app : add batched-bench, fit-params, quantize & perplexity (#23459)

* app : add batched-bench, fit-params, quantize & perplexity

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* Add missing main.cpp

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* Add EOL

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

---------

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

app/CMakeLists.txt

@@ -3,7 +3,16 @@ set(TARGET llama-app)
 add_executable(${TARGET} llama.cpp)
 set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama)
 
-target_link_libraries(${TARGET} PRIVATE llama-server-impl llama-cli-impl llama-completion-impl llama-bench-impl)
+target_link_libraries(${TARGET} PRIVATE
+    llama-server-impl
+    llama-cli-impl
+    llama-completion-impl
+    llama-bench-impl
+    llama-batched-bench-impl
+    llama-fit-params-impl
+    llama-quantize-impl
+    llama-perplexity-impl
+)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_TOOLS_INSTALL)

app/llama.cpp

@@ -4,12 +4,18 @@
 #include <string>
 #include <vector>
 
+// visible
 int llama_server(int argc, char ** argv);
 int llama_cli(int argc, char ** argv);
 
 // hidden
 int llama_completion(int argc, char ** argv);
 int llama_bench(int argc, char ** argv);
+int llama_batched_bench(int argc, char ** argv);
+int llama_fit_params(int argc, char ** argv);
+int llama_quantize(int argc, char ** argv);
+int llama_perplexity(int argc, char ** argv);
+
 static int help(int argc, char ** argv);
 static int version(int argc, char ** argv);
 
@@ -22,12 +28,16 @@ struct command {
 };
 
 static const command cmds[] = {
-    {"serve",      "HTTP API server",                    {"server"},   false, llama_server     },
+    {"serve",         "HTTP API server",                                    {"server"},   false, llama_server       },
-    {"cli",        "Command-line interactive interface", {"client"},   false, llama_cli        },
+    {"cli",           "Command-line interactive interface",                 {"client"},   false, llama_cli          },
-    {"completion", "Text completion",                    {"complete"}, true,  llama_completion },
+    {"completion",    "Text completion",                                    {"complete"}, true,  llama_completion   },
-    {"bench",      "Benchmarking tool",                  {},           true,  llama_bench      },
+    {"bench",         "Benchmark prompt processing and text generation",    {},           true,  llama_bench        },
-    {"version",    "Show version",                       {},           true,  version          },
+    {"batched-bench", "Benchmark batched decoding performance",             {},           true,  llama_batched_bench},
-    {"help",       "Show available commands",            {},           true,  help             },
+    {"fit-params",    "Compute parameters to fit a model in device memory", {},           true,  llama_fit_params   },
+    {"quantize",      "Quantize a model",                                   {},           true,  llama_quantize     },
+    {"perplexity",    "Compute model perplexity and KL divergence",         {},           true,  llama_perplexity   },
+    {"version",       "Show version",                                       {},           true,  version            },
+    {"help",          "Show available commands",                            {},           true,  help               },
 };
 
 static int version(int argc, char ** argv) {

tools/batched-bench/CMakeLists.txt

@@ -1,6 +1,18 @@
+# llama-batched-bench-impl: batched-bench logic, reusable by app
+
+set(TARGET llama-batched-bench-impl)
+
+add_library(${TARGET} STATIC batched-bench.cpp)
+
+target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+target_link_libraries(${TARGET} PUBLIC llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+
+# llama-batched-bench executable
+
 set(TARGET llama-batched-bench)
-add_executable(${TARGET} batched-bench.cpp)
+
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+add_executable(${TARGET} main.cpp)
+target_link_libraries(${TARGET} PRIVATE llama-batched-bench-impl)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_TOOLS_INSTALL)

tools/batched-bench/batched-bench.cpp

@@ -15,7 +15,10 @@ static void print_usage(int, char ** argv) {
     LOG("\n");
 }
 
-int main(int argc, char ** argv) {
+// satisfies -Wmissing-declarations
+int llama_batched_bench(int argc, char ** argv);
+
+int llama_batched_bench(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
     common_params params;

tools/batched-bench/main.cpp

@@ -0,0 +1,5 @@
+int llama_batched_bench(int argc, char ** argv);
+
+int main(int argc, char ** argv) {
+    return llama_batched_bench(argc, argv);
+}

tools/fit-params/CMakeLists.txt

@@ -1,6 +1,18 @@
+# llama-fit-params-impl: fit-params logic, reusable by app
+
+set(TARGET llama-fit-params-impl)
+
+add_library(${TARGET} STATIC fit-params.cpp)
+
+target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+target_link_libraries(${TARGET} PUBLIC llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+
+# llama-fit-params executable
+
 set(TARGET llama-fit-params)
-add_executable(${TARGET} fit-params.cpp)
+
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+add_executable(${TARGET} main.cpp)
+target_link_libraries(${TARGET} PRIVATE llama-fit-params-impl)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_TOOLS_INSTALL)

tools/fit-params/fit-params.cpp

@@ -12,7 +12,10 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-int main(int argc, char ** argv) {
+// satisfies -Wmissing-declarations
+int llama_fit_params(int argc, char ** argv);
+
+int llama_fit_params(int argc, char ** argv) {
     common_params params;
 
     common_init();

tools/fit-params/main.cpp

@@ -0,0 +1,5 @@
+int llama_fit_params(int argc, char ** argv);
+
+int main(int argc, char ** argv) {
+    return llama_fit_params(argc, argv);
+}

tools/perplexity/CMakeLists.txt

@@ -1,6 +1,18 @@
+# llama-perplexity-impl: perplexity logic, reusable by app
+
+set(TARGET llama-perplexity-impl)
+
+add_library(${TARGET} STATIC perplexity.cpp)
+
+target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+target_link_libraries(${TARGET} PUBLIC llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+
+# llama-perplexity executable
+
 set(TARGET llama-perplexity)
-add_executable(${TARGET} perplexity.cpp)
+
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+add_executable(${TARGET} main.cpp)
+target_link_libraries(${TARGET} PRIVATE llama-perplexity-impl)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_TOOLS_INSTALL)

tools/perplexity/main.cpp

@@ -0,0 +1,5 @@
+int llama_perplexity(int argc, char ** argv);
+
+int main(int argc, char ** argv) {
+    return llama_perplexity(argc, argv);
+}

tools/perplexity/perplexity.cpp

@@ -2005,7 +2005,10 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
     LOG("Same top p: %6.3lf ± %5.3lf %%\n", 100.0*same_top_p, 100.0*sqrt(same_top_p*(1.0 - same_top_p)/(kld.count - 1)));
 }
 
-int main(int argc, char ** argv) {
+// satisfies -Wmissing-declarations
+int llama_perplexity(int argc, char ** argv);
+
+int llama_perplexity(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
 
     common_params params;

tools/quantize/CMakeLists.txt

@@ -1,7 +1,18 @@
+# llama-quantize-impl: quantize logic, reusable by app
+
+set(TARGET llama-quantize-impl)
+
+add_library(${TARGET} STATIC quantize.cpp)
+
+target_include_directories(${TARGET} PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
+target_link_libraries(${TARGET} PUBLIC llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+
+# llama-quantize executable
+
 set(TARGET llama-quantize)
-add_executable(${TARGET} quantize.cpp)
+
-target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
+add_executable(${TARGET} main.cpp)
-target_include_directories(${TARGET} PRIVATE ../../common)
+target_link_libraries(${TARGET} PRIVATE llama-quantize-impl)
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_TOOLS_INSTALL)

tools/quantize/main.cpp

@@ -0,0 +1,5 @@
+int llama_quantize(int argc, char ** argv);
+
+int main(int argc, char ** argv) {
+    return llama_quantize(argc, argv);
+}

tools/quantize/quantize.cpp

@@ -490,7 +490,10 @@ static bool parse_layer_prune(const char * data, std::vector<int> & prune_layers
     return true;
 }
 
-int main(int argc, char ** argv) {
+// satisfies -Wmissing-declarations
+int llama_quantize(int argc, char ** argv);
+
+int llama_quantize(int argc, char ** argv) {
     std::setlocale(LC_NUMERIC, "C");
     if (argc < 3) {
         usage(argv[0]);