Context

sgl-project/sglang#22726 adds three new Prometheus gauges that expose the raw KV cache pool token counts:

• sglang:kv_available_tokens -- free pool slots
• sglang:kv_evictable_tokens -- radix-cached, reclaimable slots
• sglang:kv_used_tokens -- actively pinned slots

These replace the existing single sglang:token_usage metric (which excludes evictable tokens) with the full breakdown, letting operators derive any ratio in PromQL. For example, physical usage = 1 - (kv_available_tokens / (kv_available_tokens + kv_evictable_tokens + kv_used_tokens)).

What needs to happen

Once SGLang cuts a release containing sgl-project/sglang#22726, update the SGLang Grafana dashboard (deploy/observability/grafana_dashboards/sglang.json) to add panels for:

1. KV Pool Breakdown (stacked area) -- kv_used_tokens, kv_evictable_tokens, kv_available_tokens
2. KV Physical Usage % (timeseries) -- derived via PromQL from the three raw gauges
3. Consider updating the existing "GPU KV Cache Usage" panel to use the new metrics for a more complete picture

Blocked on

• SGLang release containing feat(metrics): expose raw KV cache pool token counts as prometheus gauges sgl-project/sglang#22726