Skip to content

[do not merge] GPT-OSS parser hacks #3786

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
dkalinowski wants to merge 2 commits into
base: main
Choose a base branch
from
+101 −5
Open

[do not merge] GPT-OSS parser hacks #3786

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a983ee9
unary
dkalinowski Nov 13, 2025
0521c64
save
dkalinowski Nov 13, 2025
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
2 changes: 1 addition & 1 deletion src/llm/io_processing/gptoss/harmony.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,7 +97,7 @@ ToolCalls_t Harmony::getToolCalls() {
static const std::string tool_prefix = "to=functions.";
ToolCalls_t toolCalls;
for (const auto& msg : messages) {
if (startsWith(msg.getChannel(), "commentary")) {
if (startsWith(msg.getChannel(), "commentary") || startsWith(msg.getChannel(), "analysis")) {
Copy link
Collaborator

@atobiszei atobiszei Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Just a comment. For GenAI/OV testing purposes it could be useful to be able to turn off such workarounds, to check accuracy fixes, otherwise we may just mask those issues

size_t marker = msg.getChannel().find(tool_prefix);
if (marker != std::string::npos) {
marker += tool_prefix.length();
Expand Down
9 changes: 6 additions & 3 deletions src/llm/io_processing/gptoss/tool_parser.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,9 +86,11 @@ std::optional<rapidjson::Document> GptOssToolParser::parseChunk(const std::strin
std::string chunk = newChunk;
std::optional<rapidjson::Document> result;

if (chunk.find(getParsingStartTags()[0]) != std::string::npos) {
toolCallIndex++; // starting with -1, first call will be 0
return std::nullopt;
for (const auto& parsingStartTag : getParsingStartTags()) {
if (chunk.find(parsingStartTag) != std::string::npos) {
toolCallIndex++; // starting with -1, first call will be 0
return std::nullopt;
}
}

// This should only happen during channel read if model does not produce garbage
Expand Down Expand Up @@ -158,6 +160,7 @@ std::optional<rapidjson::Document> GptOssToolParser::parseChunk(const std::strin
// Cut everything after first .
// Remove and take only remaining part
// The harmony format is: <|channel|>commentary to=functions.<function_name> <|constrain|>json<|message|>{...}<|call|>
// HACK: This does not conform to OpenAI Harmony format, but we allow for <|channel|>analysis to=... as well
std::size_t pos = chunk.find('.');
if (pos != std::string::npos) {
chunk = chunk.substr(pos + 1);
Expand Down
5 changes: 4 additions & 1 deletion src/llm/io_processing/gptoss/tool_parser.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,7 +58,10 @@ class GptOssToolParser : public BaseOutputParser {
std::optional<rapidjson::Document> parseChunk(const std::string& chunk, ov::genai::GenerationFinishReason finishReason) override;

const std::vector<std::string>& getParsingStartTags() const override {
static const std::vector<std::string> parsingStartTags{parsingStartTag};
static const std::vector<std::string> parsingStartTags{
parsingStartTag,
"<|channel|>analysis to=", // THIS IS A HACK: This does not conform to OpenAI Harmony format, however tests show that sometimes model produces tools in analysis channel
};
return parsingStartTags;
}

Expand Down
90 changes: 90 additions & 0 deletions src/test/llm/output_parsers/gptoss_output_parser_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -288,6 +288,36 @@ TEST_F(GptOssOutputUnaryParserTest, SingleToolCallWithConstrain) {
}
}

// This does not conform to OpenAI Harmony format, however tests show that sometimes model produces tools in analysis channel
TEST_F(GptOssOutputUnaryParserTest, SingleToolCallWithConstrainInAnalysisChannel) {
for (auto closureToken : std::vector<Harmony::TokenID>{
Harmony::TokenID::RETURN, // ending with <|return|>
Harmony::TokenID::END, // ending with <|end|>
Harmony::TokenID::CALL}) { // ending with <|call|>
for (auto functionDeclaration : std::vector<std::string>{
"analysis to=functions.hello", // valid channel with to=
"analysis to=functions.hello ",
"analysis to=functions.hello",
"analysis ANYTHING IN BETWEEN to=functions.hello",
}) { // spaces after hello
builder
.clear()
.add(Harmony::TokenID::CHANNEL)
.add(functionDeclaration)
.add(Harmony::TokenID::MESSAGE)
.add(R"({"Hello": "world!"})")
.add(closureToken);
Harmony harmony(*gptOssTokenizer, builder.build());
ASSERT_TRUE(harmony.parse()) << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
ASSERT_EQ(harmony.getContent(), "") << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
ASSERT_EQ(harmony.getReasoning(), "") << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
ASSERT_EQ(harmony.getToolCalls().size(), 1) << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
ASSERT_EQ(harmony.getToolCalls()[0].name, "hello") << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
ASSERT_EQ(harmony.getToolCalls()[0].arguments, R"({"Hello": "world!"})") << "Failed for closure token: " << static_cast<int64_t>(closureToken) << " function declaration: " << functionDeclaration;
}
}
}

TEST_F(GptOssOutputUnaryParserTest, InvalidSingleToolCallWithConstrain) {
for (auto closureToken : std::vector<Harmony::TokenID>{
Harmony::TokenID::RETURN, // ending with <|return|>
Expand Down Expand Up @@ -576,3 +606,63 @@ TEST_F(GptOssOutputStreamParserTest, HolisticStreamingTools) {
};
test(chunkToDeltaVec);
}

// This does not conform to OpenAI Harmony format, however tests show that sometimes model produces tools in analysis channel
TEST_F(GptOssOutputStreamParserTest, HolisticStreamingToolsInAnalysisChannel) {
std::vector<std::tuple<std::string, ov::genai::GenerationFinishReason, std::optional<std::string>>> chunkToDeltaVec{
// Reasoning
{"<|channel|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"analysis", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"<|message|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"I", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"reasoning_content":"I"}})"}},
{" will", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"reasoning_content":" will"}})"}},
{" call", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"reasoning_content":" call"}})"}},
{" fun", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"reasoning_content":" fun"}})"}},
{"ction.", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"reasoning_content":"ction."}})"}},
{"<|end|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
// Preamble
{"<|channel|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"commentary", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"<|message|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"I", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":"I"}})"}},
{" have", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":" have"}})"}},
{" to", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":" to"}})"}},
{" call", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":" call"}})"}},
{" fun", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":" fun"}})"}},
{"ction.", ov::genai::GenerationFinishReason::NONE, {R"({"delta":{"content":"ction."}})"}},
{"<|end|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
// Tool 1
{"<|channel|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"analysis", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{" to=", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"fun", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"ctions", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{".hello ", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"<|message|>", ov::genai::GenerationFinishReason::NONE, "{\"delta\":{\"tool_calls\":[{\"id\":\"XXXXXXXXX\",\"type\":\"function\",\"index\":0,\"function\":{\"name\":\"hello\"}}]}}"},
{" {\"", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":" {\""}}]}})"},
{"location", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":"location"}}]}})"},
{"\":", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\":"}}]}})"},
{" \"", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":" \""}}]}})"},
{"Paris", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":"Paris"}}]}})"},
{"\"}", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":0,"function":{"arguments":"\"}"}}]}})"},
{"<|call|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
// Tool 2 (with ignored constrain)
{"<|channel|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"analysis", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{" to=", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"fun", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"ctions", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{".world ", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"<|constrain|>", ov::genai::GenerationFinishReason::NONE, "{\"delta\":{\"tool_calls\":[{\"id\":\"XXXXXXXXX\",\"type\":\"function\",\"index\":1,\"function\":{\"name\":\"world\"}}]}}"},
{"json", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{"<|message|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
{" {\"", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":" {\""}}]}})"},
{"location", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":"location"}}]}})"},
{"\":", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":"\":"}}]}})"},
{" \"", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":" \""}}]}})"},
{"Warsaw", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":"Warsaw"}}]}})"},
{"\"}", ov::genai::GenerationFinishReason::NONE, R"({"delta":{"tool_calls":[{"index":1,"function":{"arguments":"\"}"}}]}})"},
{"<|call|>", ov::genai::GenerationFinishReason::NONE, std::nullopt},
};
test(chunkToDeltaVec);
}