fix(metrics): skip output token histogram for zero-output requests (#8085)

Signed-off-by: tmontfort <tmontfort@nvidia.com>
Signed-off-by: Thomas Montfort <tmontfort@nvidia.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

lib/llm/src/http/service/metrics.rs

@@ -1451,11 +1451,17 @@ impl Drop for ResponseMetricCollector {
                 .observe(avg_detokenize_latency_ms);
         }
 
-        // Publish final OSL when the collector is dropped
-        self.metrics
-            .output_sequence_length
-            .with_label_values(&[&self.model])
-            .observe(self.osl as f64);
+        // Publish final OSL when the collector is dropped, but only for
+        // requests that actually produced output tokens. Recording zero for
+        // failed/cancelled requests would corrupt histogram averages (sum/count)
+        // and percentiles. Failures are already tracked by requests_total with
+        // status and error_type labels.
+        if self.osl > 0 {
+            self.metrics
+                .output_sequence_length
+                .with_label_values(&[&self.model])
+                .observe(self.osl as f64);
+        }
 
         // Record request summary on the enclosing span.
         // InflightGuard::Drop and on_response logs will inherit these.

lib/llm/src/kv_router/push_router.rs

@@ -237,9 +237,15 @@ impl RequestGuard {
                     .observe(latency);
             }
         }
-        self.request_metrics
-            .output_sequence_tokens
-            .observe(self.cumulative_osl as f64);
+        // Only record output sequence length for requests that actually
+        // produced output tokens. Recording zero for failed/cancelled requests
+        // would corrupt histogram averages (sum/count) and percentiles.
+        // Failures are already tracked by requests_total.
+        if self.cumulative_osl > 0 {
+            self.request_metrics
+                .output_sequence_tokens
+                .observe(self.cumulative_osl as f64);
+        }
         self.request_metrics.requests_total.inc();
     }
 }

lib/llm/tests/http_metrics.rs

@@ -6,6 +6,7 @@ use async_stream::stream;
 use dynamo_llm::{
     http::service::{metrics::Endpoint, service_v2::HttpService},
     model_card::ModelDeploymentCard,
+    preprocessor::LLMMetricAnnotation,
     protocols::{
         Annotated,
         openai::chat_completions::{
@@ -50,10 +51,32 @@ impl
             // Simulate some processing time
             tokio::time::sleep(std::time::Duration::from_millis(10)).await;
 
-            // Generate 5 response chunks
+            // Generate 5 response chunks with LLMMetricAnnotation so that
+            // output_sequence_tokens is properly recorded (the histogram only
+            // records when osl > 0, which requires the annotation to be present).
             for i in 0..5 {
                 let output = generator.create_choice(i, Some(format!("Mock response {i}")), None, None, None);
-                yield Annotated::from_data(output);
+                let mut annotated = Annotated::from_data(output);
+                let metrics = LLMMetricAnnotation {
+                    input_tokens: 5,
+                    output_tokens: (i + 1) as usize,
+                    chunk_tokens: 1,
+                    cached_tokens: None,
+                    prefill_worker_id: None,
+                    prefill_dp_rank: None,
+                    prefill_worker_type: None,
+                    decode_worker_id: None,
+                    decode_dp_rank: None,
+                    decode_worker_type: None,
+                    tokenize_latency: None,
+                    detokenize_total_latency: None,
+                    detokenize_count: None,
+                };
+                if let Ok(ann) = metrics.to_annotation::<NvCreateChatCompletionStreamResponse>() {
+                    annotated.event = ann.event;
+                    annotated.comment = ann.comment;
+                }
+                yield annotated;
             }
         };