Add tuner config for MCCL (#282)

Author: mikethegoblin

flagcx/adaptor/tuner/mccl_tuner.cc

@@ -0,0 +1,35 @@
+#include "tuner/tuner_util.h"
+
+#ifdef USE_METAX_ADAPTOR
+static EnvVar algo("MCCL_ALGO", {"ring", "tree"}, "ring");
+
+static EnvVar proto("MCCL_PROTO", {"LL", "LL128", "Simple"}, "Simple");
+
+static EnvVar minChannels("MCCL_MIN_NCHANNELS", {"2", "4", "8", "16", "32"},
+                          "2");
+
+static EnvVar maxChannels("MCCL_MAX_NCHANNELS", {"8", "16", "32", "64"}, "32");
+
+static EnvVar rwkBuffer("MCCL_DISABLE_CACHEABLE_BUFFER", {"0", "1"}, "0");
+
+static EnvVar minP2PChannels("MCCL_MIN_P2P_NCHANNELS", {"1", "4", "8", "16"},
+                             "1");
+
+static EnvVar maxP2PChannels("MCCL_MAX_P2P_NCHANNELS", {"16", "32", "64"},
+                             "64");
+
+static EnvVar cacheFastWriteBack("MCCL_FAST_WRITE_BACK", {"-2", "1"}, "-2");
+
+static EnvVar cacheEarlyWriteBack("MCCL_EARLY_WRITE_BACK",
+                                  {"-2", "1", "4", "8", "15"}, "-2");
+
+std::vector<EnvVar> mcclTunerVars = {algo,
+                                     proto,
+                                     minChannels,
+                                     maxChannels,
+                                     rwkBuffer,
+                                     minP2PChannels,
+                                     maxP2PChannels,
+                                     cacheFastWriteBack,
+                                     cacheEarlyWriteBack};
+#endif

flagcx/adaptor/tuner/nccl_tuner.cc

@@ -12,6 +12,7 @@ static EnvVar minChannel("NCCL_MIN_NCHANNELS", {"16", "32"}, "16");
 
 static EnvVar chunkSize("NCCL_P2P_NVL_CHUNKSIZE", {"1024", "2048"}, "1024");
 
-std::vector<EnvVar> vars = {algo, proto, thread, minChannel, chunkSize};
+std::vector<EnvVar> ncclTunerVars = {algo, proto, thread, minChannel,
+                                     chunkSize};
 
 #endif

flagcx/adaptor/tuner/tuner.cc

@@ -0,0 +1,9 @@
+#include "tuner/tuner_util.h"
+
+#ifdef USE_NVIDIA_ADAPTOR
+std::vector<EnvVar> vars = ncclTunerVars;
+#elif USE_METAX_ADAPTOR
+std::vector<EnvVar> vars = mcclTunerVars;
+#else
+std::vector<EnvVar> vars = {};
+#endif

flagcx/adaptor/tuner/tuner_util.h

@@ -20,6 +20,8 @@ struct EnvVar {
 flagcxResult_t generateCandidate(std::vector<struct flagcxEnvConfig> &cfgList);
 static void safeStrCopy(char *dst, size_t dstSize, const std::string &src);
 
+extern std::vector<EnvVar> ncclTunerVars;
+extern std::vector<EnvVar> mcclTunerVars;
 extern std::vector<EnvVar> vars;
 
 #endif // end include guard