common/chat : fix LFM2/LFM2.5 reasoning round-trip and <think> leak (#24234)
* common/chat : fix LFM2 reasoning round-trip and stray <think> leak * Gate by reasoning format and whether the template supports <think>
This commit is contained in:
Tarek Dakhran
+133
-175
@@ -1825,6 +1825,104 @@ static void test_convert_responses_to_chatcmpl() {
|
||||
}
|
||||
}
|
||||
|
||||
// Shared LFM2 parser cases - all variants use one output format and parser
|
||||
static void test_lfm2_parser(const std::string & template_path, bool detailed_debug) {
|
||||
auto tst = peg_tester(template_path, detailed_debug);
|
||||
|
||||
// Basic content only
|
||||
tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
|
||||
|
||||
// Single tool call without reasoning
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.tools({ special_function_tool })
|
||||
.expect(message_assist_call)
|
||||
.run();
|
||||
|
||||
// Tool call with string argument
|
||||
tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
|
||||
.run();
|
||||
|
||||
// Python literals become JSON
|
||||
tst.test("<|tool_call_start|>[toggle(enabled=True)]<|tool_call_end|>")
|
||||
.tools({ toggle_tool })
|
||||
.expect(message_with_tool_calls("toggle", R"({"enabled": true})"))
|
||||
.run();
|
||||
|
||||
tst.test("<|tool_call_start|>[set_nullable(value=None)]<|tool_call_end|>")
|
||||
.tools({ nullable_tool })
|
||||
.expect(message_with_tool_calls("set_nullable", R"({"value": null})"))
|
||||
.run();
|
||||
|
||||
// Nested Python literal
|
||||
tst.test("<|tool_call_start|>[set_config(config={\"enabled\": True, \"count\": 3})]<|tool_call_end|>")
|
||||
.tools({ config_tool })
|
||||
.expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "count": 3}})"))
|
||||
.run();
|
||||
|
||||
// JSON literals are accepted too
|
||||
tst.test("<|tool_call_start|>[set_config(config={\"enabled\": true, \"note\": null})]<|tool_call_end|>")
|
||||
.tools({ config_tool })
|
||||
.expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "note": null}})"))
|
||||
.run();
|
||||
|
||||
// Dotted function name with structured args
|
||||
tst.test("<|tool_call_start|>[Calendar.create_event(title=\"demo\", participants=[\"Alice\", \"Bob\"], "
|
||||
"metadata={\"priority\": \"high\", \"reminder\": true})]<|tool_call_end|>")
|
||||
.tools({ calendar_create_event_tool })
|
||||
.expect(message_with_tool_calls(
|
||||
"Calendar.create_event",
|
||||
R"({"title": "demo", "participants": ["Alice", "Bob"], "metadata": {"priority": "high", "reminder": true}})"))
|
||||
.run();
|
||||
|
||||
// Markdown links stay content
|
||||
tst.test("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org).")
|
||||
.tools({ get_time_tool })
|
||||
.expect(simple_assist_msg("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org)."))
|
||||
.run();
|
||||
|
||||
// Python tool with multiline code in string
|
||||
tst.test("<|tool_call_start|>[python(code=\"def hello():\\n print('hey')\")]<|tool_call_end|>")
|
||||
.tools({ python_tool })
|
||||
.expect_tool_calls({
|
||||
{ "python", R"#({"code": "def hello():\\n print('hey')"})#", "" }
|
||||
})
|
||||
.run();
|
||||
|
||||
// Content before tool call (no reasoning)
|
||||
tst.test("Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_reasoning_content_and_multiple_tool_calls(
|
||||
"", "Let me check the time.", { { "get_time", "{\"city\":\"Paris\"}" } }
|
||||
))
|
||||
.run();
|
||||
|
||||
// Multiple tool calls (parallel)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
|
||||
.parallel_tool_calls(true)
|
||||
.tools({ special_function_tool, special_function_tool_with_optional_param })
|
||||
.expect_tool_calls({
|
||||
{ "special_function", R"({"arg1": 1})", {} },
|
||||
{ "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
|
||||
})
|
||||
.run();
|
||||
|
||||
// Partial tool call (streaming)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=")
|
||||
.tools({ special_function_tool })
|
||||
.is_partial(true)
|
||||
.expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
|
||||
.run();
|
||||
|
||||
// Tool call with empty arguments
|
||||
tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
|
||||
.tools({ empty_args_tool })
|
||||
.expect(simple_assist_msg("", "", "empty_args", "{}"))
|
||||
.run();
|
||||
|
||||
}
|
||||
|
||||
static void test_template_output_peg_parsers(bool detailed_debug) {
|
||||
LOG_DBG("%s\n", __func__);
|
||||
|
||||
@@ -4038,49 +4136,30 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
|
||||
.run();
|
||||
}
|
||||
|
||||
// LFM2-8B-A1B tests - uses <|tool_list_start|>/<|tool_list_end|> and <|tool_call_start|>[name(args)]<|tool_call_end|>
|
||||
for (const char * tmpl : {
|
||||
"models/templates/LFM2-8B-A1B.jinja",
|
||||
"models/templates/LFM2.5-Instruct.jinja",
|
||||
"models/templates/LFM2.5-8B-A1B.jinja",
|
||||
}) {
|
||||
test_lfm2_parser(tmpl, detailed_debug);
|
||||
}
|
||||
|
||||
// Thinking cases only apply to LFM2.5-8B-A1B, the one LFM2 template that emits <think>
|
||||
{
|
||||
auto tst = peg_tester("models/templates/LFM2-8B-A1B.jinja", detailed_debug);
|
||||
auto tst = peg_tester("models/templates/LFM2.5-8B-A1B.jinja", detailed_debug);
|
||||
|
||||
// Basic content only
|
||||
tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
|
||||
// Reasoning is parsed independent of enable_thinking
|
||||
|
||||
// Single tool call without reasoning
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.tools({ special_function_tool })
|
||||
.expect(message_assist_call)
|
||||
.run();
|
||||
|
||||
// Tool call with string argument
|
||||
tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
|
||||
.run();
|
||||
|
||||
// Tool call with reasoning (enable_thinking=true)
|
||||
// Tool call with reasoning
|
||||
tst.test("<think>I'm\nthinking</think><|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.enable_thinking(true)
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.tools({ special_function_tool })
|
||||
.expect(message_assist_call_thoughts)
|
||||
.run();
|
||||
|
||||
// Multiple tool calls (parallel)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
|
||||
.parallel_tool_calls(true)
|
||||
.tools({
|
||||
special_function_tool, special_function_tool_with_optional_param
|
||||
})
|
||||
.expect_tool_calls({
|
||||
{ "special_function", R"({"arg1": 1})", {} },
|
||||
{ "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
|
||||
})
|
||||
.run();
|
||||
|
||||
// Tool call with reasoning and content
|
||||
tst.test("<think>I need to call a function</think>"
|
||||
"Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
|
||||
.enable_thinking(true)
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_reasoning_content_and_multiple_tool_calls(
|
||||
@@ -4088,32 +4167,9 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
|
||||
))
|
||||
.run();
|
||||
|
||||
// Python tool with multiline code in string
|
||||
tst.test("<|tool_call_start|>[python(code=\"def hello():\\n print('hey')\")]<|tool_call_end|>")
|
||||
.tools({ python_tool })
|
||||
.expect_tool_calls({
|
||||
{ "python", R"#({"code": "def hello():\\n print('hey')"})#", "" }
|
||||
})
|
||||
.run();
|
||||
|
||||
// Partial tool call (streaming)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=")
|
||||
.tools({ special_function_tool })
|
||||
.is_partial(true)
|
||||
.expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
|
||||
.run();
|
||||
|
||||
// Tool call with empty arguments
|
||||
tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
|
||||
.tools({ empty_args_tool })
|
||||
.expect(simple_assist_msg("", "", "empty_args", "{}"))
|
||||
.run();
|
||||
|
||||
// fake tool call marker in reasoning
|
||||
tst.test(
|
||||
"<think>Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm</think>"
|
||||
"<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.enable_thinking(true)
|
||||
// Fake tool call marker inside reasoning is not parsed as a call
|
||||
tst.test("<think>Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm</think>"
|
||||
"<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.tools({ special_function_tool })
|
||||
.expect_reasoning("Let me think about <|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|> hmm")
|
||||
@@ -4122,127 +4178,21 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
|
||||
})
|
||||
.run();
|
||||
|
||||
// Continuation tests
|
||||
tst.test("world!\nWhat's up?")
|
||||
// enable_thinking=false still captures emitted reasoning
|
||||
tst.test("<think>I'm\nthinking</think>Hello, world!\nWhat's up?")
|
||||
.enable_thinking(false)
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.enable_thinking(true)
|
||||
.messages({ message_user, message_assist_prefill_content })
|
||||
.add_generation_prompt(false)
|
||||
.continue_final_message(COMMON_CHAT_CONTINUATION_CONTENT)
|
||||
.expect_reasoning("I'm thinking")
|
||||
.expect_content("Hello, world!\nWhat's up?")
|
||||
.expect(message_assist_thoughts)
|
||||
.run();
|
||||
|
||||
tst.test(" thinking</think>Hello, world!\nWhat's up?")
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.enable_thinking(true)
|
||||
.messages({ message_user, message_assist_prefill_reasoning })
|
||||
.add_generation_prompt(false)
|
||||
.continue_final_message(COMMON_CHAT_CONTINUATION_REASONING)
|
||||
.expect_reasoning("I'm thinking")
|
||||
.expect_content("Hello, world!\nWhat's up?")
|
||||
.run();
|
||||
}
|
||||
|
||||
// LFM2.5 tests - format <|tool_call_start|>[name(args)]<|tool_call_end|>
|
||||
{
|
||||
auto tst = peg_tester("models/templates/LFM2.5-Instruct.jinja", detailed_debug);
|
||||
|
||||
// Basic content only
|
||||
tst.test("Hello, world!\nWhat's up?").expect(message_assist).run();
|
||||
|
||||
// Single tool call without reasoning
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.tools({ special_function_tool })
|
||||
.expect(message_assist_call)
|
||||
.run();
|
||||
|
||||
// Tool call with string argument
|
||||
tst.test("<|tool_call_start|>[get_time(city=\"XYZCITY\")]<|tool_call_end|>")
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_tool_calls("get_time", "{\"city\":\"XYZCITY\"}"))
|
||||
.run();
|
||||
|
||||
// Python literals become JSON.
|
||||
tst.test("<|tool_call_start|>[toggle(enabled=True)]<|tool_call_end|>")
|
||||
.tools({ toggle_tool })
|
||||
.expect(message_with_tool_calls("toggle", R"({"enabled": true})"))
|
||||
.run();
|
||||
|
||||
tst.test("<|tool_call_start|>[set_nullable(value=None)]<|tool_call_end|>")
|
||||
.tools({ nullable_tool })
|
||||
.expect(message_with_tool_calls("set_nullable", R"({"value": null})"))
|
||||
.run();
|
||||
|
||||
// Nested Python literal.
|
||||
tst.test("<|tool_call_start|>[set_config(config={\"enabled\": True, \"count\": 3})]<|tool_call_end|>")
|
||||
.tools({ config_tool })
|
||||
.expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "count": 3}})"))
|
||||
.run();
|
||||
|
||||
// JSON literals are accepted too.
|
||||
tst.test("<|tool_call_start|>[set_config(config={\"enabled\": true, \"note\": null})]<|tool_call_end|>")
|
||||
.tools({ config_tool })
|
||||
.expect(message_with_tool_calls("set_config", R"({"config": {"enabled": true, "note": null}})"))
|
||||
.run();
|
||||
|
||||
// Dotted function name with structured args.
|
||||
tst.test("<|tool_call_start|>[Calendar.create_event(title=\"demo\", participants=[\"Alice\", \"Bob\"], "
|
||||
"metadata={\"priority\": \"high\", \"reminder\": true})]<|tool_call_end|>")
|
||||
.tools({ calendar_create_event_tool })
|
||||
.expect(message_with_tool_calls(
|
||||
"Calendar.create_event",
|
||||
R"({"title": "demo", "participants": ["Alice", "Bob"], "metadata": {"priority": "high", "reminder": true}})"))
|
||||
.run();
|
||||
|
||||
// Markdown links stay content.
|
||||
tst.test("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org).")
|
||||
.tools({ get_time_tool })
|
||||
.expect(simple_assist_msg("Use this format: [link text](url). Example: [Wikipedia](https://www.wikipedia.org)."))
|
||||
.run();
|
||||
|
||||
// Tool call with reasoning (enable_thinking=true)
|
||||
tst.test("<think>I'm\nthinking</think><|tool_call_start|>[special_function(arg1=1)]<|tool_call_end|>")
|
||||
.enable_thinking(true)
|
||||
.enable_thinking(false)
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.tools({ special_function_tool })
|
||||
.expect(message_assist_call_thoughts)
|
||||
.run();
|
||||
|
||||
// Multiple tool calls (parallel)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=1), special_function_with_opt(arg1=1, arg2=2)]<|tool_call_end|>")
|
||||
.parallel_tool_calls(true)
|
||||
.tools({
|
||||
special_function_tool, special_function_tool_with_optional_param
|
||||
})
|
||||
.expect_tool_calls({
|
||||
{ "special_function", R"({"arg1": 1})", {} },
|
||||
{ "special_function_with_opt", R"({"arg1": 1, "arg2": 2})", {} },
|
||||
})
|
||||
.run();
|
||||
|
||||
// Tool call with content before tool call
|
||||
tst.test("Let me check the time.<|tool_call_start|>[get_time(city=\"Paris\")]<|tool_call_end|>")
|
||||
.tools({ get_time_tool })
|
||||
.expect(message_with_reasoning_content_and_multiple_tool_calls(
|
||||
"", "Let me check the time.", { { "get_time", "{\"city\":\"Paris\"}" } }
|
||||
))
|
||||
.run();
|
||||
|
||||
// Partial tool call (streaming)
|
||||
tst.test("<|tool_call_start|>[special_function(arg1=")
|
||||
.tools({ special_function_tool })
|
||||
.is_partial(true)
|
||||
.expect(simple_assist_msg("", "", "special_function", "{\"arg1\": "))
|
||||
.run();
|
||||
|
||||
// Tool call with empty arguments
|
||||
tst.test("<|tool_call_start|>[empty_args()]<|tool_call_end|>")
|
||||
.tools({ empty_args_tool })
|
||||
.expect(simple_assist_msg("", "", "empty_args", "{}"))
|
||||
.run();
|
||||
|
||||
// Continuation tests
|
||||
// Continuation: prefill content
|
||||
tst.test("world!\nWhat's up?")
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.enable_thinking(true)
|
||||
@@ -4253,6 +4203,7 @@ static void test_template_output_peg_parsers(bool detailed_debug) {
|
||||
.expect_content("Hello, world!\nWhat's up?")
|
||||
.run();
|
||||
|
||||
// Continuation: prefill reasoning
|
||||
tst.test(" thinking</think>Hello, world!\nWhat's up?")
|
||||
.reasoning_format(COMMON_REASONING_FORMAT_AUTO)
|
||||
.enable_thinking(true)
|
||||
@@ -5478,18 +5429,25 @@ static void test_template_generation_prompt() {
|
||||
check(tmpls, continuation_reasoning(), "<|im_assistant|>assistant<|im_middle|><think>I'm");
|
||||
}
|
||||
|
||||
{
|
||||
auto tmpls = read_templates("models/templates/LFM2-8B-A1B.jinja");
|
||||
for (const char * tmpl : {
|
||||
"models/templates/LFM2-8B-A1B.jinja",
|
||||
"models/templates/LFM2.5-Instruct.jinja",
|
||||
"models/templates/LFM2.5-8B-A1B.jinja",
|
||||
}) {
|
||||
auto tmpls = read_templates(tmpl);
|
||||
check(tmpls, basic(), "<|im_start|>assistant\n");
|
||||
check(tmpls, continuation_content(), "<|im_start|>assistant\n<think>I'm thinking</think>Hello, ");
|
||||
check(tmpls, continuation_reasoning(), "<|im_start|>assistant\n<think>I'm");
|
||||
}
|
||||
|
||||
{
|
||||
auto tmpls = read_templates("models/templates/LFM2.5-Instruct.jinja");
|
||||
check(tmpls, basic(), "<|im_start|>assistant\n");
|
||||
check(tmpls, continuation_content(), "<|im_start|>assistant\n<think>I'm thinking</think>Hello, ");
|
||||
check(tmpls, continuation_reasoning(), "<|im_start|>assistant\n<think>I'm");
|
||||
// 8B-A1B renders prior-turn reasoning via the "thinking" field
|
||||
auto tmpls = read_templates("models/templates/LFM2.5-8B-A1B.jinja");
|
||||
common_chat_templates_inputs inputs;
|
||||
inputs.messages = { message_user, message_assist_call_thoughts, tool_msg };
|
||||
inputs.add_generation_prompt = true;
|
||||
auto params = common_chat_templates_apply(tmpls.get(), inputs);
|
||||
assert_contains(params.prompt, "<think>I'm\nthinking</think>");
|
||||
}
|
||||
|
||||
{
|
||||
|
||||
Reference in New Issue
Block a user