obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-20 12:40:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

[Frontend][Bugfix] Use default tool call IDs for Kimi K3 for conversation-level uniqueness (#50420)

Browse Source 
Signed-off-by: Bugen Zhao <[email protected]>
This commit is contained in:
Bugen Zhao
2026-07-31 10:02:17 +08:00
committed by GitHub
parent f1899b2ffd
commit ab98034d4c
5 changed files with 55 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files
rust/src/chat/src/renderer/kimi_k3/encoding.rs
+2 -2
View File
@@ -126,7 +126,7 @@ pub(super) fn render_request(request: &ChatRequest, tokenizer: &dyn Tokenizer) -
resolved.sort_by_key(|item| (item.0, item.1));
}
for (xtml_index, (_, _, content, name)) in resolved.into_iter().enumerate() {
for (position, _, content, name) in resolved {
let tool_name = name.as_deref().ok_or_else(|| {
Error::ChatTemplate(
"Kimi K3 tool messages need a resolvable tool name: \
@@ -135,7 +135,7 @@ pub(super) fn render_request(request: &ChatRequest, tokenizer: &dyn Tokenizer) -
.to_string(),
)
})?;
write_tool_message(out, tool_name, xtml_index + 1, &content)?;
write_tool_message(out, tool_name, position, &content)?;
}
Ok(())
};
rust/src/chat/src/renderer/kimi_k3/tests.rs
+30 -1
View File
@@ -14,11 +14,11 @@ use vllm_tokenizer::Tokenizer;
use vllm_tokenizer::test_utils::TestTokenizer;
use super::KimiK3ChatRenderer;
use crate::AssistantContentBlock;
use crate::ChatRenderer;
use crate::renderer::kimi_k3::encoding::{CLOSE, END_OF_MSG, IMAGE_PLACEHOLDER, OPEN, SEP};
use crate::renderer::test_utils::{FixtureRequestOptions, fixture_chat_request};
use crate::request::{ChatContentPart, ChatMessage, GenerationPromptMode, ReasoningEffort};
use crate::{AssistantContentBlock, AssistantToolCall};
const OPEN_ID: u32 = 256;
const CLOSE_ID: u32 = 257;
@@ -178,6 +178,35 @@ fn non_thinking_history_omits_reasoning_channel() {
));
}
#[test]
fn partial_tool_results_keep_assistant_call_position() {
let mut request = crate::request::ChatRequest::for_test();
request.messages = vec![
ChatMessage::assistant_blocks(vec![
AssistantContentBlock::ToolCall(AssistantToolCall {
id: "call-a".to_string(),
name: "weather".to_string(),
arguments: "{}".to_string(),
}),
AssistantContentBlock::ToolCall(AssistantToolCall {
id: "call-b".to_string(),
name: "search".to_string(),
arguments: "{}".to_string(),
}),
]),
ChatMessage::tool_response("result-b", "call-b"),
];
request.chat_options.generation_prompt_mode = GenerationPromptMode::NoGenerationPrompt;
let rendered = render_request(&request);
assert!(rendered.contains("<|open|>call tool=\"search\" index=\"2\"<|sep|>"));
assert!(
rendered
.contains("<|open|>message role=\"tool\" tool=\"search\" index=\"2\"<|sep|>result-b")
);
}
#[test]
fn defaults_thinking_effort_to_max() {
let rendered = render_request(&crate::request::ChatRequest::for_test());
rust/src/parser/src/unified/kimi_k3.rs
+14 -37
View File
@@ -148,9 +148,8 @@ enum KimiK3Mode {
pub struct KimiK3UnifiedParser {
buffer: String,
mode: KimiK3Mode,
/// Parser-provided tool-call IDs (`{tool}:{zero_based_index}`) by tool
/// index; its length is also the count of emitted calls.
call_ids: Vec<String>,
/// Number of calls emitted in the current response.
emitted_call_count: usize,
tokenizer: DynTokenizer,
open_token_id: u32,
sep_token_id: u32,
@@ -165,7 +164,7 @@ impl KimiK3UnifiedParser {
Ok(Self {
buffer: String::new(),
mode: KimiK3Mode::default(),
call_ids: Vec::new(),
emitted_call_count: 0,
tokenizer,
open_token_id,
sep_token_id,
@@ -226,7 +225,7 @@ impl KimiK3UnifiedParser {
}
KimiK3Event::CallComplete { arguments } => {
let mode = std::mem::replace(&mut self.mode, KimiK3Mode::Tools);
let KimiK3Mode::Call { name, index, .. } = mode else {
let KimiK3Mode::Call { name, .. } = mode else {
return Err(parsing_failed!(
"Kimi K3 call completion without an active tool call"
));
@@ -237,8 +236,8 @@ impl KimiK3UnifiedParser {
return Ok(());
}
let tool_index = self.call_ids.len();
self.call_ids.push(tool_call_id_for(&name, index.as_deref()));
let tool_index = self.emitted_call_count;
self.emitted_call_count += 1;
output.push_call(ToolCallDelta {
tool_index,
name: Some(name),
@@ -251,7 +250,7 @@ impl KimiK3UnifiedParser {
fn reset_state(&mut self) -> String {
self.mode = KimiK3Mode::Idle;
self.call_ids.clear();
self.emitted_call_count = 0;
std::mem::take(&mut self.buffer)
}
}
@@ -266,7 +265,7 @@ impl UnifiedParser for KimiK3UnifiedParser {
fn initialize(&mut self, prompt_token_ids: &[u32]) -> Result<()> {
self.buffer.clear();
self.call_ids.clear();
self.emitted_call_count = 0;
self.initialize_mode(prompt_token_ids);
Ok(())
}
@@ -279,10 +278,6 @@ impl UnifiedParser for KimiK3UnifiedParser {
Some(&KIMI_K3_STRUCTURAL_TAG_BUILDER)
}
fn tool_call_id(&self, tool_index: usize) -> Option<&str> {
self.call_ids.get(tool_index).map(String::as_str)
}
fn parse_into(&mut self, chunk: &str, output: &mut UnifiedParserOutput) -> Result<()> {
self.buffer.push_str(chunk);
@@ -313,7 +308,6 @@ impl UnifiedParser for KimiK3UnifiedParser {
}
}
// Keep call_ids so tool_call_id() stays available after the stream ends.
self.mode = KimiK3Mode::Idle;
Ok(output)
}
@@ -323,20 +317,6 @@ impl UnifiedParser for KimiK3UnifiedParser {
}
}
/// Build the API-side tool-call ID from the XTML one-based `index` attribute.
///
/// The ID uses the zero-based call ordinal; XTML's message index stays
/// one-based when rendering tool result messages.
fn tool_call_id_for(name: &str, index: Option<&str>) -> String {
match index {
None => name.to_string(),
Some(raw) => match raw.parse::<i64>() {
Ok(one_based) => format!("{name}:{}", one_based - 1),
Err(_) => format!("{name}:{raw}"),
},
}
}
/// Parse one Kimi K3 event from buffered streaming input.
fn parse_next_kimi_k3_event(
input: &mut KimiK3Input<'_>,
@@ -762,7 +742,6 @@ mod tests {
"hours": [8, 20],
})
);
assert_eq!(parser.tool_call_id(0), Some("get_weather:0"));
}
#[test]
@@ -929,8 +908,6 @@ mod tests {
assert_eq!(calls[1].tool_index, 1);
assert_eq!(calls[1].name.as_deref(), Some("get_time"));
assert_eq!(calls[1].arguments, "{}");
assert_eq!(parser.tool_call_id(0), Some("get_weather:0"));
assert_eq!(parser.tool_call_id(1), Some("get_time:1"));
}
#[test]
@@ -1018,11 +995,10 @@ mod tests {
assert_eq!(calls.len(), 1);
assert_eq!(calls[0].tool_index, 0);
assert_eq!(calls[0].name.as_deref(), Some("real"));
assert_eq!(parser.tool_call_id(0), Some("real:1"));
}
#[test]
fn kimi_k3_tool_call_id_follows_index_attribute() {
fn kimi_k3_tool_indices_ignore_xtml_index_attribute() {
let tools_body = format!(
"{}{}{}",
call("tool=\"first\" index=\"3\"", ""),
@@ -1032,11 +1008,12 @@ mod tests {
let text = thinking_output("t", "", &tools_body);
let mut parser = test_parser();
parser.parse_complete(&text).unwrap();
let output = parser.parse_complete(&text).unwrap();
assert_eq!(parser.tool_call_id(0), Some("first:2"));
assert_eq!(parser.tool_call_id(1), Some("second"));
assert_eq!(parser.tool_call_id(2), Some("third:x"));
assert_eq!(
output.calls().iter().map(|call| call.tool_index).collect::<Vec<_>>(),
[0, 1, 2]
);
}
#[test]
tests/tool_use/test_kimi_k3_tool_parser.py
+9 -3
View File
@@ -141,7 +141,6 @@ def test_extract_tool_calls_with_response_and_typed_arguments():
assert extracted.content == "answer"
assert len(extracted.tool_calls) == 1
tool_call = extracted.tool_calls[0]
assert tool_call.id == "calc:0"
assert tool_call.function.name == "calc"
assert json.loads(tool_call.function.arguments) == {
"x": 1,
@@ -168,7 +167,6 @@ def test_delegating_parser_preserves_tool_calls_after_reasoning():
assert content == "answer"
assert tool_calls is not None
assert len(tool_calls) == 1
assert tool_calls[0].id == "calc:0"
assert tool_calls[0].name == "calc"
assert json.loads(tool_calls[0].arguments) == {"x": 1}
@@ -364,11 +362,19 @@ def test_streaming_split_markers_do_not_leak():
assert OPEN not in content
assert SEP not in content
assert len(tool_deltas) == 1
assert tool_deltas[0].id == "calc:0"
assert tool_deltas[0].function.name == "calc"
assert json.loads(tool_deltas[0].function.arguments) == {"x": 1}
def test_tool_call_ids_are_unique_across_messages():
output = _tools(_call("calc", 1))
first = KimiK3ToolParser(DummyTokenizer()).extract_tool_calls(output, _request())
second = KimiK3ToolParser(DummyTokenizer()).extract_tool_calls(output, _request())
assert first.tool_calls[0].id != second.tool_calls[0].id
def test_streaming_consumed_response_prefix_no_call_keeps_content():
parser = KimiK3ToolParser(DummyTokenizer())
request = _request()
vllm/tool_parsers/kimi_k3_tool_parser.py
-10
View File
@@ -218,7 +218,6 @@ class KimiK3ToolParser(ToolParser):
"""
call_attrs = self._attrs(attrs)
tool_name = call_attrs.get("tool", "")
tool_index = call_attrs.get("index", "")
arguments: dict = {}
for arg_match in self._arg_re.finditer(body):
arg_attrs = self._attrs(arg_match["attrs"])
@@ -234,16 +233,7 @@ class KimiK3ToolParser(ToolParser):
arguments[key] = raw_value
if not tool_name:
return None
tool_call_id = tool_name
if tool_index:
try:
tool_call_id = f"{tool_name}:{int(tool_index) - 1}"
except ValueError:
tool_call_id = f"{tool_name}:{tool_index}"
# id uses the API-side zero-based call ordinal; XTML's message index
# stays one-based when rendering tool result messages.
return ToolCall(
id=tool_call_id,
type="function",
function=FunctionCall(
name=tool_name,