vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

improve markdown formatting

Browse source
This commit is contained in:
slaren 2023-08-16 02:39:15 +02:00
parent 52b94f42c8
commit f2cf01ddd2
1 changed files with 169 additions and 69 deletions
Download patch file Download diff file Expand all files Collapse all files

188
examples/llama-bench/llama-bench.cpp
View file

@ -57,7 +57,7 @@ T avg(const std::vector<T>& v) {
} }
template<typename T> template<typename T>
T stddev(const std::vector<T>& v) { T stdev(const std::vector<T> & v) {
if (v.size() <= 1) { if (v.size() <= 1) {
return 0; return 0;
} }
@ -110,7 +110,7 @@ static void print_usage(int /* argc */, char ** argv) {
fprintf(stdout, "options:\n"); fprintf(stdout, "options:\n");
fprintf(stdout, " -h, --help\n"); fprintf(stdout, " -h, --help\n");
fprintf(stdout, " -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); fprintf(stdout, " -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
fprintf(stdout, " -np, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str()); fprintf(stdout, " -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
fprintf(stdout, " -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str()); fprintf(stdout, " -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
fprintf(stdout, " -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str()); fprintf(stdout, " -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
fprintf(stdout, " --memory-f32 <0|1> (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str()); fprintf(stdout, " --memory-f32 <0|1> (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
@ -155,42 +155,48 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
} }
auto p = split<std::string>(argv[i], split_delim); auto p = split<std::string>(argv[i], split_delim);
params.model.insert(params.model.end(), p.begin(), p.end()); params.model.insert(params.model.end(), p.begin(), p.end());
} else if (arg == "-np" || arg == "--n-prompt") { } else if (arg == "-p" || arg == "--n-prompt") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.n_prompt = split<int>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.n_prompt.insert(params.n_prompt.end(), p.begin(), p.end());
} else if (arg == "-n" || arg == "--n-gen") { } else if (arg == "-n" || arg == "--n-gen") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.n_gen = split<int>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
} else if (arg == "-b" || arg == "--batch-size") { } else if (arg == "-b" || arg == "--batch-size") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.n_batch = split<int>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
} else if (arg == "--memory-f32") { } else if (arg == "--memory-f32") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.f32_kv = split<bool>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
} else if (arg == "-t" || arg == "--threads") { } else if (arg == "-t" || arg == "--threads") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.n_threads = split<int>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
} else if (arg == "-ngl" || arg == "--n-gpu-layers") { } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.n_gpu_layers = split<int>(argv[i], split_delim); auto p = split<int>(argv[i], split_delim);
params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
} else if (arg == "-mg" || arg == "--main-gpu") { } else if (arg == "-mg" || arg == "--main-gpu") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
@ -202,21 +208,23 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.low_vram = split<bool>(argv[i], split_delim); auto p = split<bool>(argv[i], split_delim);
params.low_vram.insert(params.low_vram.end(), p.begin(), p.end());
} else if (arg == "-mmq" || arg == "--mul-mat-q") { } else if (arg == "-mmq" || arg == "--mul-mat-q") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
params.mul_mat_q = split<bool>(argv[i], split_delim); auto p = split<bool>(argv[i], split_delim);
params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
} else if (arg == "-ts" || arg == "--tensor-split") { } else if (arg == "-ts" || arg == "--tensor-split") {
if (++i >= argc) { if (++i >= argc) {
invalid_param = true; invalid_param = true;
break; break;
} }
for (auto ts : split<std::string>(argv[i], split_delim)) { for (auto ts : split<std::string>(argv[i], split_delim)) {
// split string by , and / // split string by ; and /
const std::regex regex{R"([,/]+)"}; const std::regex regex{R"([;/]+)"};
std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1}; std::sregex_token_iterator it{ts.begin(), ts.end(), regex, -1};
std::vector<std::string> split_arg{it, {}}; std::vector<std::string> split_arg{it, {}};
GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES); GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
@ -371,22 +379,23 @@ static bool ggml_cpu_has_metal() {
// backend params // backend params
struct backend_params { struct backend_params {
std::string build_commit = BUILD_COMMIT; static const std::string build_commit;
int build_number = BUILD_NUMBER; static const int build_number;
bool cuda = !!ggml_cpu_has_cublas(); static const bool cuda;
bool opencl = !!ggml_cpu_has_clblast(); static const bool opencl;
bool metal = !!ggml_cpu_has_metal(); static const bool metal;
bool gpu_blas = !!ggml_cpu_has_gpublas(); static const bool gpu_blas;
bool blas = !!ggml_cpu_has_blas(); static const bool blas;
int n_batch; int n_batch;
int n_threads; int n_threads;
bool f32_kv;
int n_gpu_layers; int n_gpu_layers;
int main_gpu; int main_gpu;
bool mul_mat_q; bool mul_mat_q;
bool low_vram; bool low_vram;
std::array<float, LLAMA_MAX_DEVICES> tensor_split; std::array<float, LLAMA_MAX_DEVICES> tensor_split;
std::string get_backend() const { static std::string get_backend() {
if (cuda) { if (cuda) {
return "CUDA"; return "CUDA";
} }
@ -406,7 +415,12 @@ struct backend_params {
} }
static const std::vector<std::string> & get_fields() { static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {"build_number", "build_commit", "cuda", "opencl", "metal", "gpu_blas", "blas", "n_batch", "n_threads", "n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split"}; static const std::vector<std::string> fields = {
"build_number", "build_commit",
"cuda", "opencl", "metal", "gpu_blas", "blas",
"n_batch", "n_threads", "f16_kv",
"n_gpu_layers", "main_gpu", "mul_mat_q", "low_vram", "tensor_split"
};
return fields; return fields;
} }
@ -427,27 +441,33 @@ struct backend_params {
std::vector<std::string> values = { std::vector<std::string> values = {
std::to_string(build_number), build_commit, std::to_string(build_number), build_commit,
std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas), std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
std::to_string(n_batch), std::to_string(n_threads), std::to_string(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), std::to_string(low_vram), tensor_split_str
}; };
return values; return values;
} }
}; };
const std::string backend_params::build_commit = BUILD_COMMIT;
const int backend_params::build_number = BUILD_NUMBER;
const bool backend_params::cuda = !!ggml_cpu_has_cublas();
const bool backend_params::opencl = !!ggml_cpu_has_clblast();
const bool backend_params::metal = !!ggml_cpu_has_metal();
const bool backend_params::gpu_blas = !!ggml_cpu_has_gpublas();
const bool backend_params::blas = !!ggml_cpu_has_blas();
// benchmark params // benchmark params
struct bench_params { struct bench_params {
int n_prompt = 512; int n_prompt ;
int n_gen = 128; int n_gen;
int n_batch = 512;
bool f32_kv = true;
static const std::vector<std::string> & get_fields() { static const std::vector<std::string> & get_fields() {
static const std::vector<std::string> fields = {"n_prompt", "n_gen", "n_batch", "f16_kv"}; static const std::vector<std::string> fields = {"n_prompt", "n_gen"};
return fields; return fields;
} }
std::vector<std::string> get_values() const { std::vector<std::string> get_values() const {
return {std::to_string(n_prompt), std::to_string(n_gen), std::to_string(n_batch), std::to_string(f32_kv)}; return {std::to_string(n_prompt), std::to_string(n_gen)};
} }
}; };
@ -459,8 +479,8 @@ struct timing_samples {
return ::avg(t_ns); return ::avg(t_ns);
} }
uint64_t stddev() const { uint64_t stdev() const {
return ::stddev(t_ns); return ::stdev(t_ns);
} }
std::vector<double> get_ts(int n) const { std::vector<double> get_ts(int n) const {
@ -474,7 +494,7 @@ struct timing_samples {
} }
double stddev_ts(uint64_t n) const { double stddev_ts(uint64_t n) const {
return ::stddev(get_ts(n)); return ::stdev(get_ts(n));
} }
static const std::vector<std::string> & get_fields() { static const std::vector<std::string> & get_fields() {
@ -497,10 +517,9 @@ struct test {
bparams.n_prompt = inst.n_prompt; bparams.n_prompt = inst.n_prompt;
bparams.n_gen = inst.n_gen; bparams.n_gen = inst.n_gen;
bparams.n_batch = inst.n_batch;
bparams.f32_kv = inst.f32_kv;
bkparams.n_batch = inst.n_batch; bkparams.n_batch = inst.n_batch;
bkparams.f32_kv = inst.f32_kv;
bkparams.n_threads = inst.n_threads; bkparams.n_threads = inst.n_threads;
bkparams.n_gpu_layers = inst.n_gpu_layers; bkparams.n_gpu_layers = inst.n_gpu_layers;
bkparams.main_gpu = inst.main_gpu; bkparams.main_gpu = inst.main_gpu;
@ -567,7 +586,7 @@ struct json_printer : public printer {
fprintf(fout, " \"samples\": {\n"); fprintf(fout, " \"samples\": {\n");
fprintf(fout, " \"ns\": [ %s ],\n", join(t.tsamples.t_ns, ", ").c_str()); fprintf(fout, " \"ns\": [ %s ],\n", join(t.tsamples.t_ns, ", ").c_str());
fprintf(fout, " \"avg\": %" PRIu64 ",\n", t.tsamples.avg()); fprintf(fout, " \"avg\": %" PRIu64 ",\n", t.tsamples.avg());
fprintf(fout, " \"stddev\": %" PRIu64 "\n", t.tsamples.stddev()); fprintf(fout, " \"stddev\": %" PRIu64 "\n", t.tsamples.stdev());
fprintf(fout, " }\n"); fprintf(fout, " }\n");
fprintf(fout, "}\n"); fprintf(fout, "}\n");
} }
@ -576,20 +595,102 @@ struct json_printer : public printer {
struct markdown_printer : public printer { struct markdown_printer : public printer {
std::vector<std::string> fields; std::vector<std::string> fields;
static int get_field_width(const std::string & field) {
if (field == "model") {
return -30;
}
if (field == "t/s") {
return 15;
}
int width = std::max((int)field.length(), 10);
if (field == "backend") {
return -width;
}
return width;
}
virtual void print_header(const cmd_params & params) { virtual void print_header(const cmd_params & params) {
// TODO: print all params that have multiple values fields = { "model", "backend" };
fprintf(fout, "| model | backend | n_gpu_layers | n_prompt | n_gen | t/s |\n"); if (backend_params::get_backend() != "CPU") {
fprintf(fout, "| ----- | ------- | ------------ | -------- | ----- | --- |\n"); fields.push_back("n_gpu_layers");
}
if (params.n_batch.size() > 1) {
fields.push_back("n_batch");
}
if (params.n_threads.size() > 1 || backend_params::get_backend() == "CPU") {
fields.push_back("n_threads");
}
if (params.f32_kv.size() > 1) {
fields.push_back("f32_kv");
}
if (params.main_gpu.size() > 1) {
fields.push_back("main_gpu");
}
if (params.mul_mat_q.size() > 1) {
fields.push_back("mul_mat_q");
}
if (params.low_vram.size() > 1) {
fields.push_back("low_vram");
}
if (params.n_prompt.size() > 1 || (params.n_prompt.size() == 1 && params.n_prompt.at(0) != 0)) {
fields.push_back("n_prompt");
}
if (params.n_gen.size() > 1 || (params.n_gen.size() == 1 && params.n_gen.at(0) != 0)) {
fields.push_back("n_gen");
}
fields.push_back("t/s");
fprintf(fout, "|");
for (const auto & field : fields) {
fprintf(fout, " %*s |", get_field_width(field), field.c_str());
}
fprintf(fout, "\n");
fprintf(fout, "|");
for (const auto & field: fields) {
int width = get_field_width(field);
fprintf(fout, " %s%s%s |", width < 0 ? ":" : "", std::string(std::abs(width) - 1, '-').c_str(), width > 0 ? ":" : "");
}
fprintf(fout, "\n");
(void) params; (void) params;
} }
virtual void print_test(const test & t) { virtual void print_test(const test & t) {
int n_tokens = t.bparams.n_prompt + t.bparams.n_gen; int n_tokens = t.bparams.n_prompt + t.bparams.n_gen;
fprintf(fout, "| %s | %s | %i | %i | %i | %.2f ± %.2f |\n",
t.mparams.type.c_str(), t.bkparams.get_backend().c_str(), t.bkparams.n_gpu_layers, std::map<std::string, std::string> vmap;
t.bparams.n_prompt, t.bparams.n_gen, std::transform(model_params::get_fields().begin(), model_params::get_fields().end(), t.mparams.get_values().begin(),
t.tsamples.avg_ts(n_tokens), std::inserter(vmap, vmap.end()), std::make_pair<const std::string&, const std::string&>);
t.tsamples.stddev_ts(n_tokens)); std::transform(bench_params::get_fields().begin(), bench_params::get_fields().end(), t.bparams.get_values().begin(),
std::inserter(vmap, vmap.end()), std::make_pair<const std::string&, const std::string&>);
std::transform(backend_params::get_fields().begin(), backend_params::get_fields().end(), t.bkparams.get_values().begin(),
std::inserter(vmap, vmap.end()), std::make_pair<const std::string&, const std::string&>);
fprintf(fout, "|");
for (const auto & field : fields) {
std::string value;
if (field == "model") {
value = t.mparams.type;
} else if (field == "backend") {
value = backend_params::get_backend();
} else if (field == "t/s") {
char buf[128];
snprintf(buf, sizeof(buf), "%.2f ± %.2f", t.tsamples.avg_ts(n_tokens), t.tsamples.stddev_ts(n_tokens));
value = buf;
} else if (vmap.find(field) != vmap.end()) {
value = vmap.at(field);
} else {
assert(false);
exit(1);
}
int width = get_field_width(field);
if (field == "t/s") {
// HACK: the utf-8 character is 2 bytes
width += 1;
}
fprintf(fout, " %*s |", width, value.c_str());
}
fprintf(fout, "\n");
} }
}; };
@ -680,7 +781,6 @@ int main(int argc, char ** argv) {
p->print_test(t); p->print_test(t);
llama_print_timings(ctx); llama_print_timings(ctx);
llama_free(ctx); llama_free(ctx);