[Security] Fix chat template resource-exhaustion DoS (GHSA-4hhp-h66f-… (#50300)

Signed-off-by: jperezde <jperezde@redhat.com>
Signed-off-by: Juan Pérez de Algaba <jperezde@redhat.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Juan Pérez de Algaba
2026-09-30 08:32:00 -07:00
committed by GitHub
co-authored by Cursor
parent fc2c801ad9
commit cff08b461e
8 changed files with 230 additions and 9 deletions
+1 -1
View File
@@ -63,7 +63,7 @@ itertools = "0.14.0"
libc = "0.2.177"
llm-multimodal = { git = "https://github.com/smg-project/llm-multimodal", rev = "f0985ef65967615db2c79279aa07818499301bfd", default-features = false, features = ["native-tls"] }
mimalloc = "0.1.52"
minijinja = { version = "2.24", features = ["unstable_machinery", "json", "builtins", "loader", "loop_controls", "preserve_order"] }
minijinja = { version = "2.24", features = ["unstable_machinery", "json", "builtins", "loader", "loop_controls", "preserve_order", "fuel"] }
minijinja-contrib = { version = "2.24", features = ["pycompat"] }
native-tls-vendored = { package = "native-tls", version = "0.2.18", features = ["vendored"] }
ndarray = { version = "0.17", features = ["serde"] }
+6 -1
View File
@@ -180,7 +180,12 @@ impl ChatRequestProcessor {
async fn prepare_text_request(&self, request: ChatRequest) -> Result<TextRequest> {
// Stamp before rendering so render and tokenize count toward TTFT/e2e.
let arrival_time = vllm_llm::current_unix_timestamp_secs();
let mut rendered = self.backend.chat_renderer().render(&request)?;
let renderer = self.backend.chat_renderer();
let (mut rendered, request) = tokio::task::spawn_blocking(move || {
renderer.render(&request).map(|rendered| (rendered, request))
})
.await
.map_err(|e| Error::ChatTemplate(format!("render task join error: {e}")))??;
let chat_template_kwargs = std::mem::take(&mut rendered.effective_template_kwargs);
let (prompt, mm_features) = self.finalize_rendered_prompt(&request, rendered).await?;
Ok(TextRequest {
@@ -29,6 +29,11 @@ use crate::renderer::hf::{TemplateMessage, TemplateTool};
type Result<T> = std::result::Result<T, TemplateError>;
/// Default fuel budget for template rendering. Prevents unbounded loops and
/// string operations from exhausting CPU. Each template instruction consumes
/// one unit of fuel; when depleted, rendering aborts with an error.
const DEFAULT_TEMPLATE_FUEL: u64 = 500_000;
/// Build a pre-configured environment with the given template string.
fn build_environment(template: String) -> Result<Environment<'static>> {
let mut env = Environment::new();
@@ -36,6 +41,8 @@ fn build_environment(template: String) -> Result<Environment<'static>> {
env.set_trim_blocks(true);
env.set_lstrip_blocks(true);
env.set_fuel(Some(DEFAULT_TEMPLATE_FUEL));
env.add_template_owned("chat".to_owned(), template)?;
env.set_unknown_method_callback(minijinja_contrib::pycompat::unknown_method_callback);