[Rust Frontend] Keep --max-model-len engine-owned (#49944)

Co-authored-by: OpenAI Codex <codex@openai.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>

rust/src/cmd/src/cli.rs

@@ -148,11 +148,6 @@ pub struct SharedRuntimeArgs {
     #[arg(long)]
     #[serde(default)]
     pub language_model_only: bool,
-    /// Override the maximum model context length. When set, the frontend uses
-    /// this value instead of the model's `max_position_embeddings` from
-    /// `config.json`.
-    #[arg(long)]
-    pub max_model_len: Option<u32>,
     /// Maximum number of log probabilities to return when `logprobs` is
     /// specified in sampling parameters. `-1` means no cap.
     #[arg(long, value_parser = clap::value_parser!(i32).range(-1..), allow_negative_numbers = true)]
@@ -664,7 +659,6 @@ impl ServeArgs {
 
         self.managed_engine.clone().into_config(
             self.runtime.model.clone(),
-            self.runtime.max_model_len,
             self.runtime.max_logprobs,
             profiler_config,
             reasoning_parser.as_deref(),

rust/src/cmd/src/cli/tests.rs

@@ -58,9 +58,6 @@ fn serve_args_forward_python_flags_with_separator() {
                         reasoning_parser: Auto,
                         renderer: Auto,
                         language_model_only: false,
-                        max_model_len: Some(
-                            512,
-                        ),
                         max_logprobs: None,
                         grpc_port: None,
                         shutdown_timeout: 0,
@@ -102,6 +99,9 @@ fn serve_args_forward_python_flags_with_separator() {
                         handshake_port: None,
                         data_parallel_size: 1,
                         data_parallel_size_local: None,
+                        max_model_len: Some(
+                            "512",
+                        ),
                         python_args: [
                             "--dtype",
                             "float16",
@@ -756,7 +756,6 @@ fn frontend_args_accept_json() {
                         reasoning_parser: None,
                         renderer: Auto,
                         language_model_only: false,
-                        max_model_len: None,
                         max_logprobs: None,
                         grpc_port: None,
                         shutdown_timeout: 0,
@@ -823,11 +822,32 @@ fn frontend_args_json_applies_defaults() {
     assert_eq!(args.runtime.tool_call_parser, ParserSelection::None);
     assert_eq!(args.runtime.reasoning_parser, ParserSelection::None);
     assert_eq!(args.runtime.renderer, RendererSelection::Auto);
-    assert_eq!(args.runtime.max_model_len, None);
     assert_eq!(args.runtime.max_logprobs, None);
     assert_eq!(args.runtime.shutdown_timeout, 0);
 }
 
+#[test]
+fn frontend_args_json_ignores_engine_owned_max_model_len() {
+    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",
+        r#"{"model_tag":"Qwen/Qwen3-0.6B","max_model_len":-1}"#,
+    ])
+    .unwrap();
+
+    let Command::Frontend(args) = cli.command else {
+        panic!("expected frontend args");
+    };
+    assert_eq!(args.runtime.model, "Qwen/Qwen3-0.6B");
+}
+
 #[test]
 fn frontend_args_json_accepts_supported_non_default_fields() {
     let cli = Cli::try_parse_from([
@@ -840,7 +860,7 @@ fn frontend_args_json_accepts_supported_non_default_fields() {
         "--output-address",
         "ipc:///tmp/output.sock",
         "--args-json",
-        r#"{"model_tag":"Qwen/Qwen3-0.6B","engine_ready_timeout_secs":42,"tool_call_parser":"hermes","reasoning_parser":"qwen3_thinking","tokenizer_mode":"deepseek_v32","language_model_only":true,"max_model_len":8192,"max_logprobs":-1,"shutdown_timeout":3}"#,
+        r#"{"model_tag":"Qwen/Qwen3-0.6B","engine_ready_timeout_secs":42,"tool_call_parser":"hermes","reasoning_parser":"qwen3_thinking","tokenizer_mode":"deepseek_v32","language_model_only":true,"max_logprobs":-1,"shutdown_timeout":3}"#,
     ])
     .unwrap();
 
@@ -858,11 +878,38 @@ fn frontend_args_json_accepts_supported_non_default_fields() {
     );
     assert_eq!(args.runtime.renderer, RendererSelection::DeepSeekV32);
     assert!(args.runtime.language_model_only);
-    assert_eq!(args.runtime.max_model_len, Some(8192));
     assert_eq!(args.runtime.max_logprobs, Some(-1));
     assert_eq!(args.runtime.shutdown_timeout, 3);
 }
 
+#[test]
+fn serve_args_forward_auto_max_model_len_to_managed_engine() {
+    let cli = Cli::try_parse_from([
+        "vllm-rs",
+        "serve",
+        "Qwen/Qwen3-0.6B",
+        "--max-model-len",
+        "auto",
+    ])
+    .unwrap();
+
+    let Command::Serve(args) = cli.command else {
+        panic!("expected serve args");
+    };
+    assert_eq!(args.managed_engine.max_model_len.as_deref(), Some("auto"));
+
+    let config = args.to_managed_engine_config(5555);
+    expect![[r#"
+        [
+            "--max-model-len",
+            "auto",
+            "--reasoning-parser",
+            "qwen3",
+        ]
+    "#]]
+    .assert_debug_eq(&config.python_args);
+}
+
 #[test]
 fn serve_args_accept_none_reasoning_parser() {
     let cli = Cli::try_parse_from([
@@ -1278,7 +1325,6 @@ fn serve_args_accept_handshake_aliases() {
                         reasoning_parser: Auto,
                         renderer: Auto,
                         language_model_only: false,
-                        max_model_len: None,
                         max_logprobs: None,
                         grpc_port: None,
                         shutdown_timeout: 0,
@@ -1322,6 +1368,7 @@ fn serve_args_accept_handshake_aliases() {
                         ),
                         data_parallel_size: 4,
                         data_parallel_size_local: None,
+                        max_model_len: None,
                         python_args: [],
                     },
                 },

rust/src/managed-engine/src/cli.rs

@@ -39,6 +39,12 @@ pub struct ManagedEngineArgs {
     /// Number of data parallel replicas to run on this node.
     #[arg(long)]
     pub data_parallel_size_local: Option<usize>,
+    /// Maximum model context length forwarded to the managed Python engine.
+    ///
+    /// Rust leaves validation to Python so values such as `auto` and
+    /// human-readable integers retain their engine-owned semantics.
+    #[arg(long)]
+    pub max_model_len: Option<String>,
 
     /// Additional arguments forwarded to `python -m vllm.entrypoints.cli.main
     /// serve ...`.
@@ -78,7 +84,6 @@ impl ManagedEngineArgs {
     pub fn into_config(
         self,
         model: String,
-        max_model_len: Option<u32>,
         max_logprobs: Option<i32>,
         profiler_config: Option<String>,
         reasoning_parser: Option<&str>,
@@ -89,9 +94,9 @@ impl ManagedEngineArgs {
     ) -> ManagedEngineConfig {
         let mut python_args = self.python_args;
         // Manually forward some args to the Python engine.
-        if let Some(max_model_len) = max_model_len {
+        if let Some(max_model_len) = self.max_model_len {
             python_args.push("--max-model-len".to_string());
-            python_args.push(max_model_len.to_string());
+            python_args.push(max_model_len);
         }
         if let Some(max_logprobs) = max_logprobs {
             python_args.push("--max-logprobs".to_string());