sleepy/llama.cpp
1
0
Fork 0
Code Issues 11 Pull Requests Actions 107 Packages Projects Releases Wiki Activity

common/parser: fix reasoning whitespace bugs + extra parser tests (#21085)

Browse Source 
* fix whitespace reasoning issues + add reconstruction tests

* Proper fix

* fix Nemotron autoparser test expectations to include newline in marker
This commit is contained in:
Piotr Wilkin (ilintar)
2026-03-28 07:29:26 +01:00
committed by GitHub
parent c46758d28f
commit 1f5d15e665
3 changed files with 141 additions and 35 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/chat-diff-analyzer.cpp
+18 -10
View File
@@ -287,7 +287,7 @@ void analyze_reasoning::compare_reasoning_presence() {
return p.literal(reasoning_content) + p.space() + p.optional(p.tag("post", (p.marker() + p.space())) + p.rest());
});
auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
});
// try the more aggressive parse first, if it fails, fall back to the delimiter one
auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
@@ -297,7 +297,7 @@ void analyze_reasoning::compare_reasoning_presence() {
if (result.result.success()) {
if (!result.tags["pre"].empty() && !result.tags["post"].empty()) {
mode = reasoning_mode::TAG_BASED;
start = trim_whitespace(result.tags["pre"]);
start = trim_leading_whitespace(result.tags["pre"]);
end = trim_trailing_whitespace(result.tags["post"]);
} else if (!result.tags["post"].empty()) {
mode = reasoning_mode::TAG_BASED;
@@ -333,7 +333,7 @@ void analyze_reasoning::compare_thinking_enabled() {
if (left_trimmed.empty() && !diff.right.empty()) {
if (!right_trimmed.empty() && string_ends_with(comparison->output_B, right_trimmed)) {
if (start.empty()) {
start = right_trimmed;
start = trim_leading_whitespace(diff.right);
mode = reasoning_mode::TAG_BASED;
}
}
@@ -344,7 +344,7 @@ void analyze_reasoning::compare_thinking_enabled() {
if (seg.size() >= 2 && seg[seg.size() - 1].value == left_trimmed && seg[seg.size() - 2].type == segment_type::MARKER) {
start = seg[seg.size() - 2].value;
}
end = left_trimmed;
end = trim_trailing_whitespace(diff.left);
mode = reasoning_mode::TAG_BASED;
}
}
@@ -363,15 +363,23 @@ void analyze_reasoning::compare_thinking_enabled() {
size_t len = std::min(base.size(), anchor_len);
std::string anchor = base.substr(base.size() - len);
auto pos = extended.rfind(anchor);
if (pos == std::string::npos || pos + len >= extended.size()) continue;
if (pos == std::string::npos || pos + len >= extended.size()) {
continue;
}
std::string extra = trim_whitespace(extended.substr(pos + len));
if (extra.empty()) continue;
if (extra.empty()) {
continue;
}
auto seg = prune_whitespace_segments(segmentize_markers(extra));
if (seg.size() == 2 && seg[0].type == segment_type::MARKER && seg[1].type == segment_type::MARKER) {
if (start.empty()) start = seg[0].value;
if (end.empty()) end = seg[1].value;
if (start.empty()) {
start = seg[0].value;
}
if (end.empty()) {
end = seg[1].value;
}
mode = reasoning_mode::TAG_BASED;
break;
}
@@ -423,7 +431,7 @@ void analyze_reasoning::compare_reasoning_scope() {
LOG_DBG(ANSI_ORANGE "%s: Detected TOOLS_ONLY reasoning mode\n" ANSI_RESET, __func__);
auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
return p.tag("pre", p.marker()) + p.space() + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
return p.tag("pre", p.marker() + p.space()) + p.literal(reasoning_content) + p.space() + p.tag("post", (p.marker() + p.space()));
});
auto result = parser_wrapped.parse_anywhere_and_extract(comparison->output_B);
if (result.result.success()) {
@@ -516,7 +524,7 @@ analyze_content::analyze_content(const common_chat_template & tmpl, const analyz
// Take the more promising diff
std::string pure_content = rdiff.length() > diff_tools.left.length() ? rdiff : diff_tools.left;
auto parser_wrapped = build_tagged_peg_parser([&](common_peg_parser_builder &p) {
return p.tag("pre", p.marker()) + p.space() + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
return p.tag("pre", p.marker() + p.space()) + p.literal(response) + p.space() + p.tag("post", (p.marker() + p.space())) + p.rest();
});
auto result = parser_wrapped.parse_anywhere_and_extract(pure_content);
start = result.tags["pre"];