diff --git a/.gitignore b/.gitignore index 420e0d6d0..d288e66fc 100644 --- a/.gitignore +++ b/.gitignore @@ -55,6 +55,7 @@ models-mnt /server /simple /batched +/batched-bench /export-lora /finetune /speculative diff --git a/Makefile b/Makefile index b8b0d4b56..39290ee3b 100644 --- a/Makefile +++ b/Makefile @@ -1,8 +1,14 @@ # Define the default target now so that it is always the first target -BUILD_TARGETS = main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml simple batched save-load-state server embd-input-test gguf llama-bench baby-llama beam-search speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o +BUILD_TARGETS = \ + main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \ + simple batched batched-bench save-load-state server embd-input-test gguf llama-bench baby-llama beam-search \ + speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o # Binaries only useful for tests -TEST_TARGETS = tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe +TEST_TARGETS = \ + tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \ + tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama \ + tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe # Code coverage output files COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report @@ -554,6 +560,9 @@ simple: examples/simple/simple.cpp build-info.h ggml. batched: examples/batched/batched.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) +batched-bench: examples/batched-bench/batched-bench.cpp build-info.h ggml.o llama.o common.o $(OBJS) + $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) + quantize: examples/quantize/quantize.cpp build-info.h ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) diff --git a/examples/batched-bench/README.md b/examples/batched-bench/README.md new file mode 100644 index 000000000..fa98bf24e --- /dev/null +++ b/examples/batched-bench/README.md @@ -0,0 +1,48 @@ +# llama.cpp/example/batched-bench + +Benchmark the batched decoding performance of `llama.cpp` + +## Usage + +There are 2 modes of operation: + +- `prompt not shared` - each batch has a separate prompt of size `PP` (i.e. `N_KV = B*(PP + TG)`) +- `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`) + +```bash +./batched-bench MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] + +# LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared +./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 0 99 + +# LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared +./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 1 99 +``` + +## Sample results + +- `PP` - prompt tokens per batch +- `TG` - generated tokens per batch +- `B` - number of batches +- `N_KV` - required KV cache size +- `T_PP` - prompt processing time (i.e. time to first token) +- `S_PP` - prompt processing speed (`(B*PP)/T_PP` or `PP/T_PP`) +- `T_TG` - time to generate all batches +- `S_TG` - text generation speed (`(B*TG)/T_TG`) +- `T` - total time +- `S` - total speed (i.e. all tokens / total time) + +| PP | TG | B | N_KV | T_PP s | S_PP t/s | T_TG s | S_TG t/s | T s | S t/s | +|-------|--------|------|--------|----------|----------|----------|----------|----------|----------| +| 128 | 128 | 1 | 256 | 0.108 | 1186.64 | 3.079 | 41.57 | 3.187 | 80.32 | +| 128 | 128 | 2 | 512 | 0.198 | 1295.19 | 5.029 | 50.90 | 5.227 | 97.95 | +| 128 | 128 | 4 | 1024 | 0.373 | 1373.96 | 6.878 | 74.44 | 7.251 | 141.23 | +| 128 | 128 | 8 | 2048 | 0.751 | 1363.27 | 7.344 | 139.43 | 8.095 | 252.99 | +| 128 | 128 | 16 | 4096 | 1.570 | 1304.68 | 8.455 | 242.23 | 10.024 | 408.60 | +| 128 | 128 | 32 | 8192 | 3.408 | 1201.73 | 8.801 | 465.40 | 12.209 | 670.96 | +| 128 | 256 | 1 | 384 | 0.107 | 1196.70 | 6.329 | 40.45 | 6.436 | 59.67 | +| 128 | 256 | 2 | 768 | 0.194 | 1317.45 | 10.239 | 50.00 | 10.433 | 73.61 | +| 128 | 256 | 4 | 1536 | 0.366 | 1399.03 | 13.960 | 73.35 | 14.326 | 107.22 | +| 128 | 256 | 8 | 3072 | 0.751 | 1363.92 | 15.110 | 135.54 | 15.861 | 193.69 | +| 128 | 256 | 16 | 6144 | 1.569 | 1304.93 | 18.073 | 226.64 | 19.642 | 312.80 | +| 128 | 256 | 32 | 12288 | 3.409 | 1201.35 | 19.223 | 426.15 | 22.633 | 542.93 | diff --git a/examples/batched-bench/batched-bench.cpp b/examples/batched-bench/batched-bench.cpp index 12b38f399..58b738cac 100644 --- a/examples/batched-bench/batched-bench.cpp +++ b/examples/batched-bench/batched-bench.cpp @@ -11,10 +11,11 @@ int main(int argc, char ** argv) { gpt_params params; if (argc == 1 || argv[1][0] == '-') { - printf("usage: %s MODEL_PATH [IS_PP_SHARED] [NGL]\n" , argv[0]); + printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL]\n" , argv[0]); return 1 ; } + int n_kv_max = 2048; int is_pp_shared = 0; int n_gpu_layers = 0; @@ -23,18 +24,20 @@ int main(int argc, char ** argv) { std::vector n_pl = { 1, 2, 4, 8, 16, 32, }; //std::vector n_pl = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 32, }; - const int32_t n_ctx_max = 16*1024; - if (argc >= 2) { params.model = argv[1]; } if (argc >= 3) { - is_pp_shared = std::atoi(argv[2]); + n_kv_max = std::atoi(argv[2]); } if (argc >= 4) { - n_gpu_layers = std::atoi(argv[3]); + is_pp_shared = std::atoi(argv[3]); + } + + if (argc >= 5) { + n_gpu_layers = std::atoi(argv[4]); } // init LLM @@ -56,8 +59,8 @@ int main(int argc, char ** argv) { llama_context_params ctx_params = llama_context_default_params(); - ctx_params.seed = 1234; - ctx_params.n_ctx = n_ctx_max; + ctx_params.seed = 1234; + ctx_params.n_ctx = n_kv_max; ctx_params.n_batch = 512; ctx_params.n_threads = params.n_threads; ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch; @@ -69,7 +72,7 @@ int main(int argc, char ** argv) { return 1; } - llama_batch batch = llama_batch_init(n_ctx_max, 0); + llama_batch batch = llama_batch_init(n_kv_max, 0); // decode in batches of ctx_params.n_batch tokens auto decode_helper = [](llama_context * ctx, llama_batch & batch, int32_t n_batch) { @@ -88,7 +91,7 @@ int main(int argc, char ** argv) { const int ret = llama_decode(ctx, batch_view); if (ret != 0) { - LOG_TEE("%s : failed to decode the batch, n_batch = %d, ret = %d\n", __func__, n_batch, ret); + LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret); return false; } } @@ -117,7 +120,7 @@ int main(int argc, char ** argv) { const int n_ctx_req = is_pp_shared ? pp + pl*tg : pl*(pp + tg); - if (n_ctx_req > n_ctx_max) { + if (n_ctx_req > n_kv_max) { continue; }