{
  "id": "profiles/gemma4-e2b-int4ple-reference-export",
  "name": "gemma4-e2b-int4ple-reference-export",
  "description": "Gemma 4 E2B INT4 PLE reference-export profile for constrained Node WebGPU hosts. Keeps low-memory loading caps while preserving plain contiguous KV cache required by the execution-v1 INT4 PLE path.",
  "intent": "verify",
  "compatibleIntents": ["verify"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-04-25T00:00:00Z",
  "extends": "profiles/low-memory",
  "model": "gemma-4-e2b-it-q4k-ehf16-af32-int4ple",
  "runtime": {
    "shared": {
      "bufferPool": {
        "budget": {
          "maxTotalBytes": 13958643712,
          "highWatermarkRatio": 0.85,
          "emergencyTrimTargetRatio": 0.7,
          "hardFailOnBudgetExceeded": true
        }
      }
    },
    "loading": {
      "memoryManagement": {
        "flushIntervalLayers": 1,
        "flushThresholdBytes": 134217728,
        "budget": {
          "enabled": true,
          "maxResidentBytes": null,
          "systemMemoryFraction": 0.5,
          "reserveBytes": 4294967296,
          "minimumBudgetBytes": 2147483648
        }
      }
    },
    "inference": {
      "largeWeights": {
        "gpuResidentOverrides": []
      },
      "kvcache": {
        "maxSeqLen": 2048,
        "layout": "contiguous",
        "kvDtype": "f16",
        "pageSize": 128,
        "tiering": {
          "mode": "off"
        }
      },
      "session": {
        "kvcache": {
          "maxSeqLen": 2048,
          "layout": "contiguous",
          "kvDtype": "f16",
          "pageSize": 128,
          "tiering": {
            "mode": "off"
          }
        }
      }
    }
  }
}