Skip to content

Commit

Permalink
Add --jinja and --chat-template-file flags
Browse files Browse the repository at this point in the history
  • Loading branch information
@ochafik
ochafik committed Dec 30, 2024
1 parent 3cbdf35 commit 4ec6151
Show file tree
Hide file tree
Showing 12 changed files with 288 additions and 50 deletions.
2 changes: 2 additions & 0 deletions Makefile
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1361,7 +1361,9 @@ llama-server: \
examples/server/httplib.h \
examples/server/index.html.hpp \
examples/server/loading.html.hpp \
common/chat-template.hpp \
common/json.hpp \
common/minja.hpp \
$(OBJ_ALL)
$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
Expand Down
2 changes: 2 additions & 0 deletions common/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,7 @@ add_library(${TARGET} STATIC
arg.cpp
arg.h
base64.hpp
chat-template.hpp
common.cpp
common.h
console.cpp
Expand All @@ -64,6 +65,7 @@ add_library(${TARGET} STATIC
json.hpp
log.cpp
log.h
minja.hpp
ngram-cache.cpp
ngram-cache.h
sampling.cpp
Expand Down
43 changes: 39 additions & 4 deletions common/arg.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1889,24 +1889,59 @@ 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}));
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)) {
if (!common_chat_verify_template(value, params.use_jinja)) {
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()
"error: the supplied chat template is not supported: %s%s\n",
value.c_str(),
params.use_jinja ? "" : "\nnote: llama.cpp does not use jinja parser, we only support commonly used templates"
));
}
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",
"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"
"https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template",
[](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::string chat_template;
std::copy(
std::istreambuf_iterator<char>(file),
std::istreambuf_iterator<char>(),
std::back_inserter(chat_template)
);
if (!common_chat_verify_template(chat_template, params.use_jinja)) {
throw std::runtime_error(string_format(
"error: the supplied chat template is not supported: %s%s\n",
value.c_str(),
params.use_jinja ? "" : "\nnote: llama.cpp does not use jinja parser, we only support commonly used templates"
));
}
params.chat_template = 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),
Expand Down
68 changes: 64 additions & 4 deletions common/common.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1576,13 +1576,13 @@ std::vector<llama_token> common_tokenize(
return result;
}

std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
static std::string _common_token_to_piece(const struct llama_model * model, 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(model, 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(model, token, &piece[0], piece.size(), 0, special);
GGML_ASSERT(check == -n_chars);
}
else {
Expand All @@ -1592,6 +1592,10 @@ std::string common_token_to_piece(const struct llama_context * ctx, llama_token
return piece;
}

std::string common_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
return _common_token_to_piece(llama_get_model(ctx), token, special);
}

std::string common_detokenize(llama_context * ctx, const std::vector<llama_token> & tokens, bool special) {
std::string text;
text.resize(std::max(text.capacity(), tokens.size()));
Expand All @@ -1612,7 +1616,21 @@ 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 = minja::chat_template(tmpl, "<s>", "</s>");
chat_template.apply({{
{"role", "user"},
{"content", "test"},
}}, json(), true);
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);
return res >= 0;
Expand Down Expand Up @@ -1693,6 +1711,48 @@ std::string common_chat_format_example(const struct llama_model * model,
return common_chat_apply_template(model, tmpl, msgs, true);
}

static std::string _llama_model_meta_val_str(const struct llama_model * model, const char * key) {
int32_t tlen = llama_model_meta_val_str(model, key, nullptr, 0);
if (tlen > 0) {
std::vector<char> curr_tmpl_buf(tlen + 1, 0);
if (llama_model_meta_val_str(model, key, curr_tmpl_buf.data(), curr_tmpl_buf.size()) == tlen) {
return std::string(curr_tmpl_buf.data(), tlen);
}
}
return "";
}

llama_chat_templates llama_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override)
{
auto bos_token = _common_token_to_piece(model, llama_token_bos(model), true);
auto eos_token = _common_token_to_piece(model, llama_token_eos(model), true);
std::string default_template_src = chat_template_override;
std::string tool_use_template_src = chat_template_override;
if (chat_template_override.empty()) {
default_template_src = _llama_model_meta_val_str(model, "tokenizer.chat_template");
tool_use_template_src = _llama_model_meta_val_str(model, "tokenizer.chat_template.tool_use");
}
if (default_template_src.empty() || default_template_src == "chatml") {
if (!tool_use_template_src.empty()) {
default_template_src = tool_use_template_src;
} else {
default_template_src = R"(
{%- for message in messages -%}
{{- "<|im_start|>" + message.role + "\n" + message.content + "<|im_end|>\n" -}}
{%- endfor -%}
{%- if add_generation_prompt -%}
{{- "<|im_start|>assistant\n" -}}
{%- endif -%}
)";
}
}
return {
.default_template = { default_template_src, bos_token, eos_token },
.tool_use_template = tool_use_template_src.empty() ? std::nullopt
: std::optional<minja::chat_template>({ tool_use_template_src, bos_token, eos_token }),
};
}

//
// KV cache utils
//
Expand Down
14 changes: 12 additions & 2 deletions common/common.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
#pragma once

#include "llama.h"
#include "chat-template.hpp"

#include <string>
#include <vector>
Expand Down Expand Up @@ -324,6 +325,7 @@ 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;
Expand Down Expand Up @@ -571,8 +573,8 @@ struct common_chat_msg {
std::string content;
};

// 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);
// Check if the template is supported or not. Returns true if it's valid
bool common_chat_verify_template(const std::string & tmpl, bool use_jinja);

// CPP wrapper for llama_chat_apply_template
// If the built-in template is not supported, we default to chatml
Expand All @@ -593,6 +595,14 @@ std::string common_chat_format_single(const struct llama_model * model,
std::string common_chat_format_example(const struct llama_model * model,
const std::string & tmpl);


struct llama_chat_templates {
minja::chat_template default_template;
std::optional<minja::chat_template> tool_use_template;
};

llama_chat_templates llama_chat_templates_from_model(const struct llama_model * model, const std::string & chat_template_override);

//
// KV cache utils
//
Expand Down
2 changes: 1 addition & 1 deletion examples/server/README.md
View file
Original file line number Diff line number Diff line change
Expand Up @@ -129,7 +129,7 @@ The project is under active development, and we are [looking for feedback and co
| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
| `--grammar-file FNAME` | file to read grammar from |
| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |

| `--jinja` | Enable experimental Jinja templating engine (needed for tool use) |

**Example-specific params**

Expand Down
67 changes: 53 additions & 14 deletions examples/server/server.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1623,15 +1623,35 @@ struct server_context {
return true;
}

bool validate_model_chat_template() const {
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res >= 0) {
llama_chat_message chat[] = {{"user", "test"}};
std::string tmpl = std::string(model_template.data(), model_template.size());
int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
return chat_res > 0;
bool validate_model_chat_template(bool use_jinja) const {
llama_chat_message chat[] = {{"user", "test"}};

if (use_jinja) {
auto templates = llama_chat_templates_from_model(model, "");
try {
templates.default_template.apply({{
{"role", "user"},
{"content", "test"},
}}, json(), true);
if (templates.tool_use_template) {
templates.tool_use_template->apply({{
{"role", "user"},
{"content", "test"},
}}, json(), true);
}
return true;
} catch (const std::exception & e) {
SRV_ERR("failed to apply template: %s\n", e.what());
}
} else {
std::vector<char> model_template(2048, 0); // longest known template is about 1200 bytes
std::string template_key = "tokenizer.chat_template";
int32_t res = llama_model_meta_val_str(model, template_key.c_str(), model_template.data(), model_template.size());
if (res >= 0) {
std::string tmpl = std::string(model_template.data(), model_template.size());
int32_t chat_res = llama_chat_apply_template(model, tmpl.c_str(), chat, 1, true, nullptr, 0);
return chat_res > 0;
}
}
return false;
}
Expand Down Expand Up @@ -3476,15 +3496,30 @@ int main(int argc, char ** argv) {
}
};

const auto handle_props = [&ctx_server, &res_ok](const httplib::Request &, httplib::Response & res) {
std::mutex chat_templates_mutex;
std::optional<llama_chat_templates> chat_templates;

auto get_chat_templates = [&ctx_server, &chat_templates_mutex, &chat_templates]() -> const llama_chat_templates & {
std::lock_guard<std::mutex> lock(chat_templates_mutex);
if (!chat_templates) {
chat_templates = llama_chat_templates_from_model(ctx_server.model, ctx_server.params_base.chat_template);
}
return *chat_templates;
};

const auto handle_props = [&ctx_server, &res_ok, &get_chat_templates](const httplib::Request &, httplib::Response & res) {
// this endpoint is publicly available, please only return what is safe to be exposed
const auto & templates = get_chat_templates();
json data = {
{ "default_generation_settings", ctx_server.default_generation_settings_for_props },
{ "total_slots", ctx_server.params_base.n_parallel },
{ "model_path", ctx_server.params_base.model },
{ "chat_template", llama_get_chat_template(ctx_server.model) },
{ "chat_template", templates.default_template.source() },
{ "build_info", build_info },
};
if (ctx_server.params_base.use_jinja && templates.tool_use_template) {
data["chat_template_tool_use"] = templates.tool_use_template->source();
}

res_ok(res, data);
};
Expand Down Expand Up @@ -3685,13 +3720,17 @@ int main(int argc, char ** argv) {
return handle_completions_generic(SERVER_TASK_TYPE_INFILL, data, res);
};

const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_generic](const httplib::Request & req, httplib::Response & res) {
const auto handle_chat_completions = [&ctx_server, &params, &res_error, &handle_completions_generic, &get_chat_templates](const httplib::Request & req, httplib::Response & res) {
if (ctx_server.params_base.embedding) {
res_error(res, format_error_response("This server does not support completions. Start it without `--embeddings`", ERROR_TYPE_NOT_SUPPORTED));
return;
}

json data = oaicompat_completion_params_parse(ctx_server.model, json::parse(req.body), params.chat_template);
auto body = json::parse(req.body);
const auto & templates = get_chat_templates();
const auto & chat_template = body.contains("tools") && templates.tool_use_template ? *templates.tool_use_template : templates.default_template;
json data = oaicompat_completion_params_parse(ctx_server.model, body, chat_template, params.use_jinja);

return handle_completions_generic(
SERVER_TASK_TYPE_COMPLETION,
data,
Expand Down Expand Up @@ -4111,7 +4150,7 @@ int main(int argc, char ** argv) {

// if a custom chat template is not supplied, we will use the one that comes with the model (if any)
if (params.chat_template.empty()) {
if (!ctx_server.validate_model_chat_template()) {
if (!ctx_server.validate_model_chat_template(params.use_jinja)) {
LOG_WRN("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
params.chat_template = "chatml";
}
Expand Down
15 changes: 9 additions & 6 deletions examples/server/tests/unit/test_chat_completion.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -4,22 +4,24 @@

server = ServerPreset.tinyllama2()


@pytest.fixture(scope="module", autouse=True)
@pytest.fixture(autouse=True)
def create_server():
global server
server = ServerPreset.tinyllama2()


@pytest.mark.parametrize(
"model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason",
"model,system_prompt,user_prompt,max_tokens,re_content,n_prompt,n_predicted,finish_reason,jinja",
[
(None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length"),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length"),
(None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length", False),
(None, "Book", "What is the best book", 8, "(Suddenly)+", 77, 8, "length", True),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", False),
("codellama70b", "You are a coding assistant.", "Write the fibonacci function in c++.", 128, "(Aside|she|felter|alonger)+", 104, 64, "length", True),
]
)
def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason):
def test_chat_completion(model, system_prompt, user_prompt, max_tokens, re_content, n_prompt, n_predicted, finish_reason, jinja):
global server
server.jinja = jinja
server.start()
res = server.make_request("POST", "/chat/completions", data={
"model": model,
Expand Down Expand Up @@ -102,6 +104,7 @@ def test_chat_completion_with_openai_library():

@pytest.mark.parametrize("response_format,n_predicted,re_content", [
({"type": "json_object", "schema": {"const": "42"}}, 6, "\"42\""),
({"type": "json_schema", "json_schema": {"const": "42"}}, 6, "\"42\""),
({"type": "json_object", "schema": {"items": [{"type": "integer"}]}}, 10, "[ -3000 ]"),
({"type": "json_object"}, 10, "(\\{|John)+"),
({"type": "sound"}, 0, None),
Expand Down
Loading

0 comments on commit 4ec6151

Please sign in to comment.