{
  "id": "tiers/gemma4-24gb",
  "name": "Gemma 4 — 24 GB tier",
  "description": "Gemma 4 MoE runtime tier for 24 GB GPU memory. Moderate expert cache, contiguous KV, reduced context length.",
  "intent": "investigate",
  "compatibleIntents": ["investigate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-03-17T00:00:00Z",
  "extends": "default",
  "runtime": {
    "shared": {
      "bufferPool": {
        "budget": {
          "maxTotalBytes": 21474836480,
          "highWatermarkRatio": 0.9,
          "emergencyTrimTargetRatio": 0.75,
          "hardFailOnBudgetExceeded": true
        }
      }
    },
    "loading": {
      "expertCache": {
        "defaultSizeBytes": 3221225472,
        "maxBufferPercentage": 0.2,
        "evictionHighWatermark": 0.85,
        "emergencyTrimToRatio": 0.7
      },
      "prefetch": {
        "enabled": true,
        "layersAhead": 1,
        "maxShards": 8
      },
      "memoryManagement": {
        "flushIntervalLayers": 2,
        "flushThresholdBytes": 268435456,
        "budget": {
          "enabled": true,
          "maxResidentBytes": null,
          "systemMemoryFraction": 0.5,
          "reserveBytes": 4294967296,
          "minimumBudgetBytes": 2147483648
        }
      }
    },
    "inference": {
      "kvcache": {
        "layout": "contiguous",
        "maxSeqLen": 4096,
        "kvDtype": "f16",
        "pageSize": 256,
        "tiering": {
          "mode": "off"
        }
      },
      "moe": {
        "routing": {
          "routerDtype": "f32"
        },
        "cache": {
          "dequantCacheMaxEntries": 64
        }
      },
      "compute": {
        "activationDtype": "f32"
      },
      "session": {
        "kvcache": {
          "layout": "contiguous",
          "maxSeqLen": 4096,
          "kvDtype": "f16",
          "pageSize": 256,
          "tiering": {
            "mode": "off"
          }
        }
      }
    }
  }
}