happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

common : replace wrap_for_generation with a prefix convenience function and fix gpt-oss (#20912)

This commit is contained in:
Aldehir Rojas 2026-03-23 22:21:47 -05:00 committed by GitHub
parent 7cadbfce10
commit 312d870a89
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
6 changed files with 34 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
common/chat-auto-parser-generator.cpp
View File

@ -112,8 +112,7 @@ common_peg_arena autoparser::build_parser(const generation_params & inputs) cons
} else {
parser = content.build_parser(ctx);
}
parser = wrap_for_generation_prompt(p, parser, inputs, reasoning.start);
return parser;
return p.prefix(inputs.generation_prompt, reasoning.start) + parser;
});
}

16
common/chat-auto-parser-helpers.cpp
View File

@ -308,22 +308,6 @@ std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segm
return result;
}
common_peg_parser wrap_for_generation_prompt(common_chat_peg_builder & p,
const common_peg_parser & prs,
const autoparser::generation_params & inputs,
const std::string & reasoning_start) {
auto parser = prs;
if (!inputs.generation_prompt.empty()) {
size_t end_pos = inputs.generation_prompt.size();
if (!reasoning_start.empty() && inputs.generation_prompt.find(reasoning_start) != std::string::npos) {
end_pos = inputs.generation_prompt.find(reasoning_start);
}
std::string cut_genprompt = inputs.generation_prompt.substr(0, end_pos);
parser = p.literal(cut_genprompt) + parser;
}
return parser;
}
namespace autoparser {
std::string apply_template(const common_chat_template & tmpl, const template_params & params) {

5
common/chat-auto-parser-helpers.h
View File

@ -58,11 +58,6 @@ std::vector<segment> segmentize_markers(const std::string & text);
// (MARKER, "</function>"), (MARKER, "</tool_call>") ]
std::vector<segment> prune_whitespace_segments(const std::vector<segment> & segments);
// Wrap parser with generation prompt parser
common_peg_parser wrap_for_generation_prompt(common_chat_peg_builder & p,
const common_peg_parser & prs,
const autoparser::generation_params & inputs,
const std::string & reasoning_start = {});
namespace autoparser {
// Apply a template with the given parameters, returning the rendered string (empty on failure)

10
common/chat-peg-parser.cpp
View File

@ -802,6 +802,16 @@ common_peg_parser common_chat_peg_builder::build_json_tools_flat_keys(
return tool_choices;
}
common_peg_parser common_chat_peg_builder::prefix(const std::string & s, const std::string & delimiter) {
if (s.empty()) {
return eps();
}
if (delimiter.empty()) {
return literal(s);
}
return literal(s.substr(0, s.rfind(delimiter)));
}
common_peg_parser common_chat_peg_builder::standard_json_tools(
const std::string & section_start,
const std::string & section_end,

4
common/chat-peg-parser.h
View File

@ -82,6 +82,10 @@ class common_chat_peg_builder : public common_peg_parser_builder {
common_peg_parser tool_arg_string_value(const common_peg_parser & p) { return tag(TOOL_ARG_STRING_VALUE, p); }
common_peg_parser tool_arg_json_value(const common_peg_parser & p) { return atomic(tag(TOOL_ARG_VALUE, p)); }
// Return a parser that parses the prefix of a string, up to a given delimiter.
common_peg_parser prefix(const std::string & s, const std::string & delimiter = {});
// Legacy-compatible helper for building standard JSON tool calls
// Used by tests and manual parsers
// name_key/args_key: JSON key names for function name and arguments

43
common/chat.cpp
View File

@ -872,14 +872,14 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
};
auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
auto generation_prompt = p.prefix(inputs.generation_prompt, "[THINK]");
auto reasoning =
extract_reasoning ? p.optional("[THINK]" + p.reasoning(p.until("[/THINK]")) + "[/THINK]") : p.eps();
// Response format parser
if (inputs.json_schema.is_object() && !inputs.json_schema.empty()) {
// Ministral wants to emit json surrounded by code fences
return wrap_for_generation_prompt(p, reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```",
inputs, "[THINK]");
return generation_prompt + (reasoning << "```json" << p.content(p.schema(p.json(), "response-format", inputs.json_schema)) << "```");
}
// Tool call parser
@ -899,13 +899,12 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
auto max_calls = inputs.parallel_tool_calls ? -1 : 1;
auto tool_calls = p.trigger_rule("tool-call", p.repeat("[TOOL_CALLS]" + tool_choice, min_calls, max_calls));
return wrap_for_generation_prompt(p, reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls,
inputs, "[THINK]");
return generation_prompt + (reasoning << p.content(p.until("[TOOL_CALLS]")) << tool_calls);
}
// Content only parser
include_grammar = false;
return wrap_for_generation_prompt(p, reasoning << p.content(p.rest()), inputs, "[THINK]");
return generation_prompt + (reasoning << p.content(p.rest()));
});
data.parser = parser.save();
@ -991,8 +990,7 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
p.literal("<|channel|>final") + constraint + p.literal("<|message|>") +
p.content(p.schema(p.json(), "response-format-schema", inputs.json_schema)));
return wrap_for_generation_prompt(p, response_format | (analysis + p.zero_or_more(start + analysis) + start + response_format),
inputs, "<|channel|>");
return p.zero_or_more(start + analysis) + start + response_format;
}
if (has_tools && inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_NONE) {
@ -1021,15 +1019,13 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
auto tool_call = p.trigger_rule("tool-call", tool_choice);
if (inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_REQUIRED) {
return tool_call | ( any + p.zero_or_more(start + any) + start + tool_call);
return p.zero_or_more(start + any) + start + tool_call;
}
return wrap_for_generation_prompt(p, tool_call | final_msg | (any + p.zero_or_more(start + any) + start + (tool_call | final_msg)),
inputs, "<|channel|>");
return p.zero_or_more(start + any) + start + (tool_call | final_msg);
}
return wrap_for_generation_prompt(p, final_msg | (any + p.zero_or_more(start + any) + start + final_msg),
inputs, "<|channel|>");
return p.zero_or_more(start + any) + start + final_msg;
});
data.parser = parser.save();
@ -1080,11 +1076,12 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
// When no tools, content goes until end
auto content_until_tool = p.literal("all\n") + p.content(p.until(">>>"));
auto content_until_end = p.literal("all\n") + p.content(p.rest());
auto generation_prompt = p.literal(inputs.generation_prompt);
// If no tools or tool_choice is NONE, just parse content
if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
// When no tools, just match the prefix and capture everything after
return wrap_for_generation_prompt(p, content_until_end + p.end(), inputs);
return generation_prompt + content_until_end + p.end();
}
// Build tool call parsers for each available function
@ -1120,7 +1117,7 @@ static common_chat_params common_chat_params_init_functionary_v3_2(const common_
auto content_and_tool = content_until_tool + tool_choice;
ret = p.choice({ content_and_tool, content_only, tool_choice }) + p.end();
}
return wrap_for_generation_prompt(p, ret, inputs);
return generation_prompt + ret;
});
data.parser = parser.save();
@ -1201,12 +1198,12 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
auto reasoning = extract_reasoning ? p.optional(THINK_START + p.reasoning(
p.until_one_of({ THINK_END, "<|tool_calls_section_begin|>", "<|tool_call_begin|>" })) +
p.optional(p.literal(THINK_END))) : p.eps();
auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
// Content only parser (no tools)
if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
return wrap_for_generation_prompt(p, reasoning + p.content(p.rest()) + end,
inputs, THINK_START);
return generation_prompt + reasoning + p.content(p.rest()) + end;
}
// Build tool call parsers for each available function
@ -1242,8 +1239,7 @@ static common_chat_params common_chat_params_init_kimi_k2(const common_chat_temp
auto content_before_tools = p.content(p.until_one_of({ SECTION_BEGIN, CALL_BEGIN }));
return wrap_for_generation_prompt(p, reasoning + content_before_tools + tool_calls + end,
inputs, THINK_START);
return generation_prompt + reasoning + content_before_tools + tool_calls + end;
});
data.parser = parser.save();
@ -1301,6 +1297,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
data.thinking_end_tag = THINK_END;
auto parser = build_chat_peg_parser([&](common_chat_peg_builder & p) {
auto generation_prompt = p.prefix(inputs.generation_prompt, THINK_START);
auto end = p.end();
auto reasoning = p.eps();
@ -1309,8 +1306,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
}
if (!has_tools || inputs.tool_choice == COMMON_CHAT_TOOL_CHOICE_NONE) {
return wrap_for_generation_prompt(p, reasoning + p.content(p.rest()) + end, inputs,
THINK_START);
return generation_prompt + reasoning + p.content(p.rest()) + end;
}
auto tool_calls = p.rule("tool-calls",
@ -1322,8 +1318,7 @@ static common_chat_params common_chat_params_init_lfm2(const common_chat_templat
auto content = p.content(p.until(TOOL_CALL_START));
return wrap_for_generation_prompt(p, reasoning + content + tool_calls + end, inputs,
THINK_START);
return generation_prompt + reasoning + content + tool_calls + end;
});
data.parser = parser.save();
@ -1396,7 +1391,7 @@ static common_chat_params common_chat_params_init_gigachat_v3(
ret = p.content(p.rest());
}
return wrap_for_generation_prompt(p, ret, inputs);
return p.literal(inputs.generation_prompt) + ret;
});
data.parser = parser.save();
@ -1621,7 +1616,7 @@ static common_chat_params common_chat_templates_apply_jinja(const struct common_
data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
data.generation_prompt = params.generation_prompt;
auto parser = build_chat_peg_parser([&params](common_chat_peg_builder &p) {
return wrap_for_generation_prompt(p, p.content(p.rest()), params);
return p.prefix(params.generation_prompt) + p.content(p.rest());
});
data.parser = parser.save();
return data;