Benchmarks: Microbenchmark - Add distributed inference benchmark cpp implementation (#586)

**Description**
Add distributed inference benchmark cpp implementation.

docs/user-tutorial/benchmarks/micro-benchmarks.md

@@ -418,7 +418,7 @@ Test the performance of large scale matmul operation with multiple GPUs:
 
 #### Introduction
 
-Test the performance of distributed model inference.
+Test the performance of distributed model inference. Support both PyTorch implementation and cpp implementation.
 
 #### Metrics
 

superbench/benchmarks/micro_benchmarks/dist_inference.py

@@ -12,7 +12,7 @@ import torch.distributed as dist
 
 from superbench.common.utils import logger
 from superbench.benchmarks import DistributedImpl, DistributedBackend, BenchmarkRegistry, ReturnCode, Precision
-from superbench.benchmarks.micro_benchmarks import MicroBenchmark
+from superbench.benchmarks.micro_benchmarks import MicroBenchmarkWithInvoke
 from superbench.benchmarks.context import Enum
 from superbench.benchmarks.reducer import ReduceType
 
@@ -168,7 +168,7 @@ class DistInferenceModel(torch.nn.Module):
         return activation_out
 
 
-class DistInference(MicroBenchmark):
+class DistInference(MicroBenchmarkWithInvoke):
     """The base class of micro-benchmarks."""
     def __init__(self, name, parameters=''):
         """Constructor.
@@ -182,7 +182,9 @@ class DistInference(MicroBenchmark):
         self.__local_rank = 0
         torch.backends.cudnn.benchmark = True
         self.__device = None
-        self.__cuda_available = False
+
+        # For cpp impl path
+        self._bin_name = 'dist_inference'
 
     def __timer(self):
         """Returns the current time which ensures all previous CUDA events have been finished.
@@ -193,14 +195,19 @@ class DistInference(MicroBenchmark):
         Return:
             Current time in second.
         """
-        if self.__cuda_available:
+        torch.cuda.synchronize()
-            torch.cuda.synchronize()
         return time.time()
 
     def add_parser_arguments(self):
         """Add the specified arguments."""
         super().add_parser_arguments()
 
+        self._parser.add_argument(
+            '--use_pytorch',
+            action='store_true',
+            required=False,
+            help='Whether to use pytorch implementation. If not, cpp implementation will be used.',
+        )
         self._parser.add_argument(
             '--batch_size',
             type=int,
@@ -222,6 +229,20 @@ class DistInference(MicroBenchmark):
             required=False,
             help='Hidden size.',
         )
+        self._parser.add_argument(
+            '--alpha',
+            type=float,
+            default=1.0,
+            required=False,
+            help='Coefficient alpha in D = alpha*(A*B) + beta*(C).',
+        )
+        self._parser.add_argument(
+            '--beta',
+            type=float,
+            default=1.0,
+            required=False,
+            help='Coefficient beta in D = alpha*(A*B) + beta*(C).',
+        )
         self._parser.add_argument(
             '--num_layers',
             type=int,
@@ -285,6 +306,12 @@ class DistInference(MicroBenchmark):
             required=False,
             help='Distributed backends. E.g. {}.'.format(' '.join(DistributedBackend.get_values())),
         )
+        self._parser.add_argument(
+            '--use_cuda_graph',
+            action='store_true',
+            required=False,
+            help='Whether to launch kernels in CUDA graph mode.',
+        )
 
     def _preprocess(self):
         """Preprocess/preparation operations before the benchmarking.
@@ -295,32 +322,41 @@ class DistInference(MicroBenchmark):
         if not super()._preprocess():
             return False
 
-        if self._args.distributed_impl != DistributedImpl.DDP:
+        if self._args.use_pytorch:
-            self._result.set_return_code(ReturnCode.DISTRIBUTED_SETTING_INIT_FAILURE)
+            # Initialize PyTorch if pytorch impl path
-            logger.error(
+            if self._args.distributed_impl != DistributedImpl.DDP:
-                'Unsupported distributed implementation - model: {}, distributed implementation: {}.'.format(
+                return self._set_error_code_and_print_error_msg(
-                    self._name, self._args.distributed_impl
+                    ReturnCode.DISTRIBUTED_SETTING_INIT_FAILURE,
+                    'Unsupported distributed implementation - model: {}, distributed implementation: {}.'.format(
+                        self._name, self._args.distributed_impl
+                    )
                 )
-            )
-            return False
 
-        try:
+            try:
-            torch.distributed.init_process_group(backend=self._args.distributed_backend.value)
+                torch.distributed.init_process_group(backend=self._args.distributed_backend.value)
-            self.__world_size = int(os.environ['WORLD_SIZE'])
+                self.__world_size = int(os.environ['WORLD_SIZE'])
-            self.__local_rank = int(os.environ['LOCAL_RANK'])
+                self.__local_rank = int(os.environ['LOCAL_RANK'])
-        except BaseException as e:
+                assert (torch.cuda.is_available())
-            self._result.set_return_code(ReturnCode.DISTRIBUTED_SETTING_INIT_FAILURE)
+            except BaseException as e:
-            torch.distributed.destroy_process_group()
+                torch.distributed.destroy_process_group()
-            logger.error('Initialize distributed env failed - benchmark: {}, message: {}.'.format(self._name, str(e)))
+                return self._set_error_code_and_print_error_msg(
-            return False
+                    ReturnCode.DISTRIBUTED_SETTING_INIT_FAILURE,
+                    'Initialize distributed env failed - benchmark: {}, message: {}.'.format(self._name, str(e))
+                )
 
-        if torch.cuda.is_available():
             torch.cuda.set_device(self.__local_rank)
             self.__device = torch.device('cuda:{}'.format(self.__local_rank))
-            self.__cuda_available = True
         else:
-            self.__device = torch.device('cpu:{}'.format(self.__local_rank))
+            # Assemble commands if cpp impl path
-            self.__cuda_available = False
+            self.__bin_path = os.path.join(self._args.bin_dir, self._bin_name)
+
+            args = '-m %d -n %d -k %d' % (self._args.hidden_size, self._args.batch_size, self._args.input_size)
+            args += ' --alpha %g --beta %g' % (self._args.alpha, self._args.beta)
+            args += ' --num_layers %d --num_warmups %d --num_iters %d' % \
+                (self._args.num_layers, self._args.num_warmup, self._args.num_steps)
+            if self._args.use_cuda_graph:
+                args += ' --use_cuda_graph'
+            self._commands = ['%s %s' % (self.__bin_path, args)]
 
         return True
 
@@ -347,8 +383,7 @@ class DistInference(MicroBenchmark):
             self.__device
         )
         model = model.to(dtype=getattr(torch, precision.value))
-        if self.__cuda_available:
+        model = model.cuda()
-            model = model.cuda()
         return model
 
     def _run_model(self, model, batch_size, input_size, precision, device, num_warmup, num_steps):
@@ -401,38 +436,78 @@ class DistInference(MicroBenchmark):
         Return:
             True if _benchmark succeeds.
         """
-        batch_size = self._args.batch_size
+        if self._args.use_pytorch:
-        input_size = self._args.input_size
+            # Execute PyTorch model if pytorch impl path
-        hidden_size = self._args.hidden_size
+            batch_size = self._args.batch_size
-        num_layers = self._args.num_layers
+            input_size = self._args.input_size
-        computation = self._args.computation_kernel
+            hidden_size = self._args.hidden_size
-        communication = self._args.communication_kernel
+            num_layers = self._args.num_layers
-        activation = self._args.activation_kernel
+            computation = self._args.computation_kernel
-        precision = self._args.precision
+            communication = self._args.communication_kernel
-        num_warmup = self._args.num_warmup
+            activation = self._args.activation_kernel
-        num_steps = self._args.num_steps
+            precision = self._args.precision
+            num_warmup = self._args.num_warmup
+            num_steps = self._args.num_steps
 
-        if self.__local_rank == 0:
+            if self.__local_rank == 0:
-            logger.info(
+                logger.info(
-                'Distributed Inference - using {} GPUs: '
+                    'Distributed Inference - using {} GPUs: '
-                'batch_size={}, input_size={}, hidden_size={}, num_layers={}, '
+                    'batch_size={}, input_size={}, hidden_size={}, num_layers={}, '
-                'computation_kernel={}, communication_kernel={}, activation_kernel={}, precision={}, '
+                    'computation_kernel={}, communication_kernel={}, activation_kernel={}, precision={}, '
-                'num_warmup={} num_steps={}'.format(
+                    'num_warmup={} num_steps={}'.format(
-                    self.__world_size, batch_size, input_size, hidden_size, num_layers, computation, communication,
+                        self.__world_size, batch_size, input_size, hidden_size, num_layers, computation, communication,
-                    activation, precision, num_warmup, num_steps
+                        activation, precision, num_warmup, num_steps
+                    )
                 )
+
+            # Prepare model
+            model = self._prepare_model(
+                input_size, hidden_size, num_layers, computation, communication, activation, precision,
+                self.__world_size
             )
 
-        # Prepare model
+            # Run model
-        model = self._prepare_model(
+            step_times = self._run_model(model, batch_size, input_size, precision, self.__device, num_warmup, num_steps)
-            input_size, hidden_size, num_layers, computation, communication, activation, precision, self.__world_size
-        )
 
-        # Run model
+            # Process data and return
-        step_times = self._run_model(model, batch_size, input_size, precision, self.__device, num_warmup, num_steps)
+            return self._process_data(step_times)
+        else:
+            # Execute commands if cpp impl path
+            if not super()._benchmark():
+                return False
+            return True
 
-        # Process data and return
+    def _process_raw_result(self, cmd_idx, raw_output):
-        return self._process_data(step_times)
+        """Function to parse raw results and save the summarized results.
+
+          self._result.add_raw_data() and self._result.add_result() need to be called to save the results.
+
+        Args:
+            cmd_idx (int): the index of command corresponding with the raw_output.
+            raw_output (str): raw output string of the micro-benchmark.
+
+        Return:
+            True if the raw output string is valid and result can be extracted.
+        """
+        self._result.add_raw_data('raw_output_' + str(cmd_idx), raw_output, self._args.log_raw_data)
+
+        try:
+            output_lines = [x.strip() for x in raw_output.strip().splitlines()]
+            step_time = None
+            for output_line in output_lines:
+                if ' ms per iteration' in output_line:
+                    step_time = float(output_line.split(' ms per iteration')[0].split()[-1])
+                    break
+            return self._process_numeric_result(
+                'step_times', [step_time], reduce_type=ReduceType.MAX, cal_percentile=True
+            )
+        except BaseException as e:
+            return self._set_error_code_and_print_error_msg(
+                ReturnCode.MICROBENCHMARK_RESULT_PARSING_FAILURE,
+                'The result format is invalid - round: {}, benchmark: {}, raw output: {}, message: {}.'.format(
+                    self._curr_run_index, self._name, raw_output, str(e)
+                )
+            )
 
     def _postprocess(self):
         """Postprocess/cleanup operations after the benchmarking.
@@ -443,14 +518,26 @@ class DistInference(MicroBenchmark):
         if not super()._postprocess():
             return False
 
-        try:
+        if self._args.use_pytorch:
-            torch.distributed.destroy_process_group()
+            try:
-        except BaseException as e:
+                torch.distributed.destroy_process_group()
-            self._result.set_return_code(ReturnCode.DISTRIBUTED_SETTING_DESTROY_FAILURE)
+            except BaseException as e:
-            logger.error('Post process failed - benchmark: {}, message: {}.'.format(self._name, str(e)))
+                return self._set_error_code_and_print_error_msg(
-            return False
+                    ReturnCode.DISTRIBUTED_SETTING_DESTROY_FAILURE,
+                    'Post process failed - benchmark: {}, message: {}.'.format(self._name, str(e))
+                )
 
         return True
 
+    def _set_error_code_and_print_error_msg(self, error_code, error_msg):
+        """Set error code and print error log upon error.
+
+        Return:
+            False, representing error.
+        """
+        self._result.set_return_code(error_code)
+        logger.error(error_msg)
+        return False
+
 
 BenchmarkRegistry.register_benchmark('pytorch-dist-inference', DistInference, parameters='')

superbench/benchmarks/micro_benchmarks/dist_inference_cpp/CMakeLists.txt

@@ -0,0 +1,40 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT License.
+
+cmake_minimum_required(VERSION 3.18)
+
+project(dist_inference LANGUAGES CXX)
+
+find_package(MPI REQUIRED)
+include_directories(SYSTEM ${MPI_INCLUDE_PATH})
+
+find_package(CUDAToolkit QUIET)
+
+# Cuda environment
+if(CUDAToolkit_FOUND)
+    message(STATUS "Found CUDA: " ${CUDAToolkit_VERSION})
+
+    include(../cuda_common.cmake)
+    add_executable(dist_inference dist_inference.cu)
+    set_property(TARGET dist_inference PROPERTY CUDA_ARCHITECTURES ${NVCC_ARCHS_SUPPORTED})
+    target_link_libraries(dist_inference MPI::MPI_CXX nccl cublasLt)
+else()
+    # ROCm environment
+    include(../rocm_common.cmake)
+    find_package(hip QUIET)
+    if(hip_FOUND)
+        message(STATUS "Found ROCm: " ${HIP_VERSION})
+
+        # Convert cuda code to hip code in cpp
+        execute_process(COMMAND hipify-perl -print-stats -o dist_inference.cpp dist_inference.cu WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/)
+
+        # link hip device lib
+        add_executable(dist_inference dist_inference.cpp)
+        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2 -DROCM_USE_FLOAT16=1")
+        target_link_libraries(dist_inference MPI::MPI_CXX rccl hipblaslt hip::device)
+    else()
+        message(FATAL_ERROR "No CUDA or ROCm environment found.")
+    endif()
+endif()
+
+install(TARGETS dist_inference RUNTIME DESTINATION bin)

superbench/benchmarks/micro_benchmarks/dist_inference_cpp/dist_inference.cu

@@ -0,0 +1,455 @@
+/*******************************************************************************
+ * Copyright (c) Microsoft Corporation.
+ * Licensed under the MIT License.
+ *
+ * MIT License
+ *
+ * Copyright (C) 2022-2023 Advanced Micro Devices, Inc.
+ * Modifications Copyright (c) Microsoft Corporation. Licensed under the MIT License.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+
+#include <chrono>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <functional>
+#include <iostream>
+#include <limits>
+#include <mpi.h>
+#include <string>
+#include <unistd.h>
+#include <vector>
+
+#if defined(__HIP_PLATFORM_AMD__)
+#include <hipblaslt/hipblaslt.h>
+#include <rccl/rccl.h>
+using cublasLtHalf = hipblasLtHalf;
+#else
+#include <cublasLt.h>
+#include <cuda_fp16.h>
+#include <nccl.h>
+using cublasLtHalf = half;
+#endif
+
+#ifndef CHECK_CUDA_ERROR
+#define CHECK_CUDA_ERROR(error)                                                                                        \
+    if (error != cudaSuccess) {                                                                                        \
+        fprintf(stderr, "Cuda error: '%s'(%d) at %s:%d\n", cudaGetErrorString(error), error, __FILE__, __LINE__);      \
+        exit(-1);                                                                                                      \
+    }
+#endif
+
+#ifndef CHECK_CUBLASLT_ERROR
+#define CHECK_CUBLASLT_ERROR(error)                                                                                    \
+    if (error != CUBLAS_STATUS_SUCCESS) {                                                                              \
+        fprintf(stderr, "cuBLASLt error(Err=%d) at %s:%d\n", error, __FILE__, __LINE__);                               \
+        fprintf(stderr, "\n");                                                                                         \
+        exit(-1);                                                                                                      \
+    }
+#endif
+
+#ifndef CHECK_NCCL_ERROR
+#define CHECK_NCCL_ERROR(error)                                                                                        \
+    if (error != ncclSuccess) {                                                                                        \
+        fprintf(stderr, "NCCL error(Err=%d) at %s:%d\n", error, __FILE__, __LINE__);                                   \
+        fprintf(stderr, "\n");                                                                                         \
+        exit(-1);                                                                                                      \
+    }
+#endif
+
+static void ShowUsage(char *argv[]) {
+    std::cerr << "Usage: " << argv[0] << " <options>\n"
+              << "options:\n"
+              << "\t-h, --help\t\t\t\tShow this help message\n"
+              << "\t-m \t\t\tm\t\tGEMM_STRIDED argument m\n"
+              << "\t-n \t\t\tn\t\tGEMM_STRIDED argument n\n"
+              << "\t-k \t\t\tk \t\tGEMM_STRIDED argument k\n"
+              << "\t--alpha \t\talpha \t\tGEMM_STRIDED argument alpha\n"
+              << "\t--beta \t\t\tbeta \t\tGEMM_STRIDED argument beta\n"
+              << "\t--num_layers \t\t\tnum_layers \t\tNumber of layers in the model\n"
+              << "\t--num_warmups \t\t\tnum_warmups \t\tNumber of warmup runs\n"
+              << "\t--num_iters \t\t\tnum_iters \t\tNumber of test runs\n"
+              << "\t--use_cuda_graph \t\t\tuse_cuda_graph \t\tWhether to launch kernels in CUDA graph mode\n"
+              << std::endl;
+}
+
+static int ParseArguments(int argc, char *argv[], int64_t *m, int64_t *n, int64_t *k, float *alpha, float *beta,
+                          int32_t *num_layers, int32_t *num_warmups, int32_t *num_iters, bool *use_cuda_graph) {
+    if (argc >= 2) {
+        for (int i = 1; i < argc; ++i) {
+            std::string arg = argv[i];
+
+            if ((arg.at(0) == '-') || ((arg.at(0) == '-') && (arg.at(1) == '-'))) {
+                if ((arg == "-h") || (arg == "--help")) {
+                    return -1;
+                } else if ((arg == "-m") && (i + 1 < argc)) {
+                    *m = atoi(argv[++i]);
+                } else if ((arg == "-n") && (i + 1 < argc)) {
+                    *n = atoi(argv[++i]);
+                } else if ((arg == "-k") && (i + 1 < argc)) {
+                    *k = atoi(argv[++i]);
+                } else if ((arg == "--alpha") && (i + 1 < argc)) {
+                    *alpha = atof(argv[++i]);
+                } else if ((arg == "--beta") && (i + 1 < argc)) {
+                    *beta = atof(argv[++i]);
+                } else if ((arg == "--num_layers") && (i + 1 < argc)) {
+                    *num_layers = atoi(argv[++i]);
+                } else if ((arg == "--num_warmups") && (i + 1 < argc)) {
+                    *num_warmups = atoi(argv[++i]);
+                } else if ((arg == "--num_iters") && (i + 1 < argc)) {
+                    *num_iters = atoi(argv[++i]);
+                } else if (arg == "--use_cuda_graph") {
+#if (NCCL_MAJOR > 2 || (NCCL_MAJOR >= 2 && NCCL_MINOR >= 9)) && (CUDART_VERSION >= 11030 || HIP_VERSION >= 50221310)
+                    *use_cuda_graph = true;
+#else
+                    *use_cuda_graph = false;
+                    std::cerr << "error with " << arg << std::endl;
+                    std::cerr << "not supported by current environment" << std::endl << std::endl;
+                    return -1;
+#endif
+                } else {
+                    std::cerr << "error with " << arg << std::endl;
+                    std::cerr << "do not recognize option" << std::endl << std::endl;
+                    return -1;
+                }
+            } else {
+                std::cerr << "error with " << arg << std::endl;
+                std::cerr << "option must start with - or --" << std::endl << std::endl;
+                return -1;
+            }
+        }
+    }
+    return 0;
+}
+
+void InitializeABCDEF(std::vector<cublasLtHalf> &ha, int64_t size_a, std::vector<cublasLtHalf> &hb, int64_t size_b,
+                      std::vector<cublasLtHalf> &hc, int64_t size_c, std::vector<cublasLtHalf> &hd, int64_t size_d,
+                      std::vector<cublasLtHalf> &he, int64_t size_e, std::vector<cublasLtHalf> &hf, int64_t size_f) {
+    srand(1);
+    for (int i = 0; i < size_a; ++i) {
+        ha[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+    for (int i = 0; i < size_b; ++i) {
+        hb[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+    for (int i = 0; i < size_c; ++i) {
+        hc[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+    for (int i = 0; i < size_d; ++i) {
+        hd[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+    for (int i = 0; i < size_e; ++i) {
+        he[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+    for (int i = 0; i < size_f; ++i) {
+        hf[i] = static_cast<cublasLtHalf>((rand() % 7) - 3);
+    }
+}
+
+// B[m, k] * A[k, n] + C[m, n] = D[m, n]
+// E[k, m] * D[m, n] + F[k, n] = G[k, n]
+void TestModel(int64_t m, int64_t n, int64_t k, float alpha, float beta, int32_t num_layers, int32_t num_warmups,
+               int32_t num_iters, bool use_cuda_graph, ncclComm_t nccl_comm) {
+    const int kNcclBufAlignment = 512;
+
+    int size_a = k * n;
+    int size_b = m * k;
+    int size_c = m * n;
+    int size_d = m * n;
+    int size_e = k * m;
+    int size_f = k * n;
+    int size_g = (k * n + kNcclBufAlignment - 1) / kNcclBufAlignment * kNcclBufAlignment;
+
+    // Naming: da is in GPU (device) memory. ha is in CPU (host) memory
+    std::vector<cublasLtHalf> ha(size_a);
+    std::vector<cublasLtHalf> hb(size_b);
+    std::vector<cublasLtHalf> hc(size_c);
+    std::vector<cublasLtHalf> hd(size_d);
+    std::vector<cublasLtHalf> he(size_e);
+    std::vector<cublasLtHalf> hf(size_f);
+    std::vector<cublasLtHalf> hg(size_g);
+
+    // initial data on host
+    InitializeABCDEF(ha, size_a, hb, size_b, hc, size_c, hd, size_d, he, size_e, hf, size_f);
+
+    // allocate memory on device
+    void *da, *db, *dc, *dd, *de, *df, *dg;
+
+    // Create stream
+    cudaStream_t stream = nullptr;
+    CHECK_CUDA_ERROR(cudaStreamCreateWithFlags(&stream, cudaStreamNonBlocking));
+
+    CHECK_CUDA_ERROR(cudaMalloc(&da, size_a * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&db, size_b * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&dc, size_c * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&dd, size_d * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&de, size_e * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&df, size_f * sizeof(cublasLtHalf)));
+    CHECK_CUDA_ERROR(cudaMalloc(&dg, size_g * sizeof(cublasLtHalf)));
+    // copy matrices from host to device
+    CHECK_CUDA_ERROR(cudaMemcpy(da, ha.data(), sizeof(cublasLtHalf) * size_a, cudaMemcpyHostToDevice));
+    CHECK_CUDA_ERROR(cudaMemcpy(db, hb.data(), sizeof(cublasLtHalf) * size_b, cudaMemcpyHostToDevice));
+    CHECK_CUDA_ERROR(cudaMemcpy(dc, hc.data(), sizeof(cublasLtHalf) * size_c, cudaMemcpyHostToDevice));
+    CHECK_CUDA_ERROR(cudaMemcpy(dd, hd.data(), sizeof(cublasLtHalf) * size_d, cudaMemcpyHostToDevice));
+    CHECK_CUDA_ERROR(cudaMemcpy(de, he.data(), sizeof(cublasLtHalf) * size_e, cudaMemcpyHostToDevice));
+    CHECK_CUDA_ERROR(cudaMemcpy(df, hf.data(), sizeof(cublasLtHalf) * size_f, cudaMemcpyHostToDevice));
+
+    uint64_t workspace_size = 1024 * 1024;
+    void *d_workspace;
+    CHECK_CUDA_ERROR(cudaMalloc(&d_workspace, workspace_size));
+    int returnedAlgoCount = 0;
+
+    // cublasLt is not well supported by ROCm hipify tools, explicitly define ROCm logic instead.
+#if defined(__HIP_PLATFORM_AMD__)
+    hipblasLtHandle_t handle;
+    hipblasLtMatrixLayout_t matA, matB, matC, matD, matE, matF, matG;
+    hipblasLtMatmulDesc_t matmul1, matmul2;
+    hipblasLtMatmulPreference_t pref;
+
+    CHECK_CUBLASLT_ERROR(hipblasLtCreate(&handle));
+
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matA, HIPBLAS_R_16F, k, n, k));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matB, HIPBLAS_R_16F, m, k, m));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matC, HIPBLAS_R_16F, m, n, m));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matD, HIPBLAS_R_16F, m, n, m));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matE, HIPBLAS_R_16F, k, m, k));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matF, HIPBLAS_R_16F, k, n, k));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutCreate(&matG, HIPBLAS_R_16F, k, n, k));
+
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulDescCreate(&matmul1, HIPBLASLT_COMPUTE_F32, HIPBLAS_R_32F));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulDescCreate(&matmul2, HIPBLASLT_COMPUTE_F32, HIPBLAS_R_32F));
+
+    hipblasOperation_t trans = HIPBLAS_OP_N;
+    CHECK_CUBLASLT_ERROR(
+        hipblasLtMatmulDescSetAttribute(matmul1, HIPBLASLT_MATMUL_DESC_TRANSA, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(
+        hipblasLtMatmulDescSetAttribute(matmul1, HIPBLASLT_MATMUL_DESC_TRANSB, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(
+        hipblasLtMatmulDescSetAttribute(matmul2, HIPBLASLT_MATMUL_DESC_TRANSA, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(
+        hipblasLtMatmulDescSetAttribute(matmul2, HIPBLASLT_MATMUL_DESC_TRANSB, &trans, sizeof(int32_t)));
+
+    // Set User Preference attributes
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulPreferenceCreate(&pref));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulPreferenceSetAttribute(pref, HIPBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+                                                               &workspace_size, sizeof(workspace_size)));
+
+    // Get Heuristic results
+    hipblasLtMatmulHeuristicResult_t heuristicResult1[1] = {0};
+    hipblasLtMatmulHeuristicResult_t heuristicResult2[1] = {0};
+    // B[m, k] * A[k, n] + C[m, n] = D[m, n]
+    // E[k, m] * D[m, n] + F[k, n] = G[k, n]
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulAlgoGetHeuristic(handle, matmul1, matB, matA, matC, matD, pref, 1,
+                                                         heuristicResult1, &returnedAlgoCount));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulAlgoGetHeuristic(handle, matmul2, matE, matD, matF, matG, pref, 1,
+                                                         heuristicResult2, &returnedAlgoCount));
+#else
+    cublasLtHandle_t handle;
+    cublasLtMatrixLayout_t matA, matB, matC, matD, matE, matF, matG;
+    cublasLtMatmulDesc_t matmul1, matmul2;
+    cublasLtMatmulPreference_t pref;
+    CHECK_CUBLASLT_ERROR(cublasLtCreate(&handle));
+
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matA, CUDA_R_16F, k, n, k));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matB, CUDA_R_16F, m, k, m));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matC, CUDA_R_16F, m, n, m));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matD, CUDA_R_16F, m, n, m));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matE, CUDA_R_16F, k, m, k));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matF, CUDA_R_16F, k, n, k));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutCreate(&matG, CUDA_R_16F, k, n, k));
+
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescCreate(&matmul1, CUBLAS_COMPUTE_16F, CUDA_R_32F));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescCreate(&matmul2, CUBLAS_COMPUTE_16F, CUDA_R_32F));
+
+    cublasOperation_t trans = CUBLAS_OP_N;
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescSetAttribute(matmul1, CUBLASLT_MATMUL_DESC_TRANSA, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescSetAttribute(matmul1, CUBLASLT_MATMUL_DESC_TRANSB, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescSetAttribute(matmul2, CUBLASLT_MATMUL_DESC_TRANSA, &trans, sizeof(int32_t)));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescSetAttribute(matmul2, CUBLASLT_MATMUL_DESC_TRANSB, &trans, sizeof(int32_t)));
+
+    // Set User Preference attributes
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulPreferenceCreate(&pref));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulPreferenceSetAttribute(pref, CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES,
+                                                              &workspace_size, sizeof(workspace_size)));
+
+    // Get Heuristic results
+    cublasLtMatmulHeuristicResult_t heuristicResult1[1] = {0};
+    cublasLtMatmulHeuristicResult_t heuristicResult2[1] = {0};
+    // B[m, k] * A[k, n] + C[m, n] = D[m, n]
+    // E[k, m] * D[m, n] + F[k, n] = G[k, n]
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulAlgoGetHeuristic(handle, matmul1, matB, matA, matC, matD, pref, 1,
+                                                        heuristicResult1, &returnedAlgoCount));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulAlgoGetHeuristic(handle, matmul2, matE, matD, matF, matG, pref, 1,
+                                                        heuristicResult2, &returnedAlgoCount));
+#endif
+
+    auto model_forward = [&] {
+        for (int j = 0; j < num_layers; j++) {
+            // B[m, k] * A[k, n] + C[m, n] = D[m, n]
+            // E[k, m] * D[m, n] + F[k, n] = G[k, n]
+            // cublasLt is not well supported by ROCm hipify tools, explicitly define ROCm logic instead.
+#if defined(__HIP_PLATFORM_AMD__)
+            CHECK_CUBLASLT_ERROR(hipblasLtMatmul(handle, matmul1, &alpha, db, matB, da, matA, &beta, dc, matC, dd, matD,
+                                                 &heuristicResult1[0].algo, d_workspace, workspace_size, stream));
+            CHECK_CUBLASLT_ERROR(hipblasLtMatmul(handle, matmul1, &alpha, de, matE, dd, matD, &beta, df, matF, dg, matG,
+                                                 &heuristicResult2[0].algo, d_workspace, workspace_size, stream));
+#else
+            CHECK_CUBLASLT_ERROR(cublasLtMatmul(handle, matmul1, &alpha, db, matB, da, matA, &beta, dc, matC, dd, matD,
+                                                &heuristicResult1[0].algo, d_workspace, workspace_size, stream));
+            CHECK_CUBLASLT_ERROR(cublasLtMatmul(handle, matmul1, &alpha, de, matE, dd, matD, &beta, df, matF, dg, matG,
+                                                &heuristicResult2[0].algo, d_workspace, workspace_size, stream));
+#endif
+            CHECK_NCCL_ERROR(ncclAllReduce(dg, dg, size_g, ncclFloat16, ncclSum, nccl_comm, stream));
+        }
+    };
+
+#if (NCCL_MAJOR > 2 || (NCCL_MAJOR >= 2 && NCCL_MINOR >= 9)) && (CUDART_VERSION >= 11030 || HIP_VERSION >= 50221310)
+    cudaGraph_t graph;
+    cudaGraphExec_t instance;
+    if (use_cuda_graph) {
+        CHECK_CUDA_ERROR(cudaStreamBeginCapture(stream, cudaStreamCaptureModeGlobal));
+        model_forward();
+        CHECK_CUDA_ERROR(cudaStreamEndCapture(stream, &graph));
+        CHECK_CUDA_ERROR(cudaGraphInstantiate(&instance, graph, NULL, NULL, 0));
+    }
+#endif
+
+    std::chrono::steady_clock::time_point start_time, stop_time;
+    for (int i = 0; i < num_warmups + num_iters; ++i) {
+        if (i == num_warmups) {
+            start_time = std::chrono::steady_clock::now();
+        }
+#if (NCCL_MAJOR > 2 || (NCCL_MAJOR >= 2 && NCCL_MINOR >= 9)) && (CUDART_VERSION >= 11030 || HIP_VERSION >= 50221310)
+        if (use_cuda_graph) {
+            CHECK_CUDA_ERROR(cudaGraphLaunch(instance, stream));
+        } else {
+            model_forward();
+        }
+#else
+        model_forward();
+#endif
+        CHECK_CUDA_ERROR(cudaStreamSynchronize(stream));
+    }
+    stop_time = std::chrono::steady_clock::now();
+    double duration = std::chrono::duration_cast<std::chrono::milliseconds>(stop_time - start_time).count();
+    fprintf(stdout, "Time: %g ms in total, %g ms per iteration, %g ms per layer\n", duration, duration / num_iters,
+            duration / num_iters / num_layers);
+
+#if (NCCL_MAJOR > 2 || (NCCL_MAJOR >= 2 && NCCL_MINOR >= 9)) && (CUDART_VERSION >= 11030 || HIP_VERSION >= 50221310)
+    // Destroy graph
+    if (use_cuda_graph) {
+        CHECK_CUDA_ERROR(cudaGraphExecDestroy(instance));
+        CHECK_CUDA_ERROR(cudaGraphDestroy(graph));
+    }
+#endif
+
+    // Destroy stream
+    CHECK_CUDA_ERROR(cudaStreamDestroy(stream));
+
+    CHECK_CUDA_ERROR(cudaFree(da));
+    CHECK_CUDA_ERROR(cudaFree(db));
+    CHECK_CUDA_ERROR(cudaFree(dc));
+    CHECK_CUDA_ERROR(cudaFree(dd));
+    CHECK_CUDA_ERROR(cudaFree(de));
+    CHECK_CUDA_ERROR(cudaFree(df));
+    CHECK_CUDA_ERROR(cudaFree(dg));
+    CHECK_CUDA_ERROR(cudaFree(d_workspace));
+    // cublasLt is not well supported by ROCm hipify tools, explicitly define ROCm logic instead.
+#if defined(__HIP_PLATFORM_AMD__)
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulPreferenceDestroy(pref));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulDescDestroy(matmul1));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatmulDescDestroy(matmul2));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matA));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matB));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matC));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matD));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matE));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matF));
+    CHECK_CUBLASLT_ERROR(hipblasLtMatrixLayoutDestroy(matG));
+    CHECK_CUBLASLT_ERROR(hipblasLtDestroy(handle));
+#else
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulPreferenceDestroy(pref));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescDestroy(matmul1));
+    CHECK_CUBLASLT_ERROR(cublasLtMatmulDescDestroy(matmul2));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matA));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matB));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matC));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matD));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matE));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matF));
+    CHECK_CUBLASLT_ERROR(cublasLtMatrixLayoutDestroy(matG));
+    CHECK_CUBLASLT_ERROR(cublasLtDestroy(handle));
+#endif
+
+    return;
+}
+
+int main(int argc, char *argv[]) {
+    // Init MPI
+    int comm_rank, comm_size;
+    MPI_Init(NULL, NULL);
+    MPI_Comm_rank(MPI_COMM_WORLD, &comm_rank);
+    MPI_Comm_size(MPI_COMM_WORLD, &comm_size);
+
+    // Init NCCL
+    int num_local_ranks = 0;
+    ncclComm_t nccl_comm;
+    ncclUniqueId nccl_id;
+    if (comm_rank == 0) {
+        CHECK_NCCL_ERROR(ncclGetUniqueId(&nccl_id));
+    }
+    MPI_Bcast(&nccl_id, sizeof(ncclUniqueId), MPI_BYTE, 0, MPI_COMM_WORLD);
+    CHECK_CUDA_ERROR(cudaGetDeviceCount(&num_local_ranks))
+    CHECK_CUDA_ERROR(cudaSetDevice(comm_rank % num_local_ranks));
+    CHECK_NCCL_ERROR(ncclCommInitRank(&nccl_comm, comm_size, nccl_id, comm_rank));
+
+    // Init parameters with default values
+    int64_t m = 80;
+    int64_t n = 128;
+    int64_t k = 128;
+    float alpha = 1;
+    float beta = 1;
+    int32_t num_layers = 50;
+    int32_t num_warmups = 20;
+    int32_t num_iters = 100;
+    bool use_cuda_graph = false;
+
+    if (ParseArguments(argc, argv, &m, &n, &k, &alpha, &beta, &num_layers, &num_warmups, &num_iters, &use_cuda_graph)) {
+        ShowUsage(argv);
+        return -1;
+    }
+
+    fprintf(stdout,
+            "Parameters: m=%ld, n=%ld, k=%ld, alpha=%f, beta=%f, num_layers=%d, num_warmups=%d, num_iters=%d, "
+            "use_cuda_graph=%d\n",
+            m, n, k, alpha, beta, num_layers, num_warmups, num_iters, (int)use_cuda_graph);
+
+    TestModel(m, n, k, alpha, beta, num_layers, num_warmups, num_iters, use_cuda_graph, nccl_comm);
+
+    CHECK_NCCL_ERROR(ncclCommDestroy(nccl_comm));
+
+    MPI_Finalize();
+
+    return 0;
+}

superbench/config/azure_ndmv4.yaml

@@ -33,6 +33,8 @@ superbench:
           node_num: 1
           env:
             NCCL_ASYNC_ERROR_HANDLING: '0'
+      frameworks:
+        - pytorch
     common_model_config: &common_model_config
       duration: 0
       num_warmup: 64

superbench/config/azure_ndv4.yaml

@@ -29,6 +29,8 @@ superbench:
           node_num: 1
           env:
             NCCL_ASYNC_ERROR_HANDLING: '0'
+      frameworks:
+        - pytorch
     common_model_config: &common_model_config
       duration: 0
       num_warmup: 64

tests/benchmarks/micro_benchmarks/test_dist_inference.py

@@ -3,12 +3,15 @@
 
 """Tests for distributed inference benchmark."""
 
+import numbers
 import unittest
 
 from tests.helper import decorator
+from tests.helper.testcase import BenchmarkTestCase
 import tests.benchmarks.utils as utils
 from superbench.benchmarks \
-    import BenchmarkRegistry, Framework, BenchmarkType, ReturnCode, Precision, DistributedImpl, DistributedBackend
+    import BenchmarkRegistry, Framework, BenchmarkType, ReturnCode, Precision, DistributedImpl, DistributedBackend, \
+    Platform
 from superbench.benchmarks.micro_benchmarks.dist_inference \
     import DistInference, ComputationKernelType, CommunicationKernelType, ActivationKernelType
 from superbench.common.utils import network
@@ -20,7 +23,9 @@ from superbench.common.utils import network
 @decorator.pytorch_test
 def test_pytorch_dist_inference_normal():
     """Test pytorch-dist-inference benchmark on distributed normal case."""
-    context = BenchmarkRegistry.create_benchmark_context('dist-inference', parameters='', framework=Framework.PYTORCH)
+    context = BenchmarkRegistry.create_benchmark_context(
+        'dist-inference', parameters='--use_pytorch', framework=Framework.PYTORCH
+    )
     world_size = 2
     assert (BenchmarkRegistry.is_benchmark_context_valid(context))
     results = utils.simulated_ddp_distributed_benchmark(context, world_size)
@@ -33,9 +38,12 @@ def test_pytorch_dist_inference_normal():
         assert (benchmark.type == BenchmarkType.MICRO)
 
         # Check predefined parameters of dist-inference benchmark.
+        assert (benchmark._args.use_pytorch is True)
         assert (benchmark._args.batch_size == 64)
         assert (benchmark._args.input_size == 1024)
         assert (benchmark._args.hidden_size == 1024)
+        assert (benchmark._args.alpha == 1.0)
+        assert (benchmark._args.beta == 1.0)
         assert (benchmark._args.num_layers == 1)
         assert (benchmark._args.computation_kernel == ComputationKernelType.MATMUL)
         assert (benchmark._args.communication_kernel == CommunicationKernelType.ALLREDUCE)
@@ -45,6 +53,7 @@ def test_pytorch_dist_inference_normal():
         assert (benchmark._args.num_steps == 10000)
         assert (benchmark._args.distributed_impl == DistributedImpl.DDP)
         assert (benchmark._args.distributed_backend == DistributedBackend.NCCL)
+        assert (benchmark._args.use_cuda_graph is False)
 
         # Check results and metrics.
         assert (benchmark.run_count == 1)
@@ -52,14 +61,16 @@ def test_pytorch_dist_inference_normal():
         # step_times
         assert (len(benchmark.raw_data) == 1)
         # return code + (avg, 50th, 90th, 95th, 99th, 99.9th)
-        assert (len(benchmark.result) == 7)
+        assert (7 == len(benchmark.result))
 
 
 @decorator.cuda_test
 @decorator.pytorch_test
 def test_pytorch_dist_inference_fake_distributed():
     """Test pytorch-dist-inference benchmark on single gpu."""
-    context = BenchmarkRegistry.create_benchmark_context('dist-inference', parameters='', framework=Framework.PYTORCH)
+    context = BenchmarkRegistry.create_benchmark_context(
+        'dist-inference', parameters='--use_pytorch', framework=Framework.PYTORCH
+    )
     port = network.get_free_port()
     assert (port)
     utils.setup_simulated_ddp_distributed_env(1, 0, port)
@@ -72,9 +83,12 @@ def test_pytorch_dist_inference_fake_distributed():
     assert (benchmark.type == BenchmarkType.MICRO)
 
     # Check predefined parameters of dist-inference benchmark.
+    assert (benchmark._args.use_pytorch is True)
     assert (benchmark._args.batch_size == 64)
     assert (benchmark._args.input_size == 1024)
     assert (benchmark._args.hidden_size == 1024)
+    assert (benchmark._args.alpha == 1.0)
+    assert (benchmark._args.beta == 1.0)
     assert (benchmark._args.num_layers == 1)
     assert (benchmark._args.computation_kernel == ComputationKernelType.MATMUL)
     assert (benchmark._args.communication_kernel == CommunicationKernelType.ALLREDUCE)
@@ -84,6 +98,7 @@ def test_pytorch_dist_inference_fake_distributed():
     assert (benchmark._args.num_steps == 10000)
     assert (benchmark._args.distributed_impl == DistributedImpl.DDP)
     assert (benchmark._args.distributed_backend == DistributedBackend.NCCL)
+    assert (benchmark._args.use_cuda_graph is False)
 
     # Check results and metrics.
     assert (benchmark.run_count == 1)
@@ -94,3 +109,127 @@ def test_pytorch_dist_inference_fake_distributed():
     assert (len(benchmark.result) == 7)
 
     utils.clean_simulated_ddp_distributed_env()
+
+
+class DistInferenceCppImplTest(BenchmarkTestCase, unittest.TestCase):
+    """Test class for pytorch-dist-inference benchmark."""
+    @classmethod
+    def setUpClass(cls):
+        """Hook method for setting up class fixture before running tests in the class."""
+        super().setUpClass()
+        cls.createMockEnvs(cls)
+        cls.createMockFiles(cls, ['bin/dist_inference'])
+
+    def _test_dist_inference_command_generation(self, platform):
+        """Test pytorch-dist-inference cpp impl benchmark command generation."""
+        benchmark_name = 'pytorch-dist-inference'
+        (benchmark_class,
+         predefine_params) = BenchmarkRegistry._BenchmarkRegistry__select_benchmark(benchmark_name, platform)
+        assert (benchmark_class)
+
+        batch_size = 1
+        input_size = 2
+        hidden_size = 3
+        alpha = 4.0
+        beta = 5.0
+        num_layers = 6
+        num_warmup = 7
+        num_steps = 8
+        wrapper_params_format_str = \
+            '--batch_size %d --input_size %d --hidden_size %d ' \
+            '--alpha %g --beta %g --num_layers %d --num_warmup %d --num_steps %d --use_cuda_graph'
+        parameters = wrapper_params_format_str % (
+            batch_size, input_size, hidden_size, alpha, beta, num_layers, num_warmup, num_steps
+        )
+        benchmark = benchmark_class(benchmark_name, parameters=parameters)
+
+        # Check basic information
+        assert (benchmark)
+        ret = benchmark._preprocess()
+        assert (ret is True)
+        assert (benchmark.return_code == ReturnCode.SUCCESS)
+        assert (benchmark.name == benchmark_name)
+        assert (benchmark.type == BenchmarkType.MICRO)
+
+        # Check parameters specified in BenchmarkContext.
+        assert (benchmark._args.use_pytorch is False)
+        assert (benchmark._args.batch_size == batch_size)
+        assert (benchmark._args.input_size == input_size)
+        assert (benchmark._args.hidden_size == hidden_size)
+        assert (benchmark._args.alpha == alpha)
+        assert (benchmark._args.beta == beta)
+        assert (benchmark._args.num_layers == num_layers)
+        assert (benchmark._args.num_warmup == num_warmup)
+        assert (benchmark._args.num_steps == num_steps)
+        assert (benchmark._args.use_cuda_graph is True)
+
+        # Check command
+        assert (1 == len(benchmark._commands))
+        for cmd in benchmark._commands:
+            m, n, k = hidden_size, batch_size, input_size
+            bench_params_format_str = \
+                '%s -m %d -n %d -k %d --alpha %g --beta %g ' + \
+                '--num_layers %d --num_warmups %d --num_iters %d --use_cuda_graph'
+            assert (
+                cmd == (
+                    bench_params_format_str %
+                    (benchmark._DistInference__bin_path, m, n, k, alpha, beta, num_layers, num_warmup, num_steps)
+                )
+            )
+
+    @decorator.cuda_test
+    def test_dist_inference_command_generation_cuda(self):
+        """Test pytorch-dist-inference cpp impl benchmark command generation, CUDA case."""
+        self._test_dist_inference_command_generation(Platform.CUDA)
+
+    @decorator.rocm_test
+    def test_dist_inference_command_generation_rocm(self):
+        """Test pytorch-dist-inference cpp impl benchmark command generation, ROCm case."""
+        self._test_dist_inference_command_generation(Platform.ROCM)
+
+    @decorator.load_data('tests/data/dist_inference.log')
+    def _test_dist_inference_result_parsing(self, platform, test_raw_output):
+        """Test pytorch-dist-inference cpp impl benchmark result parsing."""
+        benchmark_name = 'pytorch-dist-inference'
+        (benchmark_class,
+         predefine_params) = BenchmarkRegistry._BenchmarkRegistry__select_benchmark(benchmark_name, platform)
+        assert (benchmark_class)
+        benchmark = benchmark_class(benchmark_name, parameters='')
+        assert (benchmark)
+        ret = benchmark._preprocess()
+        assert (ret is True)
+        assert (benchmark.return_code == ReturnCode.SUCCESS)
+        assert (benchmark.name == 'pytorch-dist-inference')
+        assert (benchmark.type == BenchmarkType.MICRO)
+
+        # Positive case - valid raw output.
+        assert (benchmark._process_raw_result(0, test_raw_output))
+        assert (benchmark.return_code == ReturnCode.SUCCESS)
+
+        # step_times
+        assert (len(benchmark.raw_data) == 2)
+        # return code + (avg, 50th, 90th, 95th, 99th, 99.9th)
+        test_latency = float(test_raw_output.splitlines()[-1].split(' ms per iteration')[0].split()[-1])
+        assert (7 == len(benchmark.result))
+        for output_key in benchmark.result:
+            if output_key == 'return_code':
+                assert (benchmark.result[output_key] == [0])
+            else:
+                assert (output_key.startswith('step_times'))
+                assert (len(benchmark.result[output_key]) == 1)
+                assert (isinstance(benchmark.result[output_key][0], numbers.Number))
+                assert (test_latency == benchmark.result[output_key][0])
+
+        # Negative case - invalid raw output.
+        assert (benchmark._process_raw_result(1, 'Invalid raw output') is False)
+        assert (benchmark.return_code == ReturnCode.MICROBENCHMARK_RESULT_PARSING_FAILURE)
+
+    @decorator.cuda_test
+    def test_dist_inference_result_parsing_cuda(self):
+        """Test pytorch-dist-inference cpp impl benchmark result parsing, CUDA case."""
+        self._test_dist_inference_result_parsing(Platform.CUDA)
+
+    @decorator.rocm_test
+    def test_dist_inference_result_parsing_rocm(self):
+        """Test pytorch-dist-inference cpp impl benchmark result parsing, ROCm case."""
+        self._test_dist_inference_result_parsing(Platform.ROCM)

tests/data/dist_inference.log

@@ -0,0 +1,2 @@
+Parameters: m=80, n=128, k=128, alpha=1.000000, beta=1.000000, num_layers=50, num_warmups=20, num_iters=100, use_cuda_graph=0
+Time: 173 ms in total, 1.73 ms per iteration, 0.0346 ms per layer