[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
@@ -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