[Rust Frontend] Improve scheduler stats logging parity (#47435)

Signed-off-by: Bugen Zhao <i@bugenzhao.com>

rust/src/engine-core-client/src/client.rs

@@ -372,6 +372,14 @@ impl EngineCoreClient {
         self.engines.len()
     }
 
+    /// Return the engine-side indices connected to this client.
+    pub fn engine_indices(&self) -> Vec<u32> {
+        self.engines
+            .iter()
+            .map(|engine| engine.engine_id.engine_index().expect("engine id must encode as u16"))
+            .collect()
+    }
+
     /// Return the engine identities of all engines connected to this client.
     pub fn engine_identities(&self) -> Vec<&[u8]> {
         self.engines.iter().map(|engine| &*engine.engine_id).collect()

rust/src/engine-core-client/src/metrics.rs

@@ -84,6 +84,10 @@ pub(crate) fn record_scheduler_stats(
             .spec_decode_num_accepted_tokens
             .get_or_create(&labels)
             .inc_by(spec_decoding_stats.num_accepted_tokens);
+        metrics.log_stats.get_or_create(&labels).observe_spec_decode(
+            spec_decoding_stats.num_drafts,
+            &spec_decoding_stats.num_accepted_tokens_per_pos,
+        );
 
         for (position, accepted_tokens) in
             spec_decoding_stats.num_accepted_tokens_per_pos.iter().copied().enumerate()
@@ -119,6 +123,15 @@ pub(crate) fn record_scheduler_stats(
             .inc_by(perf_stats.num_write_bytes_per_gpu);
     }
 
+    if let Some(cudagraph_stats) = &stats.cudagraph_stats {
+        metrics.log_stats.get_or_create(&labels).observe_cudagraph(
+            cudagraph_stats.num_unpadded_tokens,
+            cudagraph_stats.num_padded_tokens,
+            cudagraph_stats.num_paddings,
+            &cudagraph_stats.runtime_mode,
+        );
+    }
+
     // Sampled KV-cache residency histograms.
     if !stats.kv_cache_eviction_events.is_empty() {
         let kv_block_lifetime_seconds = metrics.kv_block_lifetime_seconds.get_or_create(&labels);

rust/src/engine-core-client/src/protocol/stats.rs

@@ -141,7 +141,7 @@ pub struct PerfStats {
 /// Original Python definition:
 /// <https://github.com/vllm-project/vllm/blob/bc2c0c86efb28e77677a3cfb8687e976914a313a/vllm/compilation/cuda_graph.py#L28-L33>
 #[derive(Debug, Clone, Default, PartialEq, Serialize, Deserialize)]
-pub struct CudagraphStat {
+pub struct CudagraphStats {
     /// Number of real tokens in the captured batch before padding.
     pub num_unpadded_tokens: u64,
     /// Number of padded tokens in the captured batch.
@@ -182,7 +182,7 @@ pub struct SchedulerStats {
     /// Connector-specific KV transfer stats, kept opaque for now.
     pub kv_connector_stats: Option<BTreeMap<String, OpaqueValue>>,
     /// CUDA graph runtime stats when graph metrics are enabled.
-    pub cudagraph_stats: Option<CudagraphStat>,
+    pub cudagraph_stats: Option<CudagraphStats>,
     /// Estimated MFU/performance stats, when enabled.
     pub perf_stats: Option<PerfStats>,
 }