From e849648888a11de13aaaa4cb2eda3f5a9c7b444d Mon Sep 17 00:00:00 2001 From: slaren Date: Fri, 10 May 2024 18:03:54 +0200 Subject: [PATCH] llama-bench : add pp+tg test type (#7199) --- examples/llama-bench/README.md | 18 +++++--- examples/llama-bench/llama-bench.cpp | 63 ++++++++++++++++++++++++---- scripts/compare-llama-bench.py | 8 +++- 3 files changed, 74 insertions(+), 15 deletions(-) diff --git a/examples/llama-bench/README.md b/examples/llama-bench/README.md index 10f37b441..857840564 100644 --- a/examples/llama-bench/README.md +++ b/examples/llama-bench/README.md @@ -26,16 +26,21 @@ options: -m, --model (default: models/7B/ggml-model-q4_0.gguf) -p, --n-prompt (default: 512) -n, --n-gen (default: 128) - -b, --batch-size (default: 512) - -ctk , --cache-type-k (default: f16) - -ctv , --cache-type-v (default: f16) - -t, --threads (default: 112) + -pg (default: 512,128) + -b, --batch-size (default: 2048) + -ub, --ubatch-size (default: 512) + -ctk, --cache-type-k (default: f16) + -ctv, --cache-type-v (default: f16) + -t, --threads (default: 16) -ngl, --n-gpu-layers (default: 99) -sm, --split-mode (default: layer) -mg, --main-gpu (default: 0) -nkvo, --no-kv-offload <0|1> (default: 0) + -fa, --flash-attn <0|1> (default: 0) -mmp, --mmap <0|1> (default: 1) - -ts, --tensor_split (default: 0) + --numa (default: disabled) + -embd, --embeddings <0|1> (default: 0) + -ts, --tensor-split (default: 0) -r, --repetitions (default: 5) -o, --output (default: md) -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. ``` -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`) - 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`). diff --git a/examples/llama-bench/llama-bench.cpp b/examples/llama-bench/llama-bench.cpp index 40128ec44..8b965e199 100644 --- a/examples/llama-bench/llama-bench.cpp +++ b/examples/llama-bench/llama-bench.cpp @@ -161,10 +161,17 @@ static const char * split_mode_str(llama_split_mode mode) { } } +static std::string pair_str(const std::pair & p) { + static char buf[32]; + snprintf(buf, sizeof(buf), "%d,%d", p.first, p.second); + return buf; +} + struct cmd_params { std::vector model; std::vector n_prompt; std::vector n_gen; + std::vector> n_pg; std::vector n_batch; std::vector n_ubatch; std::vector type_k; @@ -188,6 +195,7 @@ static const cmd_params cmd_params_defaults = { /* model */ {"models/7B/ggml-model-q4_0.gguf"}, /* n_prompt */ {512}, /* n_gen */ {128}, + /* n_pg */ {{512, 128}}, /* n_batch */ {2048}, /* n_ubatch */ {512}, /* type_k */ {GGML_TYPE_F16}, @@ -215,10 +223,11 @@ static void print_usage(int /* argc */, char ** argv) { printf(" -m, --model (default: %s)\n", join(cmd_params_defaults.model, ",").c_str()); printf(" -p, --n-prompt (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str()); printf(" -n, --n-gen (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str()); + printf(" -pg (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str()); printf(" -b, --batch-size (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str()); - printf(" -ub N, --ubatch-size (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str()); - printf(" -ctk , --cache-type-k (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); - printf(" -ctv , --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); + printf(" -ub, --ubatch-size (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str()); + printf(" -ctk, --cache-type-k (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str()); + printf(" -ctv, --cache-type-v (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str()); printf(" -t, --threads (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str()); printf(" -ngl, --n-gpu-layers (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str()); printf(" -sm, --split-mode (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(argv[i], split_delim); 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(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") { if (++i >= argc) { 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.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_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_ubatch.empty()) { params.n_ubatch = cmd_params_defaults.n_ubatch; } if (params.type_k.empty()) { params.type_k = cmd_params_defaults.type_k; } @@ -632,6 +653,31 @@ static std::vector get_cmd_params_instances(const cmd_param }; 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; @@ -965,6 +1011,9 @@ struct markdown_printer : public printer { if (field == "n_gpu_layers") { return 3; } + if (field == "test") { + return 13; + } int width = std::max((int)field.length(), 10); @@ -1091,12 +1140,11 @@ struct markdown_printer : public printer { value = test::get_backend(); } else if (field == "test") { if (t.n_prompt > 0 && t.n_gen == 0) { - snprintf(buf, sizeof(buf), "pp %d", t.n_prompt); + snprintf(buf, sizeof(buf), "pp%d", t.n_prompt); } 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 { - assert(false); - exit(1); + snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen); } value = buf; } else if (field == "t/s") { @@ -1297,6 +1345,7 @@ int main(int argc, char ** argv) { llama_kv_cache_clear(ctx); uint64_t t_start = get_time_ns(); + if (t.n_prompt > 0) { test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads); } diff --git a/scripts/compare-llama-bench.py b/scripts/compare-llama-bench.py index fed3c1ee3..0ede9e67c 100755 --- a/scripts/compare-llama-bench.py +++ b/scripts/compare-llama-bench.py @@ -325,8 +325,12 @@ table = [] for row in rows_show: n_prompt = int(row[-4]) n_gen = int(row[-3]) - assert n_prompt == 0 or n_gen == 0 - test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}" + if n_prompt != 0 and n_gen == 0: + 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 # VVVVVVVVVVVVV VVVVVVVVV VVVVVVVVVVVVVV VVVVVVV table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])