Vllm Observability

vLLM observability

Target audience: operators running production vLLM on H100/H200 fleets, usually containerized, usually on Kubernetes, on-call for latency and throughput SLOs.

Why this matters

nvidia-smi can show a perfectly healthy GPU while TTFT is 11 seconds. Raw throughput in tok/s can be rising while user-visible P99 TTFT is cratering. Every production incident this skill exists to catch shares one structural problem: aggregate numbers and hardware counters lie, and only the vLLM-internal per-request distributions tell the truth.

Two operator-facing outcomes matter:

1. Alerting that wakes the right person for the right reason — TTFT/ITL tail, queue depth, preemption rate, corrupted logits.
2. Diagnosis from /metrics alone — a small number of metric patterns distinguish "out of capacity" from "stuck scheduler" from "hot long-context outlier" without SSH'ing to the pod.

The core diagnostic rule

When something feels slow, read the ratio, not the absolute:

The metric surface in one paragraph

vLLM exposes a Prometheus text-format endpoint at /metrics. All series are prefixed vllm: and carry {model_name, engine} labels. Metrics fall into queue/scheduler state, KV cache pressure, per-request latency histograms (TTFT/ITL/queue/prefill/decode/e2e), throughput counters, and request outcomes (finished_reason=stop|length|abort, plus corrupted_requests for NaN-logit page-worthy events).

Full catalog with types, buckets, labels, and emission file:line anchors in references/metrics-catalog.md. The catalog is V1-first with V0 deltas noted.

Top signals to alert on

Full PromQL with multi-window burn-rate templates, SLO calibration notes, and goodput approximation in references/alerting.md.

Dashboards and stacks

The repo ships three operator-ready Grafana dashboards at examples/observability/:

• prometheus_grafana/grafana.json — 12-panel all-in-one (E2E, TTFT, ITL, KV usage, scheduler, throughput, finish-reason, queue/prefill/decode times, token-length heatmaps)
• dashboards/grafana/performance_statistics.json — 20-panel SRE dashboard (latency P50/P90/P99 over time, TPS streams)
• dashboards/grafana/query_statistics.json — 18-panel product dashboard (per-model volume, token-size distributions)

Plus a working docker-compose.yaml + prometheus.yaml for local trials. Perses YAML equivalents in dashboards/perses/. Pair with DCGM exporter (Grafana dashboard 15117) for hardware-side metrics.

Do not use GPU_UTIL as the saturation signal. It hits 100% under severe starvation. Use DCGM_FI_PROF_SM_OCCUPANCY. Full DCGM pairing catalog and external-dashboard inventory in references/dashboards.md.

Tracing

vllm serve $MODEL \
  --otlp-traces-endpoint=grpc://otel-collector:4317 \
  --collect-detailed-traces=all   # or: model, or: worker — expensive, use per-incident

Without --collect-detailed-traces, spans are emitted but the two most useful per-step metrics (model_forward_time_milliseconds, model_execute_time_milliseconds) are missing. Flag is designed to be enabled during an incident, not as baseline — docs explicitly warn about performance impact.

Protocol defaults to gRPC; HTTP/protobuf via OTEL_EXPORTER_OTLP_TRACES_PROTOCOL=http/protobuf. All OTel packages bundled with vLLM. Full stack choices (Jaeger all-in-one, OTel Collector → Tempo → Grafana, Langfuse), span catalog, and sampling patterns in references/tracing.md.

Critical pitfalls

1. Alerting on averages. sum/count hides P99 tails that are 10–50× the mean. Every latency alert must go through histogram_quantile(0.99, …).

2. Forgetting sum by (le) before histogram_quantile. Without it, per-instance quantiles mix with fleet quantiles — the most common Grafana mistake in the Prometheus world.

3. gpu_cache_usage_perc vs kv_cache_usage_perc. The new name shipped first; PR #24245 (merged 2025-09-16) then hid the old gpu_* counterparts behind --show-hidden-metrics-for-version=X.Y. The attempted revert #25392 was closed without merging (2025-09-23), so the hiding stuck — current main emits only kv_cache_usage_perc by default. Dashboards scraping pre-#24245 tags still see both; greenfield dashboards should use the new name only.

4. num_requests_swapped is deprecated on V1 and always zero. Use num_preemptions_total instead. Many copy-pasted dashboards still reference swap.

5. Multi-pod label collisions. Every pod emits identical {model_name, engine} labels. Without a Prometheus relabel adding pod/replica, counters sum across pods and hide per-replica pathology.

6. Cardinality explosion. Never add request_id or the prompt text as a Prometheus label — that path is deliberately absent. Per-request visibility lives in OTLP traces, not metrics.

7. KEDA threshold too low on num_requests_waiting. Thresholds of 1–2 per replica cause scale thrashing. Production Stack default is 5; OpenShift example is 2. Pair with cooldownPeriod: 360 — GPU pods take ~10 min to reach ready, reactive scaling fails.

8. GPU_UTIL at 100% ≠ busy GPU. The ebpfchirp "11-second TTFT" incident is canonical: util pinned high, SM occupancy was 18%, the scheduler was stalled on prefix-cache head-of-line blocking. Watch SM_OCCUPANCY.

9. Ray Serve deployments don't auto-expose /metrics. RayPrometheusStatLogger must be wired explicitly, or Ray 2.51+ ingests vLLM metrics through Ray's own endpoint (disable with log_engine_metrics: False to avoid double-scraping).

10. --collect-detailed-traces as baseline. 5–10% overhead. Toggle per-incident; leave unset by default.

Full troubleshooting matrix (dashboard-empty, metric-gone-after-upgrade, P99 NaN, histogram buckets miscalibrated for SLO) in references/alerting.md under the "When metrics lie" section.

Verify a deployment can be observed

# Basic reachability
curl -fsS http://<endpoint>/health
curl -fsS http://<endpoint>/metrics | head -30
# Confirm the load-bearing series exist
curl -s http://<endpoint>/metrics | grep -E '^vllm:(kv_cache_usage_perc|num_requests_(waiting|running)|time_to_first_token|request_success|num_preemptions|prefix_cache_(hits|queries))'

${CLAUDE_SKILL_DIR}/scripts/metrics-smoke.sh runs the full smoke check against a deployment: confirms endpoints, greps load-bearing series, warns on deprecated metric names, cross-checks DCGM availability if configured. Output is color-coded pass/warn/fail.

Version notes

Core metric surface re-verified against v0.27.0 (2026-08-11): the emitted vllm:* set in vllm/v1/metrics/loggers.py is unchanged from v0.25.1 — nothing removed, nothing renamed. The churn since then is all in the KV offload namespace, where the legacy vllm:kv_offload_total_* series are deprecated in favour of a direction-split load/store set; see references/metrics-catalog.md § KV connector / offload for the verified names.

• V1 engine is default as of late 2025. V0 metrics hidden unless --show-hidden-metrics-for-version=X.Y.
• Metric rename saga: vllm:gpu_cache_usage_perc → vllm:kv_cache_usage_perc. PR #24245 (merged 2025-09-16) hid the deprecated gpu_* names behind --show-hidden-metrics-for-version; the proposed revert PR #25392 was closed without merging (2025-09-23), so the hiding stuck. Current main emits only kv_cache_usage_perc by default.
• Deprecated on V1: num_requests_swapped, cpu_cache_usage_perc, cpu_prefix_cache_hit_rate, time_per_output_token_seconds (replaced by inter_token_latency_seconds), the model_forward_time_milliseconds / model_execute_time_milliseconds pair (now behind --collect-detailed-traces).
• New in V1: num_requests_waiting_by_reason{reason=capacity|deferred}, engine_sleep_state, prompt_tokens_by_source{source=local_compute|local_cache_hit|external_kv_transfer}, per-position spec-decode acceptance counters.
• v0.24.0 corrected two things you may already have graphed wrong:
  • vllm:cache_config_info gained group-aware kv_cache_size_tokens and kv_cache_max_concurrency (#42206). Stop deriving capacity from num_gpu_blocks * block_size — that product is wrong on hybrid models, where a request occupies blocks in several KV cache groups. The startup log was always right; Prometheus just couldn't agree with it.
  • MFU on MLA models (DeepSeek-V2/V3/R1) was computed with GQA assumptions until MLAAttentionMetrics landed (#39457) — a ~57× KV-bandwidth overestimate on DeepSeek-V3. Discard MFU history from MLA deployments on < v0.24.0.
• v0.24.0 added vllm:tool_call_parser_invocations_total{mode,outcome} (#44448) — the rollout signal for tool-calling regressions. Non-harmony path only.
• v0.25.0 added an opt-in per-request metrics block in Chat/Completions response bodies (#46768) for billing/SLA attribution: time_to_first_token_ms, generation_time_ms, queue_time_ms, mean_itl_ms, tokens_per_second. Double-gated by --enable-per-request-metrics and the include_metrics request field; suppressed when n > 1 or multi-prompt makes single-stream attribution meaningless.

External references

• Metrics design doc (canonical): https://github.com/vllm-project/vllm/blob/main/docs/design/metrics.md
• Metrics source of truth: https://github.com/vllm-project/vllm/tree/main/vllm/v1/metrics
• Example dashboards: https://github.com/vllm-project/vllm/tree/main/examples/observability
• Production metrics docs: https://docs.vllm.ai/en/stable/usage/metrics/
• OTel POC: https://docs.vllm.ai/en/latest/examples/online_serving/opentelemetry/
• Blog — Anatomy of vLLM (defines goodput, scheduler scoring): https://vllm.ai/blog/anatomy-of-vllm
• Blog — Large-Scale Serving (DeepSeek @ 2.2k tok/s/H200): https://blog.vllm.ai/2025/12/17/large-scale-serving.html
• Blog — MorIIO disagg (bimodal-ITL, goodput framing): https://vllm.ai/blog/moriio-kv-connector
• ebpfchirp — 11-Second TTFT on a Healthy Server (canonical incident): https://ebpfchirp.substack.com/p/11-second-time-to-first-token-on
• akrisanov.com — vLLM Metrics in Production (concrete alert set): https://akrisanov.com/vllm-metrics/
• DCGM exporter + Grafana dashboard 15117: https://grafana.com/grafana/dashboards/15117-nvidia-dcgm-exporter/
• Sibling skills: vllm-caching (KV tiering), vllm-benchmarking (bench + output JSON), vllm-configuration (env vars + YAML)