Skip to content

glm5.2: preserve input marking in prompt tokenization (fix output stalls on stray message tags) #148

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mp4 wants to merge 1 commit into
base: v0.4.7
Choose a base branch
from
Open

glm5.2: preserve input marking in prompt tokenization (fix output stalls on stray message tags) #148

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
31 changes: 28 additions & 3 deletions common/chat-auto-parser-generator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,7 +42,7 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
const autoparser & autoparser) {
// Create the result structure
common_chat_params data;
data.prompt = common_chat_template_direct_apply(tmpl, inputs);
data.prompt = common_chat_template_direct_apply(tmpl, inputs, &data.prompt_parts);
data.generation_prompt = common_chat_template_generation_prompt(tmpl, inputs);
data.format = COMMON_CHAT_FORMAT_PEG_NATIVE;
data.preserved_tokens = autoparser.preserved_tokens;
Expand All @@ -52,22 +52,47 @@ common_chat_params peg_generator::generate_parser(const common_chat_template &
std::string parser_generation_prompt = data.generation_prompt;

if (inputs.continue_final_message != COMMON_CHAT_CONTINUATION_NONE && !inputs.continue_msg.empty()) {
// Build up generation prompt manually
// Build up generation prompt manually, keeping the provenance of the
// template delimiters and the request-provided continuation content
// separate (the latter must not be parsed for special tokens).
const auto & msg = inputs.continue_msg;

std::vector<jinja::string_part> gen_parts;

if (!autoparser.reasoning.start.empty()) {
data.generation_prompt = data.generation_prompt.substr(0, data.generation_prompt.find(autoparser.reasoning.start));
const size_t cut = data.generation_prompt.find(autoparser.reasoning.start);

// Template-derived prefix of the generation prompt
gen_parts.push_back({false, data.generation_prompt.substr(0, cut)});
// Reasoning markers are template text
gen_parts.push_back({false, autoparser.reasoning.start});
// Request-provided reasoning content
gen_parts.push_back({true, msg.reasoning_content});
if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
gen_parts.push_back({false, autoparser.reasoning.end});
}

data.generation_prompt = data.generation_prompt.substr(0, cut);
data.generation_prompt += autoparser.reasoning.start + msg.reasoning_content;
if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
data.generation_prompt += autoparser.reasoning.end;
}
}

if (inputs.continue_final_message == COMMON_CHAT_CONTINUATION_CONTENT) {
// Request-provided message content
gen_parts.push_back({true, msg.render_content()});
data.generation_prompt += msg.render_content();
}

if (gen_parts.empty()) {
// Reasoning-only continuation without a reasoning start marker:
// nothing request-provided was appended, the prompt is template text
gen_parts.push_back({false, data.generation_prompt});
}

data.prompt += data.generation_prompt;
data.prompt_parts.insert(data.prompt_parts.end(), gen_parts.begin(), gen_parts.end());
}

auto parser = autoparser.build_parser(inputs, parser_generation_prompt);
Expand Down
Loading