lib.rs

// SPDX-FileCopyrightText: Copyright (c) 2024-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

use async_once_cell::OnceCell as AsyncOnceCell;
use libc::c_char;
use once_cell::sync::OnceCell;
use std::borrow::Cow;
use std::ffi::CStr;
use std::sync::Arc;
use std::sync::atomic::{AtomicU32, Ordering};

use dynamo_llm::{
    discovery::{KvWorkerMonitor, ModelWatcher},
    kv_router::{protocols::*, publisher::KvEventPublisher},
};
use dynamo_runtime::discovery::DiscoveryQuery;
use dynamo_runtime::{DistributedRuntime, Worker};
static WK: OnceCell<Worker> = OnceCell::new();
static DRT: AsyncOnceCell<DistributedRuntime> = AsyncOnceCell::new();
// [FIXME] shouldn't the publisher be instance passing between API calls?
static KV_PUB: OnceCell<KvEventPublisher> = OnceCell::new();

/// Convert a C string pointer to a Rust string, falling back to a default when:
/// - the pointer is NULL,
/// - the bytes are not valid UTF-8,
/// - or the resulting string is empty/whitespace.
#[inline]
unsafe fn cstr_or_default<'a>(ptr: *const c_char, default_val: &'a str) -> Cow<'a, str> {
    if ptr.is_null() {
        return Cow::from(default_val);
    }
    match unsafe { CStr::from_ptr(ptr) }
        .to_str()
        .ok()
        .map(|s| s.trim())
    {
        Some(s) if !s.is_empty() => Cow::from(s.to_owned()),
        _ => Cow::from(default_val),
    }
}

fn initialize_tracing() {
    // Sets up RUST_LOG environment variable for logging while KV Publishing
    // Example: os.environ["RUST_LOG"] = "debug"
    let subscriber = tracing_subscriber::fmt()
        .with_env_filter(tracing_subscriber::EnvFilter::from_default_env())
        .finish();

    tracing::subscriber::set_global_default(subscriber).expect("setting default subscriber failed");

    tracing::debug!("Tracing initialized");
}

#[repr(u32)]
pub enum DynamoLlmResult {
    OK = 0,
    ERR = 1,
}

/// Wait for the discovery daemon to sync and return at least one instance.
/// This ensures list() calls will have data available.
/// Returns the number of instances found, or 0 if timed out.
async fn wait_for_discovery_sync(drt: &DistributedRuntime, timeout_secs: u64) -> usize {
    tracing::info!("Waiting for discovery to sync...");
    let discovery = drt.discovery();
    let timeout = std::time::Duration::from_secs(timeout_secs);
    let start = std::time::Instant::now();

    loop {
        match discovery.list(DiscoveryQuery::AllModels).await {
            Ok(instances) if !instances.is_empty() => {
                tracing::info!(
                    "Discovery sync complete: found {} instances",
                    instances.len()
                );
                return instances.len();
            }
            Ok(_) => {
                if start.elapsed() > timeout {
                    tracing::warn!("Discovery sync timed out waiting for instances");
                    return 0;
                }
                tracing::debug!("No instances yet, waiting...");
                tokio::time::sleep(std::time::Duration::from_millis(500)).await;
            }
            Err(e) => {
                tracing::warn!("Discovery list error: {}, continuing...", e);
                return 0;
            }
        }
    }
}

/// # Safety
/// the namespace_c_str and component_c_str are passed as pointers to C strings
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_llm_init(
    namespace_c_str: *const c_char,
    component_c_str: *const c_char,
    kv_block_size: u32,
) -> DynamoLlmResult {
    initialize_tracing();
    let wk = match WK.get_or_try_init(Worker::from_settings) {
        Ok(wk) => wk.clone(),
        Err(e) => {
            tracing::error!(error = ?e, "Failed to initialize runtime (Worker::from_settings)");
            return DynamoLlmResult::ERR;
        }
    };
    let rt = wk.runtime();
    let secondary = rt.secondary().clone();
    let result = secondary.block_on(async {
        // Initialize the distributed runtime
        match DRT
            .get_or_try_init(async { DistributedRuntime::from_settings(rt.clone()).await })
            .await
        {
            Ok(drt) => {
                // Wait for discovery to sync before returning
                // This is needed because dynamo_create_worker_selection_pipeline() is called
                // immediately after, and it needs discovery.list() to return data
                // the discovery daemon takes time to query K8s and returns async, so we need to wait.
                let instance_count = wait_for_discovery_sync(drt, 10).await;
                if instance_count == 0 {
                    tracing::error!(
                        "Discovery sync failed: no worker instances found. Is the backend running?"
                    );
                    return Err(DynamoLlmResult::ERR);
                }
                Ok(())
            }
            Err(e) => {
                tracing::error!(error = ?e, "Failed to initialize distributed runtime");
                Err(DynamoLlmResult::ERR)
            }
        }
    });
    let namespace = match unsafe { CStr::from_ptr(namespace_c_str) }.to_str() {
        Ok(s) => s.to_string(),
        Err(e) => {
            tracing::error!(error = ?e, "Failed to convert C string to Rust string (namespace)");
            return DynamoLlmResult::ERR;
        }
    };

    let component_cow = unsafe { cstr_or_default(component_c_str, "backend") };
    if let Cow::Borrowed("backend") = &component_cow {
        tracing::info!("defaulting to \"backend\" for component");
    }
    let component: String = component_cow.into_owned();

    match result {
        Ok(_) => match KV_PUB.get_or_try_init(move || {
            dynamo_create_kv_publisher(namespace, component, kv_block_size)
        }) {
            Ok(_) => DynamoLlmResult::OK,
            Err(e) => {
                tracing::error!(error = ?e, "Failed to initialize distributed runtime");
                DynamoLlmResult::ERR
            }
        },
        Err(e) => e,
    }
}

#[unsafe(no_mangle)]
pub extern "C" fn dynamo_llm_shutdown() -> DynamoLlmResult {
    let wk = match WK.get() {
        Some(wk) => wk,
        None => {
            tracing::error!("Runtime not initialized");
            return DynamoLlmResult::ERR;
        }
    };

    wk.runtime().shutdown();

    DynamoLlmResult::OK
}

#[unsafe(no_mangle)]
pub extern "C" fn dynamo_llm_load_publisher_create() -> DynamoLlmResult {
    DynamoLlmResult::OK
}

// instantiate a kv publisher
// this will bring up the task to publish and the channels to await publishing events
// the [`dynamo_kv_publish_store_event`] call will use a handle to the publisher to send events
// store and the [`dynamo_kv_event_create_removed`] will create remove events
// these call mus be driving by external c++ threads that are consuming the kv events from the
// c++ executor api

fn dynamo_create_kv_publisher(
    namespace: String,
    component: String,
    kv_block_size: u32,
) -> Result<KvEventPublisher, anyhow::Error> {
    tracing::info!("Creating KV Publisher for model: {}", component);
    match DRT
        .get()
        .ok_or(anyhow::Error::msg("Could not get Distributed Runtime"))
    {
        Ok(drt) => {
            let backend = drt.namespace(namespace)?.component(component)?;
            KvEventPublisher::new(backend, kv_block_size, None)
        }
        Err(e) => Err(e),
    }
}

fn kv_event_create_stored_block_from_parts(
    block_hash: u64,
    token_ids: *const u32,
    num_tokens: usize,
    kv_block_size: u32,
    _lora_id: u64,
) -> KvCacheStoredBlockData {
    let tokens_hash = compute_block_hash_for_seq(
        unsafe { std::slice::from_raw_parts(token_ids, num_tokens) },
        kv_block_size,
        None,
    )[0];
    KvCacheStoredBlockData {
        block_hash: ExternalSequenceBlockHash(block_hash),
        tokens_hash,
        mm_extra_info: None,
    }
}
static WARN_COUNT: AtomicU32 = AtomicU32::new(0);

fn kv_event_create_stored_from_parts(
    kv_params: DynamoKvStoredEventParams,
    kv_block_size: u32,
) -> KvCacheEvent {
    let mut blocks: Vec<KvCacheStoredBlockData> = Vec::new();

    let mut token_offset: usize = 0;
    for block_idx in 0..kv_params.num_blocks {
        let block_hash = unsafe { *kv_params.block_ids.offset(block_idx.try_into().unwrap()) };
        let tokens = unsafe { kv_params.token_ids.offset(token_offset.try_into().unwrap()) };
        let num_toks = unsafe {
            *kv_params
                .num_block_tokens
                .offset(block_idx.try_into().unwrap())
        };

        if num_toks != (kv_block_size as usize) {
            if WARN_COUNT
                .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |c| {
                    if c < 3 { Some(c + 1) } else { None }
                })
                .is_ok()
            {
                tracing::warn!(
                    "Block not published. Block size must be {} tokens to be published. Block size is: {}",
                    kv_block_size,
                    num_toks
                );
            }
            break;
        }
        token_offset += num_toks;
        blocks.push(kv_event_create_stored_block_from_parts(
            block_hash,
            tokens,
            num_toks,
            kv_block_size,
            kv_params.lora_id,
        ));
    }

    KvCacheEvent {
        data: KvCacheEventData::Stored(KvCacheStoreData {
            blocks,
            parent_hash: kv_params.parent_hash.map(ExternalSequenceBlockHash),
        }),
        event_id: kv_params.event_id,
        dp_rank: 0,
    }
}

fn kv_event_create_removed_from_parts(
    event_id: u64,
    block_ids: *const u64,
    num_blocks: usize,
) -> KvCacheEvent {
    let block_hashes: Vec<ExternalSequenceBlockHash> =
        unsafe { std::slice::from_raw_parts(block_ids, num_blocks) }
            .to_vec()
            .iter()
            .map(|&v| ExternalSequenceBlockHash(v))
            .collect();
    KvCacheEvent {
        event_id,
        data: KvCacheEventData::Removed(KvCacheRemoveData { block_hashes }),
        dp_rank: 0,
    }
}

pub struct DynamoKvStoredEventParams {
    pub event_id: u64,
    pub token_ids: *const u32,
    pub num_block_tokens: *const usize,
    pub block_ids: *const u64,
    pub num_blocks: usize,
    pub parent_hash: Option<u64>,
    pub lora_id: u64,
}

/// # Safety
/// parent_hash is passed as pointer to indicate whether the blocks
/// has a parent hash or not. nullptr is used to represent no parent hash
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_kv_event_publish_stored(
    event_id: u64,
    token_ids: *const u32,
    num_block_tokens: *const usize,
    block_ids: *const u64,
    num_blocks: usize,
    parent_hash: *const u64,
    lora_id: u64,
) -> DynamoLlmResult {
    let parent_hash = {
        if parent_hash.is_null() {
            None
        } else {
            Some(unsafe { *parent_hash })
        }
    };
    let kv_params = DynamoKvStoredEventParams {
        event_id,
        token_ids,
        num_block_tokens,
        block_ids,
        num_blocks,
        parent_hash,
        lora_id,
    };
    let publisher = KV_PUB.get().unwrap();
    let event = kv_event_create_stored_from_parts(kv_params, publisher.kv_block_size());
    match publisher.publish(event) {
        Ok(_) => DynamoLlmResult::OK,
        Err(e) => {
            eprintln!("Error publishing stored kv event {:?}", e);
            DynamoLlmResult::ERR
        }
    }
}

#[unsafe(no_mangle)]
pub extern "C" fn dynamo_kv_event_publish_removed(
    event_id: u64,
    block_ids: *const u64,
    num_blocks: usize,
) -> DynamoLlmResult {
    let publisher = KV_PUB.get().unwrap();
    let event = kv_event_create_removed_from_parts(event_id, block_ids, num_blocks);
    match publisher.publish(event) {
        Ok(_) => DynamoLlmResult::OK,
        Err(e) => {
            eprintln!("Error publishing removed kv event {:?}", e);
            DynamoLlmResult::ERR
        }
    }
}

// Need to setup etcd and nats to run these tests
// #[cfg(test)]
// mod tests {
//     use super::*;
//     use std::ffi::CString;

//     #[test]
//     fn test_dynamo_llm_init() {
//         // Create C-compatible strings
//         let namespace = CString::new("test_namespace").unwrap();
//         let component = CString::new("test_component").unwrap();

//         // Call the init function
//         let result = unsafe {
//             dynamo_llm_init(
//                 namespace.as_ptr(),
//                 component.as_ptr(),
//                 1,  // worker_id
//                 32, // kv_block_size
//             )
//         };

//         assert_eq!(result as u32, DynamoLlmResult::OK as u32);

//         assert!(WK.get().is_some());

//         let shutdown_result = dynamo_llm_shutdown();
//         assert_eq!(shutdown_result as u32, DynamoLlmResult::OK as u32);
//     }
// }
/* ------------------------------------------------------------------------
 * Worker selection pipeline
 * ------------------------------------------------------------------------ */
use std::pin::Pin;

const GENERATE_ENDPOINT: &str = "generate";

use anyhow::Context;
use dynamo_runtime::{Runtime, traits::DistributedRuntimeProvider};

use dynamo_llm::discovery::ModelManager;
use dynamo_llm::entrypoint::build_routed_pipeline;
use dynamo_llm::http::service::metrics::Metrics;
use dynamo_llm::kv_router::KvRouterConfig;
use dynamo_llm::model_card::ModelDeploymentCard;
use dynamo_llm::protocols::openai::nvext::NvExt;
use dynamo_llm::types::{
    Annotated,
    openai::chat_completions::{
        NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse,
    },
};
use dynamo_runtime::{
    engine::AsyncEngineStream,
    pipeline::{ManyOut, RouterMode, ServiceEngine, SingleIn},
};
/// Opaque handle exposed to C — it owns its own Worker/runtime and engine.
pub struct WorkerSelectionPipeline {
    wk: Worker,
    engine: ServiceEngine<
        SingleIn<NvCreateChatCompletionRequest>,
        ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>,
    >,
    /// KV router for bookkeeping operations (only present when router_mode is KV)
    kv_router: Option<Arc<dynamo_llm::kv_router::KvRouter>>,
}

/// Create a worker-selection pipeline ("generate" endpoint).
///
/// # Safety
/// - `namespace_c_str`, `component_c_str`, and `model_name_c_str` must be **non-null** pointers to
///   **NUL-terminated** C strings that contain **valid UTF-8**. They must remain valid for the
///   duration of this call.
/// - `pipeline_out` must be **non-null** and point to writable memory for a `*mut WorkerSelectionPipeline`.
///   On success this function writes exactly once to `*pipeline_out`. The caller becomes the owner of
///   that pointer and **must** later free it by calling `dynamo_destroy_worker_selection_pipeline`.
/// - Must be called **after** a successful `dynamo_llm_init()`; otherwise behavior is undefined.
/// - This function is not signal-safe and must not be called from a signal handler.
/// - This function may block internally; do not call it from contexts that forbid blocking.
///
/// # Errors
/// Returns `DynamoLlmResult::ERR` on failure and does not write to `pipeline_out`.
/// # Safety
/// See detailed safety docs above. Additional parameter:
/// - `enforce_disagg`: If true, requests fail when disaggregated serving is unavailable.
///   If false, falls back to aggregated serving.
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_create_worker_selection_pipeline(
    namespace_c_str: *const c_char,
    component_c_str: *const c_char,
    model_name_c_str: *const c_char,
    use_kv_routing: bool,
    busy_threshold: f64,
    overlap_score_weight: f64,
    router_temperature: f64,
    use_kv_events: bool,
    router_replica_sync: bool,
    enforce_disagg: bool,
    pipeline_out: *mut *mut WorkerSelectionPipeline,
) -> DynamoLlmResult {
    if pipeline_out.is_null() {
        tracing::error!("pipeline_out pointer is null");
        return DynamoLlmResult::ERR;
    }

    let wk = match WK.get() {
        Some(w) => w.clone(),
        None => {
            tracing::error!("Worker not initialized. Call dynamo_llm_init first.");
            return DynamoLlmResult::ERR;
        }
    };

    let namespace = match unsafe { CStr::from_ptr(namespace_c_str) }.to_str() {
        Ok(s) => s.to_owned(),
        Err(e) => {
            tracing::error!(error = ?e, "bad namespace");
            return DynamoLlmResult::ERR;
        }
    };

    let component_cow = unsafe { cstr_or_default(component_c_str, "backend") };
    if let Cow::Borrowed("backend") = &component_cow {
        tracing::info!("defaulting to \"backend\" for component");
    }
    let component: String = component_cow.into_owned();

    let model = match unsafe { CStr::from_ptr(model_name_c_str) }.to_str() {
        Ok(s) => s.to_owned(),
        Err(e) => {
            tracing::error!(error = ?e, "bad model");
            return DynamoLlmResult::ERR;
        }
    };

    let make_engine = || async {
        let router_mode = if use_kv_routing {
            RouterMode::KV
        } else {
            RouterMode::RoundRobin
        };

        let kv_router_config = if use_kv_routing {
            Some(KvRouterConfig::new(
                (overlap_score_weight >= 0.0).then_some(overlap_score_weight),
                (router_temperature >= 0.0).then_some(router_temperature),
                Some(use_kv_events),
                Some(router_replica_sync),
                None, // track_active_blocks
                None, // assume_kv_reuse
                None, // router_snapshot_threshold
                None, // router_reset_states
                None, // router_ttl_secs
                None, // router_max_tree_size
                None, // router_prune_target_ratio
            ))
        } else {
            None
        };

        create_worker_selection_pipeline_chat(
            &namespace,
            &component,
            &model,
            router_mode,
            (busy_threshold >= 0.0).then_some(busy_threshold),
            kv_router_config,
            enforce_disagg,
        )
        .await
    };

    let (engine, kv_router) = match wk.runtime().secondary().block_on(make_engine()) {
        Ok(p) => p,
        Err(e) => {
            tracing::error!(error = ?e, "create_worker_selection_pipeline_chat failed");
            return DynamoLlmResult::ERR;
        }
    };

    let handle = Box::new(WorkerSelectionPipeline {
        wk,
        engine,
        kv_router,
    });
    unsafe {
        *pipeline_out = Box::into_raw(handle);
    }
    DynamoLlmResult::OK
}

/// Query worker selection on an existing pipeline and return:
/// - `decode_worker_id_out` (`i64`): The decode worker ID (primary worker)
/// - `prefill_worker_id_out` (`i64`): The prefill worker ID (-1 if not in disaggregated mode)
/// - `token_ids_out` (heap-allocated `*mut u32`; caller must free via
///   `dynamo_free_worker_selection_result`)
/// - `token_count_out` (`usize`)
/// - `annotated_request_json_out` (`*mut c_char` to a NUL-terminated C string;
///   caller frees via the same free function)
///
/// # Safety
/// - `pipeline`
///   - Must be a **non-null** pointer previously returned by
///     `dynamo_create_worker_selection_pipeline` and not yet passed to
///     `dynamo_destroy_worker_selection_pipeline`.
///   - Must remain valid for the entire duration of this call.
///   - **Do not** call this function concurrently on the same `pipeline` pointer
///     from multiple threads unless the surrounding code guarantees synchronization.
/// - `request_json_c_str`
///   - Must be a **non-null**, **NUL-terminated** C string containing **valid UTF-8**.
///   - The JSON must represent a valid `NvCreateChatCompletionRequest`; otherwise this
///     function returns `DynamoLlmResult::ERR`.
///   - Must remain valid for the duration of this call.
/// - Output pointers:
///   - `decode_worker_id_out`, `prefill_worker_id_out`, `token_ids_out`, `token_count_out`,
///     and `annotated_request_json_out` must each be **non-null** and point to
///     writable memory for their respective types. On success, this function
///     writes to all five outputs exactly once.
///   - On **error**, outputs are left unmodified.
/// - Ownership & deallocation:
///   - On success, if there are zero tokens, `*token_ids_out` may be set to `NULL`
///     and `*token_count_out` set to `0`.
///   - If non-null, the buffer written to `*token_ids_out` is allocated with the
///     Rust global allocator and **must** be freed by calling
///     `dynamo_free_worker_selection_result` with the same `token_count_out` value.
///   - The pointer written to `*annotated_request_json_out` is a `CString` allocated
///     by Rust and **must** be freed by calling `dynamo_free_worker_selection_result`.
///   - **Do not** free these with `free(3)` or any other allocator; doing so is
///     undefined behavior.
/// - Blocking & context:
///   - This function may **block** internally while it performs async work; do not
///     call it from contexts that forbid blocking (e.g., signal handlers).
/// - Process/ABI assumptions:
///   - The caller and callee must run in the same process and use the same Rust
///     global allocator for the paired allocation/free described above.
///   - This function is not signal-safe.
///
/// # Errors
/// Returns `DynamoLlmResult::ERR` if any precondition fails (null/invalid pointers,
/// malformed UTF-8/JSON, pipeline errors, allocation failures, etc.). On error, no
/// output pointer is written.
///
/// # Output values
/// - `decode_worker_id_out`: The decode worker ID (primary worker in aggregated mode)
/// - `prefill_worker_id_out`: The prefill worker ID (only set in disaggregated mode, -1 if not present)
/// - `token_ids_out`, `token_count_out`: Token IDs and count
/// - `annotated_request_json_out`: The annotated request JSON
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_query_worker_selection_and_annotate(
    pipeline: *mut WorkerSelectionPipeline,
    request_json_c_str: *const c_char,
    decode_worker_id_out: *mut i64,
    prefill_worker_id_out: *mut i64,
    token_ids_out: *mut *mut u32,
    token_count_out: *mut usize,
    annotated_request_json_out: *mut *mut c_char,
) -> DynamoLlmResult {
    if pipeline.is_null() {
        tracing::error!("Pipeline pointer is null");
        return DynamoLlmResult::ERR;
    }
    if decode_worker_id_out.is_null()
        || prefill_worker_id_out.is_null()
        || token_ids_out.is_null()
        || token_count_out.is_null()
        || annotated_request_json_out.is_null()
    {
        tracing::error!("One or more output pointers are null");
        return DynamoLlmResult::ERR;
    }

    let req_str = match unsafe { CStr::from_ptr(request_json_c_str) }.to_str() {
        Ok(s) => s,
        Err(e) => {
            tracing::error!(error = ?e, "bad request json");
            return DynamoLlmResult::ERR;
        }
    };
    let request: NvCreateChatCompletionRequest = match serde_json::from_str(req_str) {
        Ok(r) => r,
        Err(e) => {
            tracing::error!(error = ?e, "parse request failed");
            return DynamoLlmResult::ERR;
        }
    };

    let pl = unsafe { &*pipeline };
    let fut = async { query_worker_selection_and_annotate(&pl.engine, request).await };
    let (result, annotated_req) = match pl.wk.runtime().secondary().block_on(fut) {
        Ok(v) => v,
        Err(e) => {
            tracing::error!(error = ?e, "query_worker_selection_and_annotate failed");
            return DynamoLlmResult::ERR;
        }
    };

    let tokens_ptr = if result.tokens.is_empty() {
        std::ptr::null_mut()
    } else {
        let len = result.tokens.len();
        let layout = std::alloc::Layout::array::<u32>(len).unwrap();
        let ptr = unsafe { std::alloc::alloc(layout) as *mut u32 };
        if ptr.is_null() {
            tracing::error!("alloc tokens failed");
            return DynamoLlmResult::ERR;
        }
        unsafe {
            std::ptr::copy_nonoverlapping(result.tokens.as_ptr(), ptr, len);
        }
        ptr
    };

    let annotated_json = match serde_json::to_string(&annotated_req) {
        Ok(s) => s,
        Err(e) => {
            if !tokens_ptr.is_null() {
                let layout = std::alloc::Layout::array::<u32>(result.tokens.len()).unwrap();
                unsafe {
                    std::alloc::dealloc(tokens_ptr as *mut u8, layout);
                }
                tracing::error!(error = ?e, "serialize annotated request failed");
            }
            return DynamoLlmResult::ERR;
        }
    };
    let cjson = match std::ffi::CString::new(annotated_json) {
        Ok(c) => c,
        Err(e) => {
            tracing::error!(error = ?e, "CString::new for annotated JSON failed");
            if !tokens_ptr.is_null() {
                let layout = std::alloc::Layout::array::<u32>(result.tokens.len()).unwrap();
                unsafe {
                    std::alloc::dealloc(tokens_ptr as *mut u8, layout);
                }
            }
            return DynamoLlmResult::ERR;
        }
    };
    unsafe {
        *decode_worker_id_out = result.decode_worker_id.unwrap_or(0);
        *prefill_worker_id_out = result.prefill_worker_id.unwrap_or(-1);
        *token_ids_out = tokens_ptr;
        *token_count_out = result.tokens.len();
        *annotated_request_json_out = cjson.into_raw();
    }
    DynamoLlmResult::OK
}

/// Destroy a previously created pipeline.
///
/// # Safety
/// - `pipeline`
///   - **Must** be a non-null pointer that was **originally returned by**
///     `dynamo_create_worker_selection_pipeline` (i.e., obtained via
///     `Box::into_raw` on a `WorkerSelectionPipeline`).
///   - **Must not** have been passed to this function (or otherwise freed)
///     before. Passing the same pointer twice is a **double free** and is
///     undefined behavior.
///   - **Must not** be used by any other thread while this function runs.
///     Ensure no concurrent calls are in flight that read or write through
///     this handle (e.g., `dynamo_query_worker_selection_and_annotate`).
///   - After a successful call, the pointer is **invalid** and must not be
///     dereferenced or used again in any way.
/// - Allocator/ABI
///   - The caller and callee must be in the same process and share the same
///     allocator; this function reclaims the allocation that was created by
///     Rust for the handle.
/// - Lifetime/FFI
///   - Do not call from contexts that forbid blocking or running destructors
///     (e.g., signal handlers).
///
/// # Errors
/// - Returns `DynamoLlmResult::ERR` if `pipeline` is null.
/// - On `OK`, ownership of `pipeline` is taken and the underlying resources
///   are dropped; using the pointer after return is undefined behavior.
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_destroy_worker_selection_pipeline(
    pipeline: *mut WorkerSelectionPipeline,
) -> DynamoLlmResult {
    if pipeline.is_null() {
        tracing::error!("Pipeline pointer is null");
        return DynamoLlmResult::ERR;
    }
    let _boxed: Box<WorkerSelectionPipeline> = unsafe { Box::from_raw(pipeline) };
    DynamoLlmResult::OK
}

/// Free buffers allocated by `dynamo_query_worker_selection_and_annotate`.
///
/// # Safety
/// - `token_ids` and `annotated_request_json` **must come from this library**:
///   - `token_ids` must be the exact pointer previously returned by
///     `dynamo_query_worker_selection_and_annotate` for the tokens buffer,
///     allocated with Rust’s global allocator in this process.
///   - `annotated_request_json` must be the exact pointer previously returned by
///     `CString::into_raw` inside `dynamo_query_worker_selection_and_annotate`.
/// - **Call at most once** per pointer. Passing the same pointer again is a
///   double-free and is undefined behavior.
/// - Pointer/length invariants:
///   - If `token_ids` is non-null, `token_count` **must** be the exact length
///     originally returned. Mismatched lengths cause invalid deallocation.
///   - If `token_ids` is null, `token_count` should be `0`.
///   - Passing a non-null `token_ids` with `token_count == 0` will leak in this
///     implementation (we only dealloc when `token_count > 0`).
/// - After return, the pointers are **invalid** and must not be used again.
/// - The caller and callee must be in the same process and share the same
///   allocator/ABI (these deallocations use Rust’s global allocator).
/// - Ensure no other threads are concurrently reading/writing these buffers when
///   freeing them.
/// - Do not call from contexts that forbid running destructors (e.g., signal handlers).
///
/// Returns `DynamoLlmResult::OK` on success.
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_free_worker_selection_result(
    token_ids: *mut u32,
    token_count: usize,
    annotated_request_json: *mut c_char,
) -> DynamoLlmResult {
    if token_count > 0 {
        match std::alloc::Layout::array::<u32>(token_count) {
            Ok(layout) if !token_ids.is_null() => unsafe {
                std::alloc::dealloc(token_ids as *mut u8, layout);
            },
            _ => {}
        }
    }
    if !annotated_request_json.is_null() {
        unsafe {
            drop(std::ffi::CString::from_raw(annotated_request_json));
        }
    }
    DynamoLlmResult::OK
}

/// Default timeout for GAIE bookkeeping operations (30 seconds)
const GAIE_BOOKKEEPING_TIMEOUT_SECS: u64 = 30;

/// Helper to validate pipeline pointer and extract request_id from C string.
/// Returns `Err(DynamoLlmResult::ERR)` on validation failure, `Ok((pipeline_ref, request_id))` on success.
unsafe fn validate_pipeline_and_request_id(
    pipeline: *mut WorkerSelectionPipeline,
    request_id_c_str: *const c_char,
    operation: &str,
) -> Result<(&'static WorkerSelectionPipeline, String), DynamoLlmResult> {
    if pipeline.is_null() {
        tracing::error!("[GAIE] {} failed: pipeline pointer is null", operation);
        return Err(DynamoLlmResult::ERR);
    }

    let request_id = match unsafe { CStr::from_ptr(request_id_c_str) }.to_str() {
        Ok(s) => s.to_owned(),
        Err(e) => {
            tracing::error!(error = ?e, "[GAIE] {} failed: bad request_id", operation);
            return Err(DynamoLlmResult::ERR);
        }
    };

    // SAFETY: Caller guarantees pipeline is valid for the duration of the call
    let pl: &'static WorkerSelectionPipeline = unsafe { &*pipeline };
    Ok((pl, request_id))
}

/// Helper to run an async bookkeeping operation with timeout.
/// Returns `OK` on success or timeout, `ERR` only on validation failures (handled by caller).
fn run_bookkeeping_with_timeout<F, Fut>(
    pl: &WorkerSelectionPipeline,
    operation: &'static str,
    request_id: &str,
    f: F,
) -> DynamoLlmResult
where
    F: FnOnce() -> Fut,
    Fut: std::future::Future<Output = ()>,
{
    use std::time::Duration;

    let timeout_duration = Duration::from_secs(GAIE_BOOKKEEPING_TIMEOUT_SECS);
    let fut = f();

    let result = pl
        .wk
        .runtime()
        .secondary()
        .block_on(async { tokio::time::timeout(timeout_duration, fut).await });

    match result {
        Ok(()) => DynamoLlmResult::OK,
        Err(_elapsed) => {
            tracing::warn!(
                request_id = %request_id,
                timeout_secs = GAIE_BOOKKEEPING_TIMEOUT_SECS,
                "[GAIE] {} timed out",
                operation
            );
            // Return OK to avoid blocking the caller - the operation may still complete
            DynamoLlmResult::OK
        }
    }
}

/// Router bookkeeping functions for GAIE integration
/// Add a request to the router's bookkeeping after worker selection.
/// Call this from GAIE Stage 1 after `dynamo_query_worker_selection_and_annotate`.
///
/// This function computes the overlap_blocks internally by querying the indexer,
/// so the caller doesn't need to provide it.
///
/// # Safety
/// - `pipeline` must be a valid, non-null pointer from `dynamo_create_worker_selection_pipeline`
/// - `request_id_c_str` must be a valid NUL-terminated UTF-8 C string
/// - `token_ids` must point to at least `token_count` valid u32 values
/// - Must not be called concurrently on the same pipeline without synchronization
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_router_add_request(
    pipeline: *mut WorkerSelectionPipeline,
    request_id_c_str: *const c_char,
    token_ids: *const u32,
    token_count: usize,
    worker_id: u64,
    dp_rank: u32,
) -> DynamoLlmResult {
    let (pl, request_id) = match unsafe {
        validate_pipeline_and_request_id(pipeline, request_id_c_str, "add_request")
    } {
        Ok(v) => v,
        Err(e) => return e,
    };

    let Some(ref kv_router) = pl.kv_router else {
        tracing::debug!(
            "[GAIE] KV router not available (router_mode is not KV), skipping add_request (no-op)"
        );
        return DynamoLlmResult::OK;
    };

    // Log after kv_router check to reduce noise
    tracing::debug!(
        request_id = %request_id,
        worker_id = worker_id,
        dp_rank = dp_rank,
        token_count = token_count,
        "[GAIE] dynamo_router_add_request processing"
    );

    let tokens: Vec<u32> = if token_count > 0 && !token_ids.is_null() {
        unsafe { std::slice::from_raw_parts(token_ids, token_count) }.to_vec()
    } else {
        Vec::new()
    };

    let kv_router = kv_router.clone();
    let request_id_clone = request_id.clone();

    run_bookkeeping_with_timeout(pl, "add_request", &request_id, || async move {
        let worker = dynamo_llm::kv_router::protocols::WorkerWithDpRank::new(worker_id, dp_rank);

        // Compute overlap_blocks using the public method
        let overlap_blocks = match kv_router.get_overlap_blocks(&tokens, worker).await {
            Ok(overlap) => overlap,
            Err(e) => {
                tracing::warn!(error = ?e, "Failed to compute overlap, using 0");
                0
            }
        };

        kv_router
            .add_request(request_id_clone.clone(), &tokens, overlap_blocks, worker)
            .await;

        tracing::debug!(
            request_id = %request_id_clone,
            worker_id = worker_id,
            dp_rank = dp_rank,
            overlap_blocks = overlap_blocks,
            token_count = tokens.len(),
            "[GAIE] dynamo_router_add_request completed - request registered in router bookkeeping"
        );
    })
}

/// Mark prefill as completed for a request.
/// Call this from GAIE hook when the first token is generated.
///
/// # Safety
/// - `pipeline` must be a valid, non-null pointer from `dynamo_create_worker_selection_pipeline`
/// - `request_id_c_str` must be a valid NUL-terminated UTF-8 C string
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_router_mark_prefill_complete(
    pipeline: *mut WorkerSelectionPipeline,
    request_id_c_str: *const c_char,
) -> DynamoLlmResult {
    let (pl, request_id) = match unsafe {
        validate_pipeline_and_request_id(pipeline, request_id_c_str, "mark_prefill_complete")
    } {
        Ok(v) => v,
        Err(e) => return e,
    };

    let Some(ref kv_router) = pl.kv_router else {
        tracing::debug!(
            "[GAIE] KV router not available (router_mode is not KV), skipping mark_prefill_complete (no-op)"
        );
        return DynamoLlmResult::OK;
    };

    // Log after kv_router check to reduce noise
    tracing::debug!(
        request_id = %request_id,
        "[GAIE] dynamo_router_mark_prefill_complete processing"
    );

    let kv_router = kv_router.clone();
    let request_id_clone = request_id.clone();

    run_bookkeeping_with_timeout(pl, "mark_prefill_complete", &request_id, || async move {
        if let Err(e) = kv_router.mark_prefill_completed(&request_id_clone).await {
            tracing::warn!(
                "Failed to mark prefill completed for {}: {}",
                request_id_clone,
                e
            );
        } else {
            tracing::debug!(
                request_id = %request_id_clone,
                "[GAIE] dynamo_router_mark_prefill_complete completed - prefill tokens released"
            );
        }
    })
}

/// Free a request from the router's bookkeeping.
/// Call this from GAIE hook when the stream is closed (completed or cancelled).
///
/// # Safety
/// - `pipeline` must be a valid, non-null pointer from `dynamo_create_worker_selection_pipeline`
/// - `request_id_c_str` must be a valid NUL-terminated UTF-8 C string
#[unsafe(no_mangle)]
pub unsafe extern "C" fn dynamo_router_free_request(
    pipeline: *mut WorkerSelectionPipeline,
    request_id_c_str: *const c_char,
) -> DynamoLlmResult {
    let (pl, request_id) = match unsafe {
        validate_pipeline_and_request_id(pipeline, request_id_c_str, "free_request")
    } {
        Ok(v) => v,
        Err(e) => return e,
    };

    let Some(ref kv_router) = pl.kv_router else {
        tracing::debug!(
            "[GAIE] KV router not available (router_mode is not KV), skipping free_request (no-op)"
        );
        return DynamoLlmResult::OK;
    };

    // Log after kv_router check to reduce noise
    tracing::debug!(
        request_id = %request_id,
        "[GAIE] dynamo_router_free_request processing"
    );

    let kv_router = kv_router.clone();
    let request_id_clone = request_id.clone();

    run_bookkeeping_with_timeout(pl, "free_request", &request_id, || async move {
        if let Err(e) = kv_router.free(&request_id_clone).await {
            tracing::warn!("Failed to free request {}: {}", request_id_clone, e);
        } else {
            tracing::debug!(
                request_id = %request_id_clone,
                "[GAIE] dynamo_router_free_request completed - request removed from bookkeeping"
            );
        }
    })
}

/// Result of worker selection extraction
#[derive(Debug, Clone, Default)]
pub struct WorkerSelectionResult {
    /// Decode worker ID (primary worker for aggregated, decode-only for disaggregated)
    pub decode_worker_id: Option<i64>,
    /// Prefill worker ID (only present in disaggregated mode)
    pub prefill_worker_id: Option<i64>,
    /// Token IDs from tokenization
    pub tokens: Vec<u32>,
}

/// Helper function to extract worker selection information from the annotation stream
///
/// The response format (from disaggregated_params in nvext):
/// - worker_id: {"prefill_worker_id": 123, "decode_worker_id": 456}
/// - token_ids: [1, 2, 3, ...]
pub async fn extract_worker_selection_from_stream(
    mut stream: Pin<Box<dyn AsyncEngineStream<Annotated<NvCreateChatCompletionStreamResponse>>>>,
) -> anyhow::Result<WorkerSelectionResult> {
    use dynamo_llm::protocols::openai::nvext::WorkerIdInfo;
    use futures::StreamExt;

    let mut result = WorkerSelectionResult::default();

    while let Some(response) = stream.next().await {
        // Check for data in nvext (worker_id and token_ids are direct fields)
        // nvext is a serde_json::Value, so we access it as a JSON object
        if let Some(data) = &response.data
            && let Some(nvext) = &data.nvext
        {
            // Extract worker_id
            if let Some(worker_id_value) = nvext.get("worker_id")
                && let Ok(worker_info) =
                    serde_json::from_value::<WorkerIdInfo>(worker_id_value.clone())
            {
                result.decode_worker_id = worker_info.decode_worker_id.map(|id| id as i64);
                result.prefill_worker_id = worker_info.prefill_worker_id.map(|id| id as i64);
                tracing::debug!(
                    decode_worker_id = ?result.decode_worker_id,
                    prefill_worker_id = ?result.prefill_worker_id,
                    "Parsed worker_id from nvext"
                );
            }

            // Extract token_ids
            if let Some(token_ids_value) = nvext.get("token_ids")
                && let Ok(parsed_tokens) =
                    serde_json::from_value::<Vec<u32>>(token_ids_value.clone())
            {
                result.tokens = parsed_tokens;
                tracing::debug!(
                    "Successfully parsed {} tokens from nvext",
                    result.tokens.len()
                );
            }
        }
    }

    tracing::info!(
        decode_worker_id = ?result.decode_worker_id,
        prefill_worker_id = ?result.prefill_worker_id,
        token_count = result.tokens.len(),
        "Worker selection extraction complete"
    );
    Ok(result)
}

/// Utility function to add the "query_instance_id" annotation to an OpenAI request
///
/// This function modifies the request to include the annotation that signals the KV router
/// to return worker selection information (worker_fid and token_data) instead of
/// performing actual inference.
///
/// # Parameters
/// - `request`: Mutable reference to the OpenAI chat completion request
///
/// # Returns
/// The same request with the "query_instance_id" annotation added
pub fn add_query_instance_id(
    request: &mut NvCreateChatCompletionRequest,
) -> &mut NvCreateChatCompletionRequest {
    // Send empty value - router treats empty as aggregated / aggregated worker selection
    set_kv_annotation(request, "query_instance_id".to_string(), "")
}

/// Set worker IDs directly on the NvExt fields for GAIE Stage 2
///
/// For disaggregated mode: sets `prefill_worker_id` and `decode_worker_id`
/// For aggregated mode: sets `backend_instance_id` (when both IDs are the same)
///
/// Also sets `enable_local_updates: false` since the external caller (EPP/GAIE)
/// will handle bookkeeping via C FFI functions.
pub fn set_worker_ids_for_stage2(
    request: &mut NvCreateChatCompletionRequest,
    decode_worker_id: Option<i64>,
    prefill_worker_id: Option<i64>,
) -> &mut NvCreateChatCompletionRequest {
    let nvext = request.nvext.get_or_insert_with(|| {
        NvExt::builder()
            .build()
            .expect("NvExt builder should not fail")
    });

    // Disable local updates - external caller handles bookkeeping via C FFI
    nvext.enable_local_updates = Some(false);

    // Check if this is aggregated mode (same worker for both)
    let is_aggregated = prefill_worker_id == decode_worker_id;

    if is_aggregated {
        // Aggregated: use backend_instance_id for direct routing
        if let Some(id) = decode_worker_id {
            nvext.backend_instance_id = Some(id as u64);
            tracing::debug!(
                backend_instance_id = id,
                "GAIE Stage 2 Aggregated: Setting backend_instance_id"
            );
        }
    } else {
        // Disaggregated: use separate prefill and decode worker IDs
        if let Some(id) = prefill_worker_id {
            nvext.prefill_worker_id = Some(id as u64);
        }
        if let Some(id) = decode_worker_id {
            nvext.decode_worker_id = Some(id as u64);
        }
        tracing::debug!(
            prefill_worker_id = ?prefill_worker_id,
            decode_worker_id = ?decode_worker_id,
            "GAIE Stage 2 Disaggregated: Setting prefill and decode worker IDs"
        );
    }

    request
}

/// Set token_data directly on the NvExt field for GAIE Stage 2
pub fn set_token_data_for_stage2<'a>(
    request: &'a mut NvCreateChatCompletionRequest,
    tokens: &[u32],
) -> &'a mut NvCreateChatCompletionRequest {
    let nvext = request.nvext.get_or_insert_with(|| {
        NvExt::builder()
            .build()
            .expect("NvExt builder should not fail")
    });

    nvext.token_data = Some(tokens.to_vec());
    tracing::debug!(
        token_count = tokens.len(),
        "GAIE Stage 2: Setting token_data"
    );

    request
}

/// Ensure `nvext` exists and return a mutable slice of annotations.
fn ensure_annotations(request: &mut NvCreateChatCompletionRequest) -> &mut Vec<String> {
    let nvext = request.nvext.get_or_insert_with(|| {
        NvExt::builder()
            .build()
            .expect("NvExt builder should not fail")
    });
    nvext.annotations.get_or_insert_with(Vec::new)
}

/// Set a `key:value` annotation.
fn set_kv_annotation(
    request: &mut NvCreateChatCompletionRequest,
    key: String, // <- owned, only one borrowed param remains
    value: impl Into<String>,
) -> &mut NvCreateChatCompletionRequest {
    let prefix = format!("{}:", key);
    let kv = format!("{}{}", prefix, value.into());
    let annotations = ensure_annotations(request);
    annotations.retain(|a| !a.starts_with(&prefix));
    annotations.push(kv);
    request
}

/// Wrapper function that queries worker selection and prepares the request for GAIE Stage 2
///
/// This function performs the complete GAIE Stage 1 flow:
/// 1. Clones the original request and adds "query_instance_id:" (empty) annotation
/// 2. Calls engine.generate() with the modified request
/// 3. Extracts worker_id info and tokens from the response stream
/// 4. Sets the appropriate NvExt fields on the original request for Stage 2:
///    - Disaggregated: prefill_worker_id, decode_worker_id, token_data
///    - Aggregated: backend_instance_id, token_data
/// 5. Returns WorkerSelectionResult and the modified request ready for Stage 2
///
/// # Parameters
/// - `engine`: The worker selection pipeline engine
/// - `original_request`: The original OpenAI request to process
///
/// # Returns
/// A tuple containing (WorkerSelectionResult, modified_original_request)
/// where the modified_original_request is ready for GAIE Stage 2 execution
pub async fn query_worker_selection_and_annotate(
    engine: &ServiceEngine<
        SingleIn<NvCreateChatCompletionRequest>,
        ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>,
    >,
    mut original_request: NvCreateChatCompletionRequest,
) -> anyhow::Result<(WorkerSelectionResult, NvCreateChatCompletionRequest)> {
    // GAIE Stage 1: Query for worker selection
    let mut query_request = original_request.clone();
    add_query_instance_id(&mut query_request);
    let single_in = SingleIn::new(query_request);
    let response_stream = engine.generate(single_in).await?;
    let result = extract_worker_selection_from_stream(response_stream).await?;

    // Prepare request for GAIE Stage 2: Set NvExt fields directly
    set_worker_ids_for_stage2(
        &mut original_request,
        result.decode_worker_id,
        result.prefill_worker_id,
    );
    set_token_data_for_stage2(&mut original_request, &result.tokens);

    Ok((result, original_request))
}

/// Spawn a background task to watch for prefill models and activate prefill routers.
/// This is a lightweight watcher that only handles prefill model discovery.
fn spawn_prefill_watcher(
    drt: DistributedRuntime,
    model_manager: Arc<ModelManager>,
    target_namespace: String,
) {
    use dynamo_llm::model_card::ModelDeploymentCard;
    use dynamo_runtime::discovery::{DiscoveryEvent, DiscoveryInstance, DiscoveryQuery};
    use dynamo_runtime::protocols::EndpointId;
    use futures::StreamExt;

    tokio::spawn(async move {
        let discovery = drt.discovery();
        let mut stream = match discovery
            .list_and_watch(DiscoveryQuery::AllModels, None)
            .await
        {
            Ok(s) => s,
            Err(e) => {
                tracing::error!(error = %e, "Failed to start prefill discovery stream");
                return;
            }
        };

        while let Some(result) = stream.next().await {
            let event = match result {
                Ok(e) => e,
                Err(e) => {
                    tracing::error!(error = %e, "Error in prefill discovery stream");
                    continue;
                }
            };

            match event {
                DiscoveryEvent::Added(instance) => {
                    let (endpoint_id, card) = match &instance {
                        DiscoveryInstance::Model {
                            namespace,
                            component,
                            endpoint,
                            ..
                        } => {
                            // Filter by namespace
                            if namespace != &target_namespace {
                                continue;
                            }

                            let eid = EndpointId {
                                namespace: namespace.clone(),
                                component: component.clone(),
                                name: endpoint.clone(),
                            };

                            match instance.deserialize_model::<ModelDeploymentCard>() {
                                Ok(card) => (eid, card),
                                Err(_) => continue,
                            }
                        }
                        _ => continue,
                    };

                    // Only handle prefill models
                    if !card.model_type.supports_prefill() {
                        continue;
                    }

                    tracing::info!(
                        model_name = card.name(),
                        "Prefill model discovered, activating prefill router"
                    );

                    // Get the endpoint and activate the prefill router
                    if let Ok(ns) = drt.namespace(&endpoint_id.namespace)
                        && let Ok(comp) = ns.component(&endpoint_id.component)
                    {
                        let endpoint = comp.endpoint(&endpoint_id.name);
                        if let Err(e) = model_manager.activate_prefill_router(card.name(), endpoint)
                        {
                            tracing::warn!(
                                model_name = card.name(),
                                error = %e,
                                "Failed to activate prefill router"
                            );
                        } else {
                            tracing::info!(
                                model_name = card.name(),
                                "Prefill router activated successfully"
                            );
                        }
                    }
                }
                DiscoveryEvent::Removed(id) => {
                    // Log removal for observability
                    // Note: The PrefillRouter remains active - worker availability
                    // is handled dynamically by the underlying Client's instance tracking
                    tracing::debug!(
                        instance_id = id.instance_id(),
                        "Prefill worker instance removed from discovery"
                    );
                }
            }
        }
    });
}

/// Create a worker selection pipeline for OpenAI Chat Completion requests
///
/// This is a concrete implementation that works specifically with NvCreateChatCompletionRequest
/// and is designed for use with C bindings. Uses the "generate" endpoint by default.
///
/// # Parameters
/// - `namespace`: namespace name
/// - `component_name`: component name
/// - `model_name`: Name/slug of the model to load
/// - `router_mode`: How to route requests (KV, RoundRobin, etc.)
/// - `busy_threshold`: Optional threshold for busy worker detection
/// - `kv_router_config`: Optional KV router configuration (only used when router_mode is KV)
/// - `enforce_disagg`: If true, fail requests when disaggregated serving is unavailable
///
/// # Returns
/// A tuple of (engine, kv_router) where kv_router is Some when router_mode is KV
pub async fn create_worker_selection_pipeline_chat(
    namespace: &str,
    component_name: &str,
    model_name: &str,
    router_mode: RouterMode,
    busy_threshold: Option<f64>,
    kv_router_config: Option<KvRouterConfig>,
    enforce_disagg: bool,
) -> anyhow::Result<(
    ServiceEngine<
        SingleIn<NvCreateChatCompletionRequest>,
        ManyOut<Annotated<NvCreateChatCompletionStreamResponse>>,
    >,
    Option<Arc<dynamo_llm::kv_router::KvRouter>>,
)> {
    use dynamo_llm::kv_router::PrefillRouter;

    // Use the global DRT singleton - initialize if not already done
    // Check if already initialized (by dynamo_llm_init) to avoid redundant sync wait
    let needs_sync = DRT.get().is_none();

    let distributed_runtime = DRT
        .get_or_try_init(async {
            tracing::debug!("Initializing DistributedRuntime singleton (standalone mode)");
            DistributedRuntime::from_settings(Runtime::from_settings()?).await
        })
        .await
        .map_err(|e| anyhow::anyhow!("Failed to initialize DistributedRuntime: {}", e))?;

    // Only wait for discovery sync if we just initialized the DRT
    // (dynamo_llm_init already does this when it initializes)
    if needs_sync {
        let instance_count = wait_for_discovery_sync(distributed_runtime, 10).await;
        if instance_count == 0 {
            return Err(anyhow::anyhow!(
                "Discovery sync failed: no worker instances found. Is the backend running?"
            ));
        }
    }

    let component = distributed_runtime
        .namespace(namespace)?
        .component(component_name)?;
    let endpoint = component.endpoint(GENERATE_ENDPOINT);
    let client = endpoint.client().await?;

    // Discover the model card by searching all instances with this model name
    tracing::debug!("Looking for model: {}", model_name);
    tracing::debug!("Namespace: {}", namespace);

    let model_manager = Arc::new(ModelManager::new());
    let router_config = dynamo_llm::entrypoint::RouterConfig {
        router_mode,
        kv_router_config: kv_router_config.unwrap_or_default(),
        active_decode_blocks_threshold: busy_threshold,
        active_prefill_tokens_threshold: None,
        enforce_disagg,
    };
    // Create metrics for migration tracking (not exposed via /metrics in C bindings)
    let metrics = Arc::new(Metrics::new());
    let watcher = ModelWatcher::new(
        component.drt().clone(),
        model_manager.clone(),
        router_config,
        None,
        metrics.clone(),
    );
    let cards = watcher
        .cards_for_model(model_name, Some(namespace), false)
        .await
        .with_context(|| format!("Failed to discover model: {}", model_name))?;

    tracing::debug!("Found {} cards for model {}", cards.len(), model_name);

    let card = cards.into_iter().next().ok_or_else(|| {
        tracing::error!("No ModelDeploymentCard found for model: {}", model_name);
        anyhow::anyhow!("ModelDeploymentCard not found for model: {}", model_name)
    })?;

    let chooser = if router_mode == RouterMode::KV {
        Some(
            model_manager
                .kv_chooser_for(&endpoint, card.kv_cache_block_size, kv_router_config)
                .await?,
        )
    } else {
        None
    };

    // Create prefill chooser for dynamic disaggregation support
    // This registers the model and returns a receiver that will be activated
    // when a prefill worker is discovered
    let prefill_chooser = model_manager
        .register_prefill_router(model_name.to_string())
        .map(|rx| {
            // Create prefill-specific config with track_active_blocks disabled
            let mut prefill_config = kv_router_config.unwrap_or_default();
            prefill_config.router_track_active_blocks = false;

            PrefillRouter::new(
                rx,
                model_manager.clone(),
                router_mode,
                card.kv_cache_block_size,
                Some(prefill_config),
                enforce_disagg,
            )
        });

    // Start background watcher for prefill model discovery
    // This will activate the prefill router when prefill workers join
    spawn_prefill_watcher(
        component.drt().clone(),
        model_manager.clone(),
        namespace.to_string(),
    );

    // Download model config files from HuggingFace for EPP
    // The backend's card has NATS URLs which aren't accessible from EPP
    tracing::debug!(
        "Downloading model config files for EPP: {}",
        card.display_name
    );

    let local_path = dynamo_llm::hub::from_hf(&card.display_name, true)
        .await
        .with_context(|| {
            format!(
                "Failed to download model config files for: {}",
                card.display_name
            )
        })?;

    // Load a fresh card from local files, then copy runtime config from original card
    tracing::debug!("Loading ModelDeploymentCard from local path...");
    let mut card_with_local_files = ModelDeploymentCard::load_from_disk(&local_path, None)
        .with_context(|| format!("Failed to load card from disk: {:?}", local_path))?;

    // Copy runtime settings from the backend's card
    tracing::debug!("Copying runtime config from backend card...");
    card_with_local_files.runtime_config = card.runtime_config.clone();
    card_with_local_files.kv_cache_block_size = card.kv_cache_block_size;
    card_with_local_files.context_length = card.context_length;

    // Load the tokenizer from the downloaded files
    tracing::debug!("Loading tokenizer from local files...");
    let hf_tokenizer = card_with_local_files
        .tokenizer_hf()
        .with_context(|| format!("Failed to load tokenizer for: {}", card.display_name))?;

    // Create worker monitor if busy_threshold is set
    // Note: C bindings don't register with ModelManager, so HTTP endpoint won't see this
    let worker_monitor = busy_threshold.map(|t| KvWorkerMonitor::new(client.clone(), t, 1000000));

    // Clone chooser before passing to build_routed_pipeline (which takes ownership)
    let kv_router = chooser.clone();

    let engine = build_routed_pipeline::<
        NvCreateChatCompletionRequest,
        NvCreateChatCompletionStreamResponse,
    >(
        &card_with_local_files,
        &client,
        model_manager.clone(),
        router_mode,
        worker_monitor,
        chooser,
        hf_tokenizer,
        prefill_chooser,
        enforce_disagg,
        metrics,
    )
    .await?;

    Ok((engine, kv_router))
}