Skip to content

Commit caed709

Browse files
nicobossnicoboss
authored
Merge branch 'ggml-org:master' into mradermacher
2 parents 21a0f2c + 5fac79c commit caed709

File tree

7 files changed

+140
-32
lines changed

7 files changed

+140
-32
lines changed

‎common/arg.cpp‎

Lines changed: 42 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -1263,6 +1263,18 @@ static std::string list_builtin_chat_templates() {
12631263
return msg.str();
12641264
}
12651265

1266+
static bool is_truthy(const std::string & value) {
1267+
return value == "on" || value == "enabled" || value == "1";
1268+
}
1269+
1270+
static bool is_falsey(const std::string & value) {
1271+
return value == "off" || value == "disabled" || value == "0";
1272+
}
1273+
1274+
static bool is_autoy(const std::string & value) {
1275+
return value == "auto" || value == "-1";
1276+
}
1277+
12661278
common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
12671279
// load dynamic backends
12681280
ggml_backend_load_all();
@@ -1544,21 +1556,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
15441556
params.n_chunks = value;
15451557
}
15461558
).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_RETRIEVAL}));
1547-
add_opt(common_arg(
1548-
{"-fa", "--flash-attn"}, "FA",
1549-
string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')", llama_flash_attn_type_name(params.flash_attn_type)),
1550-
[](common_params & params, const std::string & value) {
1551-
if (value == "on" || value == "enabled" || value == "1") {
1552-
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
1553-
} else if (value == "off" || value == "disabled" || value == "0") {
1554-
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
1555-
} else if (value == "auto" || value == "-1") {
1556-
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
1557-
} else {
1558-
throw std::runtime_error(string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
1559-
}
1560-
}
1561-
).set_env("LLAMA_ARG_FLASH_ATTN"));
1559+
add_opt(common_arg({ "-fa", "--flash-attn" }, "[on|off|auto]",
1560+
string_format("set Flash Attention use ('on', 'off', or 'auto', default: '%s')",
1561+
llama_flash_attn_type_name(params.flash_attn_type)),
1562+
[](common_params & params, const std::string & value) {
1563+
if (is_truthy(value)) {
1564+
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_ENABLED;
1565+
} else if (is_falsey(value)) {
1566+
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_DISABLED;
1567+
} else if (is_autoy(value)) {
1568+
params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
1569+
} else {
1570+
throw std::runtime_error(
1571+
string_format("error: unkown value for --flash-attn: '%s'\n", value.c_str()));
1572+
}
1573+
}).set_env("LLAMA_ARG_FLASH_ATTN"));
15621574
add_opt(common_arg(
15631575
{"-p", "--prompt"}, "PROMPT",
15641576
"prompt to start generation with; for system message, use -sys",
@@ -3134,13 +3146,21 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
31343146
common_log_set_file(common_log_main(), value.c_str());
31353147
}
31363148
));
3137-
add_opt(common_arg(
3138-
{"--log-colors"},
3139-
"Enable colored logging",
3140-
[](common_params &) {
3141-
common_log_set_colors(common_log_main(), true);
3142-
}
3143-
).set_env("LLAMA_LOG_COLORS"));
3149+
add_opt(common_arg({ "--log-colors" }, "[on|off|auto]",
3150+
"Set colored logging ('on', 'off', or 'auto', default: 'auto')\n"
3151+
"'auto' enables colors when output is to a terminal",
3152+
[](common_params &, const std::string & value) {
3153+
if (is_truthy(value)) {
3154+
common_log_set_colors(common_log_main(), LOG_COLORS_ENABLED);
3155+
} else if (is_falsey(value)) {
3156+
common_log_set_colors(common_log_main(), LOG_COLORS_DISABLED);
3157+
} else if (is_autoy(value)) {
3158+
common_log_set_colors(common_log_main(), LOG_COLORS_AUTO);
3159+
} else {
3160+
throw std::invalid_argument(
3161+
string_format("error: unkown value for --log-colors: '%s'\n", value.c_str()));
3162+
}
3163+
}).set_env("LLAMA_LOG_COLORS"));
31443164
add_opt(common_arg(
31453165
{"-v", "--verbose", "--log-verbose"},
31463166
"Set verbosity level to infinity (i.e. log all messages, useful for debugging)",

‎common/chat.cpp‎

Lines changed: 13 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -163,6 +163,19 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
163163
throw std::runtime_error("Invalid tool_choice: " + tool_choice);
164164
}
165165

166+
bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates) {
167+
common_chat_templates_inputs dummy_inputs;
168+
common_chat_msg msg;
169+
msg.role = "user";
170+
msg.content = "test";
171+
dummy_inputs.messages = {msg};
172+
dummy_inputs.enable_thinking = false;
173+
const auto rendered_no_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
174+
dummy_inputs.enable_thinking = true;
175+
const auto rendered_with_thinking = common_chat_templates_apply(chat_templates, dummy_inputs);
176+
return rendered_no_thinking.prompt != rendered_with_thinking.prompt;
177+
}
178+
166179
template <>
167180
std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messages) {
168181
std::vector<common_chat_msg> msgs;

‎common/chat.h‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -199,6 +199,8 @@ common_chat_msg common_chat_parse(const std::string & input, bool is_p
199199

200200
common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::string & tool_choice);
201201

202+
bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
203+
202204
// Parses a JSON array of messages in OpenAI's chat completion API format.
203205
// T can be std::string containing JSON or nlohmann::ordered_json
204206
template <class T> std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const T & messages);

‎common/log.cpp‎

Lines changed: 53 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -4,17 +4,52 @@
44
#include <condition_variable>
55
#include <cstdarg>
66
#include <cstdio>
7+
#include <cstdlib>
8+
#include <cstring>
79
#include <mutex>
810
#include <sstream>
911
#include <thread>
1012
#include <vector>
1113

14+
#if defined(_WIN32)
15+
# include <io.h>
16+
# include <windows.h>
17+
# define isatty _isatty
18+
# define fileno _fileno
19+
#else
20+
# include <unistd.h>
21+
#endif // defined(_WIN32)
22+
1223
int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
1324

1425
void common_log_set_verbosity_thold(int verbosity) {
1526
common_log_verbosity_thold = verbosity;
1627
}
1728

29+
// Auto-detect if colors should be enabled based on terminal and environment
30+
static bool common_log_should_use_colors_auto() {
31+
// Check NO_COLOR environment variable (https://no-color.org/)
32+
if (const char * no_color = std::getenv("NO_COLOR")) {
33+
if (no_color[0] != '\0') {
34+
return false;
35+
}
36+
}
37+
38+
// Check TERM environment variable
39+
if (const char * term = std::getenv("TERM")) {
40+
if (std::strcmp(term, "dumb") == 0) {
41+
return false;
42+
}
43+
}
44+
45+
// Check if stdout and stderr are connected to a terminal
46+
// We check both because log messages can go to either
47+
bool stdout_is_tty = isatty(fileno(stdout));
48+
bool stderr_is_tty = isatty(fileno(stderr));
49+
50+
return stdout_is_tty || stderr_is_tty;
51+
}
52+
1853
static int64_t t_us() {
1954
return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
2055
}
@@ -353,6 +388,11 @@ struct common_log * common_log_init() {
353388

354389
struct common_log * common_log_main() {
355390
static struct common_log log;
391+
static std::once_flag init_flag;
392+
std::call_once(init_flag, [&]() {
393+
// Set default to auto-detect colors
394+
log.set_colors(common_log_should_use_colors_auto());
395+
});
356396

357397
return &log;
358398
}
@@ -380,8 +420,19 @@ void common_log_set_file(struct common_log * log, const char * file) {
380420
log->set_file(file);
381421
}
382422

383-
void common_log_set_colors(struct common_log * log, bool colors) {
384-
log->set_colors(colors);
423+
void common_log_set_colors(struct common_log * log, log_colors colors) {
424+
if (colors == LOG_COLORS_AUTO) {
425+
log->set_colors(common_log_should_use_colors_auto());
426+
return;
427+
}
428+
429+
if (colors == LOG_COLORS_DISABLED) {
430+
log->set_colors(false);
431+
return;
432+
}
433+
434+
GGML_ASSERT(colors == LOG_COLORS_ENABLED);
435+
log->set_colors(true);
385436
}
386437

387438
void common_log_set_prefix(struct common_log * log, bool prefix) {

‎common/log.h‎

Lines changed: 10 additions & 4 deletions
Original file line numberDiff line numberDiff line change
@@ -24,6 +24,12 @@
2424
#define LOG_DEFAULT_DEBUG 1
2525
#define LOG_DEFAULT_LLAMA 0
2626

27+
enum log_colors {
28+
LOG_COLORS_AUTO = -1,
29+
LOG_COLORS_DISABLED = 0,
30+
LOG_COLORS_ENABLED = 1,
31+
};
32+
2733
// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
2834
// set via common_log_set_verbosity()
2935
extern int common_log_verbosity_thold;
@@ -65,10 +71,10 @@ void common_log_add(struct common_log * log, enum ggml_log_level level, const ch
6571
// D - debug (stderr, V = LOG_DEFAULT_DEBUG)
6672
//
6773

68-
void common_log_set_file (struct common_log * log, const char * file); // not thread-safe
69-
void common_log_set_colors (struct common_log * log, bool colors); // not thread-safe
70-
void common_log_set_prefix (struct common_log * log, bool prefix); // whether to output prefix to each log
71-
void common_log_set_timestamps(struct common_log * log, bool timestamps); // whether to output timestamps in the prefix
74+
void common_log_set_file (struct common_log * log, const char * file); // not thread-safe
75+
void common_log_set_colors (struct common_log * log, log_colors colors); // not thread-safe
76+
void common_log_set_prefix (struct common_log * log, bool prefix); // whether to output prefix to each log
77+
void common_log_set_timestamps(struct common_log * log, bool timestamps); // whether to output timestamps in the prefix
7278

7379
// helper macros for logging
7480
// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold

‎tools/server/server.cpp‎

Lines changed: 7 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -2267,6 +2267,12 @@ struct server_context {
22672267

22682268
metrics.init();
22692269

2270+
// thinking is enabled if:
2271+
// 1. It's not explicitly disabled (reasoning_budget == 0)
2272+
// 2. The chat template supports it
2273+
const bool enable_thinking = params_base.reasoning_budget != 0 && common_chat_templates_support_enable_thinking(chat_templates.get());
2274+
SRV_INF("Enable thinking? %d\n", enable_thinking);
2275+
22702276
oai_parser_opt = {
22712277
/* use_jinja */ params_base.use_jinja,
22722278
/* prefill_assistant */ params_base.prefill_assistant,
@@ -2275,7 +2281,7 @@ struct server_context {
22752281
/* common_chat_templates */ chat_templates.get(),
22762282
/* allow_image */ mctx ? mtmd_support_vision(mctx) : false,
22772283
/* allow_audio */ mctx ? mtmd_support_audio (mctx) : false,
2278-
/* enable_thinking */ params_base.reasoning_budget != 0,
2284+
/* enable_thinking */ enable_thinking,
22792285
};
22802286
}
22812287

‎tools/server/utils.hpp‎

Lines changed: 13 additions & 3 deletions
Original file line numberDiff line numberDiff line change
@@ -54,8 +54,8 @@ static T json_value(const json & body, const std::string & key, const T & defaul
5454
if (body.contains(key) && !body.at(key).is_null()) {
5555
try {
5656
return body.at(key);
57-
} catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const &) {
58-
LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value\n", key.c_str(), json(default_value).type_name());
57+
} catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const & err) {
58+
LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value: %s\n", key.c_str(), json(default_value).type_name(), err.what());
5959
return default_value;
6060
}
6161
} else {
@@ -708,6 +708,16 @@ static json oaicompat_chat_params_parse(
708708
inputs.chat_template_kwargs[item.key()] = item.value().dump();
709709
}
710710

711+
// parse the "enable_thinking" kwarg to override the default value
712+
auto enable_thinking_kwarg = json_value(inputs.chat_template_kwargs, "enable_thinking", std::string(""));
713+
if (enable_thinking_kwarg == "true") {
714+
inputs.enable_thinking = true;
715+
} else if (enable_thinking_kwarg == "false") {
716+
inputs.enable_thinking = false;
717+
} else if (!enable_thinking_kwarg.empty() && enable_thinking_kwarg[0] == '"') {
718+
throw std::runtime_error("invalid type for \"enable_thinking\" (expected boolean, got string)");
719+
}
720+
711721
// if the assistant message appears at the end of list, we do not add end-of-turn token
712722
// for ex. this can be useful to modify the reasoning process in reasoning models
713723
bool prefill_assistant_message = !inputs.messages.empty() && inputs.messages.back().role == "assistant" && opt.prefill_assistant;
@@ -724,7 +734,7 @@ static json oaicompat_chat_params_parse(
724734
/* TODO: test this properly */
725735
inputs.reasoning_format = COMMON_REASONING_FORMAT_NONE;
726736

727-
if ( (!inputs.enable_thinking) || inputs.chat_template_kwargs.find("enable_thinking") != inputs.chat_template_kwargs.end()) {
737+
if ( inputs.enable_thinking ) {
728738
throw std::runtime_error("Assistant response prefill is incompatible with enable_thinking.");
729739
}
730740

0 commit comments

Comments
 (0)