[Rust Frontend] Simplify data-parallel size ownership (#52575)
Signed-off-by: Bugen Zhao <i@bugenzhao.com> Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
@@ -58,7 +58,6 @@ async fn main() -> Result<()> {
|
||||
local_input_address: None,
|
||||
local_output_address: None,
|
||||
},
|
||||
data_parallel_size: args.engine_count,
|
||||
coordinator_mode: CoordinatorMode::MaybeInProc,
|
||||
model: args.model,
|
||||
served_model_name: vec![],
|
||||
|
||||
@@ -161,8 +161,6 @@ impl TlsConfig {
|
||||
pub struct Config {
|
||||
/// Frontend-to-engine transport setup.
|
||||
pub transport_mode: TransportMode,
|
||||
/// Deployment-wide data-parallel size retained by the frontend.
|
||||
pub data_parallel_size: usize,
|
||||
/// Requested frontend-side coordinator behavior.
|
||||
pub coordinator_mode: CoordinatorMode,
|
||||
/// Backend model identifier used for engine-core loading.
|
||||
@@ -239,46 +237,7 @@ impl Config {
|
||||
max_logprobs
|
||||
);
|
||||
}
|
||||
if self.data_parallel_size == 0 {
|
||||
bail!("data parallel size must be at least 1");
|
||||
}
|
||||
if self.data_parallel_size > usize::from(u16::MAX) + 1 {
|
||||
bail!(
|
||||
"data parallel size ({}) exceeds the two-byte engine identity limit",
|
||||
self.data_parallel_size
|
||||
);
|
||||
}
|
||||
match &self.transport_mode {
|
||||
TransportMode::HandshakeOwner { engine_count, .. } => {
|
||||
if *engine_count != self.data_parallel_size {
|
||||
bail!(
|
||||
"managed frontend engine count ({engine_count}) must equal data parallel size ({})",
|
||||
self.data_parallel_size
|
||||
);
|
||||
}
|
||||
}
|
||||
TransportMode::Bootstrapped {
|
||||
engine_start_index,
|
||||
engine_count,
|
||||
..
|
||||
} => {
|
||||
if *engine_count == 0 {
|
||||
bail!("engine count must be at least 1");
|
||||
}
|
||||
let engine_start_index = usize::try_from(*engine_start_index)
|
||||
.map_err(|_| anyhow::anyhow!("engine start index does not fit usize"))?;
|
||||
let engine_end_index =
|
||||
engine_start_index.checked_add(*engine_count).ok_or_else(|| {
|
||||
anyhow::anyhow!("engine start index + engine count overflows")
|
||||
})?;
|
||||
if engine_end_index > self.data_parallel_size {
|
||||
bail!(
|
||||
"connected engine range [{engine_start_index}, {engine_end_index}) exceeds data parallel size {}",
|
||||
self.data_parallel_size
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
self.transport_mode.validate()?;
|
||||
|
||||
Ok(())
|
||||
}
|
||||
|
||||
@@ -43,7 +43,7 @@ impl ControlServiceImpl {
|
||||
pb::ParallelismInfo {
|
||||
tensor_parallel_size: ready.tensor_parallel_size,
|
||||
pipeline_parallel_size: ready.pipeline_parallel_size,
|
||||
data_parallel_size: self.state.data_parallel_size().min(u32::MAX as usize) as u32,
|
||||
data_parallel_size: self.client().data_parallel_size().min(u32::MAX as usize) as u32,
|
||||
data_parallel_rank: ready.data_parallel_rank,
|
||||
decode_context_parallel_size: ready.decode_context_parallel_size,
|
||||
}
|
||||
|
||||
@@ -1625,7 +1625,7 @@ async fn control_aggregates_multi_engine_capacity() {
|
||||
let mut ready_0 = default_ready_response();
|
||||
ready_0.max_model_len = 8_192;
|
||||
ready_0.num_gpu_blocks = 10;
|
||||
ready_0.data_parallel_size = 2;
|
||||
ready_0.effective_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;
|
||||
@@ -1633,7 +1633,7 @@ async fn control_aggregates_multi_engine_capacity() {
|
||||
let mut ready_1 = default_ready_response();
|
||||
ready_1.max_model_len = 4_096;
|
||||
ready_1.num_gpu_blocks = 20;
|
||||
ready_1.data_parallel_size = 2;
|
||||
ready_1.effective_data_parallel_size = 2;
|
||||
ready_1.data_parallel_rank = 1;
|
||||
|
||||
let engine_tasks = [ready_0, ready_1].map(|ready| {
|
||||
@@ -1667,7 +1667,7 @@ async fn control_aggregates_multi_engine_capacity() {
|
||||
Llm::new(client),
|
||||
Arc::new(FakeTextBackend) as Arc<dyn ChatTextBackend>,
|
||||
);
|
||||
let state = AppState::new(vec!["test-model".to_string()], chat).with_data_parallel_size(4);
|
||||
let state = AppState::new(vec!["test-model".to_string()], chat);
|
||||
let service = ControlServiceImpl::new(Arc::new(state));
|
||||
|
||||
let server = pb::control_server::Control::get_server_info(
|
||||
@@ -1684,7 +1684,7 @@ async fn control_aggregates_multi_engine_capacity() {
|
||||
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);
|
||||
assert_eq!(server.parallelism.unwrap().data_parallel_size, 2);
|
||||
|
||||
drop(engine_tasks);
|
||||
}
|
||||
|
||||
@@ -141,7 +141,6 @@ async fn build_state(config: &Config) -> Result<Arc<AppState>> {
|
||||
.with_model_path(config.model.clone())
|
||||
.with_api_server_options(config.api_server_options)
|
||||
.with_server_info(ServerInfoSnapshot::from_config(config))
|
||||
.with_data_parallel_size(config.data_parallel_size)
|
||||
.with_api_keys(config.api_keys.clone())
|
||||
.with_cors(config.cors.clone())
|
||||
.with_profiler(config.profiler.clone()),
|
||||
|
||||
@@ -704,14 +704,13 @@ async fn test_app_with_dev_mode(dev_mode_enabled: bool) -> axum::Router {
|
||||
/// stays alive for the duration of the test.
|
||||
async fn test_dev_mode_app_with_ready(
|
||||
ready_response: vllm_engine_core_client::protocol::handshake::EngineCoreReadyResponse,
|
||||
data_parallel_size: usize,
|
||||
) -> (axum::Router, MockEngineTask) {
|
||||
let ipc = IpcNamespace::new().expect("create ipc namespace");
|
||||
let handshake_address = ipc.handshake_endpoint();
|
||||
let engine_id = b"engine-world-size".to_vec();
|
||||
let engine_task = MockEngineTask::new(spawn_mock_engine_task_with_ready(
|
||||
handshake_address.clone(),
|
||||
engine_id.clone(),
|
||||
engine_id,
|
||||
ready_response,
|
||||
|_dealer, _push| boxed_test_future(async {}),
|
||||
));
|
||||
@@ -729,10 +728,10 @@ async fn test_dev_mode_app_with_ready(
|
||||
|
||||
let chat = ChatLlm::from_shared_backend(test_llm(client), Arc::new(FakeChatBackend::new()));
|
||||
let app = build_router_with_dev_mode(
|
||||
Arc::new(
|
||||
AppState::new(vec!["Qwen/Qwen1.5-0.5B-Chat".to_string()], chat)
|
||||
.with_data_parallel_size(data_parallel_size),
|
||||
),
|
||||
Arc::new(AppState::new(
|
||||
vec!["Qwen/Qwen1.5-0.5B-Chat".to_string()],
|
||||
chat,
|
||||
)),
|
||||
true,
|
||||
);
|
||||
(app, engine_task)
|
||||
@@ -6292,7 +6291,7 @@ async fn tokenize_allows_prompts_at_or_above_max_model_len() {
|
||||
max_model_len: 4,
|
||||
..default_ready_response()
|
||||
};
|
||||
let (mut app, _engine_task) = test_dev_mode_app_with_ready(ready, 1).await;
|
||||
let (mut app, _engine_task) = test_dev_mode_app_with_ready(ready).await;
|
||||
|
||||
let (completion_status, completion_json) = post_json(
|
||||
&mut app,
|
||||
@@ -6543,58 +6542,6 @@ async fn world_size_endpoint_is_dev_mode_only() {
|
||||
assert_eq!(response.status(), StatusCode::NOT_FOUND);
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
#[serial]
|
||||
async fn world_size_includes_data_parallelism_by_default() {
|
||||
let ready = vllm_engine_core_client::protocol::handshake::EngineCoreReadyResponse {
|
||||
world_size: 2,
|
||||
data_parallel_size: 1,
|
||||
..default_ready_response()
|
||||
};
|
||||
let (mut app, _engine_task) = test_dev_mode_app_with_ready(ready, 4).await;
|
||||
|
||||
let response = app
|
||||
.call(
|
||||
Request::builder()
|
||||
.uri("/get_world_size")
|
||||
.body(Body::empty())
|
||||
.expect("build request"),
|
||||
)
|
||||
.await
|
||||
.expect("call app");
|
||||
|
||||
assert_eq!(response.status(), StatusCode::OK);
|
||||
let body = to_bytes(response.into_body(), usize::MAX).await.expect("read body");
|
||||
let json: serde_json::Value = serde_json::from_slice(&body).expect("decode json");
|
||||
assert_eq!(json, json!({"world_size": 8}));
|
||||
}
|
||||
|
||||
#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
|
||||
#[serial]
|
||||
async fn world_size_excludes_data_parallelism_when_include_dp_false() {
|
||||
let ready = vllm_engine_core_client::protocol::handshake::EngineCoreReadyResponse {
|
||||
world_size: 2,
|
||||
data_parallel_size: 4,
|
||||
..default_ready_response()
|
||||
};
|
||||
let (mut app, _engine_task) = test_dev_mode_app_with_ready(ready, 4).await;
|
||||
|
||||
let response = app
|
||||
.call(
|
||||
Request::builder()
|
||||
.uri("/get_world_size?include_dp=false")
|
||||
.body(Body::empty())
|
||||
.expect("build request"),
|
||||
)
|
||||
.await
|
||||
.expect("call app");
|
||||
|
||||
assert_eq!(response.status(), StatusCode::OK);
|
||||
let body = to_bytes(response.into_body(), usize::MAX).await.expect("read body");
|
||||
let json: serde_json::Value = serde_json::from_slice(&body).expect("decode json");
|
||||
assert_eq!(json, json!({"world_size": 2}));
|
||||
}
|
||||
|
||||
// ========================= Profiler route tests =========================
|
||||
|
||||
async fn test_profiling_app_with_engine_script<F>(script: F) -> (axum::Router, MockEngineTask)
|
||||
|
||||
@@ -47,7 +47,7 @@ pub async fn get_world_size(
|
||||
let ws = client.world_size();
|
||||
|
||||
let world_size = if params.include_dp {
|
||||
let dp = state.data_parallel_size() as u64;
|
||||
let dp = client.data_parallel_size() as u64;
|
||||
ws * dp
|
||||
} else {
|
||||
ws
|
||||
|
||||
@@ -40,8 +40,6 @@ pub struct AppState {
|
||||
pub cors: CorsConfig,
|
||||
/// Runtime server information returned by `/server_info`, when available.
|
||||
server_info: Option<ServerInfoSnapshot>,
|
||||
/// Deployment-wide data-parallel size retained by the frontend.
|
||||
data_parallel_size: usize,
|
||||
/// SHA-256 hashes of API keys accepted as bearer tokens for guarded routes.
|
||||
api_key_hashes: Vec<ApiKeyHash>,
|
||||
/// Number of in-flight inference requests currently owned by this frontend.
|
||||
@@ -71,14 +69,12 @@ impl AppState {
|
||||
!served_model_names.is_empty(),
|
||||
"served_model_names must not be empty"
|
||||
);
|
||||
let data_parallel_size = chat.engine_core_client().engine_count();
|
||||
Self {
|
||||
served_model_names,
|
||||
chat,
|
||||
api_server_options: ApiServerOptions::default(),
|
||||
cors: CorsConfig::default(),
|
||||
server_info: None,
|
||||
data_parallel_size,
|
||||
api_key_hashes: Vec::new(),
|
||||
server_load: AtomicU64::new(0),
|
||||
lora_manager: LoraManager::new(),
|
||||
@@ -118,17 +114,6 @@ impl AppState {
|
||||
self
|
||||
}
|
||||
|
||||
/// Set the deployment-wide data-parallel size reported by frontend APIs.
|
||||
pub(crate) fn with_data_parallel_size(mut self, size: usize) -> Self {
|
||||
self.data_parallel_size = size;
|
||||
self
|
||||
}
|
||||
|
||||
/// Return the deployment-wide data-parallel size.
|
||||
pub(crate) fn data_parallel_size(&self) -> usize {
|
||||
self.data_parallel_size
|
||||
}
|
||||
|
||||
/// Build a `/server_info` response payload.
|
||||
pub(crate) fn server_info_response(
|
||||
&self,
|
||||
|
||||
Reference in New Issue
Block a user