feat(benchmark): Add CPU benchmark tool with context length sweep

tools/mllm-llm-benchmark/README.md

@@ -10,7 +10,6 @@ This is a benchmark tool for measuring MLLM model performance, including:
 ## Build
 
 Build from the mllm_v2 project root directory:
-
 ```bash
 mkdir -p build && cd build
 cmake ..
@@ -20,7 +19,6 @@ make mllm-llm-benchmark
 ## Usage
 
 ### Basic Usage
-
 ```bash
 ./mllm-llm-benchmark \
   -n qwen3-w4a32-kai \
@@ -32,6 +30,47 @@ make mllm-llm-benchmark
   -cl 2048
 ```
 
+### Context Sweep (New Feature)
+
+For automated benchmarking across different context lengths, use the sweep script:
+```bash
+cd tools/mllm-llm-benchmark
+chmod +x scripts/sweep_context_v2.sh
+
+# Configure paths
+export BIN=../../build/bin/mllm-llm-benchmark
+export MODEL=/path/to/your-model.mllm
+export CFG=/path/to/config.json
+
+# Run sweep
+./scripts/sweep_context_v2.sh
+```
+
+Output goes to `bench_context/context_sweep_v2.csv`. 
+
+**Configuration options:**
+- `BIN`: Path to benchmark binary (required)
+- `MODEL`: Path to model file (required)
+- `CFG`: Path to config json (default: `./examples/llama/config_tiny_llama.json`)
+- `THREADS`: Number of threads (default: 8)
+- `RUNS`: How many runs to average (default: 1)
+- `COOLDOWN`: Seconds to wait between runs (default: 0)
+- `CLS`: Context lengths to test (default: "256 512 1024 2048 4096")
+- `TG_DH`: Generate length for decode_heavy mode (default: 256)
+- `TG_TTFT`: Generate length for prefill_ttft mode (default: 2)
+- `OUTDIR`: Output directory (default: bench_context)
+
+**Test modes:**
+- `prefill_ttft`: Measures time to first token (prompt length = CL-2, generates 2 tokens)
+- `decode_heavy`: Measures decode throughput (prompt length = CL-256, generates 256 tokens)
+
+### Plot Results
+
+Visualize benchmark results:
+```bash
+python3 scripts/plot_sweep.py bench_context/context_sweep_v2.csv output_dir/
+```
+
 ### Parameters
 
 | Parameter | Long Format | Description | Example |
@@ -47,7 +86,6 @@ make mllm-llm-benchmark
 ### Examples
 
 #### Testing Qwen3-0.6B Model
-
 ```bash
 ./mllm-llm-benchmark \
   -n qwen3-w4a32-kai \
@@ -60,7 +98,6 @@ make mllm-llm-benchmark
 ```
 
 #### Quick Test (Single Configuration)
-
 ```bash
 ./mllm-llm-benchmark \
   -n qwen3-w4a32-kai \
@@ -73,7 +110,6 @@ make mllm-llm-benchmark
 ```
 
 ## Output Example
-
 ```
 MLLM Build Version : abc123def456
 ARCH               : ARM64
@@ -144,7 +180,6 @@ Each test configuration executes the following steps:
 ### 1. Create New Benchmark Class
 
 Create `YourModel_Benchmark.hpp` in the `models/` directory:
-
 ```cpp
 #include "BenchmarkTemplate.hpp"
 #include <mllm/models/yourmodel/modeling_yourmodel.hpp>
@@ -178,7 +213,6 @@ class YourModel_Benchmark final : public BenchmarkTemplate {
 ```
 
 ### 2. Register in All.hpp
-
 ```cpp
 #include "YourModel_Benchmark.hpp"
 

tools/mllm-llm-benchmark/main.cpp

@@ -1,10 +1,13 @@
 // Copyright (c) MLLM Team.
 // Licensed under the MIT License.
 
+#include <string>
+#include <fstream>
 #include <vector>
 #include <sstream>
 #include <thread>
 #include <chrono>
+#include <algorithm>
 
 #include <mllm/mllm.hpp>
 #include <mllm/utils/Argparse.hpp>
@@ -16,6 +19,9 @@
 
 #include "models/All.hpp"
 
+#define STR_HELPER(x) #x
+#define STR(x) STR_HELPER(x)
+
 MLLM_MAIN({
   auto& help = mllm::Argparse::add<bool>("-h|--help").help("Show help message");
   auto& model_name = mllm::Argparse::add<std::string>("-n|--model_name").help("Model name");
@@ -25,12 +31,21 @@ MLLM_MAIN({
   auto& pp = mllm::Argparse::add<std::string>("-pp|--prompt_length").help("Prompt length");
   auto& tg = mllm::Argparse::add<std::string>("-tg|--test_generation_length").help("Test Generation length");
   auto& cache_length = mllm::Argparse::add<int32_t>("-cl|--cache_length").help("Cache length");
+
+  auto& runs = mllm::Argparse::add<int32_t>("-r|--runs").help("Number of benchmark runs").def(3);
+  auto& cooldown_s = mllm::Argparse::add<int32_t>("-cs|--cooldown_s").help("Cooldown time between runs in seconds").def(5);
+  auto& output_csv = mllm::Argparse::add<std::string>("-oc|--output_csv").help("Output results to a CSV file").def("");
+  auto& schema_version = mllm::Argparse::add<int32_t>("-sv|--schema_version").help("Schema version for output format").def(1);
+  auto& kv_dtype_bytes =
+      mllm::Argparse::add<int32_t>("-kv|--kv_dtype_bytes").help("KV cache data type bytes (1: int8, 2: fp16, 4: fp32)").def(4);
+
   mllm::Argparse::parse(argc, argv);
 
-  // Print Build Version
+  mllm::Context::instance().setCpuOpThreads(num_threads.get());
+  mllm::setMaximumNumThreads((uint32_t)num_threads.get());
+
   mllm::print("MLLM Build Version :", STRINGIFY(MLLM_GIT_COMMIT_HASH));
 
-  // Print Device Info
   mllm::print("ARCH               :", mllm::cpu::CURRENT_ARCH_STRING);
   mllm::print("FP16               :", mllm::cpu::hasFP16());
   mllm::print("BF16               :", mllm::cpu::hasBF16());
@@ -53,12 +68,27 @@ MLLM_MAIN({
   mllm::print("AVX512VL           :", mllm::cpu::hasAVX512VL());
   mllm::print("FMA                :", mllm::cpu::hasFMA());
 
-  // Create benchmark
   mllm::print("Create Benchmark: ", model_name.get());
   auto benchmark = createBenchmark(model_name.get());
   MLLM_RT_ASSERT(benchmark != nullptr);
 
-  // Print Model Info
+  int R = runs.get();
+  if (R <= 0) {
+    mllm::print("[ERROR] --runs must be > 0, got:", R);
+    return 1;
+  }
+
+  std::ofstream csv_file;
+  if (!output_csv.get().empty()) {
+    csv_file.open(output_csv.get());
+    if (!csv_file.is_open()) {
+      mllm::print("[ERROR] Failed to open --output_csv:", output_csv.get());
+      return 1;
+    }
+    csv_file << "schema_version,git_commit,arch,model_name,pp,tg,ttft_ms,prefill_speed,decode_speed,prefill_ms,decode_ms_per_"
+                "tok,kv_est_bytes_pp,kv_est_bytes_final\n";
+  }
+
   mllm::print("Model Info");
   benchmark->init(config_path.get(), model_path.get(), cache_length.get());
   benchmark->printModelInfo();
@@ -92,7 +122,7 @@ MLLM_MAIN({
     for (size_t i = 0; i < pp_values.size(); ++i) { pp_tg_pairs.emplace_back(pp_values[i], tg_values[i]); }
   }
 
-  // Actual run for 3 turns and gives avg results. Each turn will sleep for 5 seconds to let the SoC or GPU/NPU cool down.
+  // Actual run for configurable number of turns
   mllm::print("\n========================================");
   mllm::print("Starting Benchmark Tests");
   mllm::print("========================================\n");
@@ -104,32 +134,33 @@ MLLM_MAIN({
     mllm::print("  Generation Length (TG):", tg);
     mllm::print("----------------------------------------");
 
-    // Storage for results
     std::vector<BenchmarkTemplateResult> results;
-    results.reserve(3);
+    results.reserve(static_cast<size_t>(R));
 
-    for (int i = 0; i < 3; ++i) {
-      mllm::print("  Run", i + 1, "of 3...");
+    for (int i = 0; i < R; ++i) {
+      mllm::print("  Run", i + 1, "of", R, "...");
 
-      // Clear cache before each run
       benchmark->clear();
-
-      // Run benchmark
       auto result = benchmark->run(pp, tg);
       results.push_back(result);
 
       mllm::print("    TTFT         :", result.ttft, "ms");
       mllm::print("    Prefill Speed:", result.prefill_speed, "tokens/s");
       mllm::print("    Decode Speed :", result.decode_speed, "tokens/s");
 
-      // Sleep for 5 seconds between runs to cool down
-      if (i < 2) {
-        mllm::print("    Cooling down for 5 seconds...");
-        std::this_thread::sleep_for(std::chrono::seconds(5));
+      float prefill_ms = (result.prefill_speed > 0.0f) ? (pp / result.prefill_speed) * 1000.0f : 0.0f;
+      float decode_ms_per_tok = (result.decode_speed > 0.0f) ? (1.0f / result.decode_speed) * 1000.0f : 0.0f;
+      mllm::print("    Prefill Latency   :", prefill_ms, "ms");
+      mllm::print("    Decode Latency    :", decode_ms_per_tok, "ms");
+
+      int cool = cooldown_s.get();
+      if (i + 1 < R && cool > 0) {
+        mllm::print("    Cooling down for", cool, "seconds...");
+        std::this_thread::sleep_for(std::chrono::seconds(cool));
       }
     }
 
-    // Calculate average results
+    float denom = (R > 0) ? static_cast<float>(R) : 1.0f;
     float avg_ttft = 0.0f;
     float avg_prefill_speed = 0.0f;
     float avg_decode_speed = 0.0f;
@@ -151,9 +182,35 @@ MLLM_MAIN({
     mllm::print("Average Prefill Speed:", avg_prefill_speed, "tokens/s");
     mllm::print("Average Decode Speed :", avg_decode_speed, "tokens/s");
     mllm::print("=====================================\n");
+
+    avg_ttft /= denom;
+    avg_prefill_speed /= denom;
+    avg_decode_speed /= denom;
+
+    float avg_prefill_ms = (avg_prefill_speed > 0.0f) ? (pp / avg_prefill_speed) * 1000.0f : 0.0f;
+    float avg_decode_ms_per_tok = (avg_decode_speed > 0.0f) ? (1.0f / avg_decode_speed) * 1000.0f : 0.0f;
+
+    // KV cache estimate
+    double kv_est_bytes_pp = 0.0;
+    double kv_est_bytes_final = 0.0;
+    if (auto info = benchmark->kvEstimateInfo(); info.has_value()) {
+      const int32_t bytes_per = kv_dtype_bytes.get();  // 1/2/4
+      // LLaMA-like KV: 2 * n_layers * n_kv_heads * head_dim * seq_len * bytes
+      kv_est_bytes_pp = 2.0 * info->num_layers * info->num_kv_heads * info->head_dim * (double)pp * bytes_per;
+      kv_est_bytes_final = 2.0 * info->num_layers * info->num_kv_heads * info->head_dim * (double)(pp + tg) * bytes_per;
+    }
+
+    std::stringstream ss;
+    ss << schema_version.get() << "," << STRINGIFY(MLLM_GIT_COMMIT_HASH) << "," << mllm::cpu::CURRENT_ARCH_STRING << ","
+       << model_name.get() << "," << pp << "," << tg << "," << avg_ttft << "," << avg_prefill_speed << "," << avg_decode_speed
+       << "," << avg_prefill_ms << "," << avg_decode_ms_per_tok << "," << kv_est_bytes_pp << "," << kv_est_bytes_final;
+
+    if (csv_file.is_open()) { csv_file << ss.str() << std::endl; }
   }
 
   mllm::print("\n========================================");
   mllm::print("Benchmark Tests Completed");
   mllm::print("========================================");
+
+  if (csv_file.is_open()) { csv_file.close(); }
 })

tools/mllm-llm-benchmark/models/All.hpp

@@ -4,20 +4,33 @@
 
 #include <memory>
 #include <algorithm>
+#include <string>
+#include <cctype>
 
-#include "Qwen3_W4A32_KAI.hpp"
 #include "BenchmarkTemplate.hpp"
+#include "Qwen3_W4A32_KAI.hpp"
+#include "Llama.hpp"
 
-std::shared_ptr<BenchmarkTemplate> createBenchmark(const std::string& model_name) {
+inline std::shared_ptr<BenchmarkTemplate> createBenchmark(const std::string& model_name) {
   auto tolower = [](const std::string& str) {
     std::string result = str;
-    std::transform(result.begin(), result.end(), result.begin(), ::tolower);
+    // unsigned char cast to avoid UB
+    std::transform(result.begin(), result.end(), result.begin(),
+                   [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
     return result;
   };
+
   auto normalized_model_name = tolower(model_name);
+
   if (normalized_model_name.find("qwen3") != std::string::npos && normalized_model_name.find("w4a32") != std::string::npos
       && normalized_model_name.find("kai") != std::string::npos) {
     return std::make_shared<Qwen3_W4A32_KAI_Benchmark>();
   }
+
+  if (normalized_model_name.find("llama") != std::string::npos || normalized_model_name.find("tinyllama") != std::string::npos
+      || normalized_model_name.find("tiny_llama") != std::string::npos) {
+    return std::make_shared<Llama_Benchmark>();
+  }
+
   return nullptr;
 }

tools/mllm-llm-benchmark/models/BenchmarkTemplate.hpp

@@ -3,19 +3,27 @@
 #pragma once
 
 #include <string>
+#include <optional>
+#include <cstdint>
 
 /**
  * @brief Benchmark result structure
  */
 struct BenchmarkTemplateResult {
-  float ttft;            ///< Time To First Token in milliseconds
-  float prefill_speed;   ///< Prefill phase speed in tokens/s
-  float decode_speed;    ///< Decode phase speed in tokens/s
+  float ttft;           ///< Time To First Token in milliseconds
+  float prefill_speed;  ///< Prefill phase speed in tokens/s
+  float decode_speed;   ///< Decode phase speed in tokens/s
+};
+
+struct KVCacheEstimateInfo {
+  int32_t num_layers = 0;
+  int32_t num_kv_heads = 0;
+  int32_t head_dim = 0;  // hidden_size / num_attention_heads
 };
 
 /**
  * @brief Base class for benchmark templates
- * 
+ *
  * All model benchmark implementations should inherit from this class and implement all virtual functions.
  */
 class BenchmarkTemplate {
@@ -32,21 +40,21 @@ class BenchmarkTemplate {
 
   /**
    * @brief Print model information
-   * 
+   *
    * Should output model key parameters such as number of layers, hidden size, attention heads, etc.
    */
   virtual void printModelInfo() = 0;
 
   /**
    * @brief Warmup run
-   * 
+   *
    * Run the model once with small-scale input to ensure the model enters a stable state.
    */
   virtual void warmup() = 0;
 
   /**
    * @brief Clear cache
-   * 
+   *
    * Clear KV cache and performance counters to prepare for the next test.
    */
   virtual void clear() = 0;
@@ -58,4 +66,7 @@ class BenchmarkTemplate {
    * @return Test results
    */
   virtual BenchmarkTemplateResult run(int32_t pp, int32_t tg) = 0;
+
+  // KV cache size estimation; return nullopt if unsupported
+  virtual std::optional<KVCacheEstimateInfo> kvEstimateInfo() const { return std::nullopt; }
 };