{
  "id": "profiles/qwen-3-5-0-8b-metal-throughput",
  "name": "qwen-3-5-0-8b-metal-throughput",
  "description": "Qwen 3.5 0.8B browser/Metal throughput profile. Keeps the shared Qwen Q4K execution graph and uses overlapped decode readback on Apple Metal so submit/readback-bound decode can reuse ring staging without changing Vulkan cadence.",
  "intent": "calibrate",
  "compatibleIntents": ["calibrate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-07-08T00:00:00Z",
  "extends": "profiles/throughput",
  "model": "qwen-3-5-0-8b-q4k-ehaf16",
  "runtime": {
    "inference": {
      "executionPatch": {
        "addKernels": [
          {
            "key": "attn_decode_head256",
            "kernel": {
              "kernel": "attention_decode_online_head256_f16kv_output_gate.wgsl",
              "entry": "main",
              "digest": "sha256:275a1f7cfdd40ab95c467bb4744a1509dd6eadb39ec8fdd0d1e1ba80eb34f9c5",
              "precision": {
                "kvDtype": "f16"
              }
            }
          }
        ],
        "set": [
          {
            "section": "decode",
            "op": "attention",
            "kernelKey": "attn_decode_head256"
          }
        ]
      },
      "batching": {
        "batchSize": 8,
        "readbackInterval": 8,
        "stopCheckMode": "batch",
        "readbackMode": "overlapped",
        "ringTokens": 2,
        "ringStop": 1,
        "ringStaging": 2
      },
      "session": {
        "usePostFfnNextInputRMSNormPairFusion": true,
        "useGreedyLmHeadArgmaxFusion": true,
        "lmHeadArgmaxQ4K": {
          "useFullBlockFastPath": true,
          "colsPerWorkgroup": 128,
          "threadsPerCol": 2
        },
        "attentionDecodeOnline": {
          "useDirectContiguousKVLayout": true,
          "useOutputGateFusion": true
        },
        "useFusedQKVSplitQKNormRoPE": true,
        "useLinearAttentionABProjectionFusion": true,
        "useLinearAttentionQKVZProjectionFusion": false,
        "useLinearAttentionFusedDecodeCore": true,
        "decodeLoop": {
          "batchSize": 8,
          "stopCheckMode": "batch",
          "readbackInterval": 8,
          "readbackMode": "overlapped",
          "submitLatencyThresholdMs": null,
          "maxBatchDecodeTokens": 128,
          "ringTokens": 2,
          "ringStop": 1,
          "ringStaging": 2,
          "disableCommandBatching": false
        }
      },
      "kernelPathPolicy": {
        "mode": "capability-aware",
        "sourceScope": [
          "manifest"
        ],
        "allowSources": [
          "manifest"
        ],
        "onIncompatible": "remap"
      }
    }
  }
}