[Rust Frontend] Recognize new frontend-owned serve args as unsupported or no-op (#58330)
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Bugen Zhao <i@bugenzhao.com>
Bugen Zhao committed
Sep 24, 2026 at 13:43 UTC
a5603f21ae7a8dc6f121cba3672123ca34319f43
1 file changed
+78
rust/src/cmd/src/cli/unsupported.rs
+78
@@ -246,6 +246,13 @@ pub struct EngineUnsupportedArgs {
246
#[arg(long)]
247
pub io_processor_plugin: Option<Unsupported>,
248
249
+ /// Number of worker threads in the renderer thread pool. The pool is
250
+ /// consumed by the async renderer path to parallelize tokenization, chat
251
+ /// template rendering, and multimodal preprocessing across concurrent
252
+ /// requests.
253
+ #[arg(long)]
254
+ pub renderer_num_workers: Option<Noop>,
255
+
256
/// Path to a dynamically reasoning parser plugin that can be dynamically
257
/// loaded and registered.
258
#[arg(long)]
@@ -281,6 +288,16 @@ pub struct EngineUnsupportedArgs {
288
)]
289
pub data_parallel_external_lb: Option<Unsupported>,
290
291
+ /// Enable fault tolerance for detailed error recovery, such as scaling
292
+ /// down fault DPEngineCore.
293
+ #[arg(
294
+ long,
295
+ visible_alias = "no-enable-fault-tolerance",
296
+ default_missing_value = "true",
297
+ num_args = 0..=1
298
+ )]
299
+ pub enable_fault_tolerance: Option<Unsupported>,
300
+
301
/// This feature is work in progress and no prefill optimization takes place
302
/// with this flag enabled currently.
303
#[arg(
@@ -308,6 +325,28 @@ pub struct EngineUnsupportedArgs {
325
#[arg(long)]
326
pub mm_processor_cache_type: Option<Unsupported>,
327
328
+ /// Hash algorithm to use for multi-modal input caching. Use `"sha256"` or
329
+ /// `"sha512"` for FIPS-compliant deployments.
330
+ #[arg(long)]
331
+ pub mm_hasher_algorithm: Option<Unsupported>,
332
+
333
+ /// IPC (inter-process communication) method for multimodal tensors.
334
+ /// - "direct_rpc": Use msgspec serialization via RPC
335
+ /// - "torch_shm": Use torch.multiprocessing shared memory for zero-copy IPC
336
+ #[arg(long)]
337
+ pub mm_tensor_ipc: Option<Unsupported>,
338
+
339
+ /// Device the HF multi-modal processor runs the image/video transform on.
340
+ /// Convenience for `--mm-processor-kwargs '{"device": ...}'`.
341
+ #[arg(long)]
342
+ pub mm_processor_device: Option<Unsupported>,
343
+
344
+ /// Amount of GPU memory (in GiB) sequestered on the engine's device for
345
+ /// GPU-side multimodal work in the API-server (frontend) process, such as
346
+ /// hardware video decoding.
347
+ #[arg(long)]
348
+ pub mm_ipc_gpu_memory_gb: Option<Unsupported>,
349
+
350
/// Dictionary mapping specific modalities to LoRA model paths.
351
#[arg(long)]
352
pub default_mm_loras: Option<Unsupported>,
@@ -324,6 +363,17 @@ pub struct EngineUnsupportedArgs {
363
#[arg(long)]
364
pub stream_interval: Option<Unsupported>,
365
366
+ /// Maximum number of requests that can be in-flight (waiting or running)
367
+ /// at the same time. When the limit is reached, new requests are rejected
368
+ /// with HTTP 503 so the client can retry on another instance.
369
+ #[arg(long)]
370
+ pub max_num_queued_reqs: Option<Unsupported>,
371
+
372
+ /// Maximum total prompt tokens of requests currently in the prefill phase.
373
+ /// When the limit is reached, new requests are rejected with HTTP 503.
374
+ #[arg(long)]
375
+ pub max_num_queued_tokens: Option<Unsupported>,
376
+
377
/// Structured outputs configuration.
378
#[arg(long)]
379
pub structured_outputs_config: Option<Noop>,
@@ -419,6 +469,15 @@ pub struct ServerUnsupportedArgs {
469
#[arg(long)]
470
pub max_log_len: Option<Unsupported>,
471
472
+ /// If set to True, include per-request timing metrics in API responses.
473
+ #[arg(
474
+ long,
475
+ visible_alias = "no-enable-per-request-metrics",
476
+ default_missing_value = "true",
477
+ num_args = 0..=1
478
+ )]
479
+ pub enable_per_request_metrics: Option<Unsupported>,
480
+
481
/// If set to True, enable tracking server_load_metrics in the app state.
482
#[arg(
483
long,
@@ -492,6 +551,15 @@ pub struct ServerUnsupportedArgs {
551
)]
552
pub tokens_only: Option<Unsupported>,
553
554
+ /// Controls the `system_fingerprint` field on responses: `full`, `hash`,
555
+ /// `custom`, or `none`.
556
+ #[arg(long)]
557
+ pub fingerprint_mode: Option<Unsupported>,
558
+
559
+ /// Literal fingerprint string used when `--fingerprint-mode=custom`.
560
+ #[arg(long)]
561
+ pub fingerprint_value: Option<Unsupported>,
562
+
563
/// Log level for uvicorn.
564
#[arg(long)]
565
pub uvicorn_log_level: Option<Unsupported>,
@@ -562,4 +630,14 @@ pub struct ServerUnsupportedArgs {
630
num_args = 0..=1
631
)]
632
pub enable_offline_docs: Option<Unsupported>,
633
+
634
+ /// If set, run pooling score MaxSim on GPU in the API server process.
635
+ /// Can significantly improve late-interaction scoring performance.
636
+ #[arg(
637
+ long,
638
+ visible_alias = "no-enable-flash-late-interaction",
639
+ default_missing_value = "true",
640
+ num_args = 0..=1
641
+ )]
642
+ pub enable_flash_late_interaction: Option<Noop>,
643
}