refactor: move libs to lib dir

Signed-off-by: Neelay Shah <neelays@nvidia.com>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>

llm/rust/triton-llm/src/engines/mistralrs.rs → lib/llm/src/engines/mistralrs.rs

@@ -27,10 +27,10 @@ use mistralrs::{
 };
 use tokio::sync::mpsc::channel;
 
-use triton_distributed::engine::{AsyncEngine, AsyncEngineContextProvider, ResponseStream};
-use triton_distributed::pipeline::error as pipeline_error;
-use triton_distributed::pipeline::{Error, ManyOut, SingleIn};
-use triton_distributed::protocols::annotated::Annotated;
+use triton_distributed_runtime::engine::{AsyncEngine, AsyncEngineContextProvider, ResponseStream};
+use triton_distributed_runtime::pipeline::error as pipeline_error;
+use triton_distributed_runtime::pipeline::{Error, ManyOut, SingleIn};
+use triton_distributed_runtime::protocols::annotated::Annotated;
 
 use crate::protocols::openai::chat_completions::{
     ChatCompletionChoiceDelta, ChatCompletionContent, ChatCompletionRequest,

llm/rust/triton-llm/src/http/service/discovery.rs → lib/llm/src/http/service/discovery.rs

@@ -18,7 +18,7 @@ use std::sync::Arc;
 use serde::{Deserialize, Serialize};
 use tokio::sync::mpsc::Receiver;
 
-use triton_distributed::{
+use triton_distributed_runtime::{
     protocols::{self, annotated::Annotated},
     transports::etcd::{KeyValue, WatchEvent},
     DistributedRuntime, Result,

llm/rust/triton-llm/src/http/service/openai.rs → lib/llm/src/http/service/openai.rs

@@ -48,7 +48,7 @@ use crate::types::{
     Annotated,
 };
 
-use triton_distributed::pipeline::{AsyncEngineContext, Context};
+use triton_distributed_runtime::pipeline::{AsyncEngineContext, Context};
 
 #[derive(Serialize, Deserialize)]
 pub(crate) struct ErrorResponse {
@@ -91,7 +91,7 @@ impl ErrorResponse {
         )
     }
 
-    /// The OAI endpoints call an [`triton_distributed::engine::AsyncEngine`] which are specialized to return
+    /// The OAI endpoints call an [`triton_distributed_runtime::engine::AsyncEngine`] which are specialized to return
     /// an [`anyhow::Error`]. This method will convert the [`anyhow::Error`] into an [`HttpError`].
     /// If successful, it will return the [`HttpError`] as an [`ErrorResponse::internal_server_error`]
     /// with the details of the error.

llm/rust/triton-llm/src/kv_router.rs → lib/llm/src/kv_router.rs

@@ -18,7 +18,7 @@ use futures::stream::StreamExt;
 use std::{sync::Arc, time::Duration};
 use tokio_util::sync::CancellationToken;
 use tracing as log;
-use triton_distributed::{component::Component, DistributedRuntime};
+use triton_distributed_runtime::{component::Component, DistributedRuntime};
 
 pub mod indexer;
 pub mod protocols;
@@ -65,7 +65,7 @@ impl KvRouter {
     }
 
     pub async fn new(
-        nats_client: triton_distributed::transports::nats::Client,
+        nats_client: triton_distributed_runtime::transports::nats::Client,
         service_name: String,
         kv_subject: String,
     ) -> Result<Arc<Self>> {
@@ -145,7 +145,7 @@ impl KvRouter {
 }
 
 async fn collect_endpoints(
-    nats_client: triton_distributed::transports::nats::Client,
+    nats_client: triton_distributed_runtime::transports::nats::Client,
     service_name: String,
     ep_tx: tokio::sync::mpsc::Sender<ProcessedEndpoints>,
     cancel: CancellationToken,

llm/rust/triton-llm/src/kv_router/publisher.rs → lib/llm/src/kv_router/publisher.rs

@@ -16,7 +16,7 @@
 use crate::kv_router::{indexer::RouterEvent, protocols::KvCacheEvent, KV_EVENT_SUBJECT};
 use tokio::sync::mpsc;
 use tracing as log;
-use triton_distributed::{component::Component, DistributedRuntime, Result};
+use triton_distributed_runtime::{component::Component, DistributedRuntime, Result};
 use uuid::Uuid;
 
 pub struct KvPublisher {

llm/rust/triton-llm/src/kv_router/worker.rs → lib/llm/src/kv_router/worker.rs

@@ -19,12 +19,12 @@ pub use crate::kv_router::protocols::ForwardPassMetrics;
 
 use anyhow::Result;
 use derive_builder::Builder;
-use triton_distributed::pipeline::network::{
+use triton_distributed_runtime::pipeline::network::{
     ingress::push_endpoint::PushEndpoint,
     PushWorkHandler,
 };
 
-use triton_distributed::transports::nats::{self, ServiceExt};
+use triton_distributed_runtime::transports::nats::{self, ServiceExt};
 
 use tokio::sync::watch;
 use tokio_util::sync::CancellationToken;