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

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

lib/runtime/src/storage/key_value_store/etcd.rs

@@ -149,7 +149,7 @@ impl EtcdBucket {
         tracing::trace!("etcd create: {k}");
 
         // Use atomic transaction to check and create in one operation
-        let put_options = PutOptions::new();
+        let put_options = PutOptions::new().with_lease(self.client.primary_lease().id());
 
         // Build transaction that creates key only if it doesn't exist
         let txn = Txn::new()
@@ -217,9 +217,12 @@ impl EtcdBucket {
             // So we do too in etcd.
         }
 
+        let put_options = PutOptions::new()
+            .with_lease(self.client.primary_lease().id())
+            .with_prev_key();
         let mut put_resp = self
             .client
-            .kv_put_with_options(k, value, Some(PutOptions::new().with_prev_key()))
+            .kv_put_with_options(k, value, Some(put_options))
             .await
             .map_err(|e| StorageError::EtcdError(e.to_string()))?;
         Ok(match put_resp.take_prev_key() {

lib/runtime/src/utils/typed_prefix_watcher.rs

@@ -67,12 +67,12 @@ where
 ///
 /// # Example
 /// ```ignore
-/// // Watch for ModelEntry objects and extract runtime_config field
+/// // Watch for ModelDeploymentCard objects and extract runtime_config field
 /// let watcher = watch_prefix_with_extraction(
 ///     etcd_client,
-///     "models/",
+///     "mdc/",
 ///     |kv| Some(kv.lease()),  // Use lease_id as key
-///     |entry: ModelEntry| entry.runtime_config,  // Extract runtime_config field
+///     |card: ModelDeploymentCard| card.runtime_config,  // Extract runtime_config field
 ///     cancellation_token,
 /// ).await?;
 /// ```

lib/runtime/src/utils/worker_monitor.rs

@@ -16,7 +16,6 @@ use tokio_stream::StreamExt;
 
 // Constants for monitoring configuration
 const KV_METRICS_SUBJECT: &str = "kv_metrics";
-const MODEL_ROOT_PATH: &str = "models";
 
 // Internal structs for deserializing metrics events
 #[derive(serde::Deserialize)]
@@ -35,11 +34,6 @@ struct KvStats {
     kv_active_blocks: u64,
 }
 
-#[derive(serde::Deserialize)]
-struct ModelEntry {
-    runtime_config: Option<RuntimeConfig>,
-}
-
 #[derive(serde::Deserialize)]
 struct RuntimeConfig {
     total_kv_blocks: Option<u64>,
@@ -95,11 +89,18 @@ impl WorkerMonitor {
             return Ok(());
         };
 
+        // WorkerMonitor is in the wrong crate. It deals with LLM things (KV) so it should be in
+        // dynamo-llm not dynamo-runtime.
+        // That means we cannot use ModelDeploymentCard, so use serde_json::Value for now .
         let runtime_configs_watcher = watch_prefix_with_extraction(
             etcd_client,
-            MODEL_ROOT_PATH,
+            "mdc/", // should be model_card::ROOT_PREFIX but wrong crate
             key_extractors::lease_id,
-            |entry: ModelEntry| entry.runtime_config.and_then(|rc| rc.total_kv_blocks),
+            |card: serde_json::Value| {
+                card.get("runtime_config")
+                    .and_then(|rc| rc.get("total_kv_blocks"))
+                    .and_then(|t_kv| t_kv.as_u64())
+            },
             component.drt().child_token(),
         )
         .await?;