mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2026-08-25 22:21:03 +02:00
chat: refactor handling supports_string_content / supports_typed_content (#27130)
* better supports_string_content cap detect * test: add "skip" * messages_inp_normalizer
This commit is contained in:
+72
-27
@@ -470,36 +470,80 @@ std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const json & messa
|
||||
return msgs;
|
||||
}
|
||||
|
||||
struct messages_inp_normalizer {
|
||||
const jinja::caps & caps;
|
||||
|
||||
messages_inp_normalizer(const jinja::caps & c) : caps(c) {}
|
||||
|
||||
// handle supports_string_content / supports_typed_content
|
||||
// if string=true and array=false, convert array to string
|
||||
// if string=false and array=true, convert string to array
|
||||
// if both are true, do nothing
|
||||
json normalize(const json & messages) {
|
||||
bool only_string = caps.supports_string_content && !caps.supports_typed_content;
|
||||
bool only_typed = !caps.supports_string_content && caps.supports_typed_content;
|
||||
if ((!only_string && !only_typed) || !messages.is_array()) {
|
||||
return messages;
|
||||
}
|
||||
json normalized = json::array();
|
||||
for (const auto & msg : messages) {
|
||||
json copy = msg;
|
||||
auto it = copy.find("content");
|
||||
if (it != copy.end()) {
|
||||
if (only_typed && it->is_string()) {
|
||||
*it = json::array({
|
||||
json{
|
||||
{"type", "text"},
|
||||
{"text", it->get<std::string>()},
|
||||
}
|
||||
});
|
||||
} else if (only_string && it->is_array()) {
|
||||
*it = concat_content_parts(*it);
|
||||
}
|
||||
}
|
||||
normalized.push_back(std::move(copy));
|
||||
}
|
||||
return normalized;
|
||||
}
|
||||
|
||||
// join parts with newline, do not add newline before or after media markers
|
||||
static std::string concat_content_parts(const json & parts) {
|
||||
std::string text;
|
||||
bool last_was_media_marker = false;
|
||||
for (const auto & part : parts) {
|
||||
std::string type = part.value("type", "");
|
||||
bool add_new_line = true;
|
||||
if (type == "text") {
|
||||
add_new_line = !last_was_media_marker && !text.empty();
|
||||
last_was_media_marker = false;
|
||||
} else if (type == "media_marker") {
|
||||
add_new_line = false;
|
||||
last_was_media_marker = true;
|
||||
} else {
|
||||
LOG_WRN("Ignoring content part type: %s\n", type.c_str());
|
||||
continue;
|
||||
}
|
||||
|
||||
if (add_new_line) {
|
||||
text += '\n';
|
||||
}
|
||||
|
||||
text += part.value("text", "");
|
||||
}
|
||||
return text;
|
||||
}
|
||||
};
|
||||
|
||||
static json render_message_to_json(const std::vector<common_chat_msg> & msgs, const jinja::caps & c) {
|
||||
if (!c.supports_string_content && !c.supports_typed_content) {
|
||||
LOG_WRN("%s: Neither string content nor typed content is supported by the template. This is unexpected and may lead to issues.\n", __func__);
|
||||
}
|
||||
|
||||
bool only_string_accepted = c.supports_string_content && !c.supports_typed_content;
|
||||
bool only_typed_accepted = !c.supports_string_content && c.supports_typed_content;
|
||||
|
||||
json messages = json::array();
|
||||
for (const auto & msg : msgs) {
|
||||
if (only_string_accepted) {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ true);
|
||||
messages.push_back(jmsg);
|
||||
} else if (only_typed_accepted) {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
|
||||
if (jmsg.at("content").is_string()) {
|
||||
jmsg["content"] = json::array({
|
||||
json{
|
||||
{"type", "text"},
|
||||
{"text", jmsg.at("content").get<std::string>()},
|
||||
}
|
||||
});
|
||||
}
|
||||
messages.push_back(jmsg);
|
||||
} else {
|
||||
json jmsg = msg.to_json_oaicompat(/* concat_typed_text= */ false);
|
||||
messages.push_back(jmsg);
|
||||
}
|
||||
messages.push_back(msg.to_json_oaicompat(/* concat_typed_text= */ false));
|
||||
}
|
||||
return messages;
|
||||
return messages_inp_normalizer(c).normalize(messages);
|
||||
}
|
||||
|
||||
// DEPRECATED: only used in tests
|
||||
@@ -892,8 +936,11 @@ static std::string common_chat_template_direct_apply_impl(
|
||||
const std::optional<json> & additional_context = std::nullopt) {
|
||||
jinja::context ctx(tmpl.source());
|
||||
|
||||
// messages_override is already built for this template, do not touch its content parts
|
||||
nlohmann::ordered_json inp = nlohmann::ordered_json{
|
||||
{"messages", messages_override.has_value() ? *messages_override : inputs.messages},
|
||||
{"messages", messages_override.has_value()
|
||||
? *messages_override
|
||||
: messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
|
||||
{"bos_token", tmpl.bos_token()},
|
||||
{"eos_token", tmpl.eos_token()},
|
||||
{"enable_thinking", inputs.enable_thinking},
|
||||
@@ -957,14 +1004,12 @@ static std::string common_chat_template_generation_prompt_impl(
|
||||
const std::optional<json> & tools_override = std::nullopt,
|
||||
const std::optional<json> & additional_context = std::nullopt) {
|
||||
|
||||
auto adjusted_messages = messages_override ? *messages_override : inputs.messages;
|
||||
|
||||
autoparser::generation_params params = inputs;
|
||||
params.add_generation_prompt = false;
|
||||
params.continue_final_message = COMMON_CHAT_CONTINUATION_NONE;
|
||||
std::string no_gen_prompt = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
|
||||
std::string no_gen_prompt = common_chat_template_direct_apply_impl(tmpl, params, messages_override, tools_override, additional_context);
|
||||
params.add_generation_prompt = true;
|
||||
std::string gen_prompt = common_chat_template_direct_apply_impl(tmpl, params, adjusted_messages, tools_override, additional_context);
|
||||
std::string gen_prompt = common_chat_template_direct_apply_impl(tmpl, params, messages_override, tools_override, additional_context);
|
||||
|
||||
size_t prefix_len = 0;
|
||||
size_t min_size = std::min(no_gen_prompt.size(), gen_prompt.size());
|
||||
|
||||
+10
-4
@@ -23,7 +23,7 @@ void caps_apply_preserve_reasoning(jinja::context & ctx, bool enabled) {
|
||||
ctx.set_val("preserve_thinking", mk_val<value_bool>(enabled));
|
||||
ctx.set_val("clear_thinking", mk_val<value_bool>(!enabled));
|
||||
ctx.set_val("truncate_history_thinking", mk_val<value_bool>(!enabled));
|
||||
ctx.set_val("drop_thinking", mk_val<value_bool>(!enabled));
|
||||
ctx.set_val("drop_thinking", mk_val<value_bool>(!enabled));
|
||||
}
|
||||
|
||||
void caps_apply_reasoning_effort(jinja::context & ctx, const std::string & effort) {
|
||||
@@ -117,6 +117,8 @@ caps caps_get(jinja::program & prog) {
|
||||
|
||||
JJ_DEBUG("%s\n", ">>> Running capability check: typed content");
|
||||
|
||||
static const std::string content_marker = "STRING_MARKER";
|
||||
|
||||
// case: typed content support
|
||||
caps_try_execute(
|
||||
prog,
|
||||
@@ -125,22 +127,26 @@ caps caps_get(jinja::program & prog) {
|
||||
return json::array({
|
||||
{
|
||||
{"role", "user"},
|
||||
{"content", "content"}
|
||||
{"content", content_marker}
|
||||
}
|
||||
});
|
||||
},
|
||||
nullptr, // ctx_fn
|
||||
nullptr, // tools_fn
|
||||
[&](context &, bool success, value & messages, value &, const std::string &) {
|
||||
[&](context &, bool success, value & messages, value &, const std::string & rendered) {
|
||||
auto & content = messages->at(0)->at("content");
|
||||
caps_print_stats(content, "messages[0].content");
|
||||
if (has_op(content, "selectattr") || has_op(content, "array_access")) {
|
||||
bool used_as_array = has_op(content, "selectattr") || has_op(content, "array_access");
|
||||
if (used_as_array) {
|
||||
// accessed as an array
|
||||
result.supports_typed_content = true;
|
||||
}
|
||||
if (!success) {
|
||||
// failed to execute with content as string
|
||||
result.supports_string_content = false;
|
||||
} else if (used_as_array && rendered.find(content_marker) == std::string::npos) {
|
||||
// edge case: string may be accessed for checking, but does not appear in the output
|
||||
result.supports_string_content = false;
|
||||
}
|
||||
}
|
||||
);
|
||||
|
||||
+49
-2
@@ -10,6 +10,7 @@
|
||||
#include "jinja/parser.h"
|
||||
#include "jinja/lexer.h"
|
||||
#include "jinja/utils.h"
|
||||
#include "jinja/caps.h"
|
||||
|
||||
#include "testing.h"
|
||||
|
||||
@@ -33,6 +34,7 @@ static void test_array_methods(testing & t);
|
||||
static void test_object_methods(testing & t);
|
||||
static void test_hasher(testing & t);
|
||||
static void test_stats(testing & t);
|
||||
static void test_caps(testing & t);
|
||||
static void test_string_parts(testing & t);
|
||||
static void test_fuzzing(testing & t);
|
||||
|
||||
@@ -73,6 +75,7 @@ int main(int argc, char *argv[]) {
|
||||
if (!g_python_mode) {
|
||||
t.test("hasher", test_hasher);
|
||||
t.test("stats", test_stats);
|
||||
t.test("caps", test_caps);
|
||||
t.test("string parts", test_string_parts);
|
||||
t.test("fuzzing", test_fuzzing);
|
||||
}
|
||||
@@ -2059,6 +2062,51 @@ static void test_stats(testing & t) {
|
||||
});
|
||||
}
|
||||
|
||||
static void test_caps(testing & t) {
|
||||
static auto get_caps = [](const std::string & tmpl) -> jinja::caps {
|
||||
jinja::lexer lexer;
|
||||
auto lexer_res = lexer.tokenize(tmpl);
|
||||
|
||||
jinja::program prog = jinja::parse_from_tokens(lexer_res);
|
||||
|
||||
return jinja::caps_get(prog);
|
||||
};
|
||||
|
||||
t.test("string content", [](testing & t) {
|
||||
auto caps = get_caps(
|
||||
"{% for message in messages %}"
|
||||
"{{ message['role'] + ': ' + message['content'] }}"
|
||||
"{% endfor %}"
|
||||
);
|
||||
t.assert_true("supports string content", caps.supports_string_content);
|
||||
t.assert_true("does not support typed content", !caps.supports_typed_content);
|
||||
});
|
||||
|
||||
t.test("typed content, raises on string", [](testing & t) {
|
||||
// 'selectattr' is not a String filter, so it throws
|
||||
auto caps = get_caps(
|
||||
"{% for message in messages %}"
|
||||
"{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}"
|
||||
"{{ content['text'] }}"
|
||||
"{% endfor %}"
|
||||
"{% endfor %}"
|
||||
);
|
||||
t.assert_true("does not support string content", !caps.supports_string_content);
|
||||
t.assert_true("supports typed content", caps.supports_typed_content);
|
||||
});
|
||||
|
||||
t.test("typed content, silently drops string", [](testing & t) {
|
||||
// no throw here, but content[0]['text'] is undefined for a string (MiniMax-M1 case)
|
||||
auto caps = get_caps(
|
||||
"{% for message in messages %}"
|
||||
"{{ message['content'][0]['text'] }}"
|
||||
"{% endfor %}"
|
||||
);
|
||||
t.assert_true("does not support string content", !caps.supports_string_content);
|
||||
t.assert_true("supports typed content", caps.supports_typed_content);
|
||||
});
|
||||
}
|
||||
|
||||
static void test_string_parts(testing & t) {
|
||||
static auto render = [](const std::string & tmpl, const json & vars) -> jinja::string {
|
||||
jinja::lexer lexer;
|
||||
@@ -2116,8 +2164,7 @@ static void test_template_cpp(testing & t, const std::string & name, const std::
|
||||
t.log("Actual : " + json(rendered).dump());
|
||||
}
|
||||
} catch (const jinja::not_implemented_exception & e) {
|
||||
// TODO @ngxson : remove this when the test framework supports skipping tests
|
||||
t.log("Skipped: " + std::string(e.what()));
|
||||
t.skip(e.what());
|
||||
}
|
||||
});
|
||||
}
|
||||
|
||||
+28
-3
@@ -21,6 +21,11 @@ struct testing {
|
||||
int failures = 0;
|
||||
int unnamed = 0;
|
||||
int exceptions = 0;
|
||||
int skipped = 0;
|
||||
|
||||
// set by skip(), read by the innermost test()
|
||||
bool skip_current = false;
|
||||
std::string skip_reason;
|
||||
|
||||
static constexpr std::size_t status_column = 80;
|
||||
|
||||
@@ -78,7 +83,12 @@ struct testing {
|
||||
}
|
||||
}
|
||||
|
||||
void print_result(const std::string &label, int new_failures, int new_assertions, const std::string &extra = "") const {
|
||||
void skip(const std::string &reason = "") {
|
||||
skip_current = true;
|
||||
skip_reason = reason;
|
||||
}
|
||||
|
||||
void print_result(const std::string &label, int new_failures, int new_assertions, const std::string &extra = "", bool was_skipped = false) const {
|
||||
std::string line = indent() + label;
|
||||
|
||||
std::string details;
|
||||
@@ -101,7 +111,7 @@ struct testing {
|
||||
line += " (" + details + ")";
|
||||
}
|
||||
|
||||
std::string status = (new_failures == 0) ? "[PASS]" : "[FAIL]";
|
||||
std::string status = new_failures != 0 ? "[FAIL]" : (was_skipped ? "[SKIP]" : "[PASS]");
|
||||
|
||||
if (line.size() + 1 < status_column) {
|
||||
line.append(status_column - line.size(), ' ');
|
||||
@@ -126,12 +136,26 @@ struct testing {
|
||||
int before_failures = failures;
|
||||
int before_assertions = assertions;
|
||||
|
||||
// do not let a skipped subtest also mark its parent as skipped
|
||||
bool outer_skip = skip_current;
|
||||
std::string outer_skip_reason = skip_reason;
|
||||
skip_current = false;
|
||||
skip_reason.clear();
|
||||
|
||||
run_with_exceptions([&] { f(*this); }, "test");
|
||||
|
||||
int new_failures = failures - before_failures;
|
||||
int new_assertions = assertions - before_assertions;
|
||||
|
||||
print_result(name, new_failures, new_assertions);
|
||||
bool was_skipped = skip_current && new_failures == 0;
|
||||
if (was_skipped) {
|
||||
++skipped;
|
||||
}
|
||||
|
||||
print_result(name, new_failures, new_assertions, was_skipped ? skip_reason : "", was_skipped);
|
||||
|
||||
skip_current = outer_skip;
|
||||
skip_reason = outer_skip_reason;
|
||||
|
||||
stack.pop_back();
|
||||
}
|
||||
@@ -238,6 +262,7 @@ struct testing {
|
||||
out << "assertions : " << assertions << "\n";
|
||||
out << "failures : " << failures << "\n";
|
||||
out << "exceptions : " << exceptions << "\n";
|
||||
out << "skipped : " << skipped << "\n";
|
||||
return failures == 0 ? 0 : 1;
|
||||
}
|
||||
};
|
||||
|
||||
Reference in New Issue
Block a user