[Rust Frontend] Forward per-request watermarking controls (#56338)

Co-authored-by: Codex <noreply@openai.com>
Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
Nicolò Lucchesi
2026-09-14 23:17:02 +08:00
committed by GitHub
co-authored by Codex
parent b3124a8237
commit 1713a09491
14 changed files with 133 additions and 1 deletions
+22
View File
@@ -65,6 +65,7 @@ pub fn to_text_request(
let mut sampling_params =
build_sampling_params(req.temperature, sampling, decoding, stopping, response)?;
sampling_params.watermarking = req.watermarking.unwrap_or(true);
// Thread KVCacheParameters → SamplingParams fields.
if let Some(kv) = kv {
@@ -521,6 +522,7 @@ impl ResponseOpts {
#[cfg(test)]
mod tests {
use prost::Message as _;
use vllm_engine_core_client::protocol::output::StopReason;
use vllm_text::{
FinishReason, Finished, Prompt, SamplingHints, SamplingLimits, lower_sampling_params,
@@ -539,6 +541,26 @@ mod tests {
}
}
#[test]
fn watermarking_defaults_and_opt_out_survive_protobuf_conversion() {
for watermarking in [None, Some(true), Some(false)] {
let request = pb::GenerateRequest {
watermarking,
..base_request()
};
let encoded = request.encode_to_vec();
for stream in [false, true] {
let decoded = pb::GenerateRequest::decode(encoded.as_slice()).unwrap();
let text = to_text_request(decoded, stream, &["test-model".to_string()])
.expect("convert request");
assert_eq!(
text.sampling_params.watermarking,
watermarking.unwrap_or(true)
);
}
}
}
#[test]
fn temperature_propagates_from_top_level_request_field() {
let req = pb::GenerateRequest {
@@ -164,6 +164,30 @@ mod tests {
);
}
#[test]
fn prepare_generate_request_preserves_watermarking_defaults_and_opt_out() {
for watermarking in [None, Some(true), Some(false)] {
let mut body = json!({
"token_ids": [11, 22],
"sampling_params": {}
});
if let Some(watermarking) = watermarking {
body["sampling_params"]["watermarking"] = json!(watermarking);
}
let prepared = prepare_generate_request(
serde_json::from_value(body).expect("parse request"),
&served(&["test-model"]),
ResolvedRequestContext::default(),
None,
)
.expect("prepare request");
assert_eq!(
prepared.text_request.sampling_params.watermarking,
watermarking.unwrap_or(true)
);
}
}
#[test]
fn prepare_generate_request_forwards_thinking_token_budget() {
let request: GenerateRequest = serde_json::from_value(json!({
@@ -150,6 +150,7 @@ pub(super) fn prepare_chat_request(
messages,
sampling_params: SamplingParams {
temperature: request.temperature,
watermarking: request.watermarking,
top_p: request.top_p,
top_k: request.top_k,
seed: request.seed,
@@ -116,6 +116,10 @@ pub struct ChatCompletionRequest {
pub user: Option<String>,
// -------- vLLM Sampling Parameters --------
/// Whether to apply the engine's configured watermark to this request.
#[serde(default = "default_true")]
pub watermarking: bool,
/// Use beam search instead of sampling
#[serde(default)]
pub use_beam_search: bool,
@@ -282,6 +286,7 @@ impl Default for ChatCompletionRequest {
include_reasoning: true,
parallel_tool_calls: None,
user: None,
watermarking: true,
use_beam_search: false,
top_k: None,
min_p: None,
@@ -115,6 +115,7 @@ pub(super) fn prepare_completion_request(
mm_features: None,
sampling_params: SamplingParams {
temperature: request.temperature,
watermarking: request.watermarking,
top_p: request.top_p,
top_k: request.top_k,
seed: request.seed,
@@ -90,6 +90,10 @@ pub struct CompletionRequest {
pub user: Option<String>,
// -------- vLLM Sampling Parameters --------
/// Whether to apply the engine's configured watermark to this request.
#[serde(default = "default_true")]
pub watermarking: bool,
/// Options for streaming response
pub stream_options: Option<StreamOptions>,
+43
View File
@@ -2736,6 +2736,49 @@ async fn non_stream_chat_returns_json_response() {
assert!(!message.contains_key("tool_calls"), "{json}");
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[serial]
async fn openai_watermarking_defaults_and_opt_out_reach_engine_core() {
for endpoint in ["/v1/chat/completions", "/v1/completions"] {
for stream in [false, true] {
for watermarking in [None, Some(true), Some(false)] {
let (mut app, engine_task) = test_app_with_backend_and_engine_request_check(
Arc::new(FakeChatBackend::new()),
move |request| {
let params = request.sampling_params.as_ref().expect("sampling params");
assert_eq!(params.watermarking, watermarking.unwrap_or(true));
},
)
.await;
let mut body = json!({"stream": stream, "max_tokens": 2});
if endpoint == "/v1/chat/completions" {
body["messages"] = json!([{"role": "user", "content": "hi"}]);
} else {
body["prompt"] = json!("hi");
}
if let Some(watermarking) = watermarking {
body["watermarking"] = json!(watermarking);
}
let response = app
.call(
Request::builder()
.method("POST")
.uri(endpoint)
.header("content-type", "application/json")
.body(Body::from(body.to_string()))
.expect("build request"),
)
.await
.expect("call app");
assert_eq!(response.status(), StatusCode::OK);
to_bytes(response.into_body(), usize::MAX).await.expect("drain response");
engine_task.await.expect("mock engine task");
}
}
}
}
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
#[serial]
async fn non_stream_chat_image_url_reaches_engine_mm_features() {