kvbm: offload+onboard appear to be working correctly

Signed-off-by: Ryan Olson <[email protected]>

Author: ryanolson

.sandbox/launch_vllm_with_connector.sh

@@ -43,6 +43,7 @@ source "$VENV_PATH/bin/activate"
 
 # Ensure our kvbm package is in the path
 export PYTHONPATH="$REPO_ROOT/lib/bindings/kvbm/python:$PYTHONPATH"
+export VLLM_SERVER_DEV_MODE=1
 
 # =============================================================================
 # Rust Logging Configuration

.sandbox/rebuild.sh

@@ -45,7 +45,7 @@ echo "Build mode: dev (default)"
 echo ""
 
 # Build with maturin develop
-maturin develop
+maturin develop --release
 
 echo ""
 echo "✅ Build complete!"

.sandbox/test_cmpl_1.sh

@@ -8,6 +8,6 @@ curl -X POST http://127.0.0.1:8000/v1/completions \
   -d '{
     "model": "gpt2",
     "prompt": "what is a dynamo? and how did it signify the start of the industrial revolution?",
-    "max_tokens": 16,
+    "max_tokens": 32,
     "temperature": 0
   }'

lib/bindings/kvbm/Cargo.lock

@@ -177,7 +177,7 @@ def clear_connector_metadata(self) -> None:
         """
         Clear connector metadata - no-op.
         """
-        pass
+        self.worker.clear_connector_metadata()
 
     def start_load_kv(self, forward_context: "ForwardContext", **kwargs) -> None:
         """
@@ -222,6 +222,7 @@ def get_finished(
         Returns:
             (None, None): No finished sends/receives
         """
+        print(f"SchedulerConnectorWorker.get_finished called with {len(finished_req_ids)} finished requests")
         return self.worker.get_finished()
 
     def get_block_ids_with_load_errors(self) -> set[int]:

lib/bindings/kvbm/python/kvbm/v2/vllm/schedulers/worker.py

@@ -97,7 +97,7 @@ impl Default for PresenceLfuFilterConfig {
 pub struct TierOffloadConfig {
     /// Ordered list of policies to apply (implicit AND).
     ///
-    /// If empty, defaults to pass-all behavior.
+    /// If empty, tier-specific defaults are applied by the engine.
     /// Policies are evaluated in order; a block must pass all to be transferred.
     #[serde(default)]
     pub policies: Vec<PolicyType>,
@@ -140,6 +140,7 @@ mod tests {
     #[test]
     fn test_default_config() {
         let config = OffloadConfig::default();
+        // Empty policies - engine applies tier-specific defaults
         assert!(config.g1_to_g2.policies.is_empty());
         assert!(config.g2_to_g3.policies.is_empty());
         assert_eq!(config.g2_to_g3.presence_lfu.min_lfu_count, 8);

lib/kvbm-config/src/offload.rs

@@ -314,6 +314,14 @@ impl InstanceLeader {
         &self.registry
     }
 
+    /// Get a reference to the Nova instance.
+    ///
+    /// This provides access to the Nova distributed system for features
+    /// like event coordination and cross-instance communication.
+    pub fn nova(&self) -> &Arc<Nova> {
+        &self.nova
+    }
+
     /// Get the tokio runtime handle from Nova.
     ///
     /// This handle should be used for spawning background tasks that need to