[Rust Frontend] Normalize native renderer reasoning controls (#56998)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
Bugen Zhao
2026-09-16 18:15:05 -07:00
committed by GitHub
parent 3e267bae70
commit e52da038d1
29 changed files with 1253 additions and 454 deletions
+16 -10
View File
@@ -175,16 +175,22 @@ mod tests {
}
#[test]
fn invalid_reasoning_effort_maps_to_invalid_request() {
let error = vllm_chat::Error::InvalidReasoningEffort(
"DeepSeek V4.1 reasoning_effort must be within [1, 100]".to_string(),
);
let api_error = chat_submit_error("failed to submit chat request", error);
assert_eq!(api_error.status_code(), StatusCode::BAD_REQUEST);
assert_eq!(
api_error.to_error_response().error.error_type,
"invalid_request_error"
);
fn invalid_reasoning_parameters_map_to_invalid_request() {
for error in [
vllm_chat::Error::InvalidReasoningEffort(
"DeepSeek V4.1 reasoning_effort must be within [1, 100]".to_string(),
),
vllm_chat::Error::InvalidReasoningControl {
message: "template kwarg `thinking` must be a boolean".to_string(),
},
] {
let api_error = chat_submit_error("failed to submit chat request", error);
assert_eq!(api_error.status_code(), StatusCode::BAD_REQUEST);
assert_eq!(
api_error.to_error_response().error.error_type,
"invalid_request_error"
);
}
}
#[test]
@@ -180,7 +180,7 @@ pub(super) fn prepare_chat_request(
chat_options: ChatOptions {
generation_prompt_mode,
chat_template: request.chat_template,
reasoning_effort: request.reasoning_effort,
reasoning_effort: request.reasoning_effort.map(|effort| effort.as_str().into()),
response_format,
template_kwargs,
},
@@ -488,6 +488,50 @@ mod tests {
}
}
#[test]
fn chat_http_reasoning_effort_preserves_omission_none_and_kwargs() {
for effort in [None, Some(json!(null))].into_iter().chain(
["none", "minimal", "low", "medium", "high", "xhigh", "max"]
.map(|name| Some(json!(name))),
) {
let kwargs = json!({"reasoning_effort": 37, "thinking": true});
let mut value = json!({
"messages": [{"role": "user", "content": "hello"}],
"chat_template_kwargs": kwargs,
});
if let Some(effort) = &effort {
value["reasoning_effort"] = effort.clone();
}
let request: ChatCompletionRequest = serde_json::from_value(value).unwrap();
let prepared = prepare_chat_request(
request,
&served(&["test-model"]),
request_context(&HeaderMap::new(), None),
)
.unwrap();
let options = prepared.chat_request.chat_options;
assert_eq!(
serde_json::to_value(options.reasoning_effort).unwrap(),
effort.unwrap_or(serde_json::Value::Null)
);
assert_eq!(
serde_json::to_value(options.template_kwargs).unwrap(),
kwargs
);
}
}
#[test]
fn chat_http_reasoning_effort_rejects_model_extensions_at_top_level() {
for effort in [json!(37), json!("custom")] {
let request = json!({
"messages": [{"role": "user", "content": "hello"}],
"reasoning_effort": effort,
});
assert!(serde_json::from_value::<ChatCompletionRequest>(request).is_err());
}
}
#[test]
fn chat_http_request_defaults_missing_or_null_model() {
for model in [None, Some(serde_json::Value::Null)] {
@@ -747,7 +791,7 @@ mod tests {
);
let tokenizer = Arc::new(TestTokenizer::new());
let prompt = KimiK3ChatRenderer::new(tokenizer.clone())
let prompt = KimiK3ChatRenderer::new(tokenizer.clone(), Default::default())
.render(&prepared.chat_request)
.expect("Kimi K3 rendering succeeds")
.prompt;
@@ -9,15 +9,15 @@ use serde::{Deserialize, Serialize};
use serde_json::Value;
use serde_with::SerializeDisplay;
use validator::Validate;
use vllm_chat::ReasoningEffort;
use vllm_engine_core_client::protocol::sampling::RepetitionDetectionParams;
use vllm_text::TruncationSide;
use crate::routes::openai::utils::structured_outputs::ResponseFormat;
use crate::routes::openai::utils::types::{
ChatLogProbs, ChatMessage, Normalizable, PromptLogprobs, StreamOptions, StreamResponseEnvelope,
StringOrArray, Tool, ToolCall, ToolCallDelta, ToolChoice, Usage, default_true,
deserialize_request_top_k, validate_messages, validate_stop, validate_top_p_value,
ChatLogProbs, ChatMessage, Normalizable, PromptLogprobs, ReasoningEffort, StreamOptions,
StreamResponseEnvelope, StringOrArray, Tool, ToolCall, ToolCallDelta, ToolChoice, Usage,
default_true, deserialize_request_top_k, validate_messages, validate_stop,
validate_top_p_value,
};
/// vLLM-compatible request type for the Chat Completions API.
@@ -125,8 +125,8 @@ fn validate_function_tools(tools: &[Tool], param: &'static str) -> Result<(), Ap
mod tests {
use std::collections::HashMap;
use crate::routes::openai::utils::types::ReasoningEffort;
use serde_json::json;
use vllm_chat::ReasoningEffort;
use super::validate_request_compat;
use crate::routes::openai::chat_completions::types::ChatCompletionRequest;
@@ -39,6 +39,36 @@ where
}
}
/// Effort level for reasoning models.
///
/// Fixed OpenAI HTTP request grades. Request conversion maps these names to
/// the model-independent `vllm_chat::EffortValue` before renderer validation.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
#[serde(rename_all = "lowercase")]
pub enum ReasoningEffort {
None,
Minimal,
Low,
Medium,
High,
XHigh,
Max,
}
impl ReasoningEffort {
pub fn as_str(self) -> &'static str {
match self {
Self::None => "none",
Self::Minimal => "minimal",
Self::Low => "low",
Self::Medium => "medium",
Self::High => "high",
Self::XHigh => "xhigh",
Self::Max => "max",
}
}
}
// ============================================================================
// String/Array Utilities
// ============================================================================