t*/ red) \
+  (((type==2 || type==3) && (red==0 || red==2 || red==3)) || \
+   ((type==4 || type==5) && (red==0 || red==2 || red==3)) || \
+   ((type==6 || type==9) && (red==0 || red==2 || red==3)) || \
+   (type==7 && red==0) || \
+   (type==8 && red==0))
+
 #endif