vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge branch 'ggerganov:master' into master

Browse source
This commit is contained in:
Giovanni Petrantoni 2025-02-10 11:16:22 +08:00 • committed by GitHub
commit 13763a2a9b
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
384 changed files with 54631 additions and 25978 deletions
Download patch file Download diff file Expand all files Collapse all files

32
common/CMakeLists.txt
View file

@ -56,14 +56,19 @@ add_library(${TARGET} STATIC
arg.cpp
arg.h
base64.hpp
chat.cpp
chat.hpp
chat-template.hpp
common.cpp
common.h
console.cpp
console.h
json-schema-to-grammar.cpp
json.hpp
llguidance.cpp
log.cpp
log.h
minja.hpp
ngram-cache.cpp
ngram-cache.h
sampling.cpp
@ -87,6 +92,33 @@ if (LLAMA_CURL)
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
endif ()
if (LLAMA_LLGUIDANCE)
include(ExternalProject)
set(LLGUIDANCE_SRC ${CMAKE_BINARY_DIR}/llguidance/source)
set(LLGUIDANCE_PATH ${LLGUIDANCE_SRC}/target/release)
ExternalProject_Add(llguidance_ext
GIT_REPOSITORY https://github.com/guidance-ai/llguidance
# v0.6.12:
GIT_TAG ced1c9023d47ec194fa977932d35ce65c2ebfc09
PREFIX ${CMAKE_BINARY_DIR}/llguidance
SOURCE_DIR ${LLGUIDANCE_SRC}
BUILD_IN_SOURCE TRUE
CONFIGURE_COMMAND ""
BUILD_COMMAND cargo build --release
INSTALL_COMMAND ""
BUILD_BYPRODUCTS ${LLGUIDANCE_PATH}/libllguidance.a ${LLGUIDANCE_PATH}/llguidance.h
UPDATE_COMMAND ""
)
target_compile_definitions(${TARGET} PUBLIC LLAMA_USE_LLGUIDANCE)
add_library(llguidance STATIC IMPORTED)
set_target_properties(llguidance PROPERTIES IMPORTED_LOCATION ${LLGUIDANCE_PATH}/libllguidance.a)
add_dependencies(llguidance llguidance_ext)
target_include_directories(${TARGET} PRIVATE ${LLGUIDANCE_PATH})
set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance)
endif ()
target_include_directories(${TARGET} PUBLIC .)
target_compile_features (${TARGET} PUBLIC cxx_std_17)
target_link_libraries (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)

226
common/arg.cpp
View file

@ -22,6 +22,11 @@ common_arg & common_arg::set_examples(std::initializer_list<enum llama_example>
return *this;
}
common_arg & common_arg::set_excludes(std::initializer_list<enum llama_example> excludes) {
this->excludes = std::move(excludes);
return *this;
}
common_arg & common_arg::set_env(const char * env) {
help = help + "\n(env: " + env + ")";
this->env = env;
@ -37,6 +42,10 @@ bool common_arg::in_example(enum llama_example ex) {
return examples.find(ex) != examples.end();
}
bool common_arg::is_exclude(enum llama_example ex) {
return excludes.find(ex) != excludes.end();
}
bool common_arg::get_value_from_env(std::string & output) {
if (env == nullptr) return false;
char * value = std::getenv(env);
@ -121,17 +130,27 @@ std::string common_arg::to_string() {
static void common_params_handle_model_default(
std::string & model,
std::string & model_url,
const std::string & model_url,
std::string & hf_repo,
std::string & hf_file) {
std::string & hf_file,
const std::string & hf_token,
const std::string & model_default) {
if (!hf_repo.empty()) {
// short-hand to avoid specifying --hf-file -> default it to --model
if (hf_file.empty()) {
if (model.empty()) {
throw std::invalid_argument("error: --hf-repo requires either --hf-file or --model\n");
auto auto_detected = common_get_hf_file(hf_repo, hf_token);
if (auto_detected.first.empty() || auto_detected.second.empty()) {
exit(1); // built without CURL, error message already printed
}
hf_repo = auto_detected.first;
hf_file = auto_detected.second;
} else {
hf_file = model;
}
hf_file = model;
} else if (model.empty()) {
}
// make sure model path is present (for caching purposes)
if (model.empty()) {
// this is to avoid different repo having same file name, or same file name in different subdirs
std::string filename = hf_repo + "_" + hf_file;
// to make sure we don't have any slashes in the filename
@ -145,7 +164,7 @@ static void common_params_handle_model_default(
model = fs_get_cache_file(string_split<std::string>(f, '/').back());
}
} else if (model.empty()) {
model = DEFAULT_MODEL_PATH;
model = model_default;
}
}
@ -281,8 +300,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
}
// TODO: refactor model params in a common struct
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file);
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file);
common_params_handle_model_default(params.model, params.model_url, params.hf_repo, params.hf_file, params.hf_token, DEFAULT_MODEL_PATH);
common_params_handle_model_default(params.speculative.model, params.speculative.model_url, params.speculative.hf_repo, params.speculative.hf_file, params.hf_token, "");
common_params_handle_model_default(params.vocoder.model, params.vocoder.model_url, params.vocoder.hf_repo, params.vocoder.hf_file, params.hf_token, "");
if (params.escape) {
string_process_escapes(params.prompt);
@ -305,6 +325,14 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
throw std::invalid_argument("error: either --embedding or --reranking can be specified, but not both");
}
if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
throw std::runtime_error(string_format(
"error: the supplied chat template is not supported: %s%s\n",
params.chat_template.c_str(),
params.use_jinja ? "" : "\nnote: llama.cpp was started without --jinja, we only support commonly used templates"
));
}
return true;
}
@ -358,6 +386,30 @@ static std::vector<ggml_backend_dev_t> parse_device_list(const std::string & val
return devices;
}
static void add_rpc_devices(std::string servers) {
auto rpc_servers = string_split<std::string>(servers, ',');
if (rpc_servers.empty()) {
throw std::invalid_argument("no RPC servers specified");
}
ggml_backend_reg_t rpc_reg = ggml_backend_reg_by_name("RPC");
if (!rpc_reg) {
throw std::invalid_argument("failed to find RPC backend");
}
typedef ggml_backend_dev_t (*ggml_backend_rpc_add_device_t)(const char * endpoint);
ggml_backend_rpc_add_device_t ggml_backend_rpc_add_device_fn = (ggml_backend_rpc_add_device_t) ggml_backend_reg_get_proc_address(rpc_reg, "ggml_backend_rpc_add_device");
if (!ggml_backend_rpc_add_device_fn) {
throw std::invalid_argument("failed to find RPC device add function");
}
for (const auto & server : rpc_servers) {
ggml_backend_dev_t dev = ggml_backend_rpc_add_device_fn(server.c_str());
if (dev) {
ggml_backend_device_register(dev);
} else {
throw std::invalid_argument("failed to register RPC device");
}
}
}
bool common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **)) {
auto ctx_arg = common_params_parser_init(params, ex, print_usage);
const common_params params_org = ctx_arg.params; // the example can modify the default params
@ -420,7 +472,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
* - if both {LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_*,} are set, we will prioritize the LLAMA_EXAMPLE_* matching current example
*/
auto add_opt = [&](common_arg arg) {
if (arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) {
if ((arg.in_example(ex) || arg.in_example(LLAMA_EXAMPLE_COMMON)) && !arg.is_exclude(ex)) {
ctx_arg.options.push_back(std::move(arg));
}
};
@ -649,7 +701,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) {
params.prompt = value;
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--no-perf"},
string_format("disable internal libllama performance timings (default: %s)", params.no_perf ? "true" : "false"),
@ -673,7 +725,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.prompt.pop_back();
}
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--in-file"}, "FNAME",
"an input file (repeat to specify multiple files)",
@ -700,7 +752,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.prompt = ss.str();
fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), value.c_str());
}
));
).set_excludes({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-e", "--escape"},
string_format("process escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\) (default: %s)", params.escape ? "true" : "false"),
@ -759,15 +811,19 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"-cnv", "--conversation"},
string_format(
"run in conversation mode:\n"
"- does not print special tokens and suffix/prefix\n"
"- interactive mode is also enabled\n"
"(default: %s)",
params.conversation ? "true" : "false"
),
"run in conversation mode:\n"
"- does not print special tokens and suffix/prefix\n"
"- interactive mode is also enabled\n"
"(default: auto enabled if chat template is available)",
[](common_params & params) {
params.conversation = true;
params.conversation_mode = COMMON_CONVERSATION_MODE_ENABLED;
}
).set_examples({LLAMA_EXAMPLE_MAIN}));
add_opt(common_arg(
{"-no-cnv", "--no-conversation"},
"force disable conversation mode (default: false)",
[](common_params & params) {
params.conversation_mode = COMMON_CONVERSATION_MODE_DISABLED;
}
).set_examples({LLAMA_EXAMPLE_MAIN}));
add_opt(common_arg(
@ -821,7 +877,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params) {
params.warmup = false;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}));
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_EMBEDDING}));
add_opt(common_arg(
{"--spm-infill"},
string_format(
@ -1363,7 +1419,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--rpc"}, "SERVERS",
"comma separated list of RPC servers",
[](common_params & params, const std::string & value) {
params.rpc_servers = value;
add_rpc_devices(value);
GGML_UNUSED(params);
}
).set_env("LLAMA_ARG_RPC"));
}
@ -1408,15 +1465,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--list-devices"},
"print list of available devices and exit",
[](common_params &) {
printf("Available devices:\n");
std::vector<ggml_backend_dev_t> rpc_devices;
std::vector<ggml_backend_dev_t> all_devices;
for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
auto * dev = ggml_backend_dev_get(i);
if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_GPU) {
size_t free, total;
ggml_backend_dev_memory(dev, &free, &total);
printf(" %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev);
if (ggml_backend_reg_name(reg) == std::string("RPC")) {
rpc_devices.push_back(dev);
} else {
all_devices.push_back(dev);
}
}
}
// insert RPC devices in front
all_devices.insert(all_devices.begin(), rpc_devices.begin(), rpc_devices.end());
printf("Available devices:\n");
for (size_t i = 0; i < all_devices.size(); ++i) {
auto * dev = all_devices[i];
size_t free, total;
ggml_backend_dev_memory(dev, &free, &total);
printf(" %s: %s (%zu MiB, %zu MiB free)\n", ggml_backend_dev_name(dev), ggml_backend_dev_description(dev), total / 1024 / 1024, free / 1024 / 1024);
}
exit(0);
}
));
@ -1512,7 +1582,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--lora"}, "FNAME",
"path to LoRA adapter (can be repeated to use multiple adapters)",
[](common_params & params, const std::string & value) {
params.lora_adapters.push_back({ std::string(value), 1.0 });
params.lora_adapters.push_back({ std::string(value), 1.0, nullptr });
}
// we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@ -1520,7 +1590,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
{"--lora-scaled"}, "FNAME", "SCALE",
"path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters)",
[](common_params & params, const std::string & fname, const std::string & scale) {
params.lora_adapters.push_back({ fname, std::stof(scale) });
params.lora_adapters.push_back({ fname, std::stof(scale), nullptr });
}
// we define this arg on both COMMON and EXPORT_LORA, so when showing help message of export-lora, it will be categorized as "example-specific" arg
).set_examples({LLAMA_EXAMPLE_COMMON, LLAMA_EXAMPLE_EXPORT_LORA}));
@ -1574,21 +1644,30 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_env("LLAMA_ARG_MODEL_URL"));
add_opt(common_arg(
{"-hfr", "--hf-repo"}, "REPO",
"Hugging Face model repository (default: unused)",
{"-hf", "-hfr", "--hf-repo"}, "<user>/<model>[:quant]",
"Hugging Face model repository; quant is optional, case-insensitive, default to Q4_K_M, or falls back to the first file in the repo if Q4_K_M doesn't exist.\n"
"example: unsloth/phi-4-GGUF:q4_k_m\n"
"(default: unused)",
[](common_params & params, const std::string & value) {
params.hf_repo = value;
}
).set_env("LLAMA_ARG_HF_REPO"));
add_opt(common_arg(
{"-hfd", "-hfrd", "--hf-repo-draft"}, "<user>/<model>[:quant]",
"Same as --hf-repo, but for the draft model (default: unused)",
[](common_params & params, const std::string & value) {
params.speculative.hf_repo = value;
}
).set_env("LLAMA_ARG_HFD_REPO"));
add_opt(common_arg(
{"-hff", "--hf-file"}, "FILE",
"Hugging Face model file (default: unused)",
"Hugging Face model file. If specified, it will override the quant in --hf-repo (default: unused)",
[](common_params & params, const std::string & value) {
params.hf_file = value;
}
).set_env("LLAMA_ARG_HF_FILE"));
add_opt(common_arg(
{"-hfrv", "--hf-repo-v"}, "REPO",
{"-hfv", "-hfrv", "--hf-repo-v"}, "<user>/<model>[:quant]",
"Hugging Face model repository for the vocoder model (default: unused)",
[](common_params & params, const std::string & value) {
params.vocoder.hf_repo = value;
@ -1889,24 +1968,44 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
}
).set_examples({LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--jinja"},
"use jinja template for chat (default: disabled)",
[](common_params & params) {
params.use_jinja = true;
}
).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MAIN}).set_env("LLAMA_ARG_JINJA"));
add_opt(common_arg(
{"--chat-template"}, "JINJA_TEMPLATE",
string_format(
"set custom jinja chat template (default: template taken from model's metadata)\n"
"if suffix/prefix are specified, template will be disabled\n"
"only commonly used templates are accepted (unless --jinja is set before this flag):\n"
"list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
),
[](common_params & params, const std::string & value) {
if (!common_chat_verify_template(value)) {
throw std::runtime_error(string_format(
"error: the supplied chat template is not supported: %s\n"
"note: llama.cpp does not use jinja parser, we only support commonly used templates\n",
value.c_str()
));
}
params.chat_template = value;
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
add_opt(common_arg(
{"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
string_format(
"set custom jinja chat template file (default: template taken from model's metadata)\n"
"if suffix/prefix are specified, template will be disabled\n"
"only commonly used templates are accepted (unless --jinja is set before this flag):\n"
"list of built-in templates:\n%s", list_builtin_chat_templates().c_str()
),
[](common_params & params, const std::string & value) {
std::ifstream file(value);
if (!file) {
throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
}
std::copy(
std::istreambuf_iterator<char>(file),
std::istreambuf_iterator<char>(),
std::back_inserter(params.chat_template));
}
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE_FILE"));
add_opt(common_arg(
{"-sps", "--slot-prompt-similarity"}, "SIMILARITY",
string_format("how much the prompt of a request must match the prompt of a slot in order to use that slot (default: %.2f, 0.0 = disabled)\n", params.slot_prompt_similarity),
@ -2205,6 +2304,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.vocoder.model = value;
}
).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--tts-use-guide-tokens"},
"Use guide tokens to improve TTS word recall",
[](common_params & params) {
params.vocoder.use_guide_tokens = true;
}
).set_examples({LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_SERVER}));
// model-specific
add_opt(common_arg(
@ -2218,5 +2324,47 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
}
).set_examples({LLAMA_EXAMPLE_TTS}));
add_opt(common_arg(
{"--embd-bge-small-en-default"},
string_format("use default bge-small-en-v1.5 model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/bge-small-en-v1.5-Q8_0-GGUF";
params.hf_file = "bge-small-en-v1.5-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
params.verbose_prompt = true;
params.embedding = true;
}
).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--embd-e5-small-en-default"},
string_format("use default e5-small-v2 model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/e5-small-v2-Q8_0-GGUF";
params.hf_file = "e5-small-v2-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
params.verbose_prompt = true;
params.embedding = true;
}
).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
add_opt(common_arg(
{"--embd-gte-small-default"},
string_format("use default gte-small model (note: can download weights from the internet)"),
[](common_params & params) {
params.hf_repo = "ggml-org/gte-small-Q8_0-GGUF";
params.hf_file = "gte-small-q8_0.gguf";
params.pooling_type = LLAMA_POOLING_TYPE_NONE;
params.embd_normalize = 2;
params.n_ctx = 512;
params.verbose_prompt = true;
params.embedding = true;
}
).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_SERVER}));
return ctx_arg;
}

3
common/arg.h
View file

@ -12,6 +12,7 @@
struct common_arg {
std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
std::set<enum llama_example> excludes = {};
std::vector<const char *> args;
const char * value_hint = nullptr; // help text or example for arg value
const char * value_hint_2 = nullptr; // for second arg value
@ -53,9 +54,11 @@ struct common_arg {
) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
common_arg & set_examples(std::initializer_list<enum llama_example> examples);
common_arg & set_excludes(std::initializer_list<enum llama_example> excludes);
common_arg & set_env(const char * env);
common_arg & set_sparam();
bool in_example(enum llama_example ex);
bool is_exclude(enum llama_example ex);
bool get_value_from_env(std::string & output);
bool has_value_from_env();
std::string to_string();

515
common/chat-template.hpp Normal file
View file

@ -0,0 +1,515 @@
/*
Copyright 2024 Google LLC
Use of this source code is governed by an MIT-style
license that can be found in the LICENSE file or at
https://opensource.org/licenses/MIT.
*/
// SPDX-License-Identifier: MIT
#pragma once
#include "minja.hpp"
#include <json.hpp>
#include <string>
#include <vector>
using json = nlohmann::ordered_json;
namespace minja {
struct chat_template_caps {
bool supports_tools = false;
bool supports_tool_calls = false;
bool supports_tool_responses = false;
bool supports_system_role = false;
bool supports_parallel_tool_calls = false;
bool supports_tool_call_id = false;
// meta-llama/Llama-3.1-8B-Instruct expects arguments to be an object.
// Most other templates (and OpenAI's API) expect the arguments object to be stringified.
bool requires_object_arguments = false;
// CohereForAI/c4ai-command-r-plus simple variant
bool requires_non_null_content = false;
// MiniMaxAI/MiniMax-Text-01 special
bool requires_typed_content = false;
};
struct chat_template_inputs {
nlohmann::ordered_json messages;
nlohmann::ordered_json tools;
bool add_generation_prompt = true;
nlohmann::ordered_json extra_context;
std::chrono::system_clock::time_point now = std::chrono::system_clock::now();
};
struct chat_template_options {
bool apply_polyfills = true;
bool use_bos_token = true;
bool use_eos_token = true;
bool define_strftime_now = true;
bool polyfill_tools = true;
bool polyfill_tool_call_examples = true;
bool polyfill_tool_calls = true;
bool polyfill_tool_responses = true;
bool polyfill_system_role = true;
bool polyfill_object_arguments = true;
bool polyfill_typed_content = true;
};
class chat_template {
private:
chat_template_caps caps_;
std::string source_;
std::string bos_token_;
std::string eos_token_;
std::shared_ptr<minja::TemplateNode> template_root_;
std::string tool_call_example_;
std::string try_raw_render(
const nlohmann::ordered_json & messages,
const nlohmann::ordered_json & tools,
bool add_generation_prompt,
const nlohmann::ordered_json & extra_context = nlohmann::ordered_json()) const
{
try {
chat_template_inputs inputs;
inputs.messages = messages;
inputs.tools = tools;
inputs.add_generation_prompt = add_generation_prompt;
inputs.extra_context = extra_context;
// Use fixed date for tests
inputs.now = std::chrono::system_clock::from_time_t(0);
chat_template_options opts;
opts.apply_polyfills = false;
auto prompt = apply(inputs, opts);
// fprintf(stderr, "try_raw_render: %s\n", prompt.c_str());
return prompt;
} catch (const std::exception & e) {
// fprintf(stderr, "try_raw_render error: %s\n", e.what());
return "";
}
}
public:
chat_template(const std::string & source, const std::string & bos_token, const std::string & eos_token)
: source_(source), bos_token_(bos_token), eos_token_(eos_token)
{
template_root_ = minja::Parser::parse(source_, {
/* .trim_blocks = */ true,
/* .lstrip_blocks = */ true,
/* .keep_trailing_newline = */ false,
});
auto contains = [](const std::string & haystack, const std::string & needle) {
return haystack.find(needle) != std::string::npos;
};
const std::string user_needle = "<User Needle>";
const std::string sys_needle = "<System Needle>";
const json dummy_str_user_msg = {{"role", "user"}, {"content", user_needle}};
const json dummy_typed_user_msg = {{"role", "user"}, {"content", json::array({{{"type", "text"}, {"text", user_needle}}})}};
caps_.requires_typed_content =
!contains(try_raw_render(json::array({dummy_str_user_msg}), {}, false), user_needle)
&& contains(try_raw_render(json::array({dummy_typed_user_msg}), {}, false), user_needle);
const auto dummy_user_msg = caps_.requires_typed_content
? dummy_typed_user_msg
: dummy_str_user_msg;
const json needle_system_msg = {
{"role", "system"},
{"content", caps_.requires_typed_content ? json::array({{{"type", "text"}, {"text", sys_needle}}}) : json(sys_needle)},
};
caps_.supports_system_role = contains(try_raw_render({needle_system_msg, dummy_user_msg,}, {}, false), sys_needle);
auto out = try_raw_render(json::array({
dummy_user_msg
}), json::array({
{
{"name", "some_tool"},
{"type", "function"},
{"function", {
{"name", "some_tool"},
{"description", "Some tool."},
{"parameters", {
{"type", "object"},
{"properties", {
{"arg", {
{"type", "string"},
{"description", "Some argument."},
}},
}},
{"required", json::array({ "arg" })},
}},
}},
},
}), false);
caps_.supports_tools = contains(out, "some_tool");
auto make_tool_calls_msg = [&](const json & tool_calls) {
return json {
{"role", "assistant"},
{"content", nullptr},
{"tool_calls", tool_calls},
};
};
auto make_tool_call = [](const std::string & tool_name, const json & arguments) {
return json {
{"id", "call_1___"},
{"type", "function"},
{"function", {
{"arguments", arguments},
{"name", tool_name},
}},
};
};
const json dummy_args_obj {{"argument_needle", "print('Hello, World!')"}};
// Note: the arguments are rendered in both cases, but may be double-escaped, which we don't want.
out = try_raw_render(json::array({
dummy_user_msg,
make_tool_calls_msg(json::array({make_tool_call("ipython", dummy_args_obj.dump())})),
}), {}, false);
auto tool_call_renders_str_arguments = contains(out, "\"argument_needle\":") || contains(out, "'argument_needle':");
out = try_raw_render(json::array({
dummy_user_msg,
make_tool_calls_msg(json::array({make_tool_call("ipython", dummy_args_obj)})),
}), {}, false);
auto tool_call_renders_obj_arguments = contains(out, "\"argument_needle\":") || contains(out, "'argument_needle':");
caps_.supports_tool_calls = tool_call_renders_str_arguments || tool_call_renders_obj_arguments;
caps_.requires_object_arguments = !tool_call_renders_str_arguments && tool_call_renders_obj_arguments;
auto out_empty = try_raw_render(json::array({dummy_user_msg, {{"role", "assistant"}, {"content", ""}}}), {}, false);
auto out_null = try_raw_render(json::array({dummy_user_msg, {{"role", "assistant"}, {"content", nullptr}}}), {}, false);
caps_.requires_non_null_content = contains(out_empty, user_needle) && !contains(out_null, user_needle);
if (caps_.supports_tool_calls) {
auto dummy_args = caps_.requires_object_arguments ? dummy_args_obj : json(dummy_args_obj.dump());
auto tc1 = make_tool_call("test_tool1", dummy_args);
auto tc2 = make_tool_call("test_tool2", dummy_args);
auto out = try_raw_render(json::array({
dummy_user_msg,
make_tool_calls_msg(json::array({tc1, tc2})),
}), {}, false);
caps_.supports_parallel_tool_calls = contains(out, "test_tool1") && contains(out, "test_tool2");
out = try_raw_render(json::array({
dummy_user_msg,
make_tool_calls_msg(json::array({tc1})),
{
{"role", "tool"},
{"name", "test_tool1"},
{"content", "Some response!"},
{"tool_call_id", "call_911_"},
}
}), {}, false);
caps_.supports_tool_responses = contains(out, "Some response!");
caps_.supports_tool_call_id = contains(out, "call_911_");
}
try {
if (!caps_.supports_tools) {
const json user_msg {
{"role", "user"},
{"content", "Hey"},
};
const json args {
{"arg1", "some_value"},
};
const json tool_call_msg {
{"role", "assistant"},
{"content", nullptr},
{"tool_calls", json::array({
{
// TODO: detect if requires numerical id or fixed length == 6 like Nemo
{"id", "call_1___"},
{"type", "function"},
{"function", {
{"name", "tool_name"},
{"arguments", (caps_.requires_object_arguments ? args : json(minja::Value(args).dump(-1, /* to_json= */ true)))},
}},
},
})},
};
std::string prefix, full;
{
chat_template_inputs inputs;
inputs.messages = json::array({user_msg});
inputs.add_generation_prompt = true;
prefix = apply(inputs);
}
{
chat_template_inputs inputs;
inputs.messages = json::array({user_msg, tool_call_msg});
inputs.add_generation_prompt = false;
full = apply(inputs);
}
if (full.find(prefix) != 0) {
if (prefix.rfind(eos_token_) == prefix.size() - eos_token_.size()) {
prefix = prefix.substr(0, prefix.size() - eos_token_.size());
}
}
if (full.find(prefix) != 0) {
fprintf(stderr, "Failed to infer a tool call example (possible template bug)\n");
}
tool_call_example_ = full.substr(prefix.size());
}
} catch (const std::exception & e) {
fprintf(stderr, "Failed to generate tool call example: %s\n", e.what());
}
}
const std::string & source() const { return source_; }
const std::string & bos_token() const { return bos_token_; }
const std::string & eos_token() const { return eos_token_; }
const chat_template_caps & original_caps() const { return caps_; }
// Deprecated, please use the form with chat_template_inputs and chat_template_options
std::string apply(
const nlohmann::ordered_json & messages,
const nlohmann::ordered_json & tools,
bool add_generation_prompt,
const nlohmann::ordered_json & extra_context = nlohmann::ordered_json(),
bool apply_polyfills = true)
{
fprintf(stderr, "[%s] Deprecated!\n", __func__);
chat_template_inputs inputs;
inputs.messages = messages;
inputs.tools = tools;
inputs.add_generation_prompt = add_generation_prompt;
inputs.extra_context = extra_context;
inputs.now = std::chrono::system_clock::now();
chat_template_options opts;
opts.apply_polyfills = apply_polyfills;
return apply(inputs, opts);
}
std::string apply(
const chat_template_inputs & inputs,
const chat_template_options & opts = chat_template_options()) const
{
json actual_messages;
auto has_tools = inputs.tools.is_array() && !inputs.tools.empty();
auto has_tool_calls = false;
auto has_tool_responses = false;
auto has_string_content = false;
for (const auto & message : inputs.messages) {
if (message.contains("tool_calls") && !message["tool_calls"].is_null()) {
has_tool_calls = true;
}
if (message.contains("role") && message["role"] == "tool") {
has_tool_responses = true;
}
if (message.contains("content") && message["content"].is_string()) {
has_string_content = true;
}
}
auto polyfill_system_role = opts.polyfill_system_role && !caps_.supports_system_role;
auto polyfill_tools = opts.polyfill_tools && has_tools && !caps_.supports_tools;
auto polyfill_tool_call_example = polyfill_tools && opts.polyfill_tool_call_examples;
auto polyfill_tool_calls = opts.polyfill_tool_calls && has_tool_calls && !caps_.supports_tool_calls;
auto polyfill_tool_responses = opts.polyfill_tool_responses && has_tool_responses && !caps_.supports_tool_responses;
auto polyfill_object_arguments = opts.polyfill_object_arguments && has_tool_calls && caps_.requires_object_arguments;
auto polyfill_typed_content = opts.polyfill_typed_content && has_string_content && caps_.requires_typed_content;
auto needs_polyfills = opts.apply_polyfills && (false
|| polyfill_system_role
|| polyfill_tools
|| polyfill_tool_calls
|| polyfill_tool_responses
|| polyfill_object_arguments
|| polyfill_typed_content
);
if (needs_polyfills) {
actual_messages = json::array();
auto add_message = [&](const json & msg) {
if (polyfill_typed_content && msg.contains("content") && !msg.at("content").is_null() && msg.at("content").is_string()) {
actual_messages.push_back({
{"role", msg.at("role")},
{"content", {{
{"type", "text"},
{"text", msg.at("content")},
}}},
});
} else {
actual_messages.push_back(msg);
}
};
std::string pending_system;
auto flush_sys = [&]() {
if (!pending_system.empty()) {
add_message({
{"role", "user"},
{"content", pending_system},
});
pending_system.clear();
}
};
json adjusted_messages;
if (polyfill_tools) {
adjusted_messages = add_system(inputs.messages,
"You can call any of the following tools to satisfy the user's requests: " + minja::Value(inputs.tools).dump(2, /* to_json= */ true) +
(!polyfill_tool_call_example || tool_call_example_.empty() ? "" : "\n\nExample tool call syntax:\n\n" + tool_call_example_));
} else {
adjusted_messages = inputs.messages;
}
for (const auto & message_ : adjusted_messages) {
auto message = message_;
if (!message.contains("role") || !message.contains("content")) {
throw std::runtime_error("message must have 'role' and 'content' fields: " + message.dump());
}
std::string role = message.at("role");
if (message.contains("tool_calls")) {
if (polyfill_object_arguments || polyfill_tool_calls) {
for (auto & tool_call : message.at("tool_calls")) {
if (tool_call["type"] == "function") {
auto & function = tool_call.at("function");
auto & arguments = function.at("arguments");
if (arguments.is_string()) {
try {
arguments = json::parse(arguments.get<std::string>());
} catch (const std::exception & ecvt) {
fprintf(stderr, "Failed to parse arguments: %s\n", ecvt.what());
}
}
}
}
}
if (polyfill_tool_calls) {
auto content = message.at("content");
auto tool_calls = json::array();
for (const auto & tool_call : message.at("tool_calls")) {
if (tool_call.at("type") != "function") {
continue;
}
const auto & function = tool_call.at("function");
auto tc = json {
{"name", function.at("name")},
{"arguments", function.at("arguments")},
};
if (tool_call.contains("id")) {
tc["id"] = tool_call["id"];
}
tool_calls.push_back(tc);
}
auto obj = json {
{"tool_calls", tool_calls},
};
if (!content.is_null() && content != "") {
obj["content"] = content;
}
message["content"] = obj.dump(2);
message.erase("tool_calls");
}
}
if (polyfill_tool_responses && role == "tool") {
message["role"] = "user";
auto obj = json {
{"tool_response", {
{"content", message.at("content")},
}},
};
if (message.contains("name")) {
obj["tool_response"]["name"] = message.at("name");
}
if (message.contains("tool_call_id")) {
obj["tool_response"]["tool_call_id"] = message.at("tool_call_id");
}
message["content"] = obj.dump(2);
message.erase("name");
}
if (!message["content"].is_null() && polyfill_system_role) {
std::string content = message.at("content");
if (role == "system") {
if (!pending_system.empty()) pending_system += "\n";
pending_system += content;
continue;
} else {
if (role == "user") {
if (!pending_system.empty()) {
message["content"] = pending_system + (content.empty() ? "" : "\n" + content);
pending_system.clear();
}
} else {
flush_sys();
}
}
}
add_message(message);
}
flush_sys();
} else {
actual_messages = inputs.messages;
}
auto context = minja::Context::make(json({
{"messages", actual_messages},
{"add_generation_prompt", inputs.add_generation_prompt},
}));
context->set("bos_token", opts.use_bos_token ? bos_token_ : "");
context->set("eos_token", opts.use_eos_token ? eos_token_ : "");
if (opts.define_strftime_now) {
auto now = inputs.now;
context->set("strftime_now", Value::callable([now](const std::shared_ptr<minja::Context> &, minja::ArgumentsValue & args) {
args.expectArgs("strftime_now", {1, 1}, {0, 0});
auto format = args.args[0].get<std::string>();
auto time = std::chrono::system_clock::to_time_t(now);
auto local_time = *std::localtime(&time);
std::ostringstream ss;
ss << std::put_time(&local_time, format.c_str());
return ss.str();
}));
}
if (!inputs.tools.is_null()) {
context->set("tools", minja::Value(inputs.tools));
}
if (!inputs.extra_context.is_null()) {
for (auto & kv : inputs.extra_context.items()) {
context->set(kv.key(), minja::Value(kv.value()));
}
}
auto ret = template_root_->render(context);
// fprintf(stderr, "actual_messages: %s\n", actual_messages.dump(2).c_str());
// fprintf(stderr, "apply: %s\n\n", ret.c_str());
return ret;
}
static nlohmann::ordered_json add_system(const nlohmann::ordered_json & messages, const std::string & system_prompt) {
json messages_with_system = messages;
if (messages_with_system.size() > 0 && messages_with_system[0].at("role") == "system") {
std::string existing_system = messages_with_system.at(0).at("content");
messages_with_system[0] = json {
{"role", "system"},
{"content", existing_system + "\n\n" + system_prompt},
};
} else {
messages_with_system.insert(messages_with_system.begin(), json {
{"role", "system"},
{"content", system_prompt},
});
}
return messages_with_system;
}
};
} // namespace minja

966
common/chat.cpp Normal file
View file

@ -0,0 +1,966 @@
#include "chat.hpp"
#include "chat-template.hpp"
#include "json-schema-to-grammar.h"
#include "log.h"
#include "minja.hpp"
std::string common_chat_format_name(common_chat_format format) {
switch (format) {
case COMMON_CHAT_FORMAT_CONTENT_ONLY: return "Content-only";
case COMMON_CHAT_FORMAT_GENERIC: return "Generic";
case COMMON_CHAT_FORMAT_MISTRAL_NEMO: return "Mistral Nemo";
case COMMON_CHAT_FORMAT_LLAMA_3_X: return "Llama 3.x";
case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS: return "Llama 3.x with builtin tools";
case COMMON_CHAT_FORMAT_DEEPSEEK_R1: return "DeepSeek R1";
case COMMON_CHAT_FORMAT_FIREFUNCTION_V2: return "FireFunction v2";
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2: return "Functionary v3.2";
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1: return "Functionary v3.1 Llama 3.1";
case COMMON_CHAT_FORMAT_HERMES_2_PRO: return "Hermes 2 Pro";
case COMMON_CHAT_FORMAT_COMMAND_R7B: return "Command R7B";
default:
throw std::runtime_error("Unknown chat format");
}
}
const common_grammar_options grammar_options {
/* .dotall = */ false,
/* .compact_spaces = */ false,
// /* .compact_spaces = */ true,
};
static bool parse_json(std::string::const_iterator & it, const std::string::const_iterator & end, json & out) {
// // https://json.nlohmann.me/features/parsing/sax_interface/
struct json_error_locator : public nlohmann::json_sax<json> {
std::size_t position;
bool found_error;
json_error_locator() : position(0), found_error(false) {}
bool parse_error(std::size_t position, const std::string &, const json::exception &) override {
this->position = position - 1;
this->found_error = true;
return false;
}
bool null() override { return true; }
bool boolean(bool) override { return true; }
bool number_integer(number_integer_t) override { return true; }
bool number_unsigned(number_unsigned_t) override { return true; }
bool number_float(number_float_t, const string_t &) override { return true; }
bool string(string_t &) override { return true; }
bool binary(binary_t &) override { return true; }
bool start_object(std::size_t) override { return true; }
bool key(string_t &) override { return true; }
bool end_object() override { return true; }
bool start_array(std::size_t) override { return true; }
bool end_array() override { return true; }
};
json_error_locator err_loc;
json::sax_parse(it, end, &err_loc);
std::string::const_iterator temptative_end;
if (err_loc.found_error) {
temptative_end = it + err_loc.position;
} else {
temptative_end = end;
}
std::string json_sub {it, temptative_end};
try {
out = json::parse(json_sub);
it = temptative_end;
return true;
} catch (const std::exception &) {
return false;
}
}
/**
* Takes a prefix regex that must have 1 group to capture the function name, a closing suffix, and expects json parameters in between.
* Aggregates the prefix, suffix and in-between text into the content.
*/
static common_chat_msg parse_json_tool_calls(
const std::string& input,
const std::optional<std::regex> & trigger_opt,
const std::regex & function_regex,
const std::regex & close_regex) {
std::smatch match;
common_chat_msg result;
result.role = "assistant";
auto end = input.end();
auto it = input.begin();
if (trigger_opt) {
if (!std::regex_search(it, end, match, *trigger_opt)) {
result.content = input;
return result;
}
result.content = match.prefix().str();
it = match.suffix().first;
}
while (it != end) {
std::sregex_iterator rend;
std::sregex_iterator rit(it, end, function_regex);
if (rit == rend) {
fprintf(stderr, "No more tool calls found\n");
result.content += std::string(it, end);
break;
}
auto name = rit->str(1);
result.content += std::string(it, rit->prefix().second);
it = rit->suffix().first;
json arguments;
if (!parse_json(it, end, arguments)) {
throw std::runtime_error("Failed to parse json tool call arguments");
}
if (!std::regex_search(it, end, match, close_regex)) {
throw std::runtime_error("Malformed input, missing closing pattern");
}
it = match.suffix().first;
result.tool_calls.push_back({name, arguments.is_string() ? arguments.get<std::string>() : arguments.dump(), /* id= */ ""});
}
return result;
}
static common_chat_msg parse_prefixed_json_tool_call_array(const std::string& input, const std::string & prefix, size_t rstrip_prefix = 0) {
auto content_end = input.find(prefix);
size_t tc_start = std::string::npos;
common_chat_msg result;
result.role = "assistant";
const auto process_tool_calls = [&](const json & tool_calls) {
for (const auto & tool_call : tool_calls) {
const auto & arguments = tool_call["arguments"];
result.tool_calls.push_back({
tool_call["name"],
arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
tool_call.contains("id") ? tool_call["id"] : "",
});
}
};
if (content_end == std::string::npos) {
result.content = input;
} else {
tc_start = content_end + prefix.size() - rstrip_prefix;
result.content = input.substr(0, content_end);
auto tool_calls = json::parse(input.substr(tc_start));
process_tool_calls(tool_calls);
}
return result;
}
static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
for (const auto & tool : tools) {
if (!tool.contains("type") || tool["type"] != "function" || !tool.contains("function")) {
LOG_INF("Skipping tool without function: %s", tool.dump(2).c_str());
continue;
}
fn(tool);
}
}
static std::string apply(
const common_chat_template & tmpl,
const nlohmann::ordered_json & messages,
const nlohmann::ordered_json & tools,
bool add_generation_prompt,
const nlohmann::ordered_json & extra_context = nlohmann::ordered_json())
{
minja::chat_template_inputs tmpl_inputs;
tmpl_inputs.messages = messages;
tmpl_inputs.tools = tools;
tmpl_inputs.add_generation_prompt = add_generation_prompt;
tmpl_inputs.extra_context = extra_context;
// TODO: add flag to control date/time, if only for testing purposes.
// tmpl_inputs.now = std::chrono::system_clock::now();
minja::chat_template_options tmpl_opts;
tmpl_opts.use_bos_token = false;
tmpl_opts.use_eos_token = false;
return tmpl.apply(tmpl_inputs, tmpl_opts);
}
static common_chat_params common_chat_params_init_generic(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
auto tool_call_schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
auto tool_schema = json {
{"type", "object"},
{"properties", {
{"name", {
{"type", "string"},
{"const", function["name"]},
}},
{"arguments", function["parameters"]},
}},
{"required", json::array({"name", "arguments"})},
};
if (function.contains("description")) {
tool_schema["description"] = function["description"];
}
if (inputs.parallel_tool_calls) {
tool_schema["properties"]["id"] = {
{"type", "string"},
{"minLength", 4},
};
tool_schema["required"].push_back("id");
}
tool_call_schemas.emplace_back(tool_schema);
});
const auto tool_call =
inputs.parallel_tool_calls
? json {
{"type", "object"},
{"properties", {
{"tool_calls", {
{"type", "array"},
{"items", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {
{"anyOf", tool_call_schemas},
}},
{"minItems", 1},
}},
}},
{"required", json::array({"tool_calls"})},
}
: json {
{"type", "object"},
{"properties", {
{"tool_call", tool_call_schemas.size() == 1 ? tool_call_schemas[0] : json {
{"anyOf", tool_call_schemas},
}},
}},
{"required", json::array({"tool_call"})},
};
const auto schema =
inputs.tool_choice != "required"
? json {
{"anyOf", json::array({
tool_call,
{
{"type", "object"},
{"properties", {
{"response", inputs.json_schema.is_null()
? json {{"type", "string"}}
: inputs.json_schema
},
}},
{"required", json::array({"response"})},
},
})}
}
: tool_call;
data.grammar_lazy = false;
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
builder.add_schema("root", schema);
}, grammar_options);
auto tweaked_messages = common_chat_template::add_system(
inputs.messages,
"Respond in JSON format, either with `tool_call` (a request to call tools) or with `response` reply to the user's request");
data.prompt = apply(tmpl, tweaked_messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_GENERIC;
return data;
}
static common_chat_msg common_chat_parse_generic(const std::string & input) {
json data = json::parse(input);
common_chat_msg result;
result.role = "assistant";
if (data.contains("tool_calls")) {
for (const auto & tool_call : data["tool_calls"]) {
result.tool_calls.push_back({
tool_call["name"],
tool_call["arguments"].dump(),
tool_call.contains("id") ? tool_call["id"] : "",
});
}
} else if (data.contains("tool_call")) {
result.tool_calls.push_back({
data["tool_call"]["name"],
data["tool_call"]["arguments"].dump(),
/* id= */ "",
});
} else if (data.contains("response")) {
const auto & response = data["response"];
result.content = response.is_string() ? response.get<std::string>() : response.dump(2);
}
return result;
}
static common_chat_params common_chat_params_init_mistral_nemo(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
auto schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
schemas.push_back({
{"type", "object"},
{"properties", {
// Important note: the model is probably trained to take a JSON stringified arguments value.
// It's hard to constrain that for now (while reusing the JSON schema conversion), so we're just expecting a plain object.
{"name", {
{"type", "string"},
{"const", function["name"]},
}},
{"arguments", function["parameters"]},
{"id", {
{"type", "string"},
// Nemo's template expects a 9-character alphanumeric ID.
{"pattern", "^[a-zA-Z0-9]{9}$"},
}},
}},
{"required", json::array({"name", "arguments", "id"})},
});
});
auto schema = json {
{"type", "array"},
{"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
{"minItems", 1},
};
if (!inputs.parallel_tool_calls) {
schema["maxItems"] = 1;
}
builder.add_rule("root", "\"[TOOL_CALLS]\" " + builder.add_schema("tool_calls", schema));
}, grammar_options);
data.grammar_triggers.push_back({"[TOOL_CALLS]", /* .at_start = */ true});
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_MISTRAL_NEMO;
return data;
}
static common_chat_msg common_chat_parse_mistral_nemo(const std::string & input) {
return parse_prefixed_json_tool_call_array(input, "[TOOL_CALLS]");
}
static common_chat_params common_chat_params_init_command_r7b(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
auto schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
schemas.push_back({
{"type", "object"},
{"properties", {
{"tool_call_id", {
{"type", "string"},
// Command-R's template expects an integer string.
{"pattern", "^[0-9]{1,10}$"},
}},
{"tool_name", {
{"type", "string"},
{"const", function["name"]},
}},
{"parameters", function["parameters"]},
}},
{"required", json::array({"tool_call_id", "tool_name", "parameters"})},
});
});
auto schema = json {
{"type", "array"},
{"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
{"minItems", 1},
};
if (!inputs.parallel_tool_calls) {
schema["maxItems"] = 1;
}
builder.add_rule("root", "\"<|START_ACTION|>\" " + builder.add_schema("tool_calls", schema) + " \"<|END_ACTION|>\"");
}, grammar_options);
data.grammar_triggers.push_back({"<|START_ACTION|>", /* .at_start = */ false});
data.preserved_tokens = {
"<|START_RESPONSE|>",
"<|END_RESPONSE|>",
"<|START_THINKING|>",
"<|END_THINKING|>",
"<|END_ACTION|>",
};
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_COMMAND_R7B;
return data;
}
static common_chat_msg common_chat_parse_command_r7b(const std::string & input) {
static std::regex response_regex("<\\|START_RESPONSE\\|>([\\s\\S\\n\\r]*?)<\\|END_RESPONSE\\|>");
static std::regex thought_action_regex("<\\|START_THINKING\\|>([\\s\\S\\n\\r]*?)<\\|END_THINKING\\|><\\|START_ACTION\\|>([\\s\\S\\n\\r]*?)<\\|END_ACTION\\|>");
std::smatch match;
common_chat_msg result;
result.role = "assistant";
if (std::regex_match(input, match, response_regex)) {
result.content = match[1].str();
} else if (std::regex_match(input, match, thought_action_regex)) {
result.tool_plan = match[1].str();
auto actions_str = match[2].str();
auto actions = json::parse(actions_str);
for (const auto & action : actions) {
result.tool_calls.push_back({
/* .name = */ action["tool_name"],
/* .arguments = */ action["parameters"].dump(),
/* .id = */ action["tool_call_id"],
});
}
} else {
LOG_ERR("Failed to parse command_r output");
result.content = input;
}
return result;
}
static void expect_tool_parameters(const std::string & name, const json & parameters, const std::vector<std::string> & expected_properties) {
if (!parameters.is_object() || !parameters.contains("type") || parameters["type"] != "object" || !parameters.contains("properties") || !parameters.contains("required")) {
throw std::runtime_error("Parameters of tool " + name + " must be an object w/ required properties");
}
const auto & parameters_properties = parameters.at("properties");
const auto & parameters_required = parameters.at("required");
for (const auto & prop : expected_properties) {
if (!parameters_properties.contains(prop)) {
throw std::runtime_error("Parameters of tool " + name + " is missing property: " + prop);
}
if (std::find(parameters_required.begin(), parameters_required.end(), json(prop)) == parameters_required.end()) {
throw std::runtime_error("Parameters of tool " + name + " must have property marked as required: " + prop);
}
}
if (parameters_properties.size() != expected_properties.size()) {
throw std::runtime_error("Parameters of tool " + name + " must only have these properties:" + string_join(expected_properties, ", "));
}
}
static common_chat_params common_chat_params_init_llama_3_1_tool_calls(const common_chat_template & tmpl, const struct common_chat_inputs & inputs, bool allow_python_tag_builtin_tools) {
auto builtin_tools = json::array();
common_chat_params data;
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
auto handle_builtin_tool = [&](const std::string & name, const json & parameters) {
if (name == "wolfram_alpha") {
// https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/wolfram_alpha/wolfram_alpha.py
expect_tool_parameters(name, parameters, {"query"});
} else if (name == "web_search" || name == "brave_search") {
// https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/remote/tool_runtime/brave_search/brave_search.py
expect_tool_parameters(name, parameters, {"query"});
} else if (name == "python" || name == "code_interpreter") {
// https://github.com/meta-llama/llama-stack/blob/main/llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
expect_tool_parameters(name, parameters, {"code"});
} else {
return false;
}
std::vector<std::string> kvs;
for (const auto & [key, value] : parameters.at("properties").items()) {
kvs.push_back("\"" + key + "=\" " + builder.add_schema(name + "-args-" + key, value));
}
tool_rules.push_back(
builder.add_rule(
name + "-call",
"\"<|python_tag|>" + name + ".call(\" " + string_join(kvs, " \", \" ") + " \")\""));
builtin_tools.push_back(name);
return true;
};
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
std::string name = function["name"];
auto parameters = function["parameters"];
builder.resolve_refs(parameters);
// https://github.com/meta-llama/llama-stack/tree/main/llama_stack/providers/remote/tool_runtime
if (allow_python_tag_builtin_tools) {
handle_builtin_tool(name, parameters);
}
tool_rules.push_back(
builder.add_rule(
name + "-call",
"\"{\" space "
"( \"\\\"type\\\":\" space \"\\\"function\\\",\" space )? "
"\"\\\"name\\\": \\\"" + name + "\\\", \\\"parameters\\\": \" " +
builder.add_schema(name + "-args", parameters) +
" \"}\""));
data.grammar_triggers.push_back({"{\"name\": \"" + name + "\"", /* .at_start = */ true});
});
data.grammar_triggers.push_back({"{\"name\":", /* .at_start = */ true});
data.grammar_triggers.push_back({"{\n \"name\":", /* .at_start = */ true});
data.grammar_triggers.push_back({"{\n \"name\":", /* .at_start = */ true});
data.grammar_triggers.push_back({"{\"type\": \"function\"", /* .at_start = */ true});
data.grammar_triggers.push_back({"{\n \"type\": \"function\"", /* .at_start = */ true});
data.grammar_triggers.push_back({"{\n \"type\": \"function\"", /* .at_start = */ true});
if (!builtin_tools.empty()) {
data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false});
}
builder.add_rule("root", string_join(tool_rules, " | "));
}, grammar_options);
data.additional_stops.push_back("<|eom_id|>");
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt, {
{"tools_in_user_message", false},
{"builtin_tools", builtin_tools.empty() ? json() : builtin_tools},
});
data.format = allow_python_tag_builtin_tools && !builtin_tools.empty()
? COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS
: COMMON_CHAT_FORMAT_LLAMA_3_X;
return data;
}
static common_chat_msg common_chat_parse_llama_3_1(const std::string & input, bool with_builtin_tools = false) {
// TODO: tighten & simplify the parser, don't accept leading text context.
static std::regex function_regex("\\{[\\s\\n\\r]*(?:\"type\"[\\s\\n\\r]*:[\\s\\n\\r]*\"function\"[\\s\\n\\r]*,[\\s\\n\\r]*|[\\s\\n\\r]*)\"name\"[\\s\\n\\r]*:[\\s\\n\\r]*\"([^\"]+)\"[\\s\\n\\r]*,[\\s\\n\\r]*\"parameters\": ");
static std::regex close_regex("\\}");
static std::regex builtin_call_regex("<\\|python_tag\\|>([^.(]+)\\.call\\((.*)\\)");
if (with_builtin_tools) {
std::smatch match;
if (std::regex_match(input, match, builtin_call_regex)) {
auto name = match[1].str();
auto raw_args = match[2].str();
// TODO: if/when builtin tools start accepting more than 1 argument, use parse_json for real parsing.
auto it_eq = raw_args.find('=');
auto arg_name = raw_args.substr(0, it_eq);
auto arg_value_str = raw_args.substr(it_eq + 1);
auto arg_value = json::parse(arg_value_str);
return {
/* .role = */ "assistant",
/* .content = */ match.prefix().str(),
/* .tool_calls = */ {
{
/* .name = */ match[1],
/* .arguments = */ (json {
{arg_name, arg_value},
}).dump(),
/* .id = */ "",
},
},
};
}
}
return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex);
}
static common_chat_params common_chat_params_init_deepseek_r1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
std::string name = function["name"];
auto parameters = function["parameters"];
auto args_rule = builder.add_schema(name + "-args", parameters);
tool_rules.push_back(builder.add_rule(name + "-call",
"\"<｜tool▁call▁begin｜>function<｜tool▁sep｜>" + name + "\\n```json\\n\" " + args_rule + " \"```<｜tool▁call▁end｜>\""));
});
data.grammar_triggers.push_back({"<｜tool▁calls▁begin｜>", /* .at_start = */ false});
data.preserved_tokens = {
"<｜tool▁sep｜>",
"<｜tool▁call▁end｜>",
};
builder.add_rule("root", "\"<｜tool▁calls▁begin｜>\" (" + string_join(tool_rules, " | ") + ")" + (inputs.parallel_tool_calls ? "*" : "") + " space");
}, grammar_options);
auto prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.prompt = prompt;
data.format = COMMON_CHAT_FORMAT_DEEPSEEK_R1;
return data;
}
static common_chat_msg common_chat_parse_deepseek_r1(const std::string & input) {
static std::regex trigger_regex("<｜tool▁calls▁begin｜>");
static std::regex function_regex("<｜tool▁call▁begin｜>function<｜tool▁sep｜>([^\n]+)\n```json\n");
static std::regex close_regex("```<｜tool▁call▁end｜>");
return parse_json_tool_calls(input, trigger_regex, function_regex, close_regex);
}
static common_chat_params common_chat_params_init_firefunction_v2(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
fprintf(stderr, "%s\n", __func__);
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, /* tools= */ nullptr, inputs.add_generation_prompt, {
{"datetime", "Jan 29 2025 13:00:00 GMT"},
{"functions", json(inputs.tools.empty() ? "" : inputs.tools.dump(2))},
});
if (!inputs.tools.is_null() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
auto schemas = json::array();
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
schemas.push_back({
{"type", "object"},
{"properties", {
{"name", {
{"type", "string"},
{"const", function["name"]},
}},
{"arguments", function["parameters"]},
}},
{"required", json::array({"name", "arguments", "id"})},
});
});
auto schema = json {
{"type", "array"},
{"items", schemas.size() == 1 ? schemas[0] : json {{"anyOf", schemas}}},
{"minItems", 1},
};
if (!inputs.parallel_tool_calls) {
schema["maxItems"] = 1;
}
builder.add_rule("root", "\" functools\"? " + builder.add_schema("tool_calls", schema));
}, grammar_options);
data.grammar_triggers.push_back({" functools[", /* .at_start = */ false});
data.format = COMMON_CHAT_FORMAT_FIREFUNCTION_V2;
} else {
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
}
return data;
}
static common_chat_msg common_chat_parse_firefunction_v2(const std::string & input) {
return parse_prefixed_json_tool_call_array(input, " functools[", /* rstrip_prefix= */ 1);
}
static common_chat_params common_chat_params_init_functionary_v3_2(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
// >>>all\nlet's call functions>>>fn1\n{"arg1": 1...}\n>>>fn2\n{"arg1": 1...}...
// Using ">>>f1\n", ">>>f2\n"... as trigger words for the grammar
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2;
if (!inputs.tools.is_null() && !inputs.tools.empty()) {
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> first_tool_rules;
std::vector<std::string> subsequent_tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
std::string name = function["name"];
auto parameters = function["parameters"];
auto args_rule = builder.add_schema(name + "-args", parameters);
first_tool_rules.push_back(builder.add_rule(name + "-call", "\"" + name + "\\n\" " + args_rule));
subsequent_tool_rules.push_back(builder.add_rule(name + "-call2", "\">>>" + name + "\\n\" " + args_rule));
data.grammar_triggers.push_back({name, /* .at_start = */ true});
data.grammar_triggers.push_back({">>>" + name, /* .at_start = */ false});
});
auto first_rule = first_tool_rules.empty() ? "" : builder.add_rule("first_tool_call", string_join(first_tool_rules, " | ")) + " space";
if (inputs.parallel_tool_calls) {
auto subsequent_rule = builder.add_rule("subsequent_tool_call", string_join(subsequent_tool_rules, " | ")) + " space";
builder.add_rule("root", first_rule + " (" + subsequent_rule + ")*");
} else {
builder.add_rule("root", first_rule);
}
}, grammar_options);
}
return data;
}
static bool consume(std::string::const_iterator & it, const std::string::const_iterator & end, const std::string & expected) {
auto expected_it = expected.begin();
auto tmp_it = it;
while (tmp_it != end && expected_it != expected.end() && *tmp_it == *expected_it) {
++tmp_it;
++expected_it;
}
if (expected_it == expected.end()) {
it = tmp_it;
return true;
}
return false;
}
static common_chat_msg common_chat_parse_functionary_v3_2(const std::string & input) {
static std::regex function_regex(R"((?:>>>)?(\w+)\n)");
static std::regex close_regex(R"($|(?=>>>))");
std::string content;
auto it = input.begin();
const auto end = input.end();
if (consume(it, end, "all\n")) {
std::smatch match;
if (std::regex_search(it, end, match, function_regex)) {
auto fun_it = match.prefix().second;
content = std::string(it, fun_it);
it = fun_it;
} else {
common_chat_msg res;
res.role = "assistant";
res.content = std::string(it, end);
return res;
}
}
// TODO: tighten & simplify.
try {
auto res = parse_json_tool_calls(std::string(it, end), std::nullopt, function_regex, close_regex);
res.content = content + res.content;
return res;
} catch (const std::exception & e) {
LOG_ERR("Failed to parse functionary v3.2 input: %s\n", e.what());
common_chat_msg res;
res.role = "assistant";
res.content = input;
return res;
}
}
static common_chat_params common_chat_params_init_functionary_v3_1_llama_3_1(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
// https://github.com/MeetKai/functionary/blob/main/tests/prompt_test_v3-llama3.1.txt
common_chat_params data;
json tools = inputs.tools.is_null() ? inputs.tools : json::array();
std::string python_code_argument_name;
auto has_raw_python = false;
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
const auto & parameters = function["parameters"];
std::string name = function["name"];
if (name == "python" || name == "ipython") {
if (!parameters.contains("type")) {
throw std::runtime_error("Missing type in python tool");
}
has_raw_python = true;
auto type = parameters.at("type");
if (type == "object") {
auto properties = parameters.at("properties");
for (auto it = properties.begin(); it != properties.end(); ++it) {
if (it.value().at("type") == "string") {
if (!python_code_argument_name.empty()) {
throw std::runtime_error("Multiple string arguments found in python tool");
}
python_code_argument_name = it.key();
}
}
if (python_code_argument_name.empty()) {
throw std::runtime_error("No string argument found in python tool");
}
} else if (type != "string") {
throw std::runtime_error("Invalid type in python tool: " + type.dump());
}
}
tool_rules.push_back(builder.add_rule(name + "-call", "\"<function=" + name + ">\" " + builder.add_schema(name + "-args", parameters) + " \"</function>\" space"));
});
if (has_raw_python) {
tool_rules.push_back(builder.add_rule("python-call", "\"<|python_tag|>\" .*"));
data.grammar_triggers.push_back({"<|python_tag|>", /* .at_start = */ false});
}
auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " space";
builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
data.grammar_triggers.push_back({"<function=", /* .at_start = */ false});
}, grammar_options);
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
// TODO: if (has_raw_python)
data.format = COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1;
return data;
}
static common_chat_msg common_chat_parse_functionary_v3_1_llama_3_1(const std::string & input) {
// This version of Functionary still supports the llama 3.1 tool call format for the python tool.
static std::regex python_tag_regex(R"(<\|python_tag\|>([\s\S\n]*)$)");
std::smatch match;
if (std::regex_search(input, match, python_tag_regex)) {
auto code = match[1].str();
return {
/* .role = */ "assistant",
/* .content = */ match.prefix().str(),
/* .tool_calls = */ {
{
/* .name = */ "python",
/* .arguments = */ (json {{"code", code}}).dump(),
/* .id = */ "",
},
}
};
}
static std::regex function_regex(R"(<function=(\w+)>)");
static std::regex close_regex(R"(</function>)");
// TODO: tighten & simplify.
return parse_json_tool_calls(input, std::nullopt, function_regex, close_regex);
}
static common_chat_params common_chat_params_init_hermes_2_pro(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
// (content)?(<tool_call>{"name": "foo", "arguments": {"a": 1}}</tool_call>)*
data.grammar_lazy = inputs.tool_choice != "required";
data.grammar = build_grammar([&](const common_grammar_builder & builder) {
std::vector<std::string> tool_rules;
foreach_function(inputs.tools, [&](const json & tool) {
const auto & function = tool["function"];
std::string name = function["name"];
auto parameters = function["parameters"];
builder.resolve_refs(parameters);
tool_rules.push_back(builder.add_schema(name + "-call", {
{"type", "object"},
{"properties", json {
{"name", json {{"const", name}}},
{"arguments", parameters},
}},
{"required", json::array({"name", "arguments"})},
}));
});
auto tool_call = "\"<tool_call>\" space " + builder.add_rule("tool_call", string_join(tool_rules, " | ")) + " \"</tool_call>\" space";
builder.add_rule("root", inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call);
data.grammar_triggers.push_back({"<tool_call>", /* .at_start = */ false});
data.preserved_tokens = { "</tool_call>" };
}, grammar_options);
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_HERMES_2_PRO;
return data;
}
static common_chat_msg common_chat_parse_hermes_2_pro(const std::string & input) {
try {
std::regex start_pattern(R"([\n\s]*<tool_call>)");
std::regex middle_pattern(R"([\n\s]*</tool_call>[\n\s]*<tool_call>)");
std::regex end_pattern(R"([\n\s]*</tool_call>[\n\s]*$)");
auto end = input.end();
std::sregex_iterator rend;
std::sregex_iterator rit(input.begin(), end, start_pattern);
if (rit == rend) {
return {
/* .role = */ "assistant",
/* .content = */ input,
/* .tool_calls = */ {},
};
}
common_chat_msg result;
result.role = "assistant";
result.content = rit->prefix();
auto it = rit->suffix().first;
while (it != end) {
json call;
if (!parse_json(it, end, call)) {
throw std::runtime_error("Failed to parse json tool call");
}
const auto & arguments = call["arguments"];
result.tool_calls.push_back({
call["name"],
arguments.dump(),
// arguments.is_string() ? arguments.get<std::string>() : arguments.dump(),
/* id= */ "",
});
rit = {it, end, middle_pattern};
if (rit != rend) {
it = rit->suffix().first;
} else {
rit = {it, end, end_pattern};
if (rit == rend) {
throw std::runtime_error("Malformed input, missing </tool_call>");
}
break;
}
}
return result;
} catch (const std::exception & e) {
return {
/* .role = */ "assistant",
/* .content = */ input,
/* .tool_calls = */ {},
};
}
}
static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
common_chat_params data;
data.prompt = apply(tmpl, inputs.messages, inputs.tools.empty() ? json() : inputs.tools, inputs.add_generation_prompt);
data.format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
data.grammar_lazy = false;
if (!inputs.json_schema.is_null()) {
if (!inputs.grammar.empty()) {
throw std::runtime_error("Either \"json_schema\" or \"grammar\" can be specified, but not both");
}
data.grammar = json_schema_to_grammar(inputs.json_schema);
} else {
data.grammar = inputs.grammar.empty();
}
return data;
}
common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & inputs) {
auto has_tools = !inputs.tools.is_null() && inputs.tool_choice != "none";
LOG_DBG("[%s] has_tools=%s\n", __func__, has_tools ? "true" : "false");
if (has_tools && !inputs.grammar.empty()) {
throw std::runtime_error("Cannot specify grammar with tools");
}
const auto & src = tmpl.source();
if (src.find(">>>all") != std::string::npos) {
// Functionary prepends "all\n" to plain content outputs, so we use the parser no matter when
return common_chat_params_init_functionary_v3_2(tmpl, inputs);
}
if (src.find(" functools[") != std::string::npos) {
// Firefunction v2 requires datetime and functions in the context, even w/o tools.
return common_chat_params_init_firefunction_v2(tmpl, inputs);
}
if (!has_tools) {
return common_chat_params_init_without_tools(tmpl, inputs);
}
if (src.find("<tool_call>") != std::string::npos) {
return common_chat_params_init_hermes_2_pro(tmpl, inputs);
}
if (src.find("<|start_header_id|>") != std::string::npos
&& src.find("<function=") != std::string::npos) {
return common_chat_params_init_functionary_v3_1_llama_3_1(tmpl, inputs);
}
if (src.find("<|start_header_id|>ipython<|end_header_id|>") != std::string::npos) {
auto allow_python_tag_builtin_tools = src.find("<|python_tag|>") != std::string::npos;
return common_chat_params_init_llama_3_1_tool_calls(tmpl, inputs, allow_python_tag_builtin_tools);
}
if (src.find("<｜tool▁calls▁begin｜>") != std::string::npos) {
return common_chat_params_init_deepseek_r1(tmpl, inputs);
}
if (src.find("[TOOL_CALLS]") != std::string::npos) {
return common_chat_params_init_mistral_nemo(tmpl, inputs);
}
if (src.find("<|END_THINKING|><|START_ACTION|>") != std::string::npos) {
return common_chat_params_init_command_r7b(tmpl, inputs);
}
return common_chat_params_init_generic(tmpl, inputs);
}
static common_chat_msg common_chat_parse_content_only(const std::string & input) {
return {
/* .role = */ "assistant",
/* .content = */ input,
/* .tool_calls = */ {},
};
}
common_chat_msg common_chat_parse(const std::string & input, common_chat_format format) {
switch (format) {
case COMMON_CHAT_FORMAT_CONTENT_ONLY:
return common_chat_parse_content_only(input);
case COMMON_CHAT_FORMAT_GENERIC:
return common_chat_parse_generic(input);
case COMMON_CHAT_FORMAT_MISTRAL_NEMO:
return common_chat_parse_mistral_nemo(input);
case COMMON_CHAT_FORMAT_LLAMA_3_X:
return common_chat_parse_llama_3_1(input);
case COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS:
return common_chat_parse_llama_3_1(input, /* with_builtin_tools= */ true);
case COMMON_CHAT_FORMAT_DEEPSEEK_R1:
return common_chat_parse_deepseek_r1(input);
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2:
return common_chat_parse_functionary_v3_2(input);
case COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1:
return common_chat_parse_functionary_v3_1_llama_3_1(input);
case COMMON_CHAT_FORMAT_HERMES_2_PRO:
return common_chat_parse_hermes_2_pro(input);
case COMMON_CHAT_FORMAT_FIREFUNCTION_V2:
return common_chat_parse_firefunction_v2(input);
case COMMON_CHAT_FORMAT_COMMAND_R7B:
return common_chat_parse_command_r7b(input);
default:
throw std::runtime_error("Unsupported format: " + common_chat_format_name(format));
}
}

52
common/chat.hpp Normal file
View file

@ -0,0 +1,52 @@
// Chat support (incl. tool call grammar constraining & output parsing) w/ generic & custom template handlers.
#pragma once
#include "common.h"
#include <json.hpp>
#include <optional>
#include <string>
#include <vector>
using json = nlohmann::ordered_json;
struct common_chat_inputs {
json messages;
json tools;
json tool_choice;
json json_schema;
bool parallel_tool_calls;
bool stream;
std::string grammar;
bool add_generation_prompt = true;
};
enum common_chat_format {
COMMON_CHAT_FORMAT_CONTENT_ONLY,
COMMON_CHAT_FORMAT_GENERIC,
COMMON_CHAT_FORMAT_MISTRAL_NEMO,
COMMON_CHAT_FORMAT_LLAMA_3_X,
COMMON_CHAT_FORMAT_LLAMA_3_X_WITH_BUILTIN_TOOLS,
COMMON_CHAT_FORMAT_DEEPSEEK_R1,
COMMON_CHAT_FORMAT_FIREFUNCTION_V2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_2,
COMMON_CHAT_FORMAT_FUNCTIONARY_V3_1_LLAMA_3_1,
COMMON_CHAT_FORMAT_HERMES_2_PRO,
COMMON_CHAT_FORMAT_COMMAND_R7B,
COMMON_CHAT_FORMAT_COUNT, // Not a format, just the # formats
};
struct common_chat_params {
common_chat_format format = COMMON_CHAT_FORMAT_CONTENT_ONLY;
json prompt;
std::string grammar;
bool grammar_lazy = false;
std::vector<common_grammar_trigger> grammar_triggers;
std::vector<std::string> preserved_tokens;
std::vector<std::string> additional_stops;
};
struct common_chat_params common_chat_params_init(const common_chat_template & tmpl, const struct common_chat_inputs & params);
std::string common_chat_format_name(common_chat_format format);
common_chat_msg common_chat_parse( const std::string & input, common_chat_format format);

441
common/common.cpp
View file

@ -2,6 +2,9 @@
#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING
#endif
#include "ggml.h"
#include "gguf.h"
#include "common.h"
#include "log.h"
// Change JSON_ASSERT from assert() to GGML_ASSERT:
@ -9,6 +12,8 @@
#include "json.hpp"
#include "json-schema-to-grammar.h"
#include "llama.h"
#include "chat.hpp"
#include "chat-template.hpp"
#include <algorithm>
#include <cinttypes>
@ -18,6 +23,7 @@
#include <cstdarg>
#include <cstring>
#include <ctime>
#include <filesystem>
#include <fstream>
#include <iostream>
#include <iterator>
@ -69,11 +75,29 @@ typedef unsigned short u_short;
#ifdef __linux__
#include <linux/limits.h>
#elif defined(_WIN32)
#define PATH_MAX MAX_PATH
# if !defined(PATH_MAX)
# define PATH_MAX MAX_PATH
# endif
#else
#include <sys/syslimits.h>
#endif
#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
//
// CURL utils
//
using curl_ptr = std::unique_ptr<CURL, decltype(&curl_easy_cleanup)>;
// cannot use unique_ptr for curl_slist, because we cannot update without destroying the old one
struct curl_slist_ptr {
struct curl_slist * ptr = nullptr;
~curl_slist_ptr() {
if (ptr) {
curl_slist_free_all(ptr);
}
}
};
#endif // LLAMA_USE_CURL
using json = nlohmann::ordered_json;
@ -468,6 +492,48 @@ void string_replace_all(std::string & s, const std::string & search, const std::
s = std::move(builder);
}
std::string string_join(const std::vector<std::string> & values, const std::string & separator) {
std::ostringstream result;
for (size_t i = 0; i < values.size(); ++i) {
if (i > 0) {
result << separator;
}
result << values[i];
}
return result.str();
}
std::vector<std::string> string_split(const std::string & str, const std::string & delimiter) {
std::vector<std::string> parts;
size_t start = 0;
size_t end = str.find(delimiter);
while (end != std::string::npos) {
parts.push_back(str.substr(start, end - start));
start = end + delimiter.length();
end = str.find(delimiter, start);
}
parts.push_back(str.substr(start));
return parts;
}
std::string string_repeat(const std::string & str, size_t n) {
if (n == 0) {
return "";
}
std::string result;
result.reserve(str.length() * n);
for (size_t i = 0; i < n; ++i) {
result += str;
}
return result;
}
std::string string_from(bool value) {
return value ? "true" : "false";
}
@ -850,7 +916,7 @@ struct common_init_result common_init_from_params(common_params & params) {
} else if (!params.model_url.empty()) {
model = common_load_model_from_url(params.model_url, params.model, params.hf_token, mparams);
} else {
model = llama_load_model_from_file(params.model.c_str(), mparams);
model = llama_model_load_from_file(params.model.c_str(), mparams);
}
if (model == NULL) {
@ -858,26 +924,28 @@ struct common_init_result common_init_from_params(common_params & params) {
return iparams;
}
const llama_vocab * vocab = llama_model_get_vocab(model);
if (params.reranking) {
bool ok = true;
if (llama_token_bos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have a BOS token, reranking will not work\n", __func__);
if (llama_vocab_bos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have a BOS token, reranking will not work\n", __func__);
ok = false;
}
if (llama_token_eos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have an EOS token, reranking will not work\n", __func__);
if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
ok = false;
}
if (llama_token_sep(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have a SEP token, reranking will not work\n", __func__);
if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
ok = false;
}
if (!ok) {
llama_free_model(model);
llama_model_free(model);
return iparams;
}
@ -885,40 +953,40 @@ struct common_init_result common_init_from_params(common_params & params) {
auto cparams = common_context_params_to_llama(params);
llama_context * lctx = llama_new_context_with_model(model, cparams);
llama_context * lctx = llama_init_from_model(model, cparams);
if (lctx == NULL) {
LOG_ERR("%s: failed to create context with model '%s'\n", __func__, params.model.c_str());
llama_free_model(model);
llama_model_free(model);
return iparams;
}
if (params.ctx_shift && !llama_kv_cache_can_shift(lctx)) {
LOG_ERR("%s: KV cache shifting is not supported for this model (--no-context-shift to disable)'\n", __func__);
llama_free_model(model);
return iparams;
LOG_WRN("%s: KV cache shifting is not supported for this model, disabling KV cache shifting\n", __func__);
params.ctx_shift = false;
}
if (!params.control_vectors.empty()) {
if (params.control_vector_layer_start <= 0) params.control_vector_layer_start = 1;
if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_n_layer(model);
if (params.control_vector_layer_end <= 0) params.control_vector_layer_end = llama_model_n_layer(model);
const auto cvec = common_control_vector_load(params.control_vectors);
if (cvec.n_embd == -1) {
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
int err = llama_control_vector_apply(lctx,
cvec.data.data(),
cvec.data.size(),
cvec.n_embd,
params.control_vector_layer_start,
params.control_vector_layer_end);
int err = llama_apply_adapter_cvec(
lctx,
cvec.data.data(),
cvec.data.size(),
cvec.n_embd,
params.control_vector_layer_start,
params.control_vector_layer_end);
if (err) {
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
@ -926,30 +994,31 @@ struct common_init_result common_init_from_params(common_params & params) {
// load and optionally apply lora adapters
for (auto & la : params.lora_adapters) {
common_lora_adapter_container loaded_la;
loaded_la.path = la.path;
loaded_la.scale = la.scale;
loaded_la.adapter = llama_lora_adapter_init(model, la.path.c_str());
if (loaded_la.adapter == nullptr) {
llama_adapter_lora_ptr lora;
lora.reset(llama_adapter_lora_init(model, la.path.c_str()));
if (lora == nullptr) {
LOG_ERR("%s: failed to apply lora adapter '%s'\n", __func__, la.path.c_str());
llama_free(lctx);
llama_free_model(model);
llama_model_free(model);
return iparams;
}
iparams.lora_adapters.push_back(loaded_la); // copy to list of loaded adapters
}
if (!params.lora_init_without_apply) {
common_lora_adapters_apply(lctx, iparams.lora_adapters);
la.ptr = lora.get();
iparams.lora.emplace_back(std::move(lora)); // copy to list of loaded adapters
}
if (params.sampling.ignore_eos && llama_token_eos(model) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: model does not have an EOS token, ignoring --ignore-eos\n", __func__);
if (!params.lora_init_without_apply) {
common_set_adapter_lora(lctx, params.lora_adapters);
}
if (params.sampling.ignore_eos && llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
LOG_WRN("%s: warning: vocab does not have an EOS token, ignoring --ignore-eos\n", __func__);
params.sampling.ignore_eos = false;
}
if (params.sampling.ignore_eos) {
for (llama_token i = 0; i < llama_n_vocab(model); i++) {
if (llama_token_is_eog(model, i)) {
for (llama_token i = 0; i < llama_vocab_n_tokens(vocab); i++) {
if (llama_vocab_is_eog(vocab, i)) {
LOG_INF("%s: added %s logit bias = %f\n", __func__, common_token_to_piece(lctx, i).c_str(), -INFINITY);
params.sampling.logit_bias.push_back({i, -INFINITY});
}
@ -970,8 +1039,9 @@ struct common_init_result common_init_from_params(common_params & params) {
LOG_WRN("%s: warming up the model with an empty run - please wait ... (--no-warmup to disable)\n", __func__);
std::vector<llama_token> tmp;
llama_token bos = llama_token_bos(model);
llama_token eos = llama_token_eos(model);
llama_token bos = llama_vocab_bos(vocab);
llama_token eos = llama_vocab_eos(vocab);
// some models (e.g. T5) don't have a BOS token
if (bos != LLAMA_TOKEN_NULL) {
tmp.push_back(bos);
@ -986,7 +1056,7 @@ struct common_init_result common_init_from_params(common_params & params) {
if (llama_model_has_encoder(model)) {
llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
if (decoder_start_token_id == -1) {
if (decoder_start_token_id == LLAMA_TOKEN_NULL) {
decoder_start_token_id = bos;
}
tmp.clear();
@ -1000,17 +1070,17 @@ struct common_init_result common_init_from_params(common_params & params) {
llama_perf_context_reset(lctx);
}
iparams.model = model;
iparams.context = lctx;
iparams.model.reset(model);
iparams.context.reset(lctx);
return iparams;
}
void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters) {
llama_lora_adapter_clear(ctx);
for (auto & la : lora_adapters) {
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
llama_clear_adapter_lora(ctx);
for (auto & la : lora) {
if (la.scale != 0.0f) {
llama_lora_adapter_set(ctx, la.adapter, la.scale);
llama_set_adapter_lora(ctx, la.ptr, la.scale);
}
}
}
@ -1024,7 +1094,6 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
if (params.n_gpu_layers != -1) {
mparams.n_gpu_layers = params.n_gpu_layers;
}
mparams.rpc_servers = params.rpc_servers.c_str();
mparams.main_gpu = params.main_gpu;
mparams.split_mode = params.split_mode;
mparams.tensor_split = params.tensor_split;
@ -1127,7 +1196,8 @@ static bool curl_perform_with_retry(const std::string & url, CURL * curl, int ma
static bool common_download_file(const std::string & url, const std::string & path, const std::string & hf_token) {
// Initialize libcurl
std::unique_ptr<CURL, decltype(&curl_easy_cleanup)> curl(curl_easy_init(), &curl_easy_cleanup);
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
if (!curl) {
LOG_ERR("%s: error initializing libcurl\n", __func__);
return false;
@ -1141,11 +1211,9 @@ static bool common_download_file(const std::string & url, const std::string & pa
// Check if hf-token or bearer-token was specified
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer ";
auth_header += hf_token.c_str();
struct curl_slist *http_headers = NULL;
http_headers = curl_slist_append(http_headers, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers);
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
}
#if defined(_WIN32)
@ -1155,8 +1223,7 @@ static bool common_download_file(const std::string & url, const std::string & pa
#endif
// Check if the file already exists locally
struct stat model_file_info;
auto file_exists = (stat(path.c_str(), &model_file_info) == 0);
auto file_exists = std::filesystem::exists(path);
// If the file exists, check its JSON metadata companion file.
std::string metadata_path = path + ".json";
@ -1416,7 +1483,7 @@ struct llama_model * common_load_model_from_url(
}
}
return llama_load_model_from_file(local_path.c_str(), params);
return llama_model_load_from_file(local_path.c_str(), params);
}
struct llama_model * common_load_model_from_hf(
@ -1442,6 +1509,80 @@ struct llama_model * common_load_model_from_hf(
return common_load_model_from_url(model_url, local_path, hf_token, params);
}
/**
* Allow getting the HF file from the HF repo with tag (like ollama), for example:
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q4
* - bartowski/Llama-3.2-3B-Instruct-GGUF:Q4_K_M
* - bartowski/Llama-3.2-3B-Instruct-GGUF:q5_k_s
* Tag is optional, default to "latest" (meaning it checks for Q4_K_M first, then Q4, then if not found, return the first GGUF file in repo)
*
* Return pair of <repo, file> (with "repo" already having tag removed)
*
* Note: we use the Ollama-compatible HF API, but not using the blobId. Instead, we use the special "ggufFile" field which returns the value for "hf_file". This is done to be backward-compatible with existing cache files.
*/
std::pair<std::string, std::string> common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & hf_token) {
auto parts = string_split<std::string>(hf_repo_with_tag, ':');
std::string tag = parts.size() > 1 ? parts.back() : "latest";
std::string hf_repo = parts[0];
if (string_split<std::string>(hf_repo, '/').size() != 2) {
throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
}
// fetch model info from Hugging Face Hub API
json model_info;
curl_ptr curl(curl_easy_init(), &curl_easy_cleanup);
curl_slist_ptr http_headers;
std::string res_str;
std::string url = "https://huggingface.co/v2/" + hf_repo + "/manifests/" + tag;
curl_easy_setopt(curl.get(), CURLOPT_URL, url.c_str());
curl_easy_setopt(curl.get(), CURLOPT_NOPROGRESS, 1L);
typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * ptr, size_t size, size_t nmemb, void * data);
auto write_callback = [](void * ptr, size_t size, size_t nmemb, void * data) -> size_t {
static_cast<std::string *>(data)->append((char * ) ptr, size * nmemb);
return size * nmemb;
};
curl_easy_setopt(curl.get(), CURLOPT_WRITEFUNCTION, static_cast<CURLOPT_WRITEFUNCTION_PTR>(write_callback));
curl_easy_setopt(curl.get(), CURLOPT_WRITEDATA, &res_str);
#if defined(_WIN32)
curl_easy_setopt(curl.get(), CURLOPT_SSL_OPTIONS, CURLSSLOPT_NATIVE_CA);
#endif
if (!hf_token.empty()) {
std::string auth_header = "Authorization: Bearer " + hf_token;
http_headers.ptr = curl_slist_append(http_headers.ptr, auth_header.c_str());
}
// Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
http_headers.ptr = curl_slist_append(http_headers.ptr, "Accept: application/json");
curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
CURLcode res = curl_easy_perform(curl.get());
if (res != CURLE_OK) {
throw std::runtime_error("error: cannot make GET request to HF API");
}
long res_code;
curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &res_code);
if (res_code == 200) {
model_info = json::parse(res_str);
} else if (res_code == 401) {
throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
} else {
throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
}
// check response
if (!model_info.contains("ggufFile")) {
throw std::runtime_error("error: model does not have ggufFile");
}
json & gguf_file = model_info.at("ggufFile");
if (!gguf_file.contains("rfilename")) {
throw std::runtime_error("error: ggufFile does not have rfilename");
}
return std::make_pair(hf_repo, gguf_file.at("rfilename"));
}
#else
struct llama_model * common_load_model_from_url(
@ -1463,6 +1604,11 @@ struct llama_model * common_load_model_from_hf(
return nullptr;
}
std::pair<std::string, std::string> common_get_hf_file(const std::string &, const std::string &) {
LOG_WRN("%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
return std::make_pair("", "");
}
#endif // LLAMA_USE_CURL
//
@ -1561,21 +1707,23 @@ std::vector<llama_token> common_tokenize(
const std::string & text,
bool add_special,
bool parse_special) {
return common_tokenize(llama_get_model(ctx), text, add_special, parse_special);
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_tokenize(vocab, text, add_special, parse_special);
}
std::vector<llama_token> common_tokenize(
const struct llama_model * model,
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special) {
// upper limit for the number of tokens
int n_tokens = text.length() + 2 * add_special;
std::vector<llama_token> result(n_tokens);
n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
if (n_tokens < 0) {
result.resize(-n_tokens);
int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
GGML_ASSERT(check == -n_tokens);
} else {
result.resize(n_tokens);
@ -1584,12 +1732,18 @@ std::vector<llama_token> common_tokenize(
}
std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_token_to_piece(vocab, token, special);
}
std::string common_token_to_piece(const struct llama_vocab * vocab, llama_token token, bool special) {
std::string piece;
piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
const int n_chars = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
if (n_chars < 0) {
piece.resize(-n_chars);
int check = llama_token_to_piece(llama_get_model(ctx), token, &piece[0], piece.size(), 0, special);
int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
GGML_ASSERT(check == -n_chars);
}
else {
@ -1599,13 +1753,19 @@ std::string common_token_to_piece(const struct llama_context * ctx, llama_token
return piece;
}
std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
std::string common_detokenize(const struct llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
return common_detokenize(vocab, tokens, special);
}
std::string common_detokenize(const struct llama_vocab * vocab, const std::vector<llama_token> & tokens, bool special) {
std::string text;
text.resize(std::max(text.capacity(), tokens.size()));
int32_t n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
int32_t n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
if (n_chars < 0) {
text.resize(-n_chars);
n_chars = llama_detokenize(llama_get_model(ctx), tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
n_chars = llama_detokenize(vocab, tokens.data(), (int32_t)tokens.size(), &text[0], (int32_t)text.size(), false, special);
GGML_ASSERT(n_chars <= (int32_t)text.size()); // whitespace trimming is performed after per-token detokenization
}
@ -1619,63 +1779,80 @@ std::string common_detokenize(llama_context * ctx, const std::vector<llama_token
// Chat template utils
//
bool common_chat_verify_template(const std::string & tmpl) {
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja) {
if (use_jinja) {
try {
auto chat_template = common_chat_template(tmpl, "<s>", "</s>");
common_chat_inputs inputs;
inputs.messages = json::array({{
{"role", "user"},
{"content", "test"},
}});
common_chat_params_init(chat_template, inputs);
return true;
} catch (const std::exception & e) {
LOG_ERR("%s: failed to apply template: %s\n", __func__, e.what());
return false;
}
}
llama_chat_message chat[] = {{"user", "test"}};
int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
const int res = llama_chat_apply_template(tmpl.c_str(), chat, 1, true, nullptr, 0);
return res >= 0;
}
std::string common_chat_apply_template(const struct llama_model * model,
const std::string & tmpl,
std::string common_chat_apply_template(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & msgs,
bool add_ass) {
bool add_ass,
bool use_jinja) {
if (use_jinja) {
auto messages = json::array();
for (const auto & msg : msgs) {
messages.push_back({{"role", msg.role}, {"content", msg.content}});
}
common_chat_inputs inputs;
inputs.messages = messages;
inputs.add_generation_prompt = add_ass;
return common_chat_params_init(tmpl, inputs).prompt;
}
int alloc_size = 0;
bool fallback = false; // indicate if we must fallback to default chatml
std::vector<llama_chat_message> chat;
for (auto & msg : msgs) {
for (const auto & msg : msgs) {
chat.push_back({msg.role.c_str(), msg.content.c_str()});
alloc_size += (msg.role.size() + msg.content.size()) * 1.25;
}
const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
std::vector<char> buf(alloc_size);
// run the first time to get the total output length
int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), add_ass, buf.data(), buf.size());
int32_t res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
// error: chat template is not supported
if (res < 0) {
if (ptr_tmpl != nullptr) {
// if the custom "tmpl" is not supported, we throw an error
// this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
throw std::runtime_error("this custom template is not supported");
} else {
// If the built-in template is not supported, we default to chatml
res = llama_chat_apply_template(nullptr, "chatml", chat.data(), chat.size(), add_ass, buf.data(), buf.size());
fallback = true;
}
// if the custom "tmpl" is not supported, we throw an error
// this is a bit redundant (for good), since we're not sure if user validated the custom template with llama_chat_verify_template()
throw std::runtime_error("this custom template is not supported");
}
// if it turns out that our buffer is too small, we resize it
if ((size_t) res > buf.size()) {
buf.resize(res);
res = llama_chat_apply_template(
fallback ? nullptr : model,
fallback ? "chatml" : ptr_tmpl,
chat.data(), chat.size(), add_ass, buf.data(), buf.size());
res = llama_chat_apply_template(tmpl.source().c_str(), chat.data(), chat.size(), add_ass, buf.data(), buf.size());
}
std::string formatted_chat(buf.data(), res);
return formatted_chat;
}
std::string common_chat_format_single(const struct llama_model * model,
const std::string & tmpl,
std::string common_chat_format_single(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass) {
bool add_ass,
bool use_jinja) {
std::ostringstream ss;
auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(model, tmpl, past_msg, false);
auto fmt_past_msg = past_msg.empty() ? "" : common_chat_apply_template(tmpl, past_msg, false, use_jinja);
std::vector<common_chat_msg> chat_new(past_msg);
// if the past_msg ends with a newline, we must preserve it in the formatted version
if (add_ass && !fmt_past_msg.empty() && fmt_past_msg.back() == '\n') {
@ -1683,21 +1860,87 @@ std::string common_chat_format_single(const struct llama_model * model,
};
// format chat with new_msg
chat_new.push_back(new_msg);
auto fmt_new_msg = common_chat_apply_template(model, tmpl, chat_new, add_ass);
auto fmt_new_msg = common_chat_apply_template(tmpl, chat_new, add_ass, use_jinja);
// get the diff part
ss << fmt_new_msg.substr(fmt_past_msg.size(), fmt_new_msg.size() - fmt_past_msg.size());
return ss.str();
}
std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl) {
std::string common_chat_format_example(const common_chat_template & tmpl, bool use_jinja) {
std::vector<common_chat_msg> msgs = {
{"system", "You are a helpful assistant"},
{"user", "Hello"},
{"assistant", "Hi there"},
{"user", "How are you?"},
{"system", "You are a helpful assistant", {}},
{"user", "Hello", {}},
{"assistant", "Hi there", {}},
{"user", "How are you?", {}},
};
return common_chat_apply_template(model, tmpl, msgs, true);
return common_chat_apply_template(tmpl, msgs, true, use_jinja);
}
#define CHATML_TEMPLATE_SRC \
"{%- for message in messages -%}\n" \
" {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>\n' -}}\n" \
"{%- endfor -%}\n" \
"{%- if add_generation_prompt -%}\n" \
" {{- '<|im_start|>assistant\n' -}}\n" \
"{%- endif -%}"
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
{
std::string default_template_src;
std::string template_tool_use_src;
bool has_explicit_template = !chat_template_override.empty();
if (chat_template_override.empty()) {
auto str = llama_model_chat_template(model, /* name */ nullptr);
if (str) {
default_template_src = str;
has_explicit_template = true;
}
str = llama_model_chat_template(model, /* name */ "tool_use");
if (str) {
template_tool_use_src = str;
has_explicit_template = true;
}
} else {
default_template_src = chat_template_override;
}
if (default_template_src.empty() || default_template_src == "chatml") {
if (!template_tool_use_src.empty()) {
default_template_src = template_tool_use_src;
} else {
default_template_src = CHATML_TEMPLATE_SRC;
}
}
auto vocab = llama_model_get_vocab(model);
const auto get_token = [&](llama_token token, const char * name, const char * jinja_variable_name) {
if (token == LLAMA_TOKEN_NULL) {
if (default_template_src.find(jinja_variable_name) != std::string::npos
|| template_tool_use_src.find(jinja_variable_name) != std::string::npos) {
LOG_WRN("%s: warning: vocab does not have a %s token, jinja template won't work as intended.\n", __func__, name);
}
return std::string();
} else {
return common_token_to_piece(vocab, token, true);
}
};
auto token_bos = get_token(llama_vocab_bos(vocab), "BOS", "bos_token");
auto token_eos = get_token(llama_vocab_eos(vocab), "EOS", "eos_token");
try {
return {
has_explicit_template,
std::make_unique<minja::chat_template>(default_template_src, token_bos, token_eos),
template_tool_use_src.empty()
? nullptr
: std::make_unique<minja::chat_template>(template_tool_use_src, token_bos, token_eos),
};
} catch (const std::exception & e) {
LOG_ERR("%s: failed to parse chat template: %s\n", __func__, e.what());
return {
has_explicit_template,
std::make_unique<minja::chat_template>(CHATML_TEMPLATE_SRC, token_bos, token_eos),
nullptr,
};
}
}
//

133
common/common.h
View file

@ -2,8 +2,9 @@
#pragma once
#include "llama.h"
#include "llama-cpp.h"
#include <set>
#include <string>
#include <vector>
#include <sstream>
@ -24,13 +25,11 @@
#define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
struct common_lora_adapter_info {
struct common_adapter_lora_info {
std::string path;
float scale;
};
struct common_lora_adapter_container : common_lora_adapter_info {
struct llama_lora_adapter * adapter;
struct llama_adapter_lora * ptr;
};
using llama_tokens = std::vector<llama_token>;
@ -105,6 +104,17 @@ enum dimre_method {
DIMRE_METHOD_MEAN,
};
enum common_conversation_mode {
COMMON_CONVERSATION_MODE_DISABLED = 0,
COMMON_CONVERSATION_MODE_ENABLED = 1,
COMMON_CONVERSATION_MODE_AUTO = 2,
};
struct common_grammar_trigger {
std::string word;
bool at_start;
};
// sampling parameters
struct common_params_sampling {
uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
@ -150,7 +160,11 @@ struct common_params_sampling {
COMMON_SAMPLER_TYPE_TEMPERATURE,
};
std::string grammar; // optional BNF-like grammar to constrain sampling
std::string grammar; // optional BNF-like grammar to constrain sampling
bool grammar_lazy = false;
std::vector<common_grammar_trigger> grammar_trigger_words; // optional trigger words to trigger lazy grammar
std::vector<llama_token> grammar_trigger_tokens; // optional trigger tokens to trigger lazy grammar and print trigger special tokens.
std::set<llama_token> preserved_tokens;
std::vector<llama_logit_bias> logit_bias; // logit biases to apply
@ -171,7 +185,11 @@ struct common_params_speculative {
struct cpu_params cpuparams;
struct cpu_params cpuparams_batch;
std::string model = ""; // draft model for speculative decoding // NOLINT
std::string hf_repo = ""; // HF repo // NOLINT
std::string hf_file = ""; // HF file // NOLINT
std::string model = ""; // draft model for speculative decoding // NOLINT
std::string model_url = ""; // model url to download // NOLINT
};
struct common_params_vocoder {
@ -180,6 +198,8 @@ struct common_params_vocoder {
std::string model = ""; // model path // NOLINT
std::string model_url = ""; // model url to download // NOLINT
bool use_guide_tokens = false; // enable guide tokens to improve TTS accuracy // NOLINT
};
struct common_params {
@ -242,14 +262,13 @@ struct common_params {
std::string lookup_cache_static = ""; // path of static ngram cache file for lookup decoding // NOLINT
std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding // NOLINT
std::string logits_file = ""; // file for saving *all* logits // NOLINT
std::string rpc_servers = ""; // comma separated list of RPC servers // NOLINT
std::vector<std::string> in_files; // all input files
std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
std::vector<llama_model_kv_override> kv_overrides;
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_adapter_lora_apply)
std::vector<common_adapter_lora_info> lora_adapters; // lora adapter path with user defined scale
std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
@ -277,7 +296,6 @@ struct common_params {
bool special = false; // enable special token output
bool interactive = false; // interactive mode
bool interactive_first = false; // wait for user input immediately
bool conversation = false; // conversation mode (does not print special tokens and suffix/prefix)
bool prompt_cache_all = false; // save user input and generations to prompt cache
bool prompt_cache_ro = false; // open the prompt cache read-only and do not update it
@ -303,6 +321,8 @@ struct common_params {
ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
common_conversation_mode conversation_mode = COMMON_CONVERSATION_MODE_AUTO;
// multimodal models (see examples/llava)
std::string mmproj = ""; // path to multimodal projector // NOLINT
std::vector<std::string> image; // path to image file(s)
@ -324,6 +344,7 @@ struct common_params {
std::string hostname = "127.0.0.1";
std::string public_path = ""; // NOLINT
std::string chat_template = ""; // NOLINT
bool use_jinja = false; // NOLINT
bool enable_chat_template = true;
std::vector<std::string> api_keys;
@ -418,6 +439,10 @@ std::string string_format(const char * fmt, ...);
std::string string_strip(const std::string & str);
std::string string_get_sortable_timestamp();
std::string string_join(const std::vector<std::string> & values, const std::string & separator);
std::vector<std::string> string_split(const std::string & str, const std::string & delimiter);
std::string string_repeat(const std::string & str, size_t n);
void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
template<class T>
@ -456,6 +481,11 @@ static bool string_starts_with(const std::string & str,
return str.rfind(prefix, 0) == 0;
}
static bool string_ends_with(const std::string & str,
const std::string & suffix) { // While we wait for C++20's std::string::ends_with...
return str.size() >= suffix.size() && str.compare(str.size()-suffix.size(), suffix.size(), suffix) == 0;
}
bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
void string_process_escapes(std::string & input);
@ -478,10 +508,12 @@ std::string fs_get_cache_file(const std::string & filename);
// Model utils
//
// note: defines object's lifetime
struct common_init_result {
struct llama_model * model = nullptr;
struct llama_context * context = nullptr;
std::vector<common_lora_adapter_container> lora_adapters;
llama_model_ptr model;
llama_context_ptr context;
std::vector<llama_adapter_lora_ptr> lora;
};
struct common_init_result common_init_from_params(common_params & params);
@ -495,6 +527,7 @@ struct llama_model * common_load_model_from_url(
const std::string & local_path,
const std::string & hf_token,
const struct llama_model_params & params);
struct llama_model * common_load_model_from_hf(
const std::string & repo,
const std::string & remote_path,
@ -502,8 +535,12 @@ struct llama_model * common_load_model_from_hf(
const std::string & hf_token,
const struct llama_model_params & params);
std::pair<std::string, std::string> common_get_hf_file(
const std::string & hf_repo_with_tag,
const std::string & hf_token);
// clear LoRA adapters from context, then apply new list of adapters
void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
//
// Batch utils
@ -541,7 +578,7 @@ std::vector<llama_token> common_tokenize(
bool parse_special = false);
std::vector<llama_token> common_tokenize(
const struct llama_model * model,
const struct llama_vocab * vocab,
const std::string & text,
bool add_special,
bool parse_special = false);
@ -553,11 +590,21 @@ std::string common_token_to_piece(
llama_token token,
bool special = true);
std::string common_token_to_piece(
const struct llama_vocab * vocab,
llama_token token,
bool special = true);
// detokenizes a vector of tokens into a string
// should work similar to Python's `tokenizer.decode`
// optionally renders special/control tokens
std::string common_detokenize(
llama_context * ctx,
const struct llama_context * ctx,
const std::vector<llama_token> & tokens,
bool special = true);
std::string common_detokenize(
const struct llama_vocab * vocab,
const std::vector<llama_token> & tokens,
bool special = true);
@ -565,33 +612,57 @@ std::string common_detokenize(
// Chat template utils
//
struct common_tool_call {
std::string name;
std::string arguments;
std::string id;
};
// same with llama_chat_message, but uses std::string
struct common_chat_msg {
std::string role;
std::string content;
std::vector<common_tool_call> tool_calls;
std::string tool_plan = "";
};
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
bool common_chat_verify_template(const std::string & tmpl);
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);
namespace minja {
class chat_template;
}
typedef minja::chat_template common_chat_template;
struct common_chat_templates {
bool has_explicit_template; // Model had builtin template or template overridde was specified.
std::unique_ptr<common_chat_template> template_default; // always set (defaults to chatml)
std::unique_ptr<common_chat_template> template_tool_use;
};
// CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml
// If the custom "tmpl" is not supported, we throw an error
std::string common_chat_apply_template(const struct llama_model * model,
const std::string & tmpl,
std::string common_chat_apply_template(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & chat,
bool add_ass);
bool add_ass,
bool use_jinja);
// Format single message, while taking into account the position of that message in chat history
std::string common_chat_format_single(const struct llama_model * model,
const std::string & tmpl,
std::string common_chat_format_single(
const common_chat_template & tmpl,
const std::vector<common_chat_msg> & past_msg,
const common_chat_msg & new_msg,
bool add_ass);
bool add_ass,
bool use_jinja);
// Returns an example of formatted chat
std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl);
std::string common_chat_format_example(
const common_chat_template & tmpl, bool use_jinja);
common_chat_templates common_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override);
//
// KV cache utils
@ -637,6 +708,10 @@ common_control_vector_data common_control_vector_load(const std::vector<common_c
// Split utils
//
static const char * const LLM_KV_SPLIT_NO = "split.no";
static const char * const LLM_KV_SPLIT_COUNT = "split.count";
static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
namespace {
const char * const LLM_KV_SPLIT_NO = "split.no";
const char * const LLM_KV_SPLIT_COUNT = "split.count";
const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
}

112
common/json-schema-to-grammar.cpp
View file

@ -1,4 +1,6 @@
#include "json-schema-to-grammar.h"
#include "common.h"
#include <algorithm>
#include <fstream>
#include <map>
@ -11,11 +13,6 @@
using json = nlohmann::ordered_json;
template <typename Iterator>
static std::string join(Iterator begin, Iterator end, const std::string & separator);
static std::string repeat(const std::string & str, size_t n);
static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
auto has_max = max_items != std::numeric_limits<int>::max();
@ -128,8 +125,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
if (sub_len > 0) {
auto from_sub = from.substr(i + 1);
auto to_sub = to.substr(i + 1);
auto sub_zeros = repeat("0", sub_len);
auto sub_nines = repeat("9", sub_len);
auto sub_zeros = string_repeat("0", sub_len);
auto sub_nines = string_repeat("9", sub_len);
auto to_reached = false;
out << "(";
@ -188,8 +185,8 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
auto max_digits = max_s.length();
for (auto digits = min_digits; digits < max_digits; digits++) {
uniform_range(min_s, repeat("9", digits));
min_s = "1" + repeat("0", digits);
uniform_range(min_s, string_repeat("9", digits));
min_s = "1" + string_repeat("0", digits);
out << " | ";
}
uniform_range(min_s, max_s);
@ -318,49 +315,6 @@ std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'^', '$', '.', '[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
template <typename Iterator>
std::string join(Iterator begin, Iterator end, const std::string & separator) {
std::ostringstream result;
if (begin != end) {
result << *begin;
for (Iterator it = begin + 1; it != end; ++it) {
result << separator << *it;
}
}
return result.str();
}
static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
std::vector<std::string> tokens;
size_t start = 0;
size_t end = str.find(delimiter);
while (end != std::string::npos) {
tokens.push_back(str.substr(start, end - start));
start = end + delimiter.length();
end = str.find(delimiter, start);
}
tokens.push_back(str.substr(start));
return tokens;
}
static std::string repeat(const std::string & str, size_t n) {
if (n == 0) {
return "";
}
std::string result;
result.reserve(str.length() * n);
for (size_t i = 0; i < n; ++i) {
result += str;
}
return result;
}
static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch &)> & replacement) {
std::smatch match;
std::string result;
@ -389,6 +343,7 @@ static std::string format_literal(const std::string & literal) {
class SchemaConverter {
private:
friend std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options);
std::function<json(const std::string &)> _fetch_json;
bool _dotall;
std::map<std::string, std::string> _rules;
@ -418,7 +373,7 @@ private:
for (size_t i = 0; i < alt_schemas.size(); i++) {
rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
}
return join(rules.begin(), rules.end(), " | ");
return string_join(rules, " | ");
}
std::string _visit_pattern(const std::string & pattern, const std::string & name) {
@ -481,7 +436,7 @@ private:
for (const auto & item : ret) {
results.push_back(to_rule(item));
}
return std::make_pair(join(results.begin(), results.end(), " "), false);
return std::make_pair(string_join(results, " "), false);
};
while (i < length) {
@ -539,7 +494,7 @@ private:
}
curly_brackets += '}';
i++;
auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
auto nums = string_split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
int min_times = 0;
int max_times = std::numeric_limits<int>::max();
try {
@ -809,10 +764,11 @@ private:
public:
SchemaConverter(
const std::function<json(const std::string &)> & fetch_json,
bool dotall)
bool dotall,
bool compact_spaces)
: _fetch_json(fetch_json), _dotall(dotall)
{
_rules["space"] = SPACE_RULE;
_rules["space"] = compact_spaces ? "\" \"?" : SPACE_RULE;
}
void resolve_refs(json & schema, const std::string & url) {
@ -854,7 +810,7 @@ public:
return;
}
std::string pointer = ref.substr(ref.find('#') + 1);
std::vector<std::string> tokens = split(pointer, "/");
std::vector<std::string> tokens = string_split(pointer, "/");
for (size_t i = 1; i < tokens.size(); ++i) {
std::string sel = tokens[i];
if (target.is_null() || !target.contains(sel)) {
@ -905,7 +861,7 @@ public:
for (const auto & v : schema["enum"]) {
enum_values.push_back(_generate_constant_rule(v));
}
return _add_rule(rule_name, "(" + join(enum_values.begin(), enum_values.end(), " | ") + ") space");
return _add_rule(rule_name, "(" + string_join(enum_values, " | ") + ") space");
} else if ((schema_type.is_null() || schema_type == "object")
&& (schema.contains("properties") ||
(schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
@ -1019,10 +975,10 @@ public:
void check_errors() {
if (!_errors.empty()) {
throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
throw std::runtime_error("JSON schema conversion failed:\n" + string_join(_errors, "\n"));
}
if (!_warnings.empty()) {
fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", string_join(_warnings, "; ").c_str());
}
}
@ -1035,11 +991,35 @@ public:
}
};
std::string json_schema_to_grammar(const json & schema) {
SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
auto copy = schema;
converter.resolve_refs(copy, "input");
converter.visit(copy, "");
std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
#ifdef LLAMA_USE_LLGUIDANCE
if (!force_gbnf) {
return "%llguidance {}\nstart: %json " + schema.dump();
}
#else
(void)force_gbnf;
#endif // LLAMA_USE_LLGUIDANCE
return build_grammar([&](const common_grammar_builder & callbacks) {
auto copy = schema;
callbacks.resolve_refs(copy);
callbacks.add_schema("", copy);
});
}
std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options) {
SchemaConverter converter([&](const std::string &) { return json(); }, options.dotall, options.compact_spaces);
common_grammar_builder builder {
/* .add_rule = */ [&](const std::string & name, const std::string & rule) {
return converter._add_rule(name, rule);
},
/* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
return converter.visit(schema, name == "root" ? "" : name);
},
/* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
converter.resolve_refs(schema, "");
}
};
cb(builder);
converter.check_errors();
return converter.format_grammar();
}

16
common/json-schema-to-grammar.h
View file

@ -5,4 +5,18 @@
#define JSON_ASSERT GGML_ASSERT
#include "json.hpp"
std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);
std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
bool force_gbnf = false);
struct common_grammar_builder {
std::function<std::string(const std::string &, const std::string &)> add_rule;
std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
std::function<void(nlohmann::ordered_json &)> resolve_refs;
};
struct common_grammar_options {
bool dotall = false;
bool compact_spaces = false;
};
std::string build_grammar(const std::function<void(const common_grammar_builder &)> & cb, const common_grammar_options & options = {});

270
common/llguidance.cpp Normal file
View file

@ -0,0 +1,270 @@
#include "sampling.h"
#include "log.h"
#ifdef LLAMA_USE_LLGUIDANCE
# include "llguidance.h"
# include <cmath>
struct llama_sampler_llg {
const llama_vocab * vocab;
std::string grammar_kind;
std::string grammar_data;
LlgTokenizer * tokenizer;
LlgConstraint * grammar;
LlgMaskResult llg_res;
bool has_llg_res;
};
static LlgConstraint * llama_sampler_llg_new(LlgTokenizer * tokenizer, const char * grammar_kind,
const char * grammar_data) {
LlgConstraintInit cinit;
llg_constraint_init_set_defaults(&cinit, tokenizer);
const char * log_level = getenv("LLGUIDANCE_LOG_LEVEL");
if (log_level && *log_level) {
cinit.log_stderr_level = atoi(log_level);
}
auto c = llg_new_constraint_any(&cinit, grammar_kind, grammar_data);
if (llg_get_error(c)) {
LOG_ERR("llg error: %s\n", llg_get_error(c));
llg_free_constraint(c);
return nullptr;
}
return c;
}
static const char * llama_sampler_llg_name(const llama_sampler * /*smpl*/) {
return "llguidance";
}
static void llama_sampler_llg_accept_impl(llama_sampler * smpl, llama_token token) {
auto * ctx = (llama_sampler_llg *) smpl->ctx;
if (ctx->grammar) {
LlgCommitResult res;
llg_commit_token(ctx->grammar, token, &res);
ctx->has_llg_res = false;
}
}
static void llama_sampler_llg_apply(llama_sampler * smpl, llama_token_data_array * cur_p) {
auto * ctx = (llama_sampler_llg *) smpl->ctx;
if (ctx->grammar) {
if (!ctx->has_llg_res) {
if (llg_compute_mask(ctx->grammar, &ctx->llg_res) == 0) {
ctx->has_llg_res = true;
} else {
LOG_ERR("llg error: %s\n", llg_get_error(ctx->grammar));
llg_free_constraint(ctx->grammar);
ctx->grammar = nullptr;
}
}
if (ctx->has_llg_res) {
if (ctx->llg_res.is_stop) {
for (size_t i = 0; i < cur_p->size; ++i) {
if (!llama_vocab_is_eog(ctx->vocab, cur_p->data[i].id)) {
cur_p->data[i].logit = -INFINITY;
}
}
} else {
const uint32_t * mask = ctx->llg_res.sample_mask;
for (size_t i = 0; i < cur_p->size; ++i) {
auto token = cur_p->data[i].id;
if ((mask[token / 32] & (1 << (token % 32))) == 0) {
cur_p->data[i].logit = -INFINITY;
}
}
}
}
}
}
static void llama_sampler_llg_reset(llama_sampler * smpl) {
auto * ctx = (llama_sampler_llg *) smpl->ctx;
if (!ctx->grammar) {
return;
}
auto * grammar_new = llama_sampler_llg_new(ctx->tokenizer, ctx->grammar_kind.c_str(), ctx->grammar_data.c_str());
llg_free_constraint(ctx->grammar);
ctx->grammar = grammar_new;
ctx->has_llg_res = false;
}
static llama_sampler * llama_sampler_llg_clone(const llama_sampler * smpl) {
const auto * ctx = (const llama_sampler_llg *) smpl->ctx;
auto * result = llama_sampler_init_llg(ctx->vocab, nullptr, nullptr);
// copy the state
{
auto * result_ctx = (llama_sampler_llg *) result->ctx;
if (ctx->grammar) {
result_ctx->grammar_kind = ctx->grammar_kind;
result_ctx->grammar_data = ctx->grammar_data;
result_ctx->grammar = llg_clone_constraint(ctx->grammar);
result_ctx->tokenizer = llg_clone_tokenizer(ctx->tokenizer);
}
}
return result;
}
static void llama_sampler_llg_free(llama_sampler * smpl) {
const auto * ctx = (llama_sampler_llg *) smpl->ctx;
if (ctx->grammar) {
llg_free_constraint(ctx->grammar);
llg_free_tokenizer(ctx->tokenizer);
}
delete ctx;
}
static llama_sampler_i llama_sampler_llg_i = {
/* .name = */ llama_sampler_llg_name,
/* .accept = */ llama_sampler_llg_accept_impl,
/* .apply = */ llama_sampler_llg_apply,
/* .reset = */ llama_sampler_llg_reset,
/* .clone = */ llama_sampler_llg_clone,
/* .free = */ llama_sampler_llg_free,
};
static size_t llama_sampler_llg_tokenize_fn(const void * user_data, const uint8_t * bytes, size_t bytes_len,
uint32_t * output_tokens, size_t output_tokens_len) {
const llama_vocab * vocab = (const llama_vocab *) user_data;
int r = 0;
try {
r = llama_tokenize(vocab, (const char *) bytes, bytes_len, (int32_t *) output_tokens, output_tokens_len, false,
true);
} catch (const std::exception & e) {
GGML_ABORT("llama_tokenize failed: %s\n", e.what());
}
if (r < 0) {
return -r;
}
return r;
}
static LlgTokenizer * llama_sampler_llg_new_tokenizer(const llama_vocab * vocab) {
// TODO store the tokenizer in the vocab somehow
static const llama_vocab * vocab_cache;
static LlgTokenizer * tokenizer_cache;
if (vocab_cache == vocab) {
return llg_clone_tokenizer(tokenizer_cache);
}
auto tok_eos = llama_vocab_eot(vocab);
if (tok_eos == LLAMA_TOKEN_NULL) {
tok_eos = llama_vocab_eos(vocab);
}
size_t vocab_size = llama_vocab_n_tokens(vocab);
auto token_lens = new uint32_t[vocab_size];
// we typically have ~7 bytes per token; let's go on the safe side here
auto token_bytes_size = vocab_size * 16 + 1024 * 1024;
auto token_bytes = new uint8_t[token_bytes_size];
size_t offset = 0;
for (size_t i = 0; i < vocab_size; i++) {
size_t max_token = 1024;
if (token_bytes_size - offset < max_token) {
GGML_ABORT("token_bytes buffer too small\n");
}
llama_token token = i;
auto dp = (char *) token_bytes + offset;
auto size = llama_detokenize(vocab, &token, 1, dp, max_token, false, false);
if (size < 0) {
GGML_ABORT("llama_detokenize failed\n");
}
if (size == 0) {
size = llama_detokenize(vocab, &token, 1, dp + 1, max_token - 1, false, true);
if (size < 0) {
GGML_ABORT("llama_detokenize failed\n");
}
if (size != 0) {
*dp = '\xff'; // special token prefix marker
size += 1;
}
}
token_lens[i] = size;
offset += size;
}
LlgTokenizerInit tinit = {
/* .vocab_size = */ (uint32_t) vocab_size,
/* .tok_eos = */ (uint32_t) tok_eos,
/* .token_lens = */ token_lens,
/* .token_bytes = */ token_bytes,
/* .tokenizer_json = */ nullptr,
/* .tokenize_assumes_string = */ true,
/* .tokenize_fn = */ llama_sampler_llg_tokenize_fn,
/* .use_approximate_greedy_tokenize_fn = */ false,
/* .tokenize_user_data = */ vocab,
};
char error_buffer[1024];
LlgTokenizer * tokenizer = llg_new_tokenizer(&tinit, error_buffer, sizeof(error_buffer));
delete[] token_bytes;
delete[] token_lens;
if (tokenizer == nullptr) {
LOG_ERR("llg tokenizer error: %s\n", error_buffer);
return tokenizer;
}
if (tokenizer_cache) {
llg_free_tokenizer(tokenizer_cache);
}
vocab_cache = vocab;
tokenizer_cache = tokenizer;
return llg_clone_tokenizer(tokenizer_cache);
}
llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab, const char * grammar_kind,
const char * grammar_data) {
auto * ctx = new llama_sampler_llg;
if (grammar_kind != nullptr && grammar_kind[0] != '\0') {
auto tokenizer = llama_sampler_llg_new_tokenizer(vocab);
*ctx = {
/* .vocab = */ vocab,
/* .grammar_kind = */ grammar_kind,
/* .grammar_data = */ grammar_data,
/* .tokenizer = */ tokenizer,
/* .grammar = */ llama_sampler_llg_new(tokenizer, grammar_kind, grammar_data),
/* .llg_res = */ {},
/* .has_llg_res = */ false,
};
} else {
*ctx = {
/* .vocab = */ vocab,
/* .grammar_kind = */ {},
/* .grammar_data = */ {},
/* .tokenizer = */ nullptr,
/* .grammar = */ nullptr,
/* .llg_res = */ {},
/* .has_llg_res = */ false,
};
}
return llama_sampler_init(
/* .iface = */ &llama_sampler_llg_i,
/* .ctx = */ ctx
);
}
#else
llama_sampler * llama_sampler_init_llg(const llama_vocab *, const char *, const char *) {
LOG_WRN("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
return nullptr;
}
#endif // LLAMA_USE_LLGUIDANCE

11
common/log.cpp
View file

@ -14,16 +14,6 @@ void common_log_set_verbosity_thold(int verbosity) {
common_log_verbosity_thold = verbosity;
}
#define LOG_COL_DEFAULT "\033[0m"
#define LOG_COL_BOLD "\033[1m"
#define LOG_COL_RED "\033[31m"
#define LOG_COL_GREEN "\033[32m"
#define LOG_COL_YELLOW "\033[33m"
#define LOG_COL_BLUE "\033[34m"
#define LOG_COL_MAGENTA "\033[35m"
#define LOG_COL_CYAN "\033[36m"
#define LOG_COL_WHITE "\033[37m"
static int64_t t_us() {
return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
}
@ -206,6 +196,7 @@ public:
vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
}
#endif
va_end(args_copy);
}
entry.level = level;

11
common/log.h
View file

@ -2,6 +2,17 @@
#include "ggml.h" // for ggml_log_level
#define LOG_CLR_TO_EOL "\033[K\r"
#define LOG_COL_DEFAULT "\033[0m"
#define LOG_COL_BOLD "\033[1m"
#define LOG_COL_RED "\033[31m"
#define LOG_COL_GREEN "\033[32m"
#define LOG_COL_YELLOW "\033[33m"
#define LOG_COL_BLUE "\033[34m"
#define LOG_COL_MAGENTA "\033[35m"
#define LOG_COL_CYAN "\033[36m"
#define LOG_COL_WHITE "\033[37m"
#ifndef __GNUC__
# define LOG_ATTRIBUTE_FORMAT(...)
#elif defined(__MINGW32__)

2868
common/minja.hpp Normal file
View file

File diff suppressed because it is too large Load diff

24
common/ngram-cache.cpp
View file

@ -65,13 +65,13 @@ constexpr int draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
if (part_static_it == nc_static.end()) {
return -1;
return LLAMA_TOKEN_NULL;
}
const common_ngram_cache_part part_static = part_static_it->second;
int max_count_static = 0;
int sum_count_static = 0;
llama_token max_token = -1;
llama_token max_token = LLAMA_TOKEN_NULL;
for (std::pair<llama_token, int> token_count_static : part_static) {
const llama_token token = token_count_static.first;
@ -85,10 +85,10 @@ static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram
}
if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
return -1;
return LLAMA_TOKEN_NULL;
}
if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
return -1;
return LLAMA_TOKEN_NULL;
}
return max_token;
}
@ -98,9 +98,9 @@ static llama_token try_draft(
common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
const int * min_sample_size, const int * min_percent) {
llama_token drafted_token = -1;
llama_token drafted_token = LLAMA_TOKEN_NULL;
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == LLAMA_TOKEN_NULL; --i) {
const common_ngram ngram_primary = ngrams_primary[i];
common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
@ -112,7 +112,7 @@ static llama_token try_draft(
int max_count_primary = 0;
int max_count_static = 0;
int sum_count_primary = 0;
llama_token max_token = -1;
llama_token max_token = LLAMA_TOKEN_NULL;
for (std::pair<llama_token, int> token_count_primary : part_primary) {
const llama_token token = token_count_primary.first;
@ -154,7 +154,7 @@ void common_ngram_cache_draft(
}
while ((int) draft.size()-1 < n_draft) {
llama_token drafted_token = -1;
llama_token drafted_token = LLAMA_TOKEN_NULL;
const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
common_ngram ngram_static;
@ -177,17 +177,17 @@ void common_ngram_cache_draft(
}
ngrams_cd.push_back(ngram_cd);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
drafted_token = try_draft(nc_static, ngram_static);
}
if (drafted_token == -1) {
if (drafted_token == LLAMA_TOKEN_NULL) {
break;
}

4
common/ngram-cache.h
View file

@ -17,13 +17,13 @@ struct common_ngram {
common_ngram() {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = -1;
tokens[i] = LLAMA_TOKEN_NULL;
}
}
common_ngram(const llama_token * input, const int ngram_size) {
for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
tokens[i] = i < ngram_size ? input[i] : -1;
tokens[i] = i < ngram_size ? input[i] : LLAMA_TOKEN_NULL;
}
}

38
common/sampling.cpp
View file

@ -113,7 +113,10 @@ struct common_sampler {
void set_logits(struct llama_context * ctx, int idx) {
const auto * logits = llama_get_logits_ith(ctx, idx);
const int n_vocab = llama_n_vocab(llama_get_model(ctx));
const llama_model * model = llama_get_model(ctx);
const llama_vocab * vocab = llama_model_get_vocab(model);
const int n_vocab = llama_vocab_n_tokens(vocab);
cur.resize(n_vocab);
@ -142,13 +145,36 @@ std::string common_params_sampling::print() const {
}
struct common_sampler * common_sampler_init(const struct llama_model * model, const struct common_params_sampling & params) {
const llama_vocab * vocab = llama_model_get_vocab(model);
llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
lparams.no_perf = params.no_perf;
std::vector<const char *> trigger_words;
trigger_words.reserve(params.grammar_trigger_words.size());
for (const auto & str : params.grammar_trigger_words) {
trigger_words.push_back(str.word.c_str());
}
struct llama_sampler * grmr;
if (params.grammar.compare(0, 11, "%llguidance") == 0) {
#ifdef LLAMA_USE_LLGUIDANCE
grmr = llama_sampler_init_llg(vocab, "lark", params.grammar.c_str());
#else
GGML_ABORT("llguidance (cmake -DLLAMA_LLGUIDANCE=ON) is not enabled");
#endif // LLAMA_USE_LLGUIDANCE
} else {
grmr = params.grammar_lazy
? llama_sampler_init_grammar_lazy(vocab, params.grammar.c_str(), "root",
trigger_words.data(), trigger_words.size(),
params.grammar_trigger_tokens.data(), params.grammar_trigger_tokens.size())
: llama_sampler_init_grammar(vocab, params.grammar.c_str(), "root");
}
auto * result = new common_sampler {
/* .params = */ params,
/* .grmr = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
/* .grmr = */ grmr,
/* .chain = */ llama_sampler_chain_init(lparams),
/* .prev = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
/* .cur = */ {},
@ -157,7 +183,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain,
llama_sampler_init_logit_bias(
llama_n_vocab(model),
llama_vocab_n_tokens(vocab),
params.logit_bias.size(),
params.logit_bias.data()));
@ -172,7 +198,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
c_breakers.push_back(str.c_str());
}
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
llama_sampler_chain_add(result->chain, llama_sampler_init_dry (vocab, llama_model_n_ctx_train(model), params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
}
break;
case COMMON_SAMPLER_TYPE_TOP_K:
@ -194,7 +220,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
break;
case COMMON_SAMPLER_TYPE_INFILL:
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (model));
llama_sampler_chain_add(result->chain, llama_sampler_init_infill (vocab));
break;
case COMMON_SAMPLER_TYPE_PENALTIES:
llama_sampler_chain_add(result->chain, llama_sampler_init_penalties(params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present));
@ -206,7 +232,7 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
} else if (params.mirostat == 1) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_vocab_n_tokens(vocab), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
} else if (params.mirostat == 2) {
llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));

3
common/sampling.h
View file

@ -102,3 +102,6 @@ std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);
llama_sampler * llama_sampler_init_llg(const llama_vocab * vocab,
const char * grammar_kind, const char * grammar_data);

33
common/speculative.cpp
View file

@ -79,10 +79,13 @@ bool common_speculative_are_compatible(
const struct llama_model * model_tgt = llama_get_model(ctx_tgt);
const struct llama_model * model_dft = llama_get_model(ctx_dft);
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
const struct llama_vocab * vocab_tgt = llama_model_get_vocab(model_tgt);
const struct llama_vocab * vocab_dft = llama_model_get_vocab(model_dft);
const bool vocab_type_tgt = llama_vocab_type(vocab_tgt);
LOG_DBG("%s: vocab_type tgt: %d\n", __func__, vocab_type_tgt);
const bool vocab_type_dft = llama_vocab_type(model_dft);
const bool vocab_type_dft = llama_vocab_type(vocab_dft);
LOG_DBG("%s: vocab_type dft: %d\n", __func__, vocab_type_dft);
if (vocab_type_tgt != vocab_type_dft) {
@ -91,34 +94,34 @@ bool common_speculative_are_compatible(
return false;
}
if (llama_add_bos_token(model_tgt) != llama_add_bos_token(model_dft) ||
llama_add_eos_token(model_tgt) != llama_add_eos_token(model_dft) ||
llama_token_bos(model_tgt) != llama_token_bos(model_dft) ||
llama_token_eos(model_tgt) != llama_token_eos(model_dft)) {
LOG_ERR("%s: draft model special tokens must match target model to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_tgt), llama_add_bos_token(model_tgt), llama_token_eos(model_tgt), llama_add_eos_token(model_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_token_bos(model_dft), llama_add_bos_token(model_dft), llama_token_eos(model_dft), llama_add_eos_token(model_dft));
if (llama_vocab_get_add_bos(vocab_tgt) != llama_vocab_get_add_bos(vocab_dft) ||
llama_vocab_get_add_eos(vocab_tgt) != llama_vocab_get_add_eos(vocab_dft) ||
llama_vocab_bos(vocab_tgt) != llama_vocab_bos(vocab_dft) ||
llama_vocab_eos(vocab_tgt) != llama_vocab_eos(vocab_dft)) {
LOG_ERR("%s: draft vocab special tokens must match target vocab to use speculation\n", __func__);
LOG_ERR("%s: tgt: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_tgt), llama_vocab_get_add_bos(vocab_tgt), llama_vocab_eos(vocab_tgt), llama_vocab_get_add_eos(vocab_tgt));
LOG_ERR("%s: dft: bos = %d (%d), eos = %d (%d)\n", __func__, llama_vocab_bos(vocab_dft), llama_vocab_get_add_bos(vocab_dft), llama_vocab_eos(vocab_dft), llama_vocab_get_add_eos(vocab_dft));
return false;
}
{
const int n_vocab_tgt = llama_n_vocab(model_tgt);
const int n_vocab_dft = llama_n_vocab(model_dft);
const int n_vocab_tgt = llama_vocab_n_tokens(vocab_tgt);
const int n_vocab_dft = llama_vocab_n_tokens(vocab_dft);
const int vocab_diff = std::abs(n_vocab_tgt - n_vocab_dft);
if (vocab_diff > SPEC_VOCAB_MAX_SIZE_DIFFERENCE) {
LOG_ERR("%s: draft model vocab must closely match target model to use speculation but "
"target vocab size %d does not match draft vocab size %d - difference %d, max allowed %d\n",
__func__, n_vocab_tgt, llama_n_vocab(model_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
__func__, n_vocab_tgt, llama_vocab_n_tokens(vocab_dft), vocab_diff, SPEC_VOCAB_MAX_SIZE_DIFFERENCE);
return false;
}
for (int i = SPEC_VOCAB_CHECK_START_TOKEN_ID; i < std::min(n_vocab_tgt, n_vocab_dft); ++i) {
const char * token_text_tgt = llama_token_get_text(model_tgt, i);
const char * token_text_dft = llama_token_get_text(model_dft, i);
const char * token_text_tgt = llama_vocab_get_text(vocab_tgt, i);
const char * token_text_dft = llama_vocab_get_text(vocab_dft, i);
if (std::strcmp(token_text_tgt, token_text_dft) != 0) {
LOG_ERR("%s: draft model vocab must match target model to use speculation but "
LOG_ERR("%s: draft vocab vocab must match target vocab to use speculation but "
"token %d content differs - target '%s', draft '%s'\n", __func__, i,
common_token_to_piece(ctx_tgt, i).c_str(),
common_token_to_piece(ctx_dft, i).c_str());