{
  "id": "profiles/qwen-3-5-2b-metal-throughput",
  "name": "qwen-3-5-2b-metal-throughput",
  "description": "Qwen 3.5 2B browser/Metal throughput profile. Uses 16-lane Q4_K decode projections, fused gate/up, and exact linear-attention fusions while preserving the manifest f32 compute lane.",
  "intent": "calibrate",
  "compatibleIntents": ["calibrate"],
  "stability": "experimental",
  "owner": "doppler-core",
  "createdAtUtc": "2026-07-07T00:00:00Z",
  "extends": "profiles/throughput",
  "model": "qwen-3-5-2b-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"
              }
            }
          },
          {
            "key": "q4_decode_metal_simd16",
            "kernel": {
              "kernel": "fused_matmul_q4_metal_simd16.wgsl",
              "entry": "main",
              "digest": "sha256:f3570460a332b211885baeabe7ccaf142f7234fff8f55973344157096206f4c4",
              "constants": {
                "WORKGROUP_SIZE": 256,
                "COLS_PER_WG": 16,
                "THREADS_PER_COL_GEMV": 16,
                "USE_FULL_BLOCK_FAST_PATH": true
              }
            }
          }
        ],
        "set": [
          {
            "section": "decode",
            "op": "attention",
            "kernelKey": "attn_decode_head256"
          },
          {
            "section": "decode",
            "op": "q_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "k_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "v_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "o_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "gate_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "up_proj",
            "kernelKey": "q4_decode_metal_simd16"
          },
          {
            "section": "decode",
            "op": "down_proj",
            "kernelKey": "q4_decode_metal_simd16"
          }
        ]
      },
      "batching": {
        "batchSize": 8,
        "readbackInterval": 8,
        "stopCheckMode": "batch",
        "readbackMode": "sequential",
        "ringTokens": 1,
        "ringStop": 1,
        "ringStaging": 1
      },
      "session": {
        "usePostFfnNextInputRMSNormPairFusion": true,
        "fusedFfnQ4K": {
          "decode": {
            "variant": "q4k_metal_simd16",
            "pipelineConstants": {
              "WORKGROUP_SIZE": 256,
              "COLS_PER_WG": 16,
              "THREADS_PER_COL": 16,
              "USE_FULL_BLOCK_FAST_PATH": true
            }
          }
        },
        "useGreedyLmHeadArgmaxFusion": true,
        "lmHeadArgmaxQ4K": {
          "useFullBlockFastPath": true,
          "colsPerWorkgroup": 128,
          "threadsPerCol": 2
        },
        "attentionDecodeOnline": {
          "useDirectContiguousKVLayout": true,
          "useOutputGateFusion": true
        },
        "useLinearAttentionABProjectionFusion": true,
        "useLinearAttentionQKVZProjectionFusion": true,
        "useLinearAttentionFusedDecodeCore": true,
        "useFusedQKVSplitQKNormRoPE": true,
        "decodeLoop": {
          "batchSize": 8,
          "stopCheckMode": "batch",
          "readbackInterval": 8,
          "readbackMode": "sequential",
          "submitLatencyThresholdMs": null,
          "maxBatchDecodeTokens": 64,
          "ringTokens": 1,
          "ringStop": 1,
          "ringStaging": 1,
          "disableCommandBatching": false
        }
      }
    }
  }
}