{
  "id": "profiles/gemma3-270m-q4k-throughput-overlapped-probe",
  "name": "gemma3-270m-q4k-throughput-overlapped-probe",
  "description": "Gemma 3 270M calibration profile that uses the Q4K 8x8 throughput cadence, overlapped decode readback, fused packed-QKV QK-norm/RoPE decode path, and sandwich RMSNorm-pair fusion for readback-bound browser diagnostics. Used to reduce decode readback submissions and dispatch count without changing shared prompt or sampling semantics.",
  "intent": "calibrate",
  "compatibleIntents": ["calibrate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-06-26T00:00:00Z",
  "extends": "profiles/throughput",
  "model": "gemma-3-270m-it-q4k-ehf16-af32",
  "runtime": {
    "inference": {
      "generation": {
        "disableMultiTokenDecode": false
      },
      "batching": {
        "batchSize": 8,
        "readbackInterval": 8,
        "stopCheckMode": "batch",
        "readbackMode": "overlapped"
      },
      "session": {
        "prefillChunkSubmitMode": "sync",
        "retainQ4KMaterialization": true,
        "useWideTileQ4KPrefill": true,
        "useSandwichRMSNormPairFusion": true,
        "usePostFfnNextInputRMSNormPairFusion": true,
        "useFusedQKVSplitQKNormRoPE": true,
        "decodeLoop": {
          "batchSize": 8,
          "stopCheckMode": "batch",
          "readbackInterval": 8,
          "readbackMode": "overlapped",
          "ringTokens": 2,
          "ringStop": 1,
          "ringStaging": 2,
          "disableCommandBatching": false
        }
      }
    }
  }
}