Skip to content

Commit 233f7cd

Browse files
JohannesGaesslerJohannesGaessler
committed
WIP YAML logging
1 parent 800c963 commit 233f7cd

File tree

7 files changed

+461
-15
lines changed

7 files changed

+461
-15
lines changed

‎common/common.cpp

Lines changed: 265 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -1,15 +1,21 @@
11
#include "common.h"
2+
#include "build-info.h"
3+
#include "llama.h"
24

5+
#include <algorithm>
36
#include <cassert>
4-
#include <iostream>
7+
#include <cmath>
58
#include <cstring>
9+
#include <ctime>
610
#include <fstream>
7-
#include <string>
811
#include <iterator>
9-
#include <algorithm>
12+
#include <iostream>
13+
#include <regex>
1014
#include <sstream>
15+
#include <string>
16+
#include <sys/stat.h>
1117
#include <unordered_set>
12-
#include <regex>
18+
#include <vector>
1319

1420
#if defined(__APPLE__) && defined(__MACH__)
1521
#include <sys/types.h>
@@ -93,7 +99,6 @@ void process_escapes(std::string& input) {
9399

94100
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
95101
bool invalid_param = false;
96-
bool escape_prompt = false;
97102
std::string arg;
98103
gpt_params default_params;
99104
const std::string arg_prefix = "--";
@@ -125,8 +130,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
125130
break;
126131
}
127132
params.prompt = argv[i];
128-
} else if (arg == "-e") {
129-
escape_prompt = true;
133+
} else if (arg == "-e" || arg == "--escape") {
134+
params.escape = true;
130135
} else if (arg == "--prompt-cache") {
131136
if (++i >= argc) {
132137
invalid_param = true;
@@ -415,6 +420,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
415420
break;
416421
}
417422
params.antiprompt.push_back(argv[i]);
423+
} else if (arg == "-ld" || arg == "--logdir") {
424+
if (++i >= argc) {
425+
invalid_param = true;
426+
break;
427+
}
428+
params.logdir = argv[i];
429+
430+
if (params.logdir.back() != '/') {
431+
params.logdir += "/";
432+
}
418433
} else if (arg == "--perplexity") {
419434
params.perplexity = true;
420435
} else if (arg == "--hellaswag") {
@@ -508,7 +523,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
508523
exit(1);
509524
}
510525

511-
if (escape_prompt) {
526+
if (params.escape) {
512527
process_escapes(params.prompt);
513528
process_escapes(params.input_prefix);
514529
process_escapes(params.input_suffix);
@@ -534,7 +549,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
534549
fprintf(stdout, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
535550
fprintf(stdout, " -p PROMPT, --prompt PROMPT\n");
536551
fprintf(stdout, " prompt to start generation with (default: empty)\n");
537-
fprintf(stdout, " -e process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
552+
fprintf(stdout, " -e, --escape process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
538553
fprintf(stdout, " --prompt-cache FNAME file to cache prompt state for faster startup (default: none)\n");
539554
fprintf(stdout, " --prompt-cache-all if specified, saves user input and generations to cache as well.\n");
540555
fprintf(stdout, " not supported with --interactive or other interactive options\n");
@@ -613,6 +628,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
613628
fprintf(stdout, " --lora-base FNAME optional model to use as a base for the layers modified by the LoRA adapter\n");
614629
fprintf(stdout, " -m FNAME, --model FNAME\n");
615630
fprintf(stdout, " model path (default: %s)\n", params.model.c_str());
631+
fprintf(stdout, " -ld LOGDIR, --logdir LOGDIR\n");
632+
fprintf(stdout, " path under which to save YAML logs (no logging if unset)\n");
616633
fprintf(stdout, "\n");
617634
}
618635

@@ -764,3 +781,242 @@ std::string llama_token_to_str_bpe(const struct llama_context * ctx, llama_token
764781
return std::string(result.data(), result.size());
765782
}
766783

784+
// returns true if successful, false otherwise
785+
bool create_directory_with_parents(const std::string & path) {
786+
787+
// if the path already exists, check whether it's a directory
788+
struct stat info;
789+
if (stat(path.c_str(), &info) == 0) {
790+
return S_ISDIR(info.st_mode);
791+
}
792+
793+
size_t pos_slash = 1; // skip leading slashes for directory creation
794+
795+
// process path from front to back, procedurally creating directories
796+
while ((pos_slash = path.find('/', pos_slash)) != std::string::npos) {
797+
const std::string subpath = path.substr(0, pos_slash);
798+
struct stat info;
799+
800+
// if the path already exists, ensure that it's a directory
801+
if (stat(subpath.c_str(), &info) == 0) {
802+
if (!S_ISDIR(info.st_mode)) {
803+
return false;
804+
}
805+
} else {
806+
// create parent directories
807+
const int ret = mkdir(subpath.c_str(), 0755);
808+
if (ret != 0) {
809+
return false;
810+
}
811+
}
812+
813+
pos_slash += 1;
814+
}
815+
816+
return true;
817+
}
818+
819+
void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data) {
820+
fprintf(stream, "%s: [", prop_name);
821+
for (size_t i = 0; i < data.size() - 1; ++i) {
822+
fprintf(stream, "%e, ", data[i]);
823+
}
824+
fprintf(stream, "%e]\n", data.back());
825+
}
826+
827+
void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data) {
828+
fprintf(stream, "%s: [", prop_name);
829+
for (size_t i = 0; i < data.size() - 1; ++i) {
830+
fprintf(stream, "%d, ", data[i]);
831+
}
832+
fprintf(stream, "%d]\n", data.back());
833+
}
834+
835+
void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data,
836+
const bool remove_first) {
837+
std::string data_str(data == NULL ? "" : data);
838+
839+
if (data_str.empty()) {
840+
fprintf(stream, "%s:\n", prop_name);
841+
return;
842+
}
843+
844+
if (remove_first) {
845+
data_str = data_str.substr(1);
846+
}
847+
848+
size_t pos_start = 0;
849+
size_t pos_found = 0;
850+
851+
if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) {
852+
data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
853+
data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
854+
data_str = "\"" + data_str + "\"";
855+
fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
856+
return;
857+
}
858+
859+
if (data_str.find('\n') == std::string::npos) {
860+
fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
861+
return;
862+
}
863+
864+
fprintf(stream, "%s: |\n", prop_name);
865+
while ((pos_found = data_str.find('\n', pos_start)) != std::string::npos) {
866+
fprintf(stream, " %s\n", data_str.substr(pos_start, pos_found-pos_start).c_str());
867+
pos_start = pos_found + 1;
868+
}
869+
}
870+
871+
std::string get_sortable_timestamp() {
872+
using clock = std::chrono::system_clock;
873+
874+
const clock::time_point current_time = clock::now();
875+
const time_t as_time_t = clock::to_time_t(current_time);
876+
char timestamp_no_ns[100];
877+
std::strftime(timestamp_no_ns, 100, "%Y-%m-%dT%H:%M:%S", std::localtime(&as_time_t));
878+
879+
const int64_t ns = std::chrono::duration_cast<std::chrono::nanoseconds>(
880+
current_time.time_since_epoch() % 1000000000).count();
881+
char timestamp_ns[10];
882+
snprintf(timestamp_ns, 11, "%09ld", ns);
883+
884+
return std::string(timestamp_no_ns) + "." + std::string(timestamp_ns);
885+
}
886+
887+
void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const llama_context * lctx,
888+
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model) {
889+
fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
890+
fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
891+
fprintf(stream, "cpu_has_arm_fma: %s\n", ggml_cpu_has_arm_fma() ? "true" : "false");
892+
fprintf(stream, "cpu_has_avx: %s\n", ggml_cpu_has_avx() ? "true" : "false");
893+
fprintf(stream, "cpu_has_avx2: %s\n", ggml_cpu_has_avx2() ? "true" : "false");
894+
fprintf(stream, "cpu_has_avx512: %s\n", ggml_cpu_has_avx512() ? "true" : "false");
895+
fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
896+
fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
897+
fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
898+
fprintf(stream, "cpu_has_cublas: %s\n", ggml_cpu_has_cublas() ? "true" : "false");
899+
fprintf(stream, "cpu_has_clblast: %s\n", ggml_cpu_has_clblast() ? "true" : "false");
900+
fprintf(stream, "cpu_has_fma: %s\n", ggml_cpu_has_fma() ? "true" : "false");
901+
fprintf(stream, "cpu_has_gpublas: %s\n", ggml_cpu_has_gpublas() ? "true" : "false");
902+
fprintf(stream, "cpu_has_neon: %s\n", ggml_cpu_has_neon() ? "true" : "false");
903+
fprintf(stream, "cpu_has_f16c: %s\n", ggml_cpu_has_f16c() ? "true" : "false");
904+
fprintf(stream, "cpu_has_fp16_va: %s\n", ggml_cpu_has_fp16_va() ? "true" : "false");
905+
fprintf(stream, "cpu_has_wasm_simd: %s\n", ggml_cpu_has_wasm_simd() ? "true" : "false");
906+
fprintf(stream, "cpu_has_blas: %s\n", ggml_cpu_has_blas() ? "true" : "false");
907+
fprintf(stream, "cpu_has_sse3: %s\n", ggml_cpu_has_sse3() ? "true" : "false");
908+
fprintf(stream, "cpu_has_vsx: %s\n", ggml_cpu_has_vsx() ? "true" : "false");
909+
910+
#ifdef NDEBUG
911+
fprintf(stream, "debug: false\n");
912+
#else
913+
fprintf(stream, "debug: true\n");
914+
#endif // NDEBUG
915+
916+
fprintf(stream, "model: %s\n", model);
917+
918+
#ifdef __OPTIMIZE__
919+
fprintf(stream, "optimize: true\n");
920+
#else
921+
fprintf(stream, "optimize: false\n");
922+
#endif // __OPTIMIZE__
923+
924+
fprintf(stream, "time: %s\n", timestamp.c_str());
925+
926+
fprintf(stream, "\n");
927+
fprintf(stream, "###############\n");
928+
fprintf(stream, "# User Inputs #\n");
929+
fprintf(stream, "###############\n");
930+
fprintf(stream, "\n");
931+
932+
fprintf(stream, "batch_size: %d # default: 512\n", params.n_batch);
933+
dump_string_yaml_multiline(stream, "cfg_negative_prompt", params.cfg_negative_prompt.c_str(), true);
934+
fprintf(stream, "cfg_scale: %f # default: 1.0\n", params.cfg_scale);
935+
fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
936+
fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
937+
fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
938+
fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
939+
fprintf(stream, "export: %s # default: false\n", params.export_cgraph ? "true" : "false");
940+
fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
941+
fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", params.frequency_penalty);
942+
dump_string_yaml_multiline(stream, "grammar", params.grammar.c_str(), true);
943+
fprintf(stream, "grammar-file: # never logged, see grammar instead. Can still be specified for input.\n");
944+
fprintf(stream, "hellaswag: %s # default: false\n", params.hellaswag ? "true" : "false");
945+
fprintf(stream, "hellaswag_tasks: %ld # default: 400\n", params.hellaswag_tasks);
946+
947+
const auto logit_bias_eos = params.logit_bias.find(llama_token_eos(lctx));
948+
const bool ignore_eos = logit_bias_eos != params.logit_bias.end() && logit_bias_eos->second == -INFINITY;
949+
fprintf(stream, "ignore_eos: %s # default: false\n", ignore_eos ? "true" : "false");
950+
951+
fprintf(stream, "instruct: %s # default: false\n", params.instruct ? "true" : "false");
952+
fprintf(stream, "interactive: %s # default: false\n", params.interactive ? "true" : "false");
953+
fprintf(stream, "interactive_first: %s # default: false\n", params.interactive_first ? "true" : "false");
954+
dump_string_yaml_multiline(stream, "in_prefix", params.input_prefix.c_str(), false);
955+
fprintf(stream, "in_prefix_bos: %s # default: false\n", params.input_prefix_bos ? "true" : "false");
956+
dump_string_yaml_multiline(stream, "in_suffix", params.input_prefix.c_str(), false);
957+
fprintf(stream, "keep: %d # default: 0\n", params.n_keep);
958+
fprintf(stream, "logdir: %s # default: unset (no logging)\n", params.logdir.c_str());
959+
960+
fprintf(stream, "logit_bias:\n");
961+
for (std::pair<llama_token, float> lb : params.logit_bias) {
962+
if (ignore_eos && lb.first == logit_bias_eos->first) {
963+
continue;
964+
}
965+
fprintf(stream, " %d: %f", lb.first, lb.second);
966+
}
967+
968+
fprintf(stream, "lora: %s\n", params.lora_adapter.c_str());
969+
fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
970+
fprintf(stream, "low_vram: %s # default: false\n", params.low_vram ? "true" : "false");
971+
fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
972+
fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", params.mirostat);
973+
fprintf(stream, "mirostat_ent: %f # default: 5.0\n", params.mirostat_tau);
974+
fprintf(stream, "mirostat_lr: %f # default: 0.1\n", params.mirostat_eta);
975+
fprintf(stream, "mtest: %s # default: false\n", params.mem_test ? "true" : "false");
976+
fprintf(stream, "mul_mat_q: %s # default: false\n", params.mul_mat_q ? "true" : "false");
977+
fprintf(stream, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
978+
fprintf(stream, "mlock: %s # default: false\n", params.use_mlock ? "true" : "false");
979+
fprintf(stream, "model: %s # default: models/7B/ggml-model.bin\n", params.model.c_str());
980+
fprintf(stream, "model_alias: %s # default: unknown\n", params.model_alias.c_str());
981+
fprintf(stream, "multiline_input: %s # default: false\n", params.multiline_input ? "true" : "false");
982+
fprintf(stream, "n_gpu_layers: %d # default: 0\n", params.n_gpu_layers);
983+
fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
984+
fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
985+
fprintf(stream, "no_penalize_nl: %s # default: false\n", !params.penalize_nl ? "true" : "false");
986+
fprintf(stream, "numa: %s # default: false\n", params.numa ? "true" : "false");
987+
fprintf(stream, "presence_penalty: %f # default: 0.0\n", params.presence_penalty);
988+
dump_string_yaml_multiline(stream, "prompt", params.prompt.c_str(), true);
989+
fprintf(stream, "prompt_cache: %s\n", params.path_prompt_cache.c_str());
990+
fprintf(stream, "prompt_cache_all: %s # default: false\n", params.prompt_cache_all ? "true" : "false");
991+
fprintf(stream, "prompt_cache_ro: %s # default: false\n", params.prompt_cache_ro ? "true" : "false");
992+
dump_vector_int_yaml(stream, "prompt_tokens", prompt_tokens);
993+
fprintf(stream, "random_prompt: %s # default: false\n", params.random_prompt ? "true" : "false");
994+
fprintf(stream, "repeat_penalty: %f # default: 1.1\n", params.repeat_penalty);
995+
996+
fprintf(stream, "reverse_prompt:\n");
997+
for (std::string ap : params.antiprompt) {
998+
size_t pos = 0;
999+
while ((pos = ap.find('\n', pos)) != std::string::npos) {
1000+
ap.replace(pos, 1, "\\n");
1001+
pos += 1;
1002+
}
1003+
1004+
fprintf(stream, " - %s\n", ap.c_str());
1005+
}
1006+
1007+
fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
1008+
fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
1009+
fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
1010+
fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
1011+
1012+
const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + LLAMA_MAX_DEVICES);
1013+
dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
1014+
1015+
fprintf(stream, "temp: %f # default: 0.8\n", params.temp);
1016+
fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
1017+
fprintf(stream, "tfs: %f # default: 1.0\n", params.tfs_z);
1018+
fprintf(stream, "top_k: %d # default: 40\n", params.top_k);
1019+
fprintf(stream, "top_p: %f # default: 0.95\n", params.top_p);
1020+
fprintf(stream, "typical_p: %f # default: 1.0\n", params.typical_p);
1021+
fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
1022+
}

‎common/common.h

Lines changed: 12 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -60,6 +60,7 @@ struct gpt_params {
6060
std::string input_suffix = ""; // string to suffix user inputs with
6161
std::string grammar = ""; // optional BNF-like grammar to constrain sampling
6262
std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
63+
std::string logdir = ""; // directory in which to save YAML log files
6364

6465
std::string lora_adapter = ""; // lora adapter path
6566
std::string lora_base = ""; // base model path for the lora adapter
@@ -77,6 +78,7 @@ struct gpt_params {
7778
bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it
7879

7980
bool embedding = false; // get only sentence embedding
81+
bool escape = false; // escape "\n", "\r", "\t", "\'", "\"", and "\\"
8082
bool interactive_first = false; // wait for user input immediately
8183
bool multiline_input = false; // reverse the usage of `\`
8284
bool simple_io = false; // improves compatibility with subprocesses and limited consoles
@@ -128,3 +130,13 @@ std::string llama_token_to_str(
128130
std::string llama_token_to_str_bpe(
129131
const struct llama_context * ctx,
130132
llama_token token);
133+
134+
bool create_directory_with_parents(const std::string & path);
135+
void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
136+
void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
137+
void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data, bool remove_first);
138+
std::string get_sortable_timestamp();
139+
140+
void dump_non_result_info_yaml(
141+
FILE * stream, const gpt_params & params, const llama_context * lctx,
142+
const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model);

0 commit comments

Comments
 (0)