servo/openai-harmony
1
0
Fork 0
mirror of https://github.com/openai/harmony.git synced 2025-08-22 07:17:08 -04:00
Code Issues Projects Releases Packages Wiki Activity Actions

Revert "Fix tokenization of <|constrain|> content type in rendering (#47)"

Browse source 
This reverts commit 8a4645f0f9.
This commit is contained in:
Dominik Kundel 2025-08-08 17:52:33 -07:00 • committed by GitHub
parent 8a4645f0f9
commit 488c417386
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 2 additions and 49 deletions
Download patch file Download diff file Expand all files Collapse all files

17
src/encoding.rs
View file

@ -835,22 +835,7 @@ impl Render<Message> for HarmonyEncoding {
// finally content type
if let Some(content_type) = &message.content_type {
// <|constrain|> is a unique case which needs to be tokenized as a special token
if let Some(constrain_marker) = self.mapped_format_token(FormattingToken::ConstrainedFormat) {
if content_type.starts_with(constrain_marker) {
// Render the space, then the constrain marker as a special token, then the rest as text (if any)
self.render_text_into(" ", into)?;
self.render_formatting_token_into(FormattingToken::ConstrainedFormat, into)?;
let rest = &content_type[constrain_marker.len()..];
if !rest.is_empty() {
self.render_text_into(rest, into)?;
}
} else {
self.render_text_into(format!(" {content_type}"), into)?;
}
} else {
self.render_text_into(format!(" {content_type}"), into)?;
}
self.render_text_into(format!(" {content_type}"), into)?;
}
self.render_formatting_token_into(FormattingToken::Message, into)?;

34
tests/test_harmony.py
View file

@ -233,36 +233,6 @@ def test_simple_tool_call(encoding_name):
assert parsed == expected
@pytest.mark.parametrize(
"encoding_name",
[
HarmonyEncodingName.HARMONY_GPT_OSS,
],
)
def test_tool_call_with_constrain_tokenized_correctly(encoding_name):
"""
Despite passing <|constrain|> as a string in "content_type" it has to be kept as a special token.
"""
encoding = load_harmony_encoding(encoding_name)
text = (
"<|start|>assistant to=functions.get_weather<|channel|>commentary"
' <|constrain|>json<|message|>{"location": "Tokyo"}<|call|>'
)
tokens = encoding.encode(text, allowed_special="all")
parsed = encoding.parse_messages_from_completion_tokens(tokens, role=None)
expected = [
Message.from_role_and_content(Role.ASSISTANT, '{"location": "Tokyo"}')
.with_channel("commentary")
.with_recipient("functions.get_weather")
.with_content_type("<|constrain|>json"),
]
assert parsed == expected
rendered = encoding.render_conversation(Conversation.from_messages(expected))
assert text == encoding.decode_utf8(tokens)
assert rendered == tokens
@pytest.mark.parametrize(
"encoding_name",
[
@ -278,7 +248,7 @@ def test_tool_call_with_constrain_marker_adjacent(encoding_name):
encoding = load_harmony_encoding(encoding_name)
text = (
"<|start|>assistant to=functions.get_weather<|channel|>commentary"
'<|constrain|>json<|message|>{"location": "Tokyo"}<|call|>'
'<|constrain|>json<|message|>{"location": "Tokyo"}<|end|>'
)
tokens = encoding.encode(text, allowed_special="all")
parsed = encoding.parse_messages_from_completion_tokens(tokens, role=None)
@ -732,8 +702,6 @@ def test_does_not_drop_if_ongoing_analysis():
)
assert encoding.decode_utf8(tokens) == expected_output
# ensure that <|constrain|>json part is tokenized correctly as special tokens
assert encoding.encode(expected_output, allowed_special="all") == tokens
def test_preserve_cot():