[Bugfix][Helm] Use vllm:prefix_cache_* in the GPU KV Cache Hit Rate panel - #1147

The "GPU KV Cache Hit Rate" panel in helm/dashboards/vllm-dashboard.json still queries the old prefix-cache counters:

vllm:gpu_prefix_cache_hits_total{endpoint="service-port"} / vllm:gpu_prefix_cache_queries_total

vLLM deprecated these counters in favor of vllm:prefix_cache_hits / vllm:prefix_cache_queries, hid them by default in v0.11.0 (only shown with --show-hidden-metrics-for-version), and removed them in v0.12.0 (vllm/v1/metrics/loggers.py). So with any current engine the panel shows No data. That includes the chart's default engine image, lmcache/vllm-openai:latest, which ships vLLM 0.29.0.

This PR switches the expression to the current counters:

vllm:prefix_cache_hits_total{endpoint="service-port"} / vllm:prefix_cache_queries_total

vLLM has exported vllm:prefix_cache_{hits,queries} since v0.9.2. #989 already moved this dashboard to the other renamed metrics (kv_cache_usage_perc, request_time_per_output_token_seconds) but missed this panel. The router-side parser has the same problem; that is tracked separately in #1140 / #1146. This PR only touches the Grafana dashboard.

How I checked

• vLLM v0.31.0 (vllm/vllm-openai, the current release) on a single 8 GB consumer GPU. I used a tiny randomly initialised Qwen2-style model (--load-format dummy --skip-tokenizer-init), sent 5 completions that share a 160-token prefix, then scraped /metrics:

  $ curl -s localhost:8000/metrics | grep -c gpu_prefix_cache
  0
  $ curl -s localhost:8000/metrics | grep '^vllm:prefix_cache_.*_total'
  vllm:prefix_cache_queries_total{engine="0",model_name="tiny"} 805.0
  vllm:prefix_cache_hits_total{engine="0",model_name="tiny"} 640.0
  

  The old expression matches no series. The new one gives 640 / 805 = 0.795, which is what the panel should show.

• lmcache/vllm-openai:latest (vLLM 0.29.0, LMCache 0.5.5): vllm/v1/metrics/loggers.py in the image defines vllm:prefix_cache_hits / vllm:prefix_cache_queries and has no vllm:gpu_prefix_cache* name.

• I checked that the JSON still parses. I did not import the dashboard into a live Grafana; the change is a metric-name swap inside one PromQL expression.

I have not run pre-commit locally for this one-line change; I will fix anything CI flags.

• Make sure the code changes pass the pre-commit checks.
• Sign-off your commit by using -s when doing git commit
• Try to classify PRs for easy understanding of the type of changes, such as [Bugfix], [Feat], and [CI].

This change was drafted with AI assistance and reviewed by me.

🤖 Generated with Claude Code