chore(discovery): Watch/publish ModelDeploymentCard instead of ModelEntry (#3350)

Signed-off-by: Graham King <grahamk@nvidia.com>

lib/bindings/python/examples/openai_service/server.py

@@ -69,7 +69,7 @@ async def worker(runtime: DistributedRuntime):
     host: str = "localhost"
     port: int = 8000
     service: HttpService = HttpService(port=port)
-    service.add_chat_completions_model(served_model_name, engine)
+    service.add_chat_completions_model(served_model_name, "mdcsum", engine)
 
     print("Starting service...")
     shutdown_signal = service.run(runtime.child_token())

lib/bindings/python/rust/http.rs

@@ -30,23 +30,29 @@ impl HttpService {
         Ok(Self { inner })
     }
 
-    pub fn add_completions_model(&self, model: String, engine: HttpAsyncEngine) -> PyResult<()> {
+    pub fn add_completions_model(
+        &self,
+        model: String,
+        checksum: String,
+        engine: HttpAsyncEngine,
+    ) -> PyResult<()> {
         let engine = Arc::new(engine);
         self.inner
             .model_manager()
-            .add_completions_model(&model, engine)
+            .add_completions_model(&model, &checksum, engine)
             .map_err(to_pyerr)
     }
 
     pub fn add_chat_completions_model(
         &self,
         model: String,
+        checksum: String,
         engine: HttpAsyncEngine,
     ) -> PyResult<()> {
         let engine = Arc::new(engine);
         self.inner
             .model_manager()
-            .add_chat_completions_model(&model, engine)
+            .add_chat_completions_model(&model, &checksum, engine)
             .map_err(to_pyerr)
     }
 

lib/bindings/python/tests/test_http_server.py

@@ -85,6 +85,7 @@ async def http_server(runtime: DistributedRuntime):
     model_name = "test_model"
     start_done = asyncio.Event()
     child_token = runtime.child_token()
+    checksum = "abc123"  # Checksum of ModelDeplomentCard for that model
 
     async def worker():
         """The server worker task."""
@@ -94,7 +95,7 @@ async def http_server(runtime: DistributedRuntime):
             engine = HttpAsyncEngine(python_engine.generate, loop)
 
             service = HttpService(port=port)
-            service.add_chat_completions_model(model_name, engine)
+            service.add_chat_completions_model(model_name, checksum, engine)
             service.enable_endpoint("chat", True)
 
             shutdown_signal = service.run(child_token)

lib/llm/src/common/checked_file.rs

@@ -33,7 +33,7 @@ pub struct Checksum {
     algorithm: CryptographicHashMethods,
 }
 
-#[derive(Serialize, Deserialize, Debug, Clone, Eq, PartialEq)]
+#[derive(Serialize, Deserialize, Debug, Clone, Copy, Eq, PartialEq)]
 pub enum CryptographicHashMethods {
     #[serde(rename = "blake3")]
     BLAKE3,
@@ -259,6 +259,15 @@ impl TryFrom<&str> for Checksum {
     }
 }
 
+impl Default for Checksum {
+    fn default() -> Self {
+        Self {
+            hash: "".to_string(),
+            algorithm: CryptographicHashMethods::BLAKE3,
+        }
+    }
+}
+
 impl FromStr for CryptographicHashMethods {
     type Err = String;
 

lib/llm/src/discovery.rs

@@ -4,14 +4,8 @@
 mod model_manager;
 pub use model_manager::{ModelManager, ModelManagerError};
 
-mod model_entry;
-pub use model_entry::ModelEntry;
-
 mod watcher;
 pub use watcher::{ModelUpdate, ModelWatcher};
 
-/// The root etcd path for ModelEntry
-pub const MODEL_ROOT_PATH: &str = "models";
-
 /// The root etcd path for KV Router registrations
 pub const KV_ROUTERS_ROOT_PATH: &str = "kv_routers";

lib/llm/src/discovery/model_entry.rs

-// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-use dynamo_runtime::{protocols, slug::Slug};
-use serde::{Deserialize, Serialize};
-
-use crate::local_model::runtime_config::ModelRuntimeConfig;
-
-/// [ModelEntry] contains the information to discover models
-#[derive(Debug, Clone, Serialize, Deserialize, Eq, PartialEq)]
-pub struct ModelEntry {
-    /// Public name of the model
-    /// Used to identify the model in the HTTP service from the value used in an OpenAI ChatRequest.
-    pub name: String,
-
-    /// How to address this on the network
-    #[serde(rename = "endpoint")]
-    pub endpoint_id: protocols::EndpointId,
-
-    /// Runtime configuration specific to this model instance
-    #[serde(default, skip_serializing_if = "Option::is_none")]
-    pub runtime_config: Option<ModelRuntimeConfig>,
-}
-
-impl ModelEntry {
-    /// Slugified display name for use in network storage, or URL-safe environments
-    pub fn slug(&self) -> Slug {
-        Slug::from_string(&self.name)
-    }
-}

lib/llm/src/discovery/model_manager.rs

@@ -11,7 +11,6 @@ use parking_lot::{Mutex, RwLock};
 use dynamo_runtime::component::Component;
 use dynamo_runtime::prelude::DistributedRuntimeProvider;
 
-use crate::kv_router::{KvRouterConfig, scheduler::DefaultWorkerSelector};
 use crate::{discovery::KV_ROUTERS_ROOT_PATH, model_card::ModelDeploymentCard};
 use crate::{
     kv_router::KvRouter,
@@ -21,6 +20,10 @@ use crate::{
         completions::OpenAICompletionsStreamingEngine, embeddings::OpenAIEmbeddingsStreamingEngine,
     },
 };
+use crate::{
+    kv_router::{KvRouterConfig, scheduler::DefaultWorkerSelector},
+    model_type::ModelType,
+};
 
 #[derive(Debug, thiserror::Error)]
 pub enum ModelManagerError {
@@ -39,7 +42,7 @@ pub struct ModelManager {
     embeddings_engines: RwLock<ModelEngines<OpenAIEmbeddingsStreamingEngine>>,
     tensor_engines: RwLock<ModelEngines<TensorStreamingEngine>>,
 
-    // These two are Mutex because we read and write rarely and equally
+    // These are Mutex because we read and write rarely and equally
     cards: Mutex<HashMap<String, ModelDeploymentCard>>,
     kv_choosers: Mutex<HashMap<String, Arc<KvRouter>>>,
 }
@@ -62,6 +65,43 @@ impl ModelManager {
         }
     }
 
+    pub fn is_valid_checksum(
+        &self,
+        model_type: ModelType,
+        model_name: &str,
+        candidate_checksum: &str,
+    ) -> Option<bool> {
+        let mut results = vec![];
+        for unit in model_type.units() {
+            let maybe_valid_checksum = match unit {
+                ModelType::Chat => self.chat_completion_engines.read().checksum(model_name),
+                ModelType::Completions => self.completion_engines.read().checksum(model_name),
+                ModelType::Embedding => self.embeddings_engines.read().checksum(model_name),
+                ModelType::TensorBased => self.tensor_engines.read().checksum(model_name),
+                _ => {
+                    continue;
+                }
+            };
+            if let Some(is_valid) = maybe_valid_checksum.map(|valid_checksum| {
+                tracing::debug!(
+                    model_name,
+                    valid_checksum,
+                    candidate_checksum,
+                    "is_valid_checksum: check case"
+                );
+                valid_checksum == candidate_checksum
+            }) {
+                results.push(is_valid)
+            }
+        }
+        if results.is_empty() {
+            None
+        } else {
+            // The checksum is valid if it is correct for all the ModelType in the bitflag.
+            Some(results.into_iter().all(|x| x))
+        }
+    }
+
     pub fn get_model_cards(&self) -> Vec<ModelDeploymentCard> {
         self.cards.lock().values().cloned().collect()
     }
@@ -99,37 +139,41 @@ impl ModelManager {
     pub fn add_completions_model(
         &self,
         model: &str,
+        card_checksum: &str,
         engine: OpenAICompletionsStreamingEngine,
     ) -> Result<(), ModelManagerError> {
         let mut clients = self.completion_engines.write();
-        clients.add(model, engine)
+        clients.add(model, card_checksum, engine)
     }
 
     pub fn add_chat_completions_model(
         &self,
         model: &str,
+        card_checksum: &str,
         engine: OpenAIChatCompletionsStreamingEngine,
     ) -> Result<(), ModelManagerError> {
         let mut clients = self.chat_completion_engines.write();
-        clients.add(model, engine)
+        clients.add(model, card_checksum, engine)
     }
 
     pub fn add_embeddings_model(
         &self,
         model: &str,
+        card_checksum: &str,
         engine: OpenAIEmbeddingsStreamingEngine,
     ) -> Result<(), ModelManagerError> {
         let mut clients = self.embeddings_engines.write();
-        clients.add(model, engine)
+        clients.add(model, card_checksum, engine)
     }
 
     pub fn add_tensor_model(
         &self,
         model: &str,
+        card_checksum: &str,
         engine: TensorStreamingEngine,
     ) -> Result<(), ModelManagerError> {
         let mut clients = self.tensor_engines.write();
-        clients.add(model, engine)
+        clients.add(model, card_checksum, engine)
     }
 
     pub fn remove_completions_model(&self, model: &str) -> Result<(), ModelManagerError> {
@@ -196,10 +240,11 @@ impl ModelManager {
             .ok_or(ModelManagerError::ModelNotFound(model.to_string()))
     }
 
-    /// Save a ModelDeploymentCard from an instance's etcd `models/` key so we can fetch it later when the key is
-    /// deleted from etcd.
-    pub fn save_model_card(&self, key: &str, entry: ModelDeploymentCard) {
-        self.cards.lock().insert(key.to_string(), entry);
+    /// Save a ModelDeploymentCard from an instance's ModelDeploymentCard key so we can fetch it later when the key is
+    /// deleted.
+    pub fn save_model_card(&self, key: &str, card: ModelDeploymentCard) -> anyhow::Result<()> {
+        self.cards.lock().insert(key.to_string(), card);
+        Ok(())
     }
 
     /// Remove and return model card for this instance's etcd key. We do this when the instance stops.
@@ -291,6 +336,9 @@ pub struct ModelEngines<E> {
     /// Optional default model name
     default: Option<String>,
     engines: HashMap<String, E>,
+    /// Key: Model name, value: Checksum of the ModelDeploymentCard. New instances must have the
+    /// same card.
+    checksums: HashMap<String, String>,
 }
 
 impl<E> Default for ModelEngines<E> {
@@ -298,6 +346,7 @@ impl<E> Default for ModelEngines<E> {
         Self {
             default: None,
             engines: HashMap::new(),
+            checksums: HashMap::new(),
         }
     }
 }
@@ -313,11 +362,13 @@ impl<E> ModelEngines<E> {
         self.default = None;
     }
 
-    fn add(&mut self, model: &str, engine: E) -> Result<(), ModelManagerError> {
+    fn add(&mut self, model: &str, checksum: &str, engine: E) -> Result<(), ModelManagerError> {
         if self.engines.contains_key(model) {
             return Err(ModelManagerError::ModelAlreadyExists(model.to_string()));
         }
         self.engines.insert(model.to_string(), engine);
+        self.checksums
+            .insert(model.to_string(), checksum.to_string());
         Ok(())
     }
 
@@ -325,6 +376,7 @@ impl<E> ModelEngines<E> {
         if self.engines.remove(model).is_none() {
             return Err(ModelManagerError::ModelNotFound(model.to_string()));
         }
+        let _ = self.checksums.remove(model);
         Ok(())
     }
 
@@ -339,4 +391,10 @@ impl<E> ModelEngines<E> {
     pub fn list(&self) -> Vec<String> {
         self.engines.keys().map(|k| k.to_owned()).collect()
     }
+
+    /// Returns a newly allocated String for called convenience. All the places I use
+    /// this I need a String.
+    pub fn checksum(&self, model: &str) -> Option<String> {
+        self.checksums.get(model).map(|s| s.to_string())
+    }
 }

lib/llm/src/entrypoint/input/common.rs

@@ -5,12 +5,12 @@ use std::pin::Pin;
 
 use crate::{
     backend::{Backend, ExecutionContext},
-    discovery::{MODEL_ROOT_PATH, ModelManager, ModelWatcher},
+    discovery::{ModelManager, ModelWatcher},
     engines::StreamingEngineAdapter,
     entrypoint::{self, EngineConfig},
     kv_router::{KvPushRouter, KvRouter},
     migration::Migration,
-    model_card::ModelDeploymentCard,
+    model_card::{self, ModelDeploymentCard},
     preprocessor::{OpenAIPreprocessor, prompt::PromptFormatter},
     protocols::common::llm_backend::{BackendOutput, LLMEngineOutput, PreprocessedRequest},
     request_template::RequestTemplate,
@@ -73,7 +73,9 @@ pub async fn prepare_engine(
                 None,
                 None,
             ));
-            let models_watcher = etcd_client.kv_get_and_watch_prefix(MODEL_ROOT_PATH).await?;
+            let models_watcher = etcd_client
+                .kv_get_and_watch_prefix(model_card::ROOT_PATH)
+                .await?;
             let (_prefix, _watcher, receiver) = models_watcher.dissolve();
 
             let inner_watch_obj = watch_obj.clone();

lib/llm/src/entrypoint/input/grpc.rs

@@ -4,11 +4,12 @@
 use std::sync::Arc;
 
 use crate::{
-    discovery::{MODEL_ROOT_PATH, ModelManager, ModelWatcher},
+    discovery::{ModelManager, ModelWatcher},
     engines::StreamingEngineAdapter,
     entrypoint::{self, EngineConfig, input::common},
     grpc::service::kserve,
     kv_router::KvRouterConfig,
+    model_card,
     namespace::is_global_namespace,
     types::openai::{
         chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse},
@@ -46,7 +47,7 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                         distributed_runtime,
                         grpc_service.state().manager_clone(),
                         etcd_client.clone(),
-                        MODEL_ROOT_PATH,
+                        model_card::ROOT_PATH,
                         router_config.router_mode,
                         Some(router_config.kv_router_config),
                         router_config.busy_threshold,
@@ -62,6 +63,7 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
         }
         EngineConfig::StaticRemote(local_model) => {
             let card = local_model.card();
+            let checksum = card.mdcsum();
             let router_mode = local_model.router_config().router_mode;
 
             let dst_config = DistributedConfig::from_settings(true); // true means static
@@ -103,7 +105,11 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                 tokenizer_hf.clone(),
             )
             .await?;
-            manager.add_chat_completions_model(local_model.display_name(), chat_engine)?;
+            manager.add_chat_completions_model(
+                local_model.display_name(),
+                checksum,
+                chat_engine,
+            )?;
 
             let completions_engine =
                 entrypoint::build_routed_pipeline::<
@@ -111,7 +117,11 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                     NvCreateCompletionResponse,
                 >(card, &client, router_mode, None, kv_chooser, tokenizer_hf)
                 .await?;
-            manager.add_completions_model(local_model.display_name(), completions_engine)?;
+            manager.add_completions_model(
+                local_model.display_name(),
+                checksum,
+                completions_engine,
+            )?;
 
             grpc_service
         }
@@ -119,8 +129,9 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
             let grpc_service = grpc_service_builder.build()?;
             let engine = Arc::new(StreamingEngineAdapter::new(engine));
             let manager = grpc_service.model_manager();
-            manager.add_completions_model(model.service_name(), engine.clone())?;
-            manager.add_chat_completions_model(model.service_name(), engine)?;
+            let checksum = model.card().mdcsum();
+            manager.add_completions_model(model.service_name(), checksum, engine.clone())?;
+            manager.add_chat_completions_model(model.service_name(), checksum, engine)?;
             grpc_service
         }
         EngineConfig::StaticCore {
@@ -130,6 +141,7 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
         } => {
             let grpc_service = grpc_service_builder.build()?;
             let manager = grpc_service.model_manager();
+            let checksum = model.card().mdcsum();
 
             let tokenizer_hf = model.card().tokenizer_hf()?;
             let chat_pipeline =
@@ -138,14 +150,14 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                     NvCreateChatCompletionStreamResponse,
                 >(model.card(), inner_engine.clone(), tokenizer_hf.clone())
                 .await?;
-            manager.add_chat_completions_model(model.service_name(), chat_pipeline)?;
+            manager.add_chat_completions_model(model.service_name(), checksum, chat_pipeline)?;
 
             let cmpl_pipeline = common::build_pipeline::<
                 NvCreateCompletionRequest,
                 NvCreateCompletionResponse,
             >(model.card(), inner_engine, tokenizer_hf)
             .await?;
-            manager.add_completions_model(model.service_name(), cmpl_pipeline)?;
+            manager.add_completions_model(model.service_name(), checksum, cmpl_pipeline)?;
             grpc_service
         }
     };

lib/llm/src/entrypoint/input/http.rs

@@ -4,12 +4,13 @@
 use std::sync::Arc;
 
 use crate::{
-    discovery::{MODEL_ROOT_PATH, ModelManager, ModelUpdate, ModelWatcher},
+    discovery::{ModelManager, ModelUpdate, ModelWatcher},
     endpoint_type::EndpointType,
     engines::StreamingEngineAdapter,
     entrypoint::{self, EngineConfig, input::common},
     http::service::service_v2::{self, HttpService},
     kv_router::KvRouterConfig,
+    model_card,
     namespace::is_global_namespace,
     types::openai::{
         chat_completions::{NvCreateChatCompletionRequest, NvCreateChatCompletionStreamResponse},
@@ -74,7 +75,7 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                         distributed_runtime,
                         http_service.state().manager_clone(),
                         etcd_client.clone(),
-                        MODEL_ROOT_PATH,
+                        model_card::ROOT_PATH,
                         router_config.router_mode,
                         Some(router_config.kv_router_config),
                         router_config.busy_threshold,
@@ -92,6 +93,8 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
         }
         EngineConfig::StaticRemote(local_model) => {
             let card = local_model.card();
+            let checksum = card.mdcsum();
+
             let router_mode = local_model.router_config().router_mode;
 
             let dst_config = DistributedConfig::from_settings(true); // true means static
@@ -133,7 +136,11 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                 tokenizer_hf.clone(),
             )
             .await?;
-            manager.add_chat_completions_model(local_model.display_name(), chat_engine)?;
+            manager.add_chat_completions_model(
+                local_model.display_name(),
+                checksum,
+                chat_engine,
+            )?;
 
             let completions_engine =
                 entrypoint::build_routed_pipeline::<
@@ -141,7 +148,11 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                     NvCreateCompletionResponse,
                 >(card, &client, router_mode, None, kv_chooser, tokenizer_hf)
                 .await?;
-            manager.add_completions_model(local_model.display_name(), completions_engine)?;
+            manager.add_completions_model(
+                local_model.display_name(),
+                checksum,
+                completions_engine,
+            )?;
 
             for endpoint_type in EndpointType::all() {
                 http_service.enable_model_endpoint(endpoint_type, true);
@@ -153,8 +164,9 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
             let http_service = http_service_builder.build()?;
             let engine = Arc::new(StreamingEngineAdapter::new(engine));
             let manager = http_service.model_manager();
-            manager.add_completions_model(model.service_name(), engine.clone())?;
-            manager.add_chat_completions_model(model.service_name(), engine)?;
+            let checksum = model.card().mdcsum();
+            manager.add_completions_model(model.service_name(), checksum, engine.clone())?;
+            manager.add_chat_completions_model(model.service_name(), checksum, engine)?;
 
             // Enable all endpoints
             for endpoint_type in EndpointType::all() {
@@ -169,6 +181,7 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
         } => {
             let http_service = http_service_builder.build()?;
             let manager = http_service.model_manager();
+            let checksum = model.card().mdcsum();
 
             let tokenizer_hf = model.card().tokenizer_hf()?;
             let chat_pipeline =
@@ -177,14 +190,14 @@ pub async fn run(runtime: Runtime, engine_config: EngineConfig) -> anyhow::Resul
                     NvCreateChatCompletionStreamResponse,
                 >(model.card(), inner_engine.clone(), tokenizer_hf.clone())
                 .await?;
-            manager.add_chat_completions_model(model.service_name(), chat_pipeline)?;
+            manager.add_chat_completions_model(model.service_name(), checksum, chat_pipeline)?;
 
             let cmpl_pipeline = common::build_pipeline::<
                 NvCreateCompletionRequest,
                 NvCreateCompletionResponse,
             >(model.card(), inner_engine, tokenizer_hf)
             .await?;
-            manager.add_completions_model(model.service_name(), cmpl_pipeline)?;
+            manager.add_completions_model(model.service_name(), checksum, cmpl_pipeline)?;
             // Enable all endpoints
             for endpoint_type in EndpointType::all() {
                 http_service.enable_model_endpoint(endpoint_type, true);

lib/llm/src/kv_router.rs

@@ -32,7 +32,6 @@ pub mod sequence;
 pub mod subscriber;
 
 use crate::{
-    discovery::{MODEL_ROOT_PATH, ModelEntry},
     kv_router::{
         approx::ApproxKvIndexer,
         indexer::{
@@ -45,6 +44,7 @@ use crate::{
         subscriber::start_kv_router_background,
     },
     local_model::runtime_config::ModelRuntimeConfig,
+    model_card::{self, ModelDeploymentCard},
     preprocessor::PreprocessedRequest,
     protocols::common::llm_backend::LLMEngineOutput,
 };
@@ -247,9 +247,9 @@ impl KvRouter {
 
         let runtime_configs_watcher = watch_prefix_with_extraction(
             etcd_client,
-            MODEL_ROOT_PATH,
+            model_card::ROOT_PATH,
             key_extractors::lease_id,
-            |model_entry: ModelEntry| model_entry.runtime_config,
+            |card: ModelDeploymentCard| Some(card.runtime_config),
             cancellation_token.clone(),
         )
         .await?;

lib/llm/src/local_model.rs

@@ -15,15 +15,12 @@ use dynamo_runtime::{
     storage::key_value_store::{EtcdStorage, KeyValueStore, KeyValueStoreManager},
 };
 
-use crate::discovery::ModelEntry;
 use crate::entrypoint::RouterConfig;
 use crate::mocker::protocols::MockEngineArgs;
 use crate::model_card::{self, ModelDeploymentCard};
 use crate::model_type::{ModelInput, ModelType};
 use crate::request_template::RequestTemplate;
 
-mod network_name;
-pub use network_name::ModelNetworkName;
 pub mod runtime_config;
 
 use runtime_config::ModelRuntimeConfig;
@@ -421,36 +418,13 @@ impl LocalModel {
         // Publish the Model Deployment Card to KV store
         let kvstore: Box<dyn KeyValueStore> = Box::new(EtcdStorage::new(etcd_client.clone()));
         let card_store = Arc::new(KeyValueStoreManager::new(kvstore));
-        let key = self.card.slug().to_string();
-        // TODO: Next PR will use this
-        //let lease_id = endpoint.drt().primary_lease().map(|l| l.id()).unwrap_or(0);
-        //let key = Key::from_raw(endpoint.unique_path(lease_id));
-
-        card_store
-            .publish(
-                model_card::ROOT_PATH,
-                None,
-                &Key::from_raw(key),
-                &mut self.card,
-            )
-            .await?;
+        let lease_id = endpoint.drt().primary_lease().map(|l| l.id()).unwrap_or(0);
+        let key = Key::from_raw(endpoint.unique_path(lease_id));
 
-        // Publish our ModelEntry to etcd. This allows ingress to find the model card.
-        // (Why don't we put the model card directly under this key?)
-        let network_name = ModelNetworkName::new();
-        tracing::debug!("Registering with etcd as {network_name}");
-        let model_registration = ModelEntry {
-            name: self.display_name().to_string(),
-            endpoint_id: endpoint.id(),
-            runtime_config: Some(self.runtime_config.clone()),
-        };
-        etcd_client
-            .kv_create(
-                &network_name,
-                serde_json::to_vec_pretty(&model_registration)?,
-                None, // use primary lease
-            )
-            .await
+        let _outcome = card_store
+            .publish(model_card::ROOT_PATH, None, &key, &mut self.card)
+            .await?;
+        Ok(())
     }
 }
 

lib/llm/src/local_model/network_name.rs

-// SPDX-FileCopyrightText: Copyright (c) 2024-2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-// SPDX-License-Identifier: Apache-2.0
-
-use crate::discovery::MODEL_ROOT_PATH;
-
-#[derive(Debug, Clone)]
-pub struct ModelNetworkName(String);
-
-impl ModelNetworkName {
-    pub fn new() -> Self {
-        ModelNetworkName(format!("{MODEL_ROOT_PATH}/{}", uuid::Uuid::new_v4()))
-    }
-}
-
-impl Default for ModelNetworkName {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-impl std::fmt::Display for ModelNetworkName {
-    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
-        write!(f, "{}", self.0)
-    }
-}
-
-impl AsRef<str> for ModelNetworkName {
-    fn as_ref(&self) -> &str {
-        &self.0
-    }
-}
-
-impl std::ops::Deref for ModelNetworkName {
-    type Target = str;
-    fn deref(&self) -> &Self::Target {
-        &self.0
-    }
-}

lib/llm/src/model_card.rs

@@ -15,9 +15,9 @@
 use std::fmt;
 use std::fs::File;
 use std::path::{Path, PathBuf};
-use std::sync::Arc;
+use std::sync::{Arc, OnceLock};
 
-use crate::common::checked_file::CheckedFile;
+use crate::common::checked_file::{CheckedFile, Checksum};
 use crate::local_model::runtime_config::ModelRuntimeConfig;
 use crate::model_type::{ModelInput, ModelType};
 use anyhow::{Context, Result};
@@ -43,6 +43,15 @@ pub enum ModelInfoType {
     GGUF(PathBuf),
 }
 
+impl ModelInfoType {
+    pub fn checksum(&self) -> String {
+        match self {
+            ModelInfoType::HfConfigJson(c) => c.checksum().to_string(),
+            ModelInfoType::GGUF(_) => Checksum::default().to_string(),
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Clone, Debug)]
 #[serde(rename_all = "snake_case")]
 pub enum TokenizerKind {
@@ -50,6 +59,15 @@ pub enum TokenizerKind {
     GGUF(Box<HfTokenizer>),
 }
 
+impl TokenizerKind {
+    pub fn checksum(&self) -> String {
+        match self {
+            TokenizerKind::HfTokenizerJson(c) => c.checksum().to_string(),
+            TokenizerKind::GGUF(_) => Checksum::default().to_string(),
+        }
+    }
+}
+
 /// Supported types of prompt formatters.
 ///
 /// We need a way to associate the prompt formatter template definition with an associated
@@ -70,6 +88,16 @@ pub enum PromptFormatterArtifact {
     GGUF(PathBuf),
 }
 
+impl PromptFormatterArtifact {
+    pub fn checksum(&self) -> String {
+        match self {
+            PromptFormatterArtifact::HfTokenizerConfigJson(c) => c.checksum().to_string(),
+            PromptFormatterArtifact::HfChatTemplate(c) => c.checksum().to_string(),
+            PromptFormatterArtifact::GGUF(_) => Checksum::default().to_string(),
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Clone, Debug, PartialEq, Eq, Hash)]
 #[serde(rename_all = "snake_case")]
 pub enum PromptContextMixin {
@@ -87,6 +115,15 @@ pub enum GenerationConfig {
     GGUF(PathBuf),
 }
 
+impl GenerationConfig {
+    pub fn checksum(&self) -> String {
+        match self {
+            GenerationConfig::HfGenerationConfigJson(c) => c.checksum().to_string(),
+            GenerationConfig::GGUF(_) => Checksum::default().to_string(),
+        }
+    }
+}
+
 #[derive(Serialize, Deserialize, Clone, Debug, Builder, Default)]
 pub struct ModelDeploymentCard {
     /// Human readable model name, e.g. "Meta Llama 3.1 8B Instruct"
@@ -145,6 +182,9 @@ pub struct ModelDeploymentCard {
 
     #[serde(skip)]
     cache_dir: Option<Arc<tempfile::TempDir>>,
+
+    #[serde(skip, default)]
+    checksum: OnceLock<String>,
 }
 
 impl ModelDeploymentCard {
@@ -189,6 +229,12 @@ impl ModelDeploymentCard {
         Ok(())
     }
 
+    #[inline]
+    pub fn name(&self) -> &str {
+        &self.display_name
+    }
+
+    #[inline]
     pub fn slug(&self) -> &Slug {
         &self.slug
     }
@@ -198,9 +244,45 @@ impl ModelDeploymentCard {
         Ok(serde_json::to_string(self)?)
     }
 
-    pub fn mdcsum(&self) -> String {
-        let json = self.to_json().unwrap();
-        format!("{}", blake3::hash(json.as_bytes()))
+    pub fn mdcsum(&self) -> &str {
+        self.checksum
+            .get_or_init(|| {
+                // Only include the important fields
+                let mut bytes_to_hash: Vec<u8> = Vec::with_capacity(512);
+                bytes_to_hash.extend(self.display_name.as_bytes());
+
+                // The files can be either a URL or a local path, so we ignore that and hash their
+                // checksum instead, which won't change wherever they are.
+
+                if let Some(model_info) = self.model_info.as_ref() {
+                    bytes_to_hash.extend(model_info.checksum().as_bytes());
+                }
+                if let Some(tokenizer) = self.tokenizer.as_ref() {
+                    bytes_to_hash.extend(tokenizer.checksum().as_bytes());
+                }
+                if let Some(prompt_formatter) = self.prompt_formatter.as_ref() {
+                    bytes_to_hash.extend(prompt_formatter.checksum().as_bytes());
+                }
+                if let Some(chat_template) = self.chat_template_file.as_ref() {
+                    bytes_to_hash.extend(chat_template.checksum().as_bytes());
+                }
+                if let Some(gen_config) = self.gen_config.as_ref() {
+                    bytes_to_hash.extend(gen_config.checksum().as_bytes());
+                }
+
+                if let Some(prompt_context_vec) = self.prompt_context.as_ref() {
+                    // Paste it as the bytes of the debug format. It's a Vec of enum, so this should be
+                    // fine. If the debug representation changes that only happens in a new release.
+                    bytes_to_hash.extend(format!("{prompt_context_vec:?}").as_bytes());
+                }
+                bytes_to_hash.extend(self.context_length.to_be_bytes());
+                bytes_to_hash.extend(self.kv_cache_block_size.to_be_bytes());
+
+                // TODO: Do we want any of user_data or runtime_config?
+
+                blake3::hash(&bytes_to_hash).to_string()
+            })
+            .as_ref()
     }
 
     /// Is this a full model card with tokenizer?
@@ -291,9 +373,7 @@ impl ModelDeploymentCard {
 
     /// Move the files this MDC uses from the NATS object store to local disk.
     /// Updates the URI's to point to the created files.
-    ///
-    /// The returned TempDir must be kept alive, it cleans up on drop.
-    async fn move_from_nats(&mut self, nats_client: nats::Client) -> Result<tempfile::TempDir> {
+    pub async fn move_from_nats(&mut self, nats_client: nats::Client) -> Result<()> {
         let nats_addr = nats_client.addr();
         let bucket_name = self.slug();
         let target_dir = tempfile::TempDir::with_prefix(bucket_name.to_string())?;
@@ -345,7 +425,9 @@ impl ModelDeploymentCard {
             "tokenizer.json"
         );
 
-        Ok(target_dir)
+        // This cache_dir is a tempfile::TempDir will be deleted on drop, so keep it alive.
+        self.cache_dir = Some(Arc::new(target_dir));
+        Ok(())
     }
 
     /// Delete this card from the key-value store and it's URLs from the object store
@@ -411,8 +493,7 @@ impl ModelDeploymentCard {
         else {
             return Ok(None);
         };
-        // This cache_dir is a tempfile::TempDir will be deleted on drop, so keep it alive.
-        card.cache_dir = Some(Arc::new(card.move_from_nats(drt.nats_client()).await?));
+        card.move_from_nats(drt.nats_client()).await?;
         Ok(Some(card))
     }
 
@@ -487,6 +568,7 @@ impl ModelDeploymentCard {
             user_data: None,
             runtime_config: ModelRuntimeConfig::default(),
             cache_dir: None,
+            checksum: OnceLock::new(),
         })
     }
 
@@ -551,6 +633,7 @@ impl ModelDeploymentCard {
             user_data: None,
             runtime_config: ModelRuntimeConfig::default(),
             cache_dir: None,
+            checksum: OnceLock::new(),
         })
     }
 }

lib/llm/src/model_type.rs

@@ -74,6 +74,25 @@ impl ModelType {
         result
     }
 
+    /// Decompose the bitflag into it's component units:
+    /// Chat | Completion -> [Chat, Completion]
+    pub fn units(&self) -> Vec<ModelType> {
+        let mut result = Vec::new();
+        if self.supports_chat() {
+            result.push(ModelType::Chat);
+        }
+        if self.supports_completions() {
+            result.push(ModelType::Completions);
+        }
+        if self.supports_embedding() {
+            result.push(ModelType::Embedding);
+        }
+        if self.supports_tensor() {
+            result.push(ModelType::TensorBased);
+        }
+        result
+    }
+
     /// Returns all endpoint types supported by this model type.
     /// This properly handles combinations like Chat | Completions.
     pub fn as_endpoint_types(&self) -> Vec<crate::endpoint_type::EndpointType> {

lib/llm/src/preprocessor.rs

@@ -124,7 +124,7 @@ impl OpenAIPreprocessor {
         formatter: Arc<dyn OAIPromptFormatter>,
         hf_tokenizer: tokenizers::Tokenizer,
     ) -> Result<Arc<Self>> {
-        let mdcsum = mdc.mdcsum();
+        let mdcsum = mdc.mdcsum().to_string();
         let tokenizer = Arc::new(HuggingFaceTokenizer::from_tokenizer(hf_tokenizer));
         let Some(model_info) = mdc.model_info else {
             anyhow::bail!(

lib/llm/tests/http-service.rs

@@ -4,17 +4,6 @@
 use anyhow::Error;
 use async_stream::stream;
 use dynamo_async_openai::config::OpenAIConfig;
-use dynamo_llm::http::{
-    client::{
-        GenericBYOTClient, HttpClientConfig, HttpRequestContext, NvCustomClient, PureOpenAIClient,
-    },
-    service::{
-        Metrics,
-        error::HttpError,
-        metrics::{Endpoint, RequestType, Status},
-        service_v2::HttpService,
-    },
-};
 use dynamo_llm::protocols::{
     Annotated,
     codec::SseLineCodec,
@@ -24,6 +13,21 @@ use dynamo_llm::protocols::{
         completions::{NvCreateCompletionRequest, NvCreateCompletionResponse},
     },
 };
+use dynamo_llm::{
+    http::{
+        client::{
+            GenericBYOTClient, HttpClientConfig, HttpRequestContext, NvCustomClient,
+            PureOpenAIClient,
+        },
+        service::{
+            Metrics,
+            error::HttpError,
+            metrics::{Endpoint, RequestType, Status},
+            service_v2::HttpService,
+        },
+    },
+    model_card::ModelDeploymentCard,
+};
 use dynamo_runtime::metrics::prometheus_names::{frontend_service, name_prefix};
 use dynamo_runtime::{
     CancellationToken,
@@ -275,15 +279,18 @@ async fn test_http_service() {
 
     let registry = Registry::new();
 
+    // TODO: Shouldn't this test know the card before it registers a model?
+    let card = ModelDeploymentCard::with_name_only("foo");
     let counter = Arc::new(CounterEngine {});
-    let result = manager.add_chat_completions_model("foo", counter);
+    let result = manager.add_chat_completions_model("foo", card.mdcsum(), counter);
     assert!(result.is_ok());
 
     let failure = Arc::new(AlwaysFailEngine {});
-    let result = manager.add_chat_completions_model("bar", failure.clone());
+    let card = ModelDeploymentCard::with_name_only("bar");
+    let result = manager.add_chat_completions_model("bar", card.mdcsum(), failure.clone());
     assert!(result.is_ok());
 
-    let result = manager.add_completions_model("bar", failure);
+    let result = manager.add_completions_model("bar", card.mdcsum(), failure);
     assert!(result.is_ok());
 
     let metrics = state.metrics_clone();
@@ -578,14 +585,16 @@ async fn service_with_engines() -> (HttpService, Arc<CounterEngine>, Arc<AlwaysF
     let counter = Arc::new(CounterEngine {});
     let failure = Arc::new(AlwaysFailEngine {});
 
+    let card = ModelDeploymentCard::with_name_only("foo");
     manager
-        .add_chat_completions_model("foo", counter.clone())
+        .add_chat_completions_model("foo", card.mdcsum(), counter.clone())
         .unwrap();
+    let card = ModelDeploymentCard::with_name_only("bar");
     manager
-        .add_chat_completions_model("bar", failure.clone())
+        .add_chat_completions_model("bar", card.mdcsum(), failure.clone())
         .unwrap();
     manager
-        .add_completions_model("bar", failure.clone())
+        .add_completions_model("bar", card.mdcsum(), failure.clone())
         .unwrap();
 
     (service, counter, failure, port)
@@ -977,9 +986,10 @@ async fn test_client_disconnect_cancellation_unary() {
     wait_for_service_ready(port).await;
 
     // Create a long-running engine (10 seconds)
+    let card = ModelDeploymentCard::with_name_only("slow-model");
     let long_running_engine = Arc::new(LongRunningEngine::new(10_000));
     manager
-        .add_chat_completions_model("slow-model", long_running_engine.clone())
+        .add_chat_completions_model("slow-model", card.mdcsum(), long_running_engine.clone())
         .unwrap();
 
     let client = reqwest::Client::new();
@@ -1068,9 +1078,14 @@ async fn test_client_disconnect_cancellation_streaming() {
     wait_for_service_ready(port).await;
 
     // Create a long-running engine (10 seconds)
+    let card = ModelDeploymentCard::with_name_only("slow-stream-model");
     let long_running_engine = Arc::new(LongRunningEngine::new(10_000));
     manager
-        .add_chat_completions_model("slow-stream-model", long_running_engine.clone())
+        .add_chat_completions_model(
+            "slow-stream-model",
+            card.mdcsum(),
+            long_running_engine.clone(),
+        )
         .unwrap();
 
     let client = reqwest::Client::new();
@@ -1166,9 +1181,10 @@ async fn test_request_id_annotation() {
     wait_for_service_ready(port).await;
 
     // Add a counter engine for this test
+    let card = ModelDeploymentCard::with_name_only("test-model");
     let counter_engine = Arc::new(CounterEngine {});
     manager
-        .add_chat_completions_model("test-model", counter_engine)
+        .add_chat_completions_model("test-model", card.mdcsum(), counter_engine)
         .unwrap();
 
     // Create reqwest client directly

lib/llm/tests/http_metrics.rs

@@ -4,8 +4,8 @@
 use anyhow::Error;
 use async_stream::stream;
 use dynamo_llm::{
-    http::service::metrics::Endpoint,
-    http::service::service_v2::HttpService,
+    http::service::{metrics::Endpoint, service_v2::HttpService},
+    model_card::ModelDeploymentCard,
     protocols::{
         Annotated,
         openai::chat_completions::{
@@ -206,9 +206,10 @@ async fn test_metrics_with_mock_model() {
         let task = tokio::spawn(async move { service.run(token.clone()).await });
 
         // Add mock model engine
+        let card = ModelDeploymentCard::with_name_only("mockmodel");
         let mock_engine = Arc::new(MockModelEngine {});
         manager
-            .add_chat_completions_model("mockmodel", mock_engine)
+            .add_chat_completions_model("mockmodel", card.mdcsum(), mock_engine)
             .unwrap();
 
         // Wait for service to be ready
@@ -293,10 +294,8 @@ async fn test_metrics_with_mock_model() {
 mod integration_tests {
     use super::*;
     use dynamo_llm::{
-        discovery::{MODEL_ROOT_PATH, ModelEntry, ModelWatcher},
-        engines::make_echo_engine,
-        entrypoint::EngineConfig,
-        local_model::LocalModelBuilder,
+        discovery::ModelWatcher, engines::make_echo_engine, entrypoint::EngineConfig,
+        local_model::LocalModelBuilder, model_card,
     };
     use dynamo_runtime::DistributedRuntime;
     use dynamo_runtime::pipeline::RouterMode;
@@ -348,7 +347,7 @@ mod integration_tests {
         // Start watching etcd for model registrations
         if let Some(etcd_client) = distributed_runtime.etcd_client() {
             let models_watcher = etcd_client
-                .kv_get_and_watch_prefix(MODEL_ROOT_PATH)
+                .kv_get_and_watch_prefix(model_card::ROOT_PATH)
                 .await
                 .unwrap();
             let (_prefix, _watcher, receiver) = models_watcher.dissolve();
@@ -364,10 +363,11 @@ mod integration_tests {
             panic!("Expected StaticFull config");
         };
 
+        let card = local_model.card().clone();
         let engine = Arc::new(dynamo_llm::engines::StreamingEngineAdapter::new(engine));
         let manager = service.model_manager();
         manager
-            .add_chat_completions_model(model.service_name(), engine.clone())
+            .add_chat_completions_model(model.service_name(), card.mdcsum(), engine.clone())
             .unwrap();
 
         // Now do the proper MDC registration via LocalModel::attach()
@@ -376,7 +376,7 @@ mod integration_tests {
         let test_component = namespace.component("test-mdc-component").unwrap();
         let test_endpoint = test_component.endpoint("test-mdc-endpoint");
 
-        // This will store the MDC in etcd and create the ModelEntry for discovery
+        // This will store the MDC in etcd for discovery
         local_model
             .attach(
                 &test_endpoint,
@@ -388,8 +388,7 @@ mod integration_tests {
 
         // Manually save the model card and update metrics
         // This simulates what the ModelWatcher polling task would do in production
-        let card = local_model.card().clone();
-        manager.save_model_card("test-mdc-key", card.clone());
+        let _ = manager.save_model_card("test-mdc-key", card.clone());
 
         if let Err(e) = service
             .state()
@@ -500,8 +499,13 @@ mod integration_tests {
         assert!(metrics_body.contains("request_type=\"stream\""));
         assert!(metrics_body.contains("status=\"success\""));
 
+        // etcd lease will ensure we everything is deleted from etcd
+
         // Now test the complete lifecycle: remove the model from etcd
+        // We don't need to cleanup model manager because it's local to this test
 
+        /*
+        // Clean up
         // Remove the model using the cleaner ModelWatcher approach
         if let Some(etcd_client) = distributed_runtime.etcd_client() {
             // Use ModelWatcher to find and remove the model (following ModelWatcher::handle_delete pattern)
@@ -514,10 +518,7 @@ mod integration_tests {
             );
 
             // Get all model entries for our test model
-            let model_entries = watcher
-                .entries_for_model("test-mdc-model", None, true)
-                .await
-                .unwrap();
+            let model_entries = watcher.entries_for_model("test-mdc-model").await.unwrap();
 
             if !model_entries.is_empty() {
                 // For each model entry, we need to find its etcd key and remove it
@@ -566,8 +567,8 @@ mod integration_tests {
                 }
             }
         }
+        */
 
-        // Clean up
         cancel_token.cancel();
         task.await.unwrap().unwrap();
     }

lib/llm/tests/kserve_service.rs

@@ -280,31 +280,32 @@ pub mod kserve_test {
         let failure = Arc::new(AlwaysFailEngine {});
         let long_running = Arc::new(LongRunningEngine::new(1_000));
 
-        manager
-            .add_completions_model("split", split.clone())
-            .unwrap();
         let mut card = ModelDeploymentCard::with_name_only("split");
         card.model_type = ModelType::Completions;
         card.model_input = ModelInput::Text;
-        manager.save_model_card("split", card);
-
         manager
-            .add_chat_completions_model("failure", failure.clone())
-            .unwrap();
-        manager
-            .add_completions_model("failure", failure.clone())
+            .add_completions_model("split", card.mdcsum(), split.clone())
             .unwrap();
+        let _ = manager.save_model_card("split", card.clone());
+
         let mut card = ModelDeploymentCard::with_name_only("failure");
         card.model_type = ModelType::Completions | ModelType::Chat;
         card.model_input = ModelInput::Text;
-        manager.save_model_card("failure", card);
         manager
-            .add_completions_model("long_running", long_running.clone())
+            .add_chat_completions_model("failure", card.mdcsum(), failure.clone())
             .unwrap();
+        manager
+            .add_completions_model("failure", card.mdcsum(), failure.clone())
+            .unwrap();
+        let _ = manager.save_model_card("failure", card);
+
         let mut card = ModelDeploymentCard::with_name_only("long_running");
         card.model_type = ModelType::Completions;
         card.model_input = ModelInput::Text;
-        manager.save_model_card("long_running", card);
+        manager
+            .add_completions_model("long_running", card.mdcsum(), long_running.clone())
+            .unwrap();
+        let _ = manager.save_model_card("long_running", card);
 
         (service, split, failure, long_running)
     }
@@ -1130,11 +1131,16 @@ pub mod kserve_test {
         text_input: inference::model_infer_request::InferInputTensor,
     ) {
         // add tensor model
+
+        // Failure, model registered as Tensor but does not provide model config (in runtime config)
+        let mut card = ModelDeploymentCard::with_name_only("tensor");
+        card.model_type = ModelType::TensorBased;
+        card.model_input = ModelInput::Tensor;
         let tensor = Arc::new(TensorEngine {});
         service_with_engines
             .0
             .model_manager()
-            .add_tensor_model("tensor", tensor.clone())
+            .add_tensor_model("tensor", card.mdcsum(), tensor.clone())
             .unwrap();
 
         // start server
@@ -1147,11 +1153,7 @@ pub mod kserve_test {
             version: "".into(),
         });
 
-        // Failure, model registered as Tensor but does not provide model config (in runtime config)
-        let mut card = ModelDeploymentCard::with_name_only("tensor");
-        card.model_type = ModelType::TensorBased;
-        card.model_input = ModelInput::Tensor;
-        service_with_engines
+        let _ = service_with_engines
             .0
             .model_manager()
             .save_model_card("key", card);
@@ -1217,7 +1219,7 @@ pub mod kserve_test {
             }),
             ..Default::default()
         };
-        service_with_engines
+        let _ = service_with_engines
             .0
             .model_manager()
             .save_model_card("key", card);