{
  "id": "tiers/gemma4-128gb",
  "name": "Gemma 4 — 128 GB tier (unified memory APU)",
  "description": "Gemma 4 MoE runtime tier for 128 GB unified memory systems (e.g., Ryzen AI MAX+ 395). Generous expert cache, contiguous KV, extended context. Budget leaves headroom for system and replay-prefill intermediates.",
  "intent": "investigate",
  "compatibleIntents": ["investigate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-04-04T00:00:00Z",
  "extends": "default",
  "runtime": {
    "shared": {
      "bufferPool": {
        "budget": {
          "maxTotalBytes": 107374182400,
          "highWatermarkRatio": 0.9,
          "emergencyTrimTargetRatio": 0.75,
          "hardFailOnBudgetExceeded": true
        }
      }
    },
    "loading": {
      "expertCache": {
        "defaultSizeBytes": 8589934592,
        "maxBufferPercentage": 0.25,
        "evictionHighWatermark": 0.9,
        "emergencyTrimToRatio": 0.75
      },
      "prefetch": {
        "enabled": true,
        "layersAhead": 2,
        "maxShards": 16
      },
      "memoryManagement": {
        "flushIntervalLayers": 4,
        "flushThresholdBytes": 536870912,
        "budget": {
          "enabled": true,
          "maxResidentBytes": null,
          "systemMemoryFraction": 0.5,
          "reserveBytes": 4294967296,
          "minimumBudgetBytes": 2147483648
        }
      }
    },
    "inference": {
      "kvcache": {
        "layout": "contiguous",
        "maxSeqLen": 16384,
        "kvDtype": "f16",
        "pageSize": 256,
        "tiering": {
          "mode": "off"
        }
      },
      "moe": {
        "routing": {
          "routerDtype": "f32"
        },
        "cache": {
          "dequantCacheMaxEntries": 128
        }
      },
      "compute": {
        "activationDtype": "f32"
      },
      "session": {
        "kvcache": {
          "layout": "contiguous",
          "maxSeqLen": 16384,
          "kvDtype": "f16",
          "pageSize": 256,
          "tiering": {
            "mode": "off"
          }
        }
      }
    }
  }
}