Skip to content

Commit

Permalink
cont : llama-cli + common [no ci]
Browse files Browse the repository at this point in the history
  • Loading branch information
@ggerganov
ggerganov committed Sep 11, 2024
1 parent c1845a9 commit d206f87
Show file tree
Hide file tree
Showing 10 changed files with 299 additions and 207 deletions.
3 changes: 0 additions & 3 deletions common/arg.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1950,8 +1950,6 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
else { std::invalid_argument("invalid value"); }
}
).set_examples({LLAMA_EXAMPLE_BENCH}));
#ifndef LOG_DISABLE_LOGS
// TODO: make this looks less weird
add_opt(llama_arg(
{"--log-disable"},
"Log disable",
Expand All @@ -1966,7 +1964,6 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
gpt_log_set_file(gpt_log_main(), value.c_str());
}
));
#endif // LOG_DISABLE_LOGS

return ctx_arg;
}
Expand Down
169 changes: 117 additions & 52 deletions common/common.cpp
View file

Large diffs are not rendered by default.

3 changes: 3 additions & 0 deletions common/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -374,6 +374,9 @@ static std::vector<T> string_split(const std::string & str, char delim) {
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);

std::string string_from_tokens(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
std::string string_from_batch (const struct llama_context * ctx, const struct llama_batch & batch);

//
// Filesystem utils
//
Expand Down
81 changes: 56 additions & 25 deletions common/log.cpp
View file
Original file line number Diff line number Diff line change
@@ -1,9 +1,11 @@
#include "log.h"

#include <thread>
#include <mutex>
#include <cstdio>
#include <condition_variable>
#include <cstdio>
#include <mutex>
#include <thread>

int gpt_log_verbosity_env = getenv("LLAMA_LOG") ? atoi(getenv("LLAMA_LOG")) : LOG_DEFAULT_LLAMA;

#define LOG_COLORS // TMP

Expand Down Expand Up @@ -36,46 +38,56 @@ static int64_t t_us() {
struct gpt_log_entry {
enum ggml_log_level level;

int verbosity;
int64_t timestamp;

std::vector<char> msg;

// signals the worker thread to stop
bool is_end;

void print(FILE * file) {
void print(FILE * file = nullptr) const {
FILE * fcur = file;
if (!fcur) {
// stderr displays DBG messages only when the verbosity is high
// these messages can still be logged to a file
if (level == GGML_LOG_LEVEL_DEBUG && gpt_log_verbosity_env < LOG_DEFAULT_DEBUG) {
return;
}

fcur = stdout;

if (level != GGML_LOG_LEVEL_NONE) {
fcur = stderr;
}
}

if (level != GGML_LOG_LEVEL_NONE) {
if (timestamp) {
// [M.s.ms.us]
fprintf(file, "[%04d.%02d.%03d.%03d] ",
fprintf(fcur, "" LOG_COL_BLUE "%05d.%02d.%03d.%03d" LOG_COL_DEFAULT " ",
(int) (timestamp / 1000000 / 60),
(int) (timestamp / 1000000 % 60),
(int) (timestamp / 1000 % 1000),
(int) (timestamp % 1000));
}

switch (level) {
case GGML_LOG_LEVEL_INFO:
fprintf(file, LOG_COL_GREEN "INF " LOG_COL_DEFAULT);
break;
case GGML_LOG_LEVEL_WARN:
fprintf(file, LOG_COL_MAGENTA "WRN " LOG_COL_DEFAULT);
break;
case GGML_LOG_LEVEL_ERROR:
fprintf(file, LOG_COL_RED "ERR " LOG_COL_DEFAULT);
break;
case GGML_LOG_LEVEL_DEBUG:
fprintf(file, LOG_COL_YELLOW "DBG " LOG_COL_DEFAULT);
break;
case GGML_LOG_LEVEL_INFO: fprintf(fcur, LOG_COL_GREEN "I " LOG_COL_DEFAULT); break;
case GGML_LOG_LEVEL_WARN: fprintf(fcur, LOG_COL_MAGENTA "W " ); break;
case GGML_LOG_LEVEL_ERROR: fprintf(fcur, LOG_COL_RED "E " ); break;
case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, LOG_COL_YELLOW "D " ); break;
default:
break;
}
}

fprintf(file, "%s", msg.data());
fprintf(fcur, "%s", msg.data());

fflush(file);
if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
fprintf(fcur, LOG_COL_DEFAULT);
}

fflush(fcur);
}
};

Expand Down Expand Up @@ -120,7 +132,7 @@ struct gpt_log {
gpt_log_entry cur;

public:
void add(enum ggml_log_level level, int verbosity, const char * fmt, va_list args) {
void add(enum ggml_log_level level, const char * fmt, va_list args) {
std::lock_guard<std::mutex> lock(mtx);

if (!running) {
Expand All @@ -130,15 +142,34 @@ struct gpt_log {
auto & entry = entries[tail];

{
#if 1
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
}
#else
// hack for bolding arguments
std::stringstream ss;
for (int i = 0; fmt[i] != 0; i++) {
if (fmt[i] == '%') {
ss << LOG_COL_BOLD;
while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
ss << LOG_COL_DEFAULT;
if (fmt[i] == 0) break;
}
ss << fmt[i];
}
const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
if (n >= entry.msg.size()) {
entry.msg.resize(n + 1);
vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
}
#endif
}

entry.level = level;
entry.verbosity = verbosity;
entry.timestamp = 0;
if (timestamps) {
entry.timestamp = t_us() - t_start;
Expand Down Expand Up @@ -192,7 +223,7 @@ struct gpt_log {
break;
}

cur.print(stdout);
cur.print(); // stdout and stderr

if (file) {
cur.print(file);
Expand Down Expand Up @@ -267,10 +298,10 @@ void gpt_log_free(struct gpt_log * log) {
delete log;
}

void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, int verbosity, const char * fmt, ...) {
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...) {
va_list args;
va_start(args, fmt);
log->add(level, verbosity, fmt, args);
log->add(level, fmt, args);
va_end(args);
}

Expand Down
28 changes: 13 additions & 15 deletions common/log.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,8 +2,6 @@

#include "ggml.h"

#include <cstdarg>

#ifndef __GNUC__
# define LOG_ATTRIBUTE_FORMAT(...)
#elif defined(__MINGW32__)
Expand All @@ -12,9 +10,11 @@
# define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
#endif

#ifndef LOG_VERBOSITY
#define LOG_VERBOSITY 10
#endif
#define LOG_DEFAULT_DEBUG 10
#define LOG_DEFAULT_LLAMA 5

// intialized in log.cpp from environment variable LLAMA_LOG
extern int gpt_log_verbosity_env;

struct gpt_log;

Expand All @@ -24,30 +24,28 @@ void gpt_log_pause (struct gpt_log * log);
void gpt_log_resume(struct gpt_log * log);
void gpt_log_free (struct gpt_log * log);

LOG_ATTRIBUTE_FORMAT(4, 5)
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, int verbosity, const char * fmt, ...);
LOG_ATTRIBUTE_FORMAT(3, 4)
void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...);

void gpt_log_set_file (struct gpt_log * log, const char * file); // not thread-safe
void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps);

#define LOG_TMPL(level, verbosity, ...) \
do { \
if ((verbosity) <= LOG_VERBOSITY) { \
gpt_log_add(gpt_log_main(), (level), (verbosity), __VA_ARGS__); \
if ((verbosity) <= gpt_log_verbosity_env) { \
gpt_log_add(gpt_log_main(), (level), __VA_ARGS__); \
} \
} while (0)

#define LOG(...) LOG_TMPL(GGML_LOG_LEVEL_NONE, 0, __VA_ARGS__)
#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)

#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, 0, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, 0, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, 0, __VA_ARGS__)
#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, 0, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, 0, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)

#define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO, verbosity, __VA_ARGS__)
#define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN, verbosity, __VA_ARGS__)
#define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
#define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)

#define LOG_TOKENS_TOSTR_PRETTY(...) std::string("dummy")
2 changes: 1 addition & 1 deletion common/sampling.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -325,7 +325,7 @@ llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
}

std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
std::string result = "\tlogits ";
std::string result = "logits ";

for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
Expand Down
Loading

0 comments on commit d206f87

Please sign in to comment.