Merge branch 'main' into user/dongfengy/fix_ref

Author: dongfengy

cpp/micro_benchmarks/mixtureOfExpertsBackendBenchmarkFixture.h

@@ -1074,10 +1074,10 @@ void MixtureOfExpertsBenchmark<TypeTuple_>::runBenchmark(benchmark::State& state
         state.SkipWithMessage("Out of range tactic");
         return;
     }
+    auto tactics1 = mMoERunner.getTactics(MoeGemmId::GEMM_1);
+    auto tactics2 = mMoERunner.getTactics(MoeGemmId::GEMM_2);
     if (LOG_LEVEL >= INFO)
     {
-        auto tactics1 = mMoERunner.getTactics(MoeGemmId::GEMM_1);
-        auto tactics2 = mMoERunner.getTactics(MoeGemmId::GEMM_2);
         std::cout << "Selected tactic #1: " << tactic_idx1 << "/" << tactics1.size() << "\n"
                   << tactics1[tactic_idx1].toString() << std::endl;
         std::cout << "Selected tactic #2: " << tactic_idx2 << "/" << tactics2.size() << "\n"
@@ -1086,6 +1086,20 @@ void MixtureOfExpertsBenchmark<TypeTuple_>::runBenchmark(benchmark::State& state
     state.counters["tactic_idx1"] = tactic_idx1;
     state.counters["tactic_idx2"] = tactic_idx2;
 
+    state.counters["t1_sm_version"] = tactics1[tactic_idx1].sm_version;
+    state.counters["t1_tile_shape"] = tactics1[tactic_idx1].getTileConfigAsInt();
+    state.counters["t1_cluster_shape"] = (int) tactics1[tactic_idx1].cluster_shape;
+    state.counters["t1_dynamic_cluster_shape"] = (int) tactics1[tactic_idx1].dynamic_cluster_shape;
+    state.counters["t1_fallback_cluster_shape"] = (int) tactics1[tactic_idx1].fallback_cluster_shape;
+    state.counters["t1_epilogue_schedule"] = (int) tactics1[tactic_idx1].epilogue_schedule;
+
+    state.counters["t2_sm_version"] = tactics2[tactic_idx2].sm_version;
+    state.counters["t2_tile_shape"] = tactics2[tactic_idx2].getTileConfigAsInt();
+    state.counters["t2_cluster_shape"] = (int) tactics2[tactic_idx2].cluster_shape;
+    state.counters["t2_dynamic_cluster_shape"] = (int) tactics2[tactic_idx2].dynamic_cluster_shape;
+    state.counters["t2_fallback_cluster_shape"] = (int) tactics2[tactic_idx2].fallback_cluster_shape;
+    state.counters["t2_epilogue_schedule"] = (int) tactics2[tactic_idx2].epilogue_schedule;
+
     createGraph(parallelism_config, gemm_to_profile);
 
     {

cpp/micro_benchmarks/mixtureOfExpertsBackendBenchmarkLauncher.cu

@@ -160,6 +160,10 @@ void argGenLoadFile(benchmark::internal::Benchmark* benchmark)
      */
 
     std::ifstream file{workloadFile};
+    if (!file.is_open())
+    {
+        throw std::invalid_argument("Failed to open benchmark file: " + std::string(workloadFile));
+    }
     std::stringstream buffer;
     buffer << file.rdbuf();
     auto file_contents = buffer.str();
@@ -294,7 +298,7 @@ void argGenLoadFile(benchmark::internal::Benchmark* benchmark)
         int gemm_to_profile = get_or("gemm_to_profile", (int) GemmToProfile::LAYER);
         TLLM_CHECK_WITH_INFO(world_rank < tp_size * ep_size, "Rank is out of bounds of tp*ep");
 
-        if (gemm_to_profile != (int) GemmToProfile::LAYER && routing_config != UNIFORM_ROUTING_CONFIG)
+        if (gemm_to_profile != (int) GemmToProfile::LAYER)
         {
             static bool info_printed = false;
             if (!info_printed && LOG_LEVEL >= INFO)
@@ -304,13 +308,14 @@ void argGenLoadFile(benchmark::internal::Benchmark* benchmark)
             }
 
             static bool printed = false;
-            if (LOG_LEVEL >= ERROR && !printed)
+            if (routing_config != UNIFORM_ROUTING_CONFIG && LOG_LEVEL >= ERROR && !printed)
             {
                 std::cerr << "Warning: Profiling a specific GEMM will always use uniform random token distribution"
                           << std::endl;
                 printed = true;
             }
             routing_config = UNIFORM_ROUTING_CONFIG;
+
             if (gemm_to_profile == (int) GemmToProfile::GEMM_1)
             {
                 tactic_ids2 = {-1};