[Rust Frontend] Coerce completion max_tokens: null to default (#47166)

Signed-off-by: Blas Rodriguez Irizar <rodrigblas@gmail.com>
This commit is contained in:
Blas Rodriguez Irizar
2026-07-01 06:41:33 +00:00
committed by GitHub
parent 89e99202f2
commit 5b431b905c
2 changed files with 34 additions and 1 deletions
@@ -199,6 +199,7 @@ mod tests {
use super::prepare_completion_request;
use crate::lora::LoraModelResolution;
use crate::routes::openai::completions::types::CompletionRequest;
use crate::routes::openai::utils::types::Normalizable;
use crate::utils::{ResolvedRequestContext, resolve_request_context};
fn request_context(headers: &HeaderMap, request_id: Option<&str>) -> ResolvedRequestContext {
@@ -249,6 +250,28 @@ mod tests {
assert!(request.ignore_eos);
}
#[test]
fn normalize_coerces_null_max_tokens_to_default() {
// An absent `max_tokens` already gets the serde default.
let absent: CompletionRequest =
serde_json::from_value(base_request_json()).expect("parse request");
assert_eq!(absent.max_tokens, Some(16));
// An explicit `null` deserializes to `None`, bypassing the default;
// `normalize` must coerce it back to match Python vLLM.
let mut request: CompletionRequest = serde_json::from_value(json!({
"model": "Qwen/Qwen1.5-0.5B-Chat",
"prompt": "hello",
"stream": true,
"max_tokens": null
}))
.expect("parse request");
assert_eq!(request.max_tokens, None);
request.normalize();
assert_eq!(request.max_tokens, Some(16));
}
#[test]
fn prepare_completion_request_maps_sampling_fields() {
let request: CompletionRequest = serde_json::from_value(json!({
@@ -179,7 +179,17 @@ pub struct CompletionRequest {
pub other: Map<String, Value>,
}
impl Normalizable for CompletionRequest {}
impl Normalizable for CompletionRequest {
/// Normalize the request by applying defaults.
fn normalize(&mut self) {
// An explicit `"max_tokens": null` deserializes to `None`, bypassing the
// serde field default. Coerce it back to the default so it behaves like
// an absent field, matching Python vLLM's `normalize_null_max_tokens`.
if self.max_tokens.is_none() {
self.max_tokens = default_completion_max_tokens();
}
}
}
/// Mirrors the Python vLLM `CompletionResponse` class.
#[serde_with::skip_serializing_none]