chore(kv-router): drop easy llm facade reexports (#7474)

Signed-off-by: PeaBrane <yanrpei@gmail.com>

lib/llm/src/kv_router/sequence.rs

@@ -11,6 +11,7 @@ pub use dynamo_kv_router::multi_worker_sequence::{
     ActiveSequencesMultiWorker, SequenceError, SequencePublisher, SequenceRequest,
     SequenceSubscriber,
 };
+use dynamo_kv_router::protocols::{ActiveLoad, ActiveSequenceEvent, WorkerWithDpRank};
 pub use dynamo_kv_router::sequence::{ActiveSequences, RequestId};
 
 use anyhow::Result;
@@ -21,7 +22,6 @@ use std::collections::HashMap;
 use std::sync::Arc;
 
 use super::metrics::WORKER_LOAD_METRICS;
-use super::protocols::{ActiveLoad, ActiveSequenceEvent, WorkerWithDpRank};
 use crate::kv_router::{ACTIVE_SEQUENCES_SUBJECT, KV_METRICS_SUBJECT};
 use crate::local_model::runtime_config::ModelRuntimeConfig;
 

lib/llm/src/kv_router/subscriber.rs

 // SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 // SPDX-License-Identifier: Apache-2.0
 
-use crate::kv_router::{
-    Indexer, KV_EVENT_SUBJECT, KvRouterConfig, protocols::RouterEvent,
-    worker_query::WorkerQueryClient,
-};
+use crate::kv_router::{Indexer, worker_query::WorkerQueryClient};
 use anyhow::Result;
+use dynamo_kv_router::{
+    config::KvRouterConfig,
+    protocols::{KV_EVENT_SUBJECT, RouterEvent},
+};
 use dynamo_runtime::{
     component::Component, discovery::EventTransportKind, prelude::*,
     transports::event_plane::EventSubscriber,

lib/llm/src/kv_router/worker_query.rs

@@ -21,9 +21,11 @@ use futures::StreamExt;
 use tokio::sync::{Mutex, Semaphore};
 
 use crate::kv_router::Indexer;
-use crate::kv_router::indexer::{LocalKvIndexer, WorkerKvQueryRequest, WorkerKvQueryResponse};
-use crate::kv_router::protocols::{DpRank, KvCacheEventData, RouterEvent, WorkerId};
 use crate::kv_router::worker_kv_indexer_query_endpoint;
+use dynamo_kv_router::{
+    indexer::{LocalKvIndexer, WorkerKvQueryRequest, WorkerKvQueryResponse},
+    protocols::{DpRank, KvCacheEventData, RouterEvent, WorkerId},
+};
 
 // Recovery retry configuration
 const RECOVERY_MAX_RETRIES: u32 = 8;
@@ -763,11 +765,10 @@ impl AsyncEngine<SingleIn<WorkerKvQueryRequest>, ManyOut<WorkerKvQueryResponse>,
 mod tests {
     use super::*;
     use crate::kv_router::Indexer;
-    use crate::kv_router::RouterEvent;
-    use crate::kv_router::indexer::{KvIndexer, KvIndexerInterface, KvIndexerMetrics};
-    use crate::kv_router::protocols::{
+    use dynamo_kv_router::indexer::{KvIndexer, KvIndexerInterface, KvIndexerMetrics};
+    use dynamo_kv_router::protocols::{
         ExternalSequenceBlockHash, KvCacheEvent, KvCacheEventData, KvCacheStoreData,
-        KvCacheStoredBlockData, LocalBlockHash,
+        KvCacheStoredBlockData, LocalBlockHash, RouterEvent,
     };
     use dynamo_runtime::{DistributedRuntime, Runtime, distributed::DistributedConfig};
     use std::collections::VecDeque;

lib/llm/src/lora/load_estimator.rs

@@ -11,12 +11,12 @@ use std::sync::Arc;
 use std::time::{Duration, Instant};
 
 use dashmap::DashMap;
+use dynamo_kv_router::protocols::{ActiveSequenceEvent, ActiveSequenceEventData};
 use dynamo_runtime::component::Component;
 use dynamo_runtime::traits::DistributedRuntimeProvider;
 use dynamo_runtime::transports::event_plane::EventSubscriber;
 
 use crate::kv_router::ACTIVE_SEQUENCES_SUBJECT;
-use crate::kv_router::protocols::{ActiveSequenceEvent, ActiveSequenceEventData};
 use crate::kv_router::scheduler::KvScheduler;
 
 /// Time-series sample of LORA load

lib/llm/src/lora/routing/hrw.rs

@@ -2,7 +2,7 @@
 // SPDX-License-Identifier: Apache-2.0
 
 use super::LoraAllocator;
-use crate::kv_router::protocols::WorkerWithDpRank;
+use dynamo_kv_router::protocols::WorkerWithDpRank;
 
 /// Rendezvous (HRW) hashing implementation for LoRA allocation
 pub struct RendezvousHasher;

lib/llm/src/lora/routing/mod.rs

@@ -3,7 +3,7 @@
 
 //! LoRA Allocation Algorithms - HRW and Random
 
-use crate::kv_router::protocols::WorkerWithDpRank;
+use dynamo_kv_router::protocols::WorkerWithDpRank;
 use std::str::FromStr;
 
 pub mod hrw;

lib/llm/src/lora/routing/table.rs

@@ -4,11 +4,10 @@
 //! LoRA Routing Table - Thread-safe data structure for storing LoRA allocation decisions.
 
 use dashmap::DashMap;
+use dynamo_kv_router::protocols::WorkerWithDpRank;
 use std::sync::Arc;
 use std::time::Instant;
 
-use crate::kv_router::protocols::WorkerWithDpRank;
-
 /// Configuration for a single LoRA's allocation
 #[derive(Debug, Clone)]
 pub struct LoraReplicaConfig {

lib/llm/src/protocols/common/preprocessor.rs

@@ -5,12 +5,14 @@ use std::collections::HashSet;
 use std::sync::Arc;
 
 use derive_builder::Builder;
+use dynamo_kv_router::{
+    config::RouterConfigOverride,
+    protocols::{BlockExtraInfo, WorkerId},
+};
 use serde::{Deserialize, Serialize};
 
 use super::timing::RequestTracker;
 use super::{OutputOptions, SamplingOptions, StopConditions};
-use crate::kv_router::RouterConfigOverride;
-use crate::kv_router::protocols::{BlockExtraInfo, WorkerId};
 use crate::preprocessor::media::RdmaMediaDataDescriptor;
 use crate::protocols::TokenIdType;