diff --git a/docs/training/weight_transfer/README.md b/docs/training/weight_transfer/README.md index d8e6c7e409..b394e41415 100644 --- a/docs/training/weight_transfer/README.md +++ b/docs/training/weight_transfer/README.md @@ -158,6 +158,8 @@ When running vLLM as an HTTP server, the following endpoints are available for w !!! note The HTTP weight transfer endpoints require `VLLM_SERVER_DEV_MODE=1` to be set. +The Rust frontend's optional gRPC `Control` service exposes the same pause, sleep, weight-transfer, and weight-version lifecycle for trusted sidecars. The `ServerInfo.rl_capabilities` response reports whether weight transfer and sleep mode were configured. Backend-specific `init_info` and `update_info` remain JSON metadata; model tensors continue to move over the configured NCCL, IPC, or sparse-NCCL transport. + ## Extending the System Every piece of the system is replaceable: the weights you send (`WeightSource`), diff --git a/docs/usage/security.md b/docs/usage/security.md index 68c05a6842..5f7b0ce753 100644 --- a/docs/usage/security.md +++ b/docs/usage/security.md @@ -345,7 +345,7 @@ vLLM supports loading out-of-tree HTTP routes via the `vllm.endpoint_plugins` en ## gRPC Interface -vLLM provides an optional gRPC Generate service on a separate TCP port, enabled via the `--grpc-port` flag. When not specified, no gRPC server is started. The gRPC listener binds to the same host address as the HTTP server. +vLLM provides optional gRPC `Inference` and `Control` services on a separate TCP port, enabled via the `--grpc-port` flag. When not specified, no gRPC server is started. The gRPC listener binds to the same host address as the HTTP server. **Warning:** The gRPC interface is **insecure by default** — it does not implement authentication, authorization, or encryption. It should be considered a private, internal interface intended for use only between co-located services within a trusted network. Do not expose the gRPC port to the public internet or untrusted clients. If you enable the gRPC interface, protect it via network-level access controls such as firewall rules, network segmentation, or deployment on an isolated private network. @@ -354,8 +354,9 @@ vLLM provides an optional gRPC Generate service on a separate TCP port, enabled An attacker who can reach the gRPC port can: 1. **Run arbitrary inference** via the `Generate` and `GenerateStream` RPCs without any credentials -2. **Consume GPU and compute resources** by submitting unbounded generation requests -3. **Cause Denial of Service** by exploiting bugs in the gRPC interface that can crash vLLM. +2. **Mutate engine state** by pausing generation, sleeping the engine, or initiating configured RL weight updates through the `Control` service +3. **Consume GPU and compute resources** by submitting unbounded generation requests +4. **Cause Denial of Service** by exploiting bugs in the gRPC interface that can crash vLLM. ### Recommendations diff --git a/rust/proto/control.proto b/rust/proto/control.proto index 0e25aea264..428ad64604 100644 --- a/rust/proto/control.proto +++ b/rust/proto/control.proto @@ -9,6 +9,21 @@ service Control { rpc GetModelInfo (GetModelInfoRequest) returns (ModelInfo) {} rpc Abort (AbortRequest) returns (AbortResponse) {} rpc GetKvEventSources (GetKvEventSourcesRequest) returns (GetKvEventSourcesResponse) {} + + // Reinforcement-learning lifecycle and weight updates. + rpc PauseGeneration (PauseGenerationRequest) returns (PauseGenerationResponse) {} + rpc ResumeGeneration (ResumeGenerationRequest) returns (ResumeGenerationResponse) {} + rpc IsPaused (IsPausedRequest) returns (IsPausedResponse) {} + rpc Sleep (SleepRequest) returns (SleepResponse) {} + rpc WakeUp (WakeUpRequest) returns (WakeUpResponse) {} + rpc IsSleeping (IsSleepingRequest) returns (IsSleepingResponse) {} + rpc InitWeightTransferEngine (InitWeightTransferEngineRequest) returns (InitWeightTransferEngineResponse) {} + rpc StartWeightUpdate (StartWeightUpdateRequest) returns (StartWeightUpdateResponse) {} + rpc StartDraftWeightUpdate (StartDraftWeightUpdateRequest) returns (StartDraftWeightUpdateResponse) {} + rpc UpdateWeights (UpdateWeightsRequest) returns (UpdateWeightsResponse) {} + rpc FinishWeightUpdate (FinishWeightUpdateRequest) returns (FinishWeightUpdateResponse) {} + rpc UpdateWeightVersion (UpdateWeightVersionRequest) returns (UpdateWeightVersionResponse) {} + rpc GetWeightVersion (GetWeightVersionRequest) returns (GetWeightVersionResponse) {} } message GetServerInfoRequest {} @@ -23,6 +38,14 @@ message ServerInfo { uint64 total_kv_blocks = 7; uint64 max_running_requests = 8; uint64 max_batched_tokens = 9; + RlCapabilities rl_capabilities = 11; +} + +message RlCapabilities { + bool weight_transfer_enabled = 1; + string weight_transfer_backend = 2; + bool sleep_mode_enabled = 3; + bool draft_weight_updates_enabled = 4; } message ParallelismInfo { @@ -53,6 +76,64 @@ message AbortRequest { message AbortResponse {} +// ====================================================================================== +// Reinforcement-learning control +// ====================================================================================== + +enum PauseMode { + PAUSE_MODE_UNSPECIFIED = 0; + PAUSE_MODE_ABORT = 1; + PAUSE_MODE_WAIT = 2; + PAUSE_MODE_KEEP = 3; +} + +message PauseGenerationRequest { + PauseMode mode = 1; + optional bool clear_cache = 2; +} +message PauseGenerationResponse {} + +message ResumeGenerationRequest {} +message ResumeGenerationResponse {} + +message IsPausedRequest {} +message IsPausedResponse { bool paused = 1; } + +message SleepRequest { + optional uint32 level = 1; + PauseMode mode = 2; +} +message SleepResponse {} + +message WakeUpRequest { repeated string tags = 1; } +message WakeUpResponse {} + +message IsSleepingRequest {} +message IsSleepingResponse { bool sleeping = 1; } + +// The payloads are backend-specific JSON objects. Tensor data remains on the +// configured NCCL, IPC, or sparse-NCCL transport rather than crossing gRPC. +message InitWeightTransferEngineRequest { bytes init_info_json = 1; } +message InitWeightTransferEngineResponse {} + +message StartWeightUpdateRequest {} +message StartWeightUpdateResponse {} + +message StartDraftWeightUpdateRequest {} +message StartDraftWeightUpdateResponse {} + +message UpdateWeightsRequest { bytes update_info_json = 1; } +message UpdateWeightsResponse {} + +message FinishWeightUpdateRequest { optional string weight_version = 1; } +message FinishWeightUpdateResponse {} + +message UpdateWeightVersionRequest { string weight_version = 1; } +message UpdateWeightVersionResponse {} + +message GetWeightVersionRequest {} +message GetWeightVersionResponse { string weight_version = 1; } + // ====================================================================================== // KV discovery // ====================================================================================== diff --git a/rust/src/engine-core-client/src/client.rs b/rust/src/engine-core-client/src/client.rs index d1c20ad582..7f8663f852 100644 --- a/rust/src/engine-core-client/src/client.rs +++ b/rust/src/engine-core-client/src/client.rs @@ -1,12 +1,14 @@ // SPDX-License-Identifier: Apache-2.0 // SPDX-FileCopyrightText: Copyright contributors to the vLLM project +use std::collections::BTreeMap; use std::sync::Arc; use std::time::Duration; use futures::future::{join_all, try_join_all}; use itertools::Itertools; use serde::Serialize; +use serde_json::Value as JsonValue; use tokio::sync::mpsc; use tokio_util::task::AbortOnDropHandle; use tracing::{debug, info, trace}; @@ -674,6 +676,77 @@ impl EngineCoreClient { .collect()) } + /// Initialize the configured RL weight-transfer backend. + pub async fn init_weight_transfer_engine(&self, init_info: JsonValue) -> Result<()> { + self.collective_rpc( + "init_weight_transfer_engine", + None, + Vec::::new(), + BTreeMap::from([("init_info".to_string(), init_info)]), + ) + .await?; + Ok(()) + } + + /// Start a weight update for the base model. + pub async fn start_weight_update(&self) -> Result<()> { + self.collective_rpc( + "start_weight_update", + None, + Vec::::new(), + BTreeMap::::new(), + ) + .await?; + Ok(()) + } + + /// Start a weight update for the speculative draft model. + pub async fn start_draft_weight_update(&self) -> Result<()> { + self.collective_rpc( + "start_draft_weight_update", + None, + Vec::::new(), + BTreeMap::::new(), + ) + .await?; + Ok(()) + } + + /// Apply one backend-specific weight metadata chunk. + pub async fn update_weights(&self, update_info: JsonValue) -> Result<()> { + self.collective_rpc( + "update_weights", + None, + Vec::::new(), + BTreeMap::from([("update_info".to_string(), update_info)]), + ) + .await?; + Ok(()) + } + + /// Finish the current weight update. + pub async fn finish_weight_update(&self) -> Result<()> { + self.collective_rpc( + "finish_weight_update", + None, + Vec::::new(), + BTreeMap::::new(), + ) + .await?; + Ok(()) + } + + /// Set the committed weight version on every connected engine. + pub async fn set_weight_version(&self, weight_version: &str) -> Result<()> { + self.call_utility::<(), _>("set_weight_version", (weight_version,)).await?; + Ok(()) + } + + /// Return the committed weight version agreed on by every connected engine. + pub async fn get_weight_version(&self) -> Result { + self.call_utility_consensus("get_weight_version", ()).await + } + /// Return whether the engine is currently sleeping at any level. pub async fn is_sleeping(&self) -> Result { self.call_utility_consensus("is_sleeping", ()).await diff --git a/rust/src/engine-core-client/src/mock_engine.rs b/rust/src/engine-core-client/src/mock_engine.rs index ee066635ff..28fb4d84fe 100644 --- a/rust/src/engine-core-client/src/mock_engine.rs +++ b/rust/src/engine-core-client/src/mock_engine.rs @@ -67,6 +67,9 @@ pub fn default_ready_response() -> EngineCoreReadyResponse { kv_cache_size_tokens: None, kv_cache_max_concurrency: None, kv_events_config: None, + weight_transfer_backend: None, + enable_sleep_mode: false, + supports_draft_weight_updates: false, } } diff --git a/rust/src/engine-core-client/src/protocol/handshake.rs b/rust/src/engine-core-client/src/protocol/handshake.rs index a4e1cebde6..b4214a06f2 100644 --- a/rust/src/engine-core-client/src/protocol/handshake.rs +++ b/rust/src/engine-core-client/src/protocol/handshake.rs @@ -87,6 +87,15 @@ pub struct EngineCoreReadyResponse { /// KV-event publisher configuration, if configured. #[serde(default)] pub kv_events_config: Option, + /// Configured RL weight-transfer backend, if weight transfer is enabled. + #[serde(default)] + pub weight_transfer_backend: Option, + /// Whether the engine was started with sleep mode enabled. + #[serde(default)] + pub enable_sleep_mode: bool, + /// Whether the engine has a speculative draft model that can be updated. + #[serde(default)] + pub supports_draft_weight_updates: bool, } /// Frontend-owned ZMQ addresses that are sent to the engine during startup diff --git a/rust/src/engine-core-client/src/tests/client.rs b/rust/src/engine-core-client/src/tests/client.rs index 1d8d6e37cc..8f07a8059e 100644 --- a/rust/src/engine-core-client/src/tests/client.rs +++ b/rust/src/engine-core-client/src/tests/client.rs @@ -2686,6 +2686,12 @@ fn python_msgpack_fixtures_match_rust_encoding() { let ready_response: EngineCoreReadyResponse = rmp_serde::from_slice(&hex::decode(ready_response_hex).unwrap()).unwrap(); + assert_eq!( + ready_response.weight_transfer_backend.as_deref(), + Some("nccl") + ); + assert!(ready_response.enable_sleep_mode); + assert!(ready_response.supports_draft_weight_updates); let kv_events_config = ready_response.kv_events_config.expect("KV events config should decode"); assert!(kv_events_config.enable_kv_cache_events); assert_eq!(kv_events_config.publisher, "zmq"); diff --git a/rust/src/engine-core-client/src/tests/python_compat.py b/rust/src/engine-core-client/src/tests/python_compat.py index 7883c7150c..389e97cd99 100755 --- a/rust/src/engine-core-client/src/tests/python_compat.py +++ b/rust/src/engine-core-client/src/tests/python_compat.py @@ -412,6 +412,9 @@ class EngineCoreReadyResponse: kv_cache_size_tokens: int | None = None kv_cache_max_concurrency: float | None = None kv_events_config: KVEventsConfig | None = None + weight_transfer_backend: str | None = None + enable_sleep_mode: bool = False + supports_draft_weight_updates: bool = False ready_response = EngineCoreReadyResponse( @@ -430,6 +433,9 @@ ready_response = EngineCoreReadyResponse( max_num_seqs=256, max_num_batched_tokens=8192, instance_id="test-instance", + weight_transfer_backend="nccl", + enable_sleep_mode=True, + supports_draft_weight_updates=True, kv_events_config=KVEventsConfig( enable_kv_cache_events=True, publisher="zmq", diff --git a/rust/src/server/src/grpc/control.rs b/rust/src/server/src/grpc/control.rs index 201d8ace95..c7d9af88c8 100644 --- a/rust/src/server/src/grpc/control.rs +++ b/rust/src/server/src/grpc/control.rs @@ -3,9 +3,13 @@ use std::sync::Arc; +use serde_json::Value as JsonValue; use thiserror_ext::AsReport as _; +use tokio::sync::Mutex; use tonic::{Request, Response, Status}; +use vllm_engine_core_client::EngineCoreClient; use vllm_engine_core_client::protocol::handshake::EngineCoreReadyResponse; +use vllm_engine_core_client::protocol::utility::PauseMode as EnginePauseMode; use super::{ControlServer, pb}; use crate::state::AppState; @@ -15,17 +19,25 @@ pub(crate) type ControlGrpcService = ControlServer; /// gRPC control service backed by the shared application state. pub struct ControlServiceImpl { state: Arc, + rl_lock: Mutex<()>, } impl ControlServiceImpl { pub fn new(state: Arc) -> Self { - Self { state } + Self { + state, + rl_lock: Mutex::new(()), + } } fn ready(&self) -> &EngineCoreReadyResponse { self.state.engine_core_client().ready_response() } + fn client(&self) -> &EngineCoreClient { + self.state.engine_core_client() + } + fn parallelism_info(&self) -> pb::ParallelismInfo { let ready = self.ready(); pb::ParallelismInfo { @@ -36,10 +48,91 @@ impl ControlServiceImpl { decode_context_parallel_size: ready.decode_context_parallel_size, } } + + fn weight_transfer_backend(&self) -> Option<&str> { + let responses = self.client().ready_responses(); + let backend = responses.first()?.weight_transfer_backend.as_deref()?; + responses + .iter() + .all(|ready| ready.weight_transfer_backend.as_deref() == Some(backend)) + .then_some(backend) + } + + fn sleep_mode_enabled(&self) -> bool { + self.client().ready_responses().iter().all(|ready| ready.enable_sleep_mode) + } + + fn draft_weight_updates_enabled(&self) -> bool { + self.client() + .ready_responses() + .iter() + .all(|ready| ready.supports_draft_weight_updates) + } + + fn rl_capabilities(&self) -> pb::RlCapabilities { + let backend = self.weight_transfer_backend(); + pb::RlCapabilities { + weight_transfer_enabled: backend.is_some(), + weight_transfer_backend: backend.unwrap_or_default().to_string(), + sleep_mode_enabled: self.sleep_mode_enabled(), + draft_weight_updates_enabled: self.draft_weight_updates_enabled(), + } + } + + fn require_weight_transfer(&self) -> Result<(), Status> { + self.weight_transfer_backend().map(|_| ()).ok_or_else(|| { + Status::failed_precondition( + "weight transfer is not configured; start vLLM with --weight-transfer-config", + ) + }) + } + + fn require_sleep_mode(&self) -> Result<(), Status> { + self.sleep_mode_enabled().then_some(()).ok_or_else(|| { + Status::failed_precondition( + "sleep mode is not configured; start vLLM with --enable-sleep-mode", + ) + }) + } } const GRPC_API_VERSION: &str = "vllm"; +fn utility_status(method: &'static str, error: vllm_engine_core_client::Error) -> Status { + Status::internal(format!("{method} failed: {}", error.to_report_string())) +} + +fn pause_mode(mode: i32) -> Result { + match pb::PauseMode::try_from(mode) { + Ok(pb::PauseMode::Unspecified | pb::PauseMode::Abort) => Ok(EnginePauseMode::Abort), + Ok(pb::PauseMode::Wait) => Ok(EnginePauseMode::Wait), + Ok(pb::PauseMode::Keep) => Ok(EnginePauseMode::Keep), + Err(_) => Err(Status::invalid_argument("invalid pause mode")), + } +} + +fn json_object(bytes: &[u8], field: &'static str) -> Result { + let value = serde_json::from_slice::(bytes).map_err(|error| { + Status::invalid_argument(format!( + "{field} must contain valid JSON: {}", + error.to_report_string() + )) + })?; + if !value.is_object() { + return Err(Status::invalid_argument(format!( + "{field} must contain a JSON object" + ))); + } + Ok(value) +} + +fn weight_version(value: String) -> Result { + if value.trim().is_empty() { + return Err(Status::invalid_argument("weight_version must not be empty")); + } + Ok(value) +} + #[tonic::async_trait] impl pb::control_server::Control for ControlServiceImpl { async fn get_server_info( @@ -57,6 +150,7 @@ impl pb::control_server::Control for ControlServiceImpl { total_kv_blocks: self.state.engine_core_client().total_num_gpu_blocks(), max_running_requests: ready.max_num_seqs, max_batched_tokens: ready.max_num_batched_tokens, + rl_capabilities: Some(self.rl_capabilities()), })) } @@ -112,6 +206,194 @@ impl pb::control_server::Control for ControlServiceImpl { let sources = client.ready_responses().into_iter().filter_map(kv_event_source).collect(); Ok(Response::new(pb::GetKvEventSourcesResponse { sources })) } + + async fn pause_generation( + &self, + request: Request, + ) -> Result, Status> { + let request = request.into_inner(); + let mode = pause_mode(request.mode)?; + let clear_cache = request.clear_cache.unwrap_or(true); + let _guard = self.rl_lock.lock().await; + self.client() + .pause_scheduler(mode, clear_cache) + .await + .map_err(|error| utility_status("pause_generation", error))?; + Ok(Response::new(pb::PauseGenerationResponse {})) + } + + async fn resume_generation( + &self, + _request: Request, + ) -> Result, Status> { + let _guard = self.rl_lock.lock().await; + self.client() + .resume_scheduler() + .await + .map_err(|error| utility_status("resume_generation", error))?; + Ok(Response::new(pb::ResumeGenerationResponse {})) + } + + async fn is_paused( + &self, + _request: Request, + ) -> Result, Status> { + let paused = self + .client() + .is_scheduler_paused() + .await + .map_err(|error| utility_status("is_paused", error))?; + Ok(Response::new(pb::IsPausedResponse { paused })) + } + + async fn sleep( + &self, + request: Request, + ) -> Result, Status> { + self.require_sleep_mode()?; + let request = request.into_inner(); + let mode = pause_mode(request.mode)?; + let level = request.level.unwrap_or(1); + let _guard = self.rl_lock.lock().await; + self.client() + .sleep(level, mode) + .await + .map_err(|error| utility_status("sleep", error))?; + Ok(Response::new(pb::SleepResponse {})) + } + + async fn wake_up( + &self, + request: Request, + ) -> Result, Status> { + self.require_sleep_mode()?; + let tags = request.into_inner().tags; + let tags = (!tags.is_empty()).then_some(tags); + let _guard = self.rl_lock.lock().await; + self.client() + .wake_up(tags) + .await + .map_err(|error| utility_status("wake_up", error))?; + Ok(Response::new(pb::WakeUpResponse {})) + } + + async fn is_sleeping( + &self, + _request: Request, + ) -> Result, Status> { + let sleeping = self + .client() + .is_sleeping() + .await + .map_err(|error| utility_status("is_sleeping", error))?; + Ok(Response::new(pb::IsSleepingResponse { sleeping })) + } + + async fn init_weight_transfer_engine( + &self, + request: Request, + ) -> Result, Status> { + self.require_weight_transfer()?; + let init_info = json_object(&request.into_inner().init_info_json, "init_info_json")?; + let _guard = self.rl_lock.lock().await; + self.client() + .init_weight_transfer_engine(init_info) + .await + .map_err(|error| utility_status("init_weight_transfer_engine", error))?; + Ok(Response::new(pb::InitWeightTransferEngineResponse {})) + } + + async fn start_weight_update( + &self, + _request: Request, + ) -> Result, Status> { + self.require_weight_transfer()?; + let _guard = self.rl_lock.lock().await; + self.client() + .start_weight_update() + .await + .map_err(|error| utility_status("start_weight_update", error))?; + Ok(Response::new(pb::StartWeightUpdateResponse {})) + } + + async fn start_draft_weight_update( + &self, + _request: Request, + ) -> Result, Status> { + self.require_weight_transfer()?; + if !self.draft_weight_updates_enabled() { + return Err(Status::failed_precondition( + "draft weight updates require a configured speculative draft model", + )); + } + let _guard = self.rl_lock.lock().await; + self.client() + .start_draft_weight_update() + .await + .map_err(|error| utility_status("start_draft_weight_update", error))?; + Ok(Response::new(pb::StartDraftWeightUpdateResponse {})) + } + + async fn update_weights( + &self, + request: Request, + ) -> Result, Status> { + self.require_weight_transfer()?; + let update_info = json_object(&request.into_inner().update_info_json, "update_info_json")?; + let _guard = self.rl_lock.lock().await; + self.client() + .update_weights(update_info) + .await + .map_err(|error| utility_status("update_weights", error))?; + Ok(Response::new(pb::UpdateWeightsResponse {})) + } + + async fn finish_weight_update( + &self, + request: Request, + ) -> Result, Status> { + self.require_weight_transfer()?; + let version = request.into_inner().weight_version.map(weight_version).transpose()?; + let _guard = self.rl_lock.lock().await; + self.client() + .finish_weight_update() + .await + .map_err(|error| utility_status("finish_weight_update", error))?; + if let Some(version) = version { + self.client() + .set_weight_version(&version) + .await + .map_err(|error| utility_status("update_weight_version", error))?; + } + Ok(Response::new(pb::FinishWeightUpdateResponse {})) + } + + async fn update_weight_version( + &self, + request: Request, + ) -> Result, Status> { + let version = weight_version(request.into_inner().weight_version)?; + let _guard = self.rl_lock.lock().await; + self.client() + .set_weight_version(&version) + .await + .map_err(|error| utility_status("update_weight_version", error))?; + Ok(Response::new(pb::UpdateWeightVersionResponse {})) + } + + async fn get_weight_version( + &self, + _request: Request, + ) -> Result, Status> { + let weight_version = self + .client() + .get_weight_version() + .await + .map_err(|error| utility_status("get_weight_version", error))?; + Ok(Response::new(pb::GetWeightVersionResponse { + weight_version, + })) + } } pub(super) fn kv_event_source(response: &EngineCoreReadyResponse) -> Option { diff --git a/rust/src/server/src/grpc/tests.rs b/rust/src/server/src/grpc/tests.rs index 7eabd27014..d5ee87a87f 100644 --- a/rust/src/server/src/grpc/tests.rs +++ b/rust/src/server/src/grpc/tests.rs @@ -12,6 +12,7 @@ use std::time::Duration; use futures::StreamExt as _; use hyper_util::rt::TokioIo; use openssl::ssl::{SslConnector, SslFiletype, SslMethod}; +use rmpv::Value; use serial_test::serial; use tokio::io::{AsyncReadExt as _, AsyncWriteExt as _}; use tokio::net::TcpStream; @@ -30,14 +31,15 @@ use vllm_engine_core_client::mock_engine::{ DEFAULT_MOCK_BLOCK_SIZE, DEFAULT_MOCK_MAX_MODEL_LEN, DEFAULT_MOCK_NUM_GPU_BLOCKS, default_ready_response, }; -use vllm_engine_core_client::protocol::handshake::KvEventsConfig; +use vllm_engine_core_client::protocol::decode_value; +use vllm_engine_core_client::protocol::handshake::{EngineCoreReadyResponse, KvEventsConfig}; use vllm_engine_core_client::protocol::output::{ EngineCoreFinishReason, EngineCoreOutput, EngineCoreOutputs, RequestBatchOutputs, + UtilityCallOutput, }; use vllm_engine_core_client::protocol::request::EngineCoreRequest; -use vllm_engine_core_client::test_utils::{ - IpcNamespace, spawn_mock_engine_task, spawn_mock_engine_task_with_ready, -}; +use vllm_engine_core_client::protocol::utility::{UtilityOutput, UtilityResultEnvelope}; +use vllm_engine_core_client::test_utils::{IpcNamespace, spawn_mock_engine_task_with_ready}; use vllm_engine_core_client::{EngineCoreClient, EngineCoreClientConfig, EngineId, TransportMode}; use vllm_llm::Llm; use vllm_text::tokenizer::DynTokenizer; @@ -291,13 +293,10 @@ async fn setup_grpc_service_with_backend( where F: FnOnce(&EngineCoreRequest) + Send + 'static, { - let ipc = IpcNamespace::new().expect("create ipc namespace"); - let handshake_address = ipc.handshake_endpoint(); - let engine_id = engine_id.into(); - - let engine_task = MockEngineTask::new(spawn_mock_engine_task( - handshake_address.clone(), - engine_id.clone(), + setup_grpc_service_with_engine_script( + engine_id, + default_ready_response(), + backend, move |dealer, push| { boxed_test_future(async move { let add = recv_engine_message(dealer).await; @@ -311,6 +310,33 @@ where .await; }) }, + ) + .await +} + +async fn setup_grpc_service_with_engine_script( + engine_id: impl Into, + ready: EngineCoreReadyResponse, + backend: Arc, + script: F, +) -> ( + InferenceServer, + ControlServer, + tokio::sync::watch::Receiver, + MockEngineTask, +) +where + F: for<'a> FnOnce(&'a mut DealerSocket, &'a mut PushSocket) -> TestFuture<'a> + Send + 'static, +{ + let ipc = IpcNamespace::new().expect("create ipc namespace"); + let handshake_address = ipc.handshake_endpoint(); + let engine_id = engine_id.into(); + + let engine_task = MockEngineTask::new(spawn_mock_engine_task_with_ready( + handshake_address.clone(), + engine_id.clone(), + ready, + script, )); let client = EngineCoreClient::connect( @@ -1509,6 +1535,11 @@ async fn control_reports_server_and_model_info() { assert_eq!(parallelism.data_parallel_size, 1); assert_eq!(parallelism.data_parallel_rank, 0); assert_eq!(parallelism.decode_context_parallel_size, 1); + let rl = server.rl_capabilities.expect("RL capabilities"); + assert!(!rl.weight_transfer_enabled); + assert!(rl.weight_transfer_backend.is_empty()); + assert!(!rl.sleep_mode_enabled); + assert!(!rl.draft_weight_updates_enabled); let model = client .get_model_info(pb::GetModelInfoRequest {}) @@ -1527,6 +1558,65 @@ async fn control_reports_server_and_model_info() { server_task.abort(); } +#[tokio::test(flavor = "multi_thread", worker_threads = 2)] +#[serial] +async fn control_forwards_weight_update_without_pause_guard() { + let mut ready = default_ready_response(); + ready.weight_transfer_backend = Some("nccl".to_string()); + let (inference_service, control_service, engine_health, engine_task) = + setup_grpc_service_with_engine_script( + b"engine-grpc-rl".to_vec(), + ready, + Arc::new(FakeTextBackend), + |dealer, push| { + boxed_test_future(async move { + let frames = recv_engine_message(dealer).await; + assert_eq!(frames[0].as_ref(), &[0x03]); + let payload = decode_value(&frames[1]).expect("decode utility payload"); + let fields = payload.as_array().expect("utility payload array"); + let call_id = fields[1].as_u64().expect("utility call id"); + assert_eq!(fields[2].as_str(), Some("collective_rpc")); + let args = fields[3].as_array().expect("collective_rpc arguments"); + assert_eq!(args[0].as_str(), Some("update_weights")); + send_outputs( + push, + UtilityCallOutput { + output: UtilityOutput { + call_id: call_id.into(), + failure_message: None, + result: Some(UtilityResultEnvelope::without_type_info( + Value::Array(vec![Value::Nil]), + )), + }, + ..Default::default() + } + .into(), + ) + .await; + }) + }, + ) + .await; + let (channel, server_task) = start_grpc_test_server( + inference_service, + control_service, + engine_health, + tokio_util::sync::CancellationToken::new(), + ) + .await; + let mut client = ControlClient::new(channel); + + client + .update_weights(pb::UpdateWeightsRequest { + update_info_json: br#"{"names":["model.weight"]}"#.to_vec(), + }) + .await + .expect("forward weight update without a pause probe"); + + engine_task.await.expect("mock engine task"); + server_task.abort(); +} + #[tokio::test(flavor = "multi_thread", worker_threads = 2)] async fn control_aggregates_multi_engine_capacity() { let ipc = IpcNamespace::new().expect("create ipc namespace"); @@ -1536,6 +1626,9 @@ async fn control_aggregates_multi_engine_capacity() { ready_0.max_model_len = 8_192; ready_0.num_gpu_blocks = 10; ready_0.data_parallel_size = 2; + ready_0.weight_transfer_backend = Some("nccl".to_string()); + ready_0.enable_sleep_mode = true; + ready_0.supports_draft_weight_updates = true; let mut ready_1 = default_ready_response(); ready_1.max_model_len = 4_096; @@ -1586,6 +1679,11 @@ async fn control_aggregates_multi_engine_capacity() { .into_inner(); assert_eq!(server.max_model_len, 4_096); assert_eq!(server.total_kv_blocks, 30); + let rl = server.rl_capabilities.expect("RL capabilities"); + assert!(!rl.weight_transfer_enabled); + assert!(rl.weight_transfer_backend.is_empty()); + assert!(!rl.sleep_mode_enabled); + assert!(!rl.draft_weight_updates_enabled); assert_eq!(server.parallelism.unwrap().data_parallel_size, 4); drop(engine_tasks); diff --git a/rust/src/server/src/lib.rs b/rust/src/server/src/lib.rs index ca19d0ab9e..4423287e51 100644 --- a/rust/src/server/src/lib.rs +++ b/rust/src/server/src/lib.rs @@ -217,7 +217,8 @@ where health_reporter.set_serving::().await; health_reporter.set_serving::().await; let control_service = - grpc::ControlGrpcService::new(grpc::ControlServiceImpl::new(state.clone())); + grpc::ControlGrpcService::new(grpc::ControlServiceImpl::new(state.clone())) + .max_decoding_message_size(DEFAULT_REQUEST_BODY_LIMIT_BYTES); let inference_service = grpc::InferenceGrpcService::new(grpc::InferenceServiceImpl::new(state.clone())) .max_decoding_message_size(DEFAULT_REQUEST_BODY_LIMIT_BYTES); diff --git a/tests/v1/engine/test_engine_core_client.py b/tests/v1/engine/test_engine_core_client.py index 64adf7a8b3..5a943e8d60 100644 --- a/tests/v1/engine/test_engine_core_client.py +++ b/tests/v1/engine/test_engine_core_client.py @@ -1292,6 +1292,7 @@ def test_engine_core_proc_instantiation_cuda_empty(monkeypatch: pytest.MonkeyPat mock_executor.get_kv_cache_specs.return_value = [{"default": mock_spec}] mock_executor.determine_available_memory.return_value = [1024 * 1024 * 1024] mock_executor.initialize_from_config.return_value = None + mock_executor.supports_draft_weight_updates.return_value = False return mock_executor diff --git a/vllm/v1/engine/__init__.py b/vllm/v1/engine/__init__.py index d89c9a3f5d..d70778eb51 100644 --- a/vllm/v1/engine/__init__.py +++ b/vllm/v1/engine/__init__.py @@ -92,6 +92,9 @@ class EngineCoreReadyResponse: kv_cache_size_tokens: int | None = None kv_cache_max_concurrency: float | None = None kv_events_config: KVEventsConfig | None = None + weight_transfer_backend: str | None = None + enable_sleep_mode: bool = False + supports_draft_weight_updates: bool = False class EngineCoreRequest( diff --git a/vllm/v1/engine/core.py b/vllm/v1/engine/core.py index 956c800c7f..55fe5b90c5 100644 --- a/vllm/v1/engine/core.py +++ b/vllm/v1/engine/core.py @@ -1635,6 +1635,15 @@ class EngineCoreProc(EngineCore): max_num_batched_tokens=scheduler_config.max_num_batched_tokens, instance_id=self.vllm_config.instance_id, kv_events_config=self.scheduler.get_kv_event_publisher_config(), + weight_transfer_backend=( + self.vllm_config.weight_transfer_config.backend + if self.vllm_config.weight_transfer_config is not None + else None + ), + enable_sleep_mode=self.vllm_config.model_config.enable_sleep_mode, + supports_draft_weight_updates=( + self.model_executor.supports_draft_weight_updates() + ), ) def process_input_sockets( diff --git a/vllm/v1/executor/abstract.py b/vllm/v1/executor/abstract.py index 404acd50de..864989e6ac 100644 --- a/vllm/v1/executor/abstract.py +++ b/vllm/v1/executor/abstract.py @@ -289,6 +289,12 @@ class Executor(ABC): output = self.collective_rpc("get_supported_tasks") return output[0] + def supports_draft_weight_updates(self) -> bool: + worker_support: list[bool] = self.collective_rpc( + "supports_draft_weight_updates" + ) + return all(worker_support) + def add_lora(self, lora_request: LoRARequest) -> bool: assert lora_request.lora_int_id > 0, "lora_id must be greater than 0." return all(self.collective_rpc("add_lora", args=(lora_request,))) diff --git a/vllm/v1/worker/gpu_worker.py b/vllm/v1/worker/gpu_worker.py index 2340b64aba..c0fb12a948 100644 --- a/vllm/v1/worker/gpu_worker.py +++ b/vllm/v1/worker/gpu_worker.py @@ -869,6 +869,19 @@ class Worker(WorkerBase): def get_draft_model(self) -> nn.Module | None: return self.model_runner.get_draft_model() + def supports_draft_weight_updates(self) -> bool: + engine = self.weight_transfer_engine + speculative_config = self.speculative_config + get_draft_model = getattr(self.model_runner, "get_draft_model", None) + return ( + engine is not None + and engine.supports_draft_weight_update + and callable(get_draft_model) + and get_draft_model() is not None + and speculative_config is not None + and speculative_config.draft_model_config is not None + ) + def _set_draft_weight_update_target(self) -> None: assert self.weight_transfer_engine is not None diff --git a/vllm/v1/worker/worker_base.py b/vllm/v1/worker/worker_base.py index 9381d71913..998380a877 100644 --- a/vllm/v1/worker/worker_base.py +++ b/vllm/v1/worker/worker_base.py @@ -125,6 +125,10 @@ class WorkerBase: def get_model(self) -> nn.Module: raise NotImplementedError + def supports_draft_weight_updates(self) -> bool: + """Whether this worker can update its configured speculative model.""" + return False + def apply_model(self, fn: Callable[[nn.Module], _R]) -> _R: """Apply a function on the model inside this worker.""" return fn(self.get_model())