llama-bench : add pp+tg test type (#7199)

This commit is contained in:
slaren 2024-05-10 18:03:54 +02:00 committed by GitHub
parent 18e437665c
commit e849648888
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 74 additions and 15 deletions

View File

@ -26,16 +26,21 @@ options:
-m, --model <filename> (default: models/7B/ggml-model-q4_0.gguf) -m, --model <filename> (default: models/7B/ggml-model-q4_0.gguf)
-p, --n-prompt <n> (default: 512) -p, --n-prompt <n> (default: 512)
-n, --n-gen <n> (default: 128) -n, --n-gen <n> (default: 128)
-b, --batch-size <n> (default: 512) -pg <pp,tg> (default: 512,128)
-ctk <t>, --cache-type-k <t> (default: f16) -b, --batch-size <n> (default: 2048)
-ctv <t>, --cache-type-v <t> (default: f16) -ub, --ubatch-size <n> (default: 512)
-t, --threads <n> (default: 112) -ctk, --cache-type-k <t> (default: f16)
-ctv, --cache-type-v <t> (default: f16)
-t, --threads <n> (default: 16)
-ngl, --n-gpu-layers <n> (default: 99) -ngl, --n-gpu-layers <n> (default: 99)
-sm, --split-mode <none|layer|row> (default: layer) -sm, --split-mode <none|layer|row> (default: layer)
-mg, --main-gpu <i> (default: 0) -mg, --main-gpu <i> (default: 0)
-nkvo, --no-kv-offload <0|1> (default: 0) -nkvo, --no-kv-offload <0|1> (default: 0)
-fa, --flash-attn <0|1> (default: 0)
-mmp, --mmap <0|1> (default: 1) -mmp, --mmap <0|1> (default: 1)
-ts, --tensor_split <ts0/ts1/..> (default: 0) --numa <distribute|isolate|numactl> (default: disabled)
-embd, --embeddings <0|1> (default: 0)
-ts, --tensor-split <ts0/ts1/..> (default: 0)
-r, --repetitions <n> (default: 5) -r, --repetitions <n> (default: 5)
-o, --output <csv|json|md|sql> (default: md) -o, --output <csv|json|md|sql> (default: md)
-v, --verbose (default: 0) -v, --verbose (default: 0)
@ -43,10 +48,11 @@ options:
Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times. Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
``` ```
llama-bench can perform two types of tests: llama-bench can perform three types of tests:
- Prompt processing (pp): processing a prompt in batches (`-p`) - Prompt processing (pp): processing a prompt in batches (`-p`)
- Text generation (tg): generating a sequence of tokens (`-n`) - Text generation (tg): generating a sequence of tokens (`-n`)
- Prompt processing + text generation (pg): processing a prompt followed by generating a sequence of tokens (`-pg`)
With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`). With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).

View File

@ -161,10 +161,17 @@ static const char * split_mode_str(llama_split_mode mode) {
} }
} }
static std::string pair_str(const std::pair<int, int> & p) {
static char buf[32];
snprintf(buf, sizeof(buf), "%d,%d", p.first, p.second);
return buf;
}
struct cmd_params { struct cmd_params {
std::vector<std::string> model; std::vector<std::string> model;
std::vector<int> n_prompt; std::vector<int> n_prompt;
std::vector<int> n_gen; std::vector<int> n_gen;
std::vector<std::pair<int, int>> n_pg;
std::vector<int> n_batch; std::vector<int> n_batch;
std::vector<int> n_ubatch; std::vector<int> n_ubatch;
std::vector<ggml_type> type_k; std::vector<ggml_type> type_k;
@ -188,6 +195,7 @@ static const cmd_params cmd_params_defaults = {
/* model */ {"models/7B/ggml-model-q4_0.gguf"}, /* model */ {"models/7B/ggml-model-q4_0.gguf"},
/* n_prompt */ {512}, /* n_prompt */ {512},
/* n_gen */ {128}, /* n_gen */ {128},
/* n_pg */ {{512, 128}},
/* n_batch */ {2048}, /* n_batch */ {2048},
/* n_ubatch */ {512}, /* n_ubatch */ {512},
/* type_k */ {GGML_TYPE_F16}, /* type_k */ {GGML_TYPE_F16},
@ -215,10 +223,11 @@ static void print_usage(int /* argc */, char ** argv) {
printf(" -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); printf(" -m, --model <filename> (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
printf(" -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str()); printf(" -p, --n-prompt <n> (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
printf(" -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str()); printf(" -n, --n-gen <n> (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
printf(" -pg <pp,tg> (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
printf(" -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str()); printf(" -b, --batch-size <n> (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
printf(" -ub N, --ubatch-size <n> (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str()); printf(" -ub, --ubatch-size <n> (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
printf(" -ctk <t>, --cache-type-k <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); printf(" -ctk, --cache-type-k <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
printf(" -ctv <t>, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); printf(" -ctv, --cache-type-v <t> (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); printf(" -t, --threads <n> (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); printf(" -ngl, --n-gpu-layers <n> (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
printf(" -sm, --split-mode <none|layer|row> (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str()); printf(" -sm, --split-mode <none|layer|row> (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
@ -304,6 +313,17 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
} }
auto p = 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()); params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
} else if (arg == "-pg") {
if (++i >= argc) {
invalid_param = true;
break;
}
auto p = split<std::string>(argv[i], ',');
if (p.size() != 2) {
invalid_param = true;
break;
}
params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])});
} 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;
@ -493,6 +513,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
if (params.model.empty()) { params.model = cmd_params_defaults.model; } if (params.model.empty()) { params.model = cmd_params_defaults.model; }
if (params.n_prompt.empty()) { params.n_prompt = cmd_params_defaults.n_prompt; } if (params.n_prompt.empty()) { params.n_prompt = cmd_params_defaults.n_prompt; }
if (params.n_gen.empty()) { params.n_gen = cmd_params_defaults.n_gen; } if (params.n_gen.empty()) { params.n_gen = cmd_params_defaults.n_gen; }
if (params.n_pg.empty()) { params.n_pg = cmd_params_defaults.n_pg; }
if (params.n_batch.empty()) { params.n_batch = cmd_params_defaults.n_batch; } if (params.n_batch.empty()) { params.n_batch = cmd_params_defaults.n_batch; }
if (params.n_ubatch.empty()) { params.n_ubatch = cmd_params_defaults.n_ubatch; } if (params.n_ubatch.empty()) { params.n_ubatch = cmd_params_defaults.n_ubatch; }
if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; } if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; }
@ -632,6 +653,31 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
}; };
instances.push_back(instance); instances.push_back(instance);
} }
for (const auto & n_pg : params.n_pg) {
if (n_pg.first == 0 && n_pg.second == 0) {
continue;
}
cmd_params_instance instance = {
/* .model = */ m,
/* .n_prompt = */ n_pg.first,
/* .n_gen = */ n_pg.second,
/* .n_batch = */ nb,
/* .n_ubatch = */ nub,
/* .type_k = */ tk,
/* .type_v = */ tv,
/* .n_threads = */ nt,
/* .n_gpu_layers = */ nl,
/* .split_mode = */ sm,
/* .main_gpu = */ mg,
/* .no_kv_offload= */ nkvo,
/* .flash_attn = */ fa,
/* .tensor_split = */ ts,
/* .use_mmap = */ mmp,
/* .embeddings = */ embd,
};
instances.push_back(instance);
}
} }
return instances; return instances;
@ -965,6 +1011,9 @@ struct markdown_printer : public printer {
if (field == "n_gpu_layers") { if (field == "n_gpu_layers") {
return 3; return 3;
} }
if (field == "test") {
return 13;
}
int width = std::max((int)field.length(), 10); int width = std::max((int)field.length(), 10);
@ -1095,8 +1144,7 @@ struct markdown_printer : public printer {
} else if (t.n_gen > 0 && t.n_prompt == 0) { } else if (t.n_gen > 0 && t.n_prompt == 0) {
snprintf(buf, sizeof(buf), "tg%d", t.n_gen); snprintf(buf, sizeof(buf), "tg%d", t.n_gen);
} else { } else {
assert(false); snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
exit(1);
} }
value = buf; value = buf;
} else if (field == "t/s") { } else if (field == "t/s") {
@ -1297,6 +1345,7 @@ int main(int argc, char ** argv) {
llama_kv_cache_clear(ctx); llama_kv_cache_clear(ctx);
uint64_t t_start = get_time_ns(); uint64_t t_start = get_time_ns();
if (t.n_prompt > 0) { if (t.n_prompt > 0) {
test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads); test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
} }

View File

@ -325,8 +325,12 @@ table = []
for row in rows_show: for row in rows_show:
n_prompt = int(row[-4]) n_prompt = int(row[-4])
n_gen = int(row[-3]) n_gen = int(row[-3])
assert n_prompt == 0 or n_gen == 0 if n_prompt != 0 and n_gen == 0:
test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}" test_name = f"pp{n_prompt}"
elif n_prompt == 0 and n_gen != 0:
test_name = f"tg{n_gen}"
else:
test_name = f"pp{n_prompt}+tg{n_gen}"
# Regular columns test name avg t/s values Speedup # Regular columns test name avg t/s values Speedup
# VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV # VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV
table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])]) table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])