[Benchmark] Add --use_random_tokens flag to C benchmark

.github/copilot-instructions.md

@@ -149,7 +149,7 @@ The repo supports specific model architectures - check `src/models/model_type.h`
 
 ### Performance Considerations
 
-- **KV caching** is automatically managed but can be configured via `runtime_settings.cpp`
+- **KV caching** is automatically managed but can be configured via `src/runtime_settings.cpp`
 - **Continuous decoding** (session continuation) requires careful state management
 - **Multi-LoRA** adapters use separate weight loading in `src/models/adapters.cpp`
 

benchmark/c/main.cpp

@@ -4,11 +4,14 @@
 #include <cmath>
 #include <algorithm>
 #include <chrono>
+#include <cstdint>
+#include <random>
 #include <iostream>
 #include <numeric>
 #include <stdexcept>
 #include <string>
 #include <string_view>
+#include <variant>
 #include <vector>
 
 #include "ort_genai.h"
@@ -158,7 +161,10 @@ void RunBenchmark(const benchmark::Options& opts) {
   bool need_generate_prompt = false;
   std::string prompt;
 
-  if (const size_t* npt = std::get_if<size_t>(&opts.prompt_num_tokens_or_content)) {
+  if (opts.use_random_tokens) {
+    num_prompt_tokens = std::get<size_t>(opts.prompt_num_tokens_or_content);
+    need_generate_prompt = false;
+  } else if (const size_t* npt = std::get_if<size_t>(&opts.prompt_num_tokens_or_content)) {
     num_prompt_tokens = *npt;
     need_generate_prompt = true;
   } else {
@@ -203,8 +209,21 @@ void RunBenchmark(const benchmark::Options& opts) {
   }
 
   auto prompt_sequences = OgaSequences::Create();
-  for (size_t i = 0; i < opts.batch_size; ++i) {
-    tokenizer->Encode(prompt.c_str(), *prompt_sequences);
+  if (opts.use_random_tokens) {
+    std::random_device rd;
+    std::mt19937 rng(rd());
+    std::uniform_int_distribution<int32_t> dist(0, 99);
+    for (size_t i = 0; i < opts.batch_size; ++i) {
+      std::vector<int32_t> random_tokens(num_prompt_tokens);
+      std::generate(random_tokens.begin(), random_tokens.end(), [&]() {
+        return dist(rng);
+      });
+      prompt_sequences->Append(random_tokens.data(), random_tokens.size());
+    }
+  } else {
+    for (size_t i = 0; i < opts.batch_size; ++i) {
+      tokenizer->Encode(prompt.c_str(), *prompt_sequences);
+    }
   }
 
   // warmup
@@ -226,7 +245,12 @@ void RunBenchmark(const benchmark::Options& opts) {
 
     if (opts.verbose && i == 0) {
       // show prompt and output on first iteration
-      std::cout << "[PROMPT BEGIN]" << prompt << "[PROMPT END]\n";
+      if (opts.use_random_tokens) {
+        std::cout << "[PROMPT] random token IDs in [0, 99], batch_size=" << opts.batch_size
+                  << ", tokens per sequence=" << num_prompt_tokens << "\n";
+      } else {
+        std::cout << "[PROMPT BEGIN]" << prompt << "[PROMPT END]\n";
+      }
       const auto output_sequence_length = gen->TokenCount();
       const auto* output_sequence_data = gen->GetSequenceData(0);
       const auto output = tokenizer->Decode(output_sequence_data, output_sequence_length);

benchmark/c/options.cpp

@@ -40,7 +40,12 @@ namespace {
     << "        Prompt text to use. Default: See --prompt_length.\n"
     << "      --prompt_file <file containing prompt text>\n"
     << "        Path to file containing prompt text to use. Default: See --prompt_length.\n"
-    << "      Note: --prompt_length, --prompt, and --prompt_file are mutually exclusive.\n"
+    << "      --use_random_tokens\n"
+    << "        Use random token IDs in [0, 99] per position instead of generating or encoding\n"
+    << "        a text prompt. Requires -l/--prompt_length (cannot be used with --prompt or\n"
+    << "        --prompt_file).\n"
+    << "      Note: --prompt, --prompt_file, and --use_random_tokens are mutually exclusive;\n"
+    << "        --use_random_tokens requires --prompt_length.\n"
     << "    -g,--generation_length <number>\n"
     << "      Number of tokens to generate. Default: " << defaults.num_tokens_to_generate << "\n"
     << "    -r,--repetitions <number>\n"
@@ -95,6 +100,19 @@ void VerifyOptions(const Options& opts) {
     throw std::runtime_error("ONNX model directory path must be provided.");
   }
 
+  const int prompt_source_count = opts.prompt_length_specified + opts.prompt_specified + opts.prompt_file_specified;
+  if (prompt_source_count > 1) {
+    throw std::runtime_error("--prompt_length, --prompt, and --prompt_file are mutually exclusive.");
+  }
+
+  if (opts.use_random_tokens && !opts.prompt_length_specified) {
+    throw std::runtime_error("--use_random_tokens requires -l/--prompt_length.");
+  }
+
+  if (opts.use_random_tokens && (opts.prompt_specified || opts.prompt_file_specified)) {
+    throw std::runtime_error("--use_random_tokens cannot be used with --prompt or --prompt_file.");
+  }
+
   // validate execution provider since it has a valid value
   ValidateExecutionProvider(opts.execution_provider);
 }
@@ -126,10 +144,13 @@ Options ParseOptionsFromCommandLine(int argc, const char* const* argv) {
         opts.batch_size = ParseNumber<size_t>(next_arg(i));
       } else if (arg == "-l" || arg == "--prompt_length") {
         prompt_num_tokens_or_content = ParseNumber<size_t>(next_arg(i));
+        opts.prompt_length_specified = true;
       } else if (arg == "--prompt") {
         prompt_num_tokens_or_content = std::string{next_arg(i)};
+        opts.prompt_specified = true;
       } else if (arg == "--prompt_file") {
         prompt_num_tokens_or_content = ReadFileContent(next_arg(i));
+        opts.prompt_file_specified = true;
       } else if (arg == "-g" || arg == "--generation_length") {
         opts.num_tokens_to_generate = ParseNumber<size_t>(next_arg(i));
       } else if (arg == "-r" || arg == "--repetitions") {
@@ -140,6 +161,8 @@ Options ParseOptionsFromCommandLine(int argc, const char* const* argv) {
         opts.max_length = ParseNumber<int64_t>(next_arg(i));
       } else if (arg == "--reuse_generator") {
         opts.reuse_generator = true;
+      } else if (arg == "--use_random_tokens") {
+        opts.use_random_tokens = true;
       } else if (arg == "-v" || arg == "--verbose") {
         opts.verbose = true;
       } else if (arg == "-h" || arg == "--help") {

benchmark/c/options.h

@@ -20,8 +20,12 @@ struct Options {
   size_t num_iterations{5};
   size_t num_warmup_iterations{1};
   int64_t max_length{0};
+  bool prompt_length_specified{};
+  bool prompt_specified{};
+  bool prompt_file_specified{};
   bool verbose{};
   bool reuse_generator{};
+  bool use_random_tokens{};
 };
 
 Options ParseOptionsFromCommandLine(int argc, const char* const* argv);