[Rust Frontend] Support HF config overrides --hf-overrides (#56931)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
Bugen Zhao
2026-09-16 02:52:55 +00:00
committed by GitHub
parent 0673de9800
commit bfd713bf8b
16 changed files with 298 additions and 19 deletions
+1
View File
@@ -5813,6 +5813,7 @@ dependencies = [
"vllm-engine-core-client",
"vllm-managed-engine",
"vllm-server",
"vllm-text",
"vllm-tracing",
]
+9 -13
View File
@@ -129,7 +129,8 @@ pub(super) async fn load_model_backends(
model_id: &str,
options: LoadModelBackendsOptions,
) -> Result<LoadedModelBackends> {
let files = ResolvedModelFiles::new(model_id, options.revision.as_deref()).await?;
let mut files = ResolvedModelFiles::new(model_id, options.revision.as_deref()).await?;
files.apply_overrides(&options.hf_overrides)?;
let text_backend = HfTextBackend::from_resolved_model_files(
files.clone(),
model_id.to_string(),
@@ -166,13 +167,11 @@ fn resolve_multimodal_render_info(
#[cfg(test)]
mod tests {
use std::collections::HashMap;
use std::path::PathBuf;
use std::sync::Arc;
use tempfile::tempdir;
use thiserror_ext::AsReport as _;
use vllm_text::Prompt;
use vllm_text::backend::hf::TokenizerSource;
use vllm_text::tokenizer::DynTokenizer;
use vllm_tokenizer::test_utils::TestTokenizer;
@@ -206,16 +205,12 @@ mod tests {
write_json(&config_path, config_json);
write_json(&tokenizer_config_path, tokenizer_config_json);
vllm_text::backend::hf::ResolvedModelFiles {
tokenizer: TokenizerSource::HuggingFace(PathBuf::from("/tmp/unused-tokenizer.json")),
tokenizer_config_path: Some(tokenizer_config_path),
generation_config_path: None,
preprocessor_config_path: None,
video_preprocessor_config_path: None,
processor_config_path: None,
chat_template_path: None,
config_path: Some(config_path),
}
write_json(&root.join("tokenizer.json"), "{}");
futures::executor::block_on(vllm_text::backend::hf::ResolvedModelFiles::new(
root.to_str().unwrap(),
None,
))
.unwrap()
}
fn test_tokenizer() -> DynTokenizer {
@@ -232,6 +227,7 @@ mod tests {
"test-model".to_string(),
LoadModelBackendsOptions {
revision: None,
hf_overrides: Default::default(),
generation_config: Default::default(),
renderer,
language_model_only: false,
+3
View File
@@ -5,6 +5,7 @@ use std::collections::HashMap;
use std::sync::Arc;
use serde_json::Value;
use vllm_text::backend::hf::HfOverrides;
use vllm_text::{DynTextBackend, GenerationConfigMode, TextBackend};
use crate::error::Result;
@@ -63,6 +64,8 @@ pub type DynChatTextBackend = Arc<dyn ChatTextBackend>;
pub struct LoadModelBackendsOptions {
/// Model revision on the Hugging Face Hub (branch, tag, or commit SHA).
pub revision: Option<String>,
/// JSON Merge Patch applied to the model config before loading any backend.
pub hf_overrides: HfOverrides,
/// Which generation-config sampling defaults to inherit.
pub generation_config: GenerationConfigMode,
/// Which chat renderer implementation to use.
+1
View File
@@ -33,6 +33,7 @@ vllm-chat.workspace = true
vllm-engine-core-client.workspace = true
vllm-managed-engine.workspace = true
vllm-server.workspace = true
vllm-text.workspace = true
vllm-tracing.workspace = true
[dev-dependencies]
+17
View File
@@ -33,6 +33,7 @@ use vllm_server::{
DEFAULT_KEEP_ALIVE_TIMEOUT, HttpListenerMode, LoraModulePath, ParserSelection, RenderConfig,
RendererSelection,
};
use vllm_text::backend::hf::HfOverrides;
use crate::cli::ssl::SslArgs;
use crate::cli::unsupported::UnsupportedArgs;
@@ -113,6 +114,9 @@ pub struct RenderArgs {
/// Model revision on the Hugging Face Hub (branch, tag, or commit SHA).
#[arg(long)]
revision: Option<String>,
/// JSON Merge Patch (RFC 7396) for config.json; null removes a field.
#[arg(long, value_parser = parse_json::<HfOverrides>, default_value = "{}", value_name = "JSON")]
hf_overrides: HfOverrides,
/// HTTP bind host.
#[arg(long, default_value = "127.0.0.1")]
host: String,
@@ -161,6 +165,7 @@ impl RenderArgs {
RenderConfig {
model: self.model,
revision: self.revision,
hf_overrides: self.hf_overrides,
served_model_name: self.served_model_name,
host: self.host,
port: self.port,
@@ -204,6 +209,12 @@ pub struct SharedRuntimeArgs {
#[serde(default)]
pub revision: Option<String>,
/// JSON Merge Patch (RFC 7396) for config.json; null removes a field.
/// Objects merge recursively and arrays/scalars replace existing values.
#[arg(long, value_parser = parse_json::<HfOverrides>, default_value = "{}", value_name = "JSON")]
#[serde(default)]
pub hf_overrides: HfOverrides,
/// The source of generation-config sampling defaults. `"auto"` loads the
/// model's defaults, while `"vllm"` uses vLLM's neutral defaults.
#[arg(long, default_value_t)]
@@ -500,6 +511,7 @@ impl SharedRuntimeArgs {
},
model: self.model,
revision: self.revision,
hf_overrides: self.hf_overrides,
generation_config: self.generation_config,
served_model_name: self.served_model_name,
listener_mode: HttpListenerMode::InheritedFd { fd: listen_fd },
@@ -556,6 +568,7 @@ impl SharedRuntimeArgs {
coordinator_mode: CoordinatorMode::MaybeInProc,
model: self.model,
revision: self.revision,
hf_overrides: self.hf_overrides,
generation_config: self.generation_config,
served_model_name: self.served_model_name,
listener_mode,
@@ -761,6 +774,9 @@ impl ServeArgs {
let reasoning_parser =
effective_engine_reasoning_parser(&self.runtime.reasoning_parser, &self.runtime.model);
let profiler_config = self.runtime.profiler_config_json();
let hf_overrides = (!self.runtime.hf_overrides.is_empty()).then(|| {
serde_json::to_string(&self.runtime.hf_overrides).expect("JSON object serializes")
});
self.managed_engine.clone().into_config(
self.runtime.model.clone(),
@@ -773,6 +789,7 @@ impl ServeArgs {
self.runtime.shutdown_timeout,
handshake_port,
self.runtime.limit_mm_per_prompt_json(),
hf_overrides,
)
}
}
+85
View File
@@ -177,6 +177,9 @@ fn serve_args_forward_python_flags_with_separator() {
runtime: SharedRuntimeArgs {
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
engine_ready_timeout_secs: 600,
tool_call_parser: Auto,
@@ -1013,6 +1016,9 @@ fn frontend_args_accept_json() {
runtime: SharedRuntimeArgs {
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
engine_ready_timeout_secs: 600,
tool_call_parser: None,
@@ -1120,6 +1126,73 @@ fn frontend_args_json_ignores_engine_owned_max_model_len() {
assert_eq!(args.runtime.model, "Qwen/Qwen3-0.6B");
}
#[test]
fn hf_overrides_preserve_merge_patch_in_cli_and_python_bootstrap() {
let patch = r#"{"text_config":{"rope_parameters":{"factor":4}},"sliding_window":null}"#;
let cli = Cli::try_parse_from([
"vllm-rs",
"serve",
"Qwen/Qwen3-0.6B",
"--hf-overrides",
patch,
])
.unwrap();
let Command::Serve(args) = cli.command else {
panic!("expected serve args")
};
let expected: serde_json::Value = serde_json::from_str(patch).unwrap();
assert_eq!(
serde_json::to_value(&args.runtime.hf_overrides).unwrap(),
expected
);
let engine = args.to_managed_engine_config(1234);
let index = engine.python_args.iter().position(|arg| arg == "--hf-overrides").unwrap();
assert_eq!(
serde_json::from_str::<serde_json::Value>(&engine.python_args[index + 1]).unwrap(),
expected
);
assert!(!engine.python_args.iter().any(|arg| arg == "--hf-config-path"));
let config = args.to_frontend_config("tcp://127.0.0.1:1234".to_string());
assert_eq!(serde_json::to_value(config.hf_overrides).unwrap(), expected);
let payload = serde_json::json!({"model_tag":"Qwen/Qwen3-0.6B", "hf_overrides":expected});
let cli = Cli::try_parse_from([
"vllm-rs",
"frontend",
"--listen-fd",
"3",
"--input-address",
"ipc:///tmp/input.sock",
"--output-address",
"ipc:///tmp/output.sock",
"--args-json",
&payload.to_string(),
])
.unwrap();
let Command::Frontend(args) = cli.command else {
panic!("expected frontend args")
};
assert_eq!(
serde_json::to_value(args.into_config().hf_overrides).unwrap(),
expected
);
}
#[test]
fn hf_overrides_reject_non_object_inputs() {
for patch in ["null", "[]", "1", r#""callable""#] {
let error = Cli::try_parse_from([
"vllm-rs",
"serve",
"Qwen/Qwen3-0.6B",
"--hf-overrides",
patch,
])
.unwrap_err();
assert_eq!(error.kind(), clap::error::ErrorKind::ValueValidation);
}
}
#[test]
fn frontend_args_json_accepts_supported_non_default_fields() {
let cli = Cli::try_parse_from([
@@ -1613,6 +1686,9 @@ fn serve_args_accept_handshake_aliases() {
runtime: SharedRuntimeArgs {
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
engine_ready_timeout_secs: 600,
tool_call_parser: Auto,
@@ -1763,6 +1839,9 @@ fn serve_frontend_config_uses_dp_address_as_advertised_host() {
coordinator_mode: MaybeInProc,
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
served_model_name: [],
listener_mode: BindTcp {
@@ -1852,6 +1931,9 @@ fn serve_frontend_config_keeps_tcp_transport_for_non_local_only_topology() {
coordinator_mode: MaybeInProc,
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
served_model_name: [],
listener_mode: BindTcp {
@@ -1963,6 +2045,9 @@ fn frontend_config_uses_external_coordinator_when_coordinator_address_is_present
},
model: "Qwen/Qwen3-0.6B",
revision: None,
hf_overrides: HfOverrides(
{},
),
generation_config: Auto,
served_model_name: [],
listener_mode: InheritedFd {
-5
View File
@@ -235,11 +235,6 @@ pub struct EngineUnsupportedArgs {
#[arg(long, default_missing_value = "true", num_args = 0..=1)]
pub hf_token: Option<Unsupported>,
/// If a dictionary, contains arguments to be forwarded to the Hugging Face
/// config. If a callable, it is called to update the HuggingFace config.
#[arg(long)]
pub hf_overrides: Option<Unsupported>,
/// Overrides or sets generation config. e.g. `{"temperature": 0.5}`. If
/// used with `--generation-config auto`, the override parameters will be
/// merged with the default config from the model. If used with
+5
View File
@@ -93,6 +93,7 @@ impl ManagedEngineArgs {
shutdown_timeout: u64,
handshake_port: u16,
limit_mm_per_prompt: Option<String>,
hf_overrides: Option<String>,
) -> ManagedEngineConfig {
let mut python_args = self.python_args;
// Manually forward some args to the Python engine.
@@ -136,6 +137,10 @@ impl ManagedEngineArgs {
python_args.push("--limit-mm-per-prompt".to_string());
python_args.push(limit_mm_per_prompt);
}
if let Some(hf_overrides) = hf_overrides {
python_args.push("--hf-overrides".to_string());
python_args.push(hf_overrides);
}
ManagedEngineConfig {
python: self.python,
@@ -61,6 +61,7 @@ async fn main() -> Result<()> {
coordinator_mode: CoordinatorMode::MaybeInProc,
model: args.model,
revision: None,
hf_overrides: Default::default(),
generation_config: Default::default(),
served_model_name: vec![],
listener_mode: HttpListenerMode::BindTcp {
+3
View File
@@ -16,6 +16,7 @@ use vllm_chat::{
ChatTemplateContentFormatOption, GenerationConfigMode, ParserSelection, RendererSelection,
};
use vllm_engine_core_client::{CoordinatorMode as EngineCoreCoordinatorMode, TransportMode};
use vllm_text::backend::hf::HfOverrides;
/// Default keep-alive idle timeout (seconds); also the head-read bound
/// when keep-alive is disabled (`0`).
@@ -214,6 +215,8 @@ pub struct Config {
pub model: String,
/// Model revision on the Hugging Face Hub (branch, tag, or commit SHA).
pub revision: Option<String>,
/// JSON Merge Patch applied to the model config before backend initialization.
pub hf_overrides: HfOverrides,
/// Which generation-config sampling defaults to inherit.
pub generation_config: GenerationConfigMode,
/// Model name(s) exposed to clients via the OpenAI API. When non-empty,
+1
View File
@@ -99,6 +99,7 @@ async fn build_state(config: &Config) -> Result<Arc<AppState>> {
&config.model,
LoadModelBackendsOptions {
revision: config.revision.clone(),
hf_overrides: config.hf_overrides.clone(),
generation_config: config.generation_config,
renderer: config.renderer,
language_model_only: config.language_model_only,
+4
View File
@@ -13,6 +13,7 @@ use vllm_chat::{
ParserSelection, RendererSelection, load_model_backends,
};
use vllm_text::TextRequestProcessor;
use vllm_text::backend::hf::HfOverrides;
use crate::{
HttpListenerMode, TlsConfig,
@@ -25,6 +26,7 @@ use crate::{
pub struct RenderConfig {
pub model: String,
pub revision: Option<String>,
pub hf_overrides: HfOverrides,
pub served_model_name: Vec<String>,
pub host: String,
pub port: u16,
@@ -69,6 +71,7 @@ async fn build_state(config: &RenderConfig) -> Result<Arc<RenderState>> {
&config.model,
LoadModelBackendsOptions {
revision: config.revision.clone(),
hf_overrides: config.hf_overrides.clone(),
generation_config: Default::default(),
renderer: config.renderer,
language_model_only: true,
@@ -153,6 +156,7 @@ mod tests {
RenderConfig {
model: "test-model".to_string(),
revision: None,
hf_overrides: Default::default(),
served_model_name: Vec::new(),
host: "127.0.0.1".to_string(),
port: 8000,
+1 -1
View File
@@ -16,6 +16,7 @@ reqwest.workspace = true
serde.workspace = true
serde_json.workspace = true
serde_with.workspace = true
tempfile.workspace = true
thiserror.workspace = true
thiserror-ext.workspace = true
tracing.workspace = true
@@ -28,7 +29,6 @@ vllm-tokenizer.workspace = true
expect-test.workspace = true
futures.workspace = true
serial_test.workspace = true
tempfile.workspace = true
tokio.workspace = true
vllm-llm = { workspace = true, features = ["test-util"] }
vllm-tokenizer = { workspace = true, features = ["test-utils"] }
+2
View File
@@ -3,6 +3,7 @@
mod config;
mod model_files;
mod overrides;
use std::collections::BTreeSet;
use std::sync::Arc;
@@ -18,6 +19,7 @@ pub use self::config::{
load_tokenizer_config,
};
pub use self::model_files::{ResolvedModelFiles, TokenizerSource};
pub use self::overrides::HfOverrides;
use crate::backend::{GenerationConfigMode, SamplingHints, TextBackend};
use crate::error::Result;
@@ -2,11 +2,13 @@
// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
use std::path::{Path, PathBuf};
use std::sync::Arc;
use hf_hub::api::tokio::{Api, ApiBuilder, ApiRepo};
use hf_hub::{Cache, Repo, RepoType};
use thiserror_ext::AsReport as _;
use super::HfOverrides;
use super::config::{HfTokenizerConfig, load_tokenizer_config};
use crate::error::{Error, Result};
@@ -50,9 +52,27 @@ pub struct ResolvedModelFiles {
pub processor_config_path: Option<PathBuf>,
pub chat_template_path: Option<PathBuf>,
pub config_path: Option<PathBuf>,
/// Keep a patched config alive across clones and backend ownership.
config_temp_path: Option<Arc<tempfile::TempPath>>,
}
impl ResolvedModelFiles {
/// Materialize a merge-patched config owned by these resolved model files.
pub fn apply_overrides(&mut self, overrides: &HfOverrides) -> Result<()> {
if overrides.is_empty() {
return Ok(());
}
let path = materialize_config(self.config_path.as_deref(), overrides).map_err(|error| {
Error::Tokenizer(format!(
"failed to materialize HF overrides: {}",
error.as_report()
))
})?;
self.config_path = Some(path.to_path_buf());
self.config_temp_path = Some(Arc::new(path));
Ok(())
}
/// Resolve tokenizer/config files from a local model directory first when
/// `model_id` points to one, otherwise consult the local HF cache and
/// finally the Hub.
@@ -84,6 +104,24 @@ impl ResolvedModelFiles {
}
}
fn materialize_config(
path: Option<&Path>,
overrides: &HfOverrides,
) -> anyhow::Result<tempfile::TempPath> {
let mut config: serde_json::Value = match path {
Some(path) => serde_json::from_slice(&std::fs::read(path)?)?,
None => serde_json::json!({}),
};
anyhow::ensure!(config.is_object(), "HF model config must be a JSON object");
overrides.apply(&mut config);
let file = tempfile::Builder::new()
.prefix("vllm-rs-hf-config-")
.suffix(".json")
.tempfile()?;
serde_json::to_writer(file.as_file(), &config)?;
Ok(file.into_temp_path())
}
fn resolve_local_model_files(model_dir: &Path) -> Result<ResolvedModelFiles> {
let tokenizer_config_path = local_file_if_exists(model_dir, "tokenizer_config.json");
let tokenizer_config = load_tokenizer_config(tokenizer_config_path.as_deref())?;
@@ -101,6 +139,7 @@ fn resolve_local_model_files(model_dir: &Path) -> Result<ResolvedModelFiles> {
processor_config_path: local_file_if_exists(model_dir, "processor_config.json"),
chat_template_path: discover_chat_template_in_dir(model_dir),
config_path: local_file_if_exists(model_dir, "config.json"),
config_temp_path: None,
})
}
@@ -164,6 +203,7 @@ async fn resolve_remote_model_files(
processor_config_path,
chat_template_path,
config_path,
config_temp_path: None,
})
}
@@ -204,6 +244,7 @@ fn resolve_cached_model_files(cache: &Cache, repo: &Repo) -> Result<Option<Resol
processor_config_path,
chat_template_path,
config_path,
config_temp_path: None,
}))
}
@@ -434,6 +475,37 @@ mod tests {
use super::{ResolvedModelFiles, TokenizerSource};
#[tokio::test]
async fn overrides_keep_configs_alive_across_clones_and_isolate_instances() {
let source = tempdir().unwrap();
let original = r#"{"vocab_size":10}"#;
fs::write(source.path().join("config.json"), original).unwrap();
fs::write(source.path().join("tokenizer.json"), "{}").unwrap();
let mut first =
ResolvedModelFiles::new(source.path().to_str().unwrap(), None).await.unwrap();
let patch = serde_json::from_value(serde_json::json!({"vocab_size":20})).unwrap();
first.apply_overrides(&patch).unwrap();
let first_path = first.config_path.clone().unwrap();
let mut second = first.clone();
second.apply_overrides(&patch).unwrap();
let second_path = second.config_path.clone().unwrap();
assert_ne!(first_path, second_path);
let retained = first.clone();
drop(first);
assert!(first_path.exists());
drop(second);
assert!(!second_path.exists());
assert!(first_path.exists());
drop(retained);
assert!(!first_path.exists());
assert_eq!(
fs::read_to_string(source.path().join("config.json")).unwrap(),
original
);
}
#[tokio::test]
async fn resolved_model_files_prefers_absolute_local_model_dir() {
let dir = tempdir().expect("create temp dir");
+93
View File
@@ -0,0 +1,93 @@
// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright contributors to the vLLM project
use serde::{Deserialize, Serialize};
use serde_json::{Map, Value};
/// An object-shaped JSON Merge Patch (RFC 7396) for the model's HF config.
/// Objects merge recursively, arrays/scalars replace, and null removes a key.
#[derive(Debug, Clone, Default, PartialEq, Eq, Deserialize, Serialize)]
#[serde(transparent)]
pub struct HfOverrides(pub Map<String, Value>);
impl HfOverrides {
pub fn is_empty(&self) -> bool {
self.0.is_empty()
}
/// Apply the patch to an in-memory config, preserving the source file.
pub fn apply(&self, config: &mut Value) {
merge_object(config, &self.0);
}
}
/// Apply RFC 7396 object-member updates at every nesting level.
fn merge_object(target: &mut Value, patch: &Map<String, Value>) {
if !target.is_object() {
*target = Value::Object(Map::new());
}
let target = target.as_object_mut().unwrap();
for (key, value) in patch {
match value {
Value::Null => {
target.remove(key);
}
Value::Object(object) => {
merge_object(target.entry(key.clone()).or_insert(Value::Null), object)
}
_ => {
target.insert(key.clone(), value.clone());
}
}
}
}
#[cfg(test)]
mod tests {
use serde_json::json;
use super::HfOverrides;
#[test]
fn merge_patch_preserves_siblings_replaces_arrays_and_deletes_null_members() {
let mut config = json!({
"rope_parameters": {"rope_type": "yarn", "factor": 2},
"text_config": {"rope_parameters": {"rope_type": "yarn", "factor": 2}},
"architectures": ["Old", "Other"],
"sliding_window": 4096
});
let patch: HfOverrides = serde_json::from_value(json!({
"rope_parameters": {"factor": 4},
"text_config": {"rope_parameters": {"factor": 4}},
"architectures": ["New"],
"sliding_window": null,
"missing": null
}))
.unwrap();
patch.apply(&mut config);
assert_eq!(
config,
json!({
"rope_parameters": {"rope_type": "yarn", "factor": 4},
"text_config": {"rope_parameters": {"rope_type": "yarn", "factor": 4}},
"architectures": ["New"]
})
);
}
#[test]
fn merge_patch_creates_objects_and_preserves_nulls_inside_replacement_arrays() {
let mut config = json!({"a": 1, "b": [1, 2]});
let patch: HfOverrides = serde_json::from_value(json!({
"a": {"nested": {"removed": null}},
"b": {},
"c": [null, {"literal": null}]
}))
.unwrap();
patch.apply(&mut config);
assert_eq!(
config,
json!({"a": {"nested": {}}, "b": {}, "c": [null, {"literal": null}]})
);
}
}