{
  "op": "com.microsoft.MatMulNBits",
  "fixtureArrays": {
    "q4_prefill_tiled_reg_tailk_m32_k33_n4096_input_bT": [16, 50, 84, 118, 152, 186, 220, 254, 135, 119, 102, 85, 68, 51, 34, 17],
    "q4_weight_cycle_b_t": [16, 50, 84, 118, 152, 186, 220, 254, 33, 67, 101, 135, 169, 203, 237, 15],
    "quant_scale_cycle_t": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07, 0.085],
    "mixed_weight_cycle_b_t": [17, 200, 91, 45, 233, 128, 7, 176, 250, 33, 142, 99, 210, 64, 188, 121],
    "q8_zero_bias_gemv_m1_tail_n5_input_bT": [19, 56, 93, 130, 167, 204, 241, 22, 59, 96, 133, 170, 207, 244, 25, 62]
  },
  "cases": [
    {
      "name": "q4_zero_bias_prefill_tile4x4_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "block_size": 32 }
    },
    {
      "name": "q4_zero_bias_prefill_tile4x4_small_m8",
      "provenance": {
        "notes": "Small-M (M=8) q4 prefill with bias and zero points. M<64 excludes prefill_tiled_zero_bias, while the row-guarded prefill_tile4x4_zero_bias route admits M>=2 when N is divisible by 4. This pins the tile4x4 lower-bound contract and its bias/zero-point arithmetic."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_tile4x4_small_m8",
      "provenance": {
        "notes": "q8 companion to q4_zero_bias_prefill_tile4x4_small_m8: the relaxed tile4x4 bias floor (M>=2) also makes prefill_tile4x4_zero_bias selectable for q8 prefill+bias at M in [2,64). Validates the q8 unpack branch + bias epilogue on the tile4x4 path at small M."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "ort_projection_q4_block16_tail_zero_bias",
      "provenance": {
        "source": "onnxruntime/test/contrib_ops/matmul_4bits_test.cc",
        "test": "MatMulNBits.Float32_4b_Accuracy0",
        "notes": "Rank-2 projection of ORT q4 block-size coverage with per-block zero points and bias."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [2, 17],
          "data": { "kind": "fillFloat32", "sinStep": 0.19, "cosStep": 0.37 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [2, 2, 8],
          "data": {
            "kind": "values",
            "values": [16, 50, 84, 118, 152, 186, 220, 254, 33, 0, 0, 0, 0, 0, 0, 0, 255, 238, 221, 204, 187, 170, 153, 136, 240, 0, 0, 0, 0, 0, 0, 0]
          }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [2, 2],
          "data": { "kind": "values", "values": [0.04, 0.08, 0.06, 0.12] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [2, 2],
          "data": { "kind": "values", "values": [7.0, 1.0, 8.0, 0.0] }
        },
        "biasT": { "dtype": "float32", "shape": [2], "data": { "kind": "values", "values": [0.5, -1.0] } }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.00001 } },
      "attrs": { "K": 17, "N": 2, "bits": 4, "block_size": 16 }
    },
    {
      "name": "ort_projection_q4_block32_tail_no_zero",
      "provenance": {
        "source": "onnxruntime/test/contrib_ops/matmul_4bits_test.cc",
        "test": "MatMulNBits.Float32_4b_Accuracy0",
        "notes": "Rank-2 deterministic projection of ORT q4 coverage with blockSize=32 and tail K."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [2, 33],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.75 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [2, 2, 16],
          "data": {
            "kind": "values",
            "values": [16, 50, 84, 118, 152, 186, 220, 254, 16, 50, 84, 118, 152, 186, 220, 254, 135, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 255, 238, 221, 204, 187, 170, 153, 136, 119, 102, 85, 68, 51, 34, 17, 0, 248, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
          }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [2, 2],
          "data": { "kind": "values", "values": [0.03, 0.06, 0.04, 0.07] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [2, 2], "tolerance": 0.00001 } },
      "attrs": { "K": 33, "N": 2, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_block512_schema_legal_cpu_provider_gap_exact",
      "provenance": {
        "source": "onnxruntime/test/contrib_ops/matmul_4bits_test.cc",
        "test": "MatMulNBits.UnsupportedBlockSize_512",
        "notes": "The public schema permits every power-of-two block size of at least 16, and this package implements block_size=512. ORT's installed, pinned, and inspected current CPU providers instead restrict execution to 16, 32, 64, 128, or 256. With A=1, packed q4 codes=9, default zero point=8, and scale=0.25, the exact result is 512 * (9 - 8) * 0.25 = 128."
      },
      "inputs": {
        "aT": { "dtype": "float32", "shape": [1, 512], "data": { "kind": "constant", "value": 1.0 } },
        "bT": { "dtype": "uint8", "shape": [1, 1, 256], "data": { "kind": "constant", "value": 153 } },
        "scalesT": { "dtype": "float32", "shape": [1, 1], "data": { "kind": "constant", "value": 0.25 } }
      },
      "outputs": {
        "yT": { "dtype": "float32", "shape": [1, 1], "tolerance": 0, "data": { "kind": "values", "values": [128.0] } }
      },
      "attrs": { "K": 512, "N": 1, "bits": 4, "block_size": 512 }
    },
    {
      "name": "q4_no_zero_prefill_sgmat_large",
      "inputs": {
        "aT": {
          "shape": [128, 128],
          "dtype": "float32",
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "shape": [64, 4, 16],
          "dtype": "uint8",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "shape": [64, 4],
          "dtype": "float32",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_no_zero_prefill_sgmat_m32_storage_floor",
      "provenance": {
        "notes": "The 32-row subgroup-matrix tile fits the WebGPU minimum 16 KiB workgroup-storage limit."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 32, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_zero_bias_prefill_sgmat_m32_storage_floor",
      "provenance": { "notes": "Zero-point/bias sibling of the 32-row subgroup-matrix storage-floor regression." },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [7.0, 8.0, 6.0, 9.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 32, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_no_zero_prefill_g64_sgmat",
      "provenance": {
        "notes": "Compact blockSize=64 q4 default-zero prefill sibling for model g64 benchmarks. Full tier routes through prefill_sgmat_default_zero; no-MMA/no-subgroups/baseline route through prefill_tiled_default_zero."
      },
      "inputs": {
        "aT": {
          "shape": [64, 128],
          "dtype": "float32",
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "shape": [64, 2, 32],
          "dtype": "uint8",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "shape": [64, 2],
          "dtype": "float32",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [64, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 64 }
    },
    {
      "name": "q8_no_zero_prefill_sgmat_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_sgmat_m32_storage_floor",
      "provenance": {
        "notes": "q8 sibling proving the 32-row subgroup-matrix tile remains available at the 16 KiB storage floor."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.001 } },
      "attrs": { "K": 32, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_sgmat_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_sgmat_m32_storage_floor",
      "provenance": { "notes": "q8 zero-point/bias sibling of the 32-row subgroup-matrix storage-floor regression." },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.001 } },
      "attrs": { "K": 32, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_tile4x4_unaligned",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 64],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [12, 2, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [12, 2],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 12], "tolerance": 0.001 } },
      "attrs": { "K": 64, "N": 12, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_f16_zero_points_prefill_sgmat",
      "inputs": {
        "aT": {
          "dtype": "float16",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float16",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float16",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [7.0, 8.0, 9.0, 6.0] }
        },
        "biasT": { "dtype": "float16", "shape": [64], "data": { "kind": "constant", "value": 0.0 } }
      },
      "outputs": { "yT": { "dtype": "float16", "shape": [128, 64], "tolerance": 0.05 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_no_zero_prefill_sgmat_unaligned_n72",
      "inputs": {
        "aT": {
          "shape": [128, 128],
          "dtype": "float32",
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "shape": [72, 4, 16],
          "dtype": "uint8",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "shape": [72, 4],
          "dtype": "float32",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 72], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 72, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_sgmat_unaligned_n72",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [72, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [72, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 72], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 72, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_sgmat_unaligned_n72_m32",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [72, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [72, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 72], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 72, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_sgmat_unaligned_n72",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [72, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [72, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [72, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [72],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 72], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 72, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_sgmat_unaligned_n65_m33",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [33, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [65, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [65, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [65, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [65],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [33, 65], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 65, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_sgmat_unaligned_n72_m32_bs16",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [72, 8, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [72, 8],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [72, 8],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [72],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 72], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 72, "bits": 8, "block_size": 16 }
    },
    {
      "name": "empty_zero_dim",
      "provenance": {
        "notes": "Zero-length M (rows) axis of A (empty input). With M=0 the N/K-driven packed uint8 B[4,1,16] and scales[4,1] stay normal and Y=[0,N] is empty. K=32,N=4,bits=4,block_size=32."
      },
      "inputs": {
        "aT": { "dtype": "float32", "shape": [0, 32], "data": { "kind": "values", "values": [] } },
        "bT": { "dtype": "uint8", "shape": [4, 1, 16], "data": { "kind": "constant", "value": 17 } },
        "scalesT": { "dtype": "float32", "shape": [4, 1], "data": { "kind": "constant", "value": 0.05 } }
      },
      "outputs": {
        "yT": { "dtype": "float32", "shape": [0, 4], "tolerance": 0, "data": { "kind": "values", "values": [] } }
      },
      "attrs": { "K": 32, "N": 4, "bits": 4, "block_size": 32 }
    },
    {
      "name": "empty_zero_dim_f16",
      "provenance": { "notes": "f16 variant of the empty zero-M-axis case." },
      "inputs": {
        "aT": { "dtype": "float16", "shape": [0, 32], "data": { "kind": "values", "values": [] } },
        "bT": { "dtype": "uint8", "shape": [4, 1, 16], "data": { "kind": "constant", "value": 17 } },
        "scalesT": { "dtype": "float16", "shape": [4, 1], "data": { "kind": "constant", "value": 0.05 } }
      },
      "outputs": {
        "yT": { "dtype": "float16", "shape": [0, 4], "tolerance": 0, "data": { "kind": "values", "values": [] } }
      },
      "attrs": { "K": 32, "N": 4, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_no_zero_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "shape": [512, 128],
          "dtype": "float32",
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "shape": [1024, 4, 16],
          "dtype": "uint8",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "shape": [1024, 4],
          "dtype": "float32",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_f16_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "shape": [512, 128],
          "dtype": "float16",
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "shape": [1024, 4, 16],
          "dtype": "uint8",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "shape": [1024, 4],
          "dtype": "float16",
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float16", "shape": [512, 1024], "tolerance": 0.0001, "relTolerance": 0.001 } },
      "provenance": {
        "notes": "f16 activations on the register-tiled prefill path: this is the only shape class where the shared A/B tiles are staged in f16, so it is the fixture that covers that staging."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_zero_bias_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_f16_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float16",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float16",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float16", "shape": [512, 1024], "tolerance": 0.0001, "relTolerance": 0.001 } },
      "provenance": {
        "notes": "q8 sibling of q4_f16_prefill_reg_m512_n1024. 8-bit codes give the widest dequantized range, so the f16 tile store rounds hardest here."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_no_zero_prefill_tile4x4_partial_row_tile_m6",
      "provenance": {
        "notes": "M=6 is not a multiple of TILE_M=4, so the tile4x4 kernel's second row-tile (row_base=4) has valid rows 4,5 and guarded rows 6,7. Verifies the store_row partial-row-tile guard writes rows 4,5 correctly and does not corrupt/OOB rows 6,7. N=8 (%4==0), K=64, blockSize=32 routes to prefill_tile4x4_default_zero."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [6, 64],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [8, 2, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [8, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [6, 8], "tolerance": 0.0001 } },
      "attrs": { "K": 64, "N": 8, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_gemv_default_zero_m1_tail_n7",
      "provenance": {
        "notes": "M=1 GEMV with N=7 (not a multiple of 4). Exercises the vec4 tail-column store guards (col_base+1/2/3 < N) in matmul-nbits-gemv-q4.wgsl.jinja; existing M==1 tests all use N=4. Routes to gemv_default_zero (prio 20)."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [7, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [7, 1],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 7], "tolerance": 0.00001 } },
      "attrs": { "K": 32, "N": 7, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_gemv_default_zero_m1_odd_k33_n5",
      "provenance": {
        "notes": "K=33 is odd, so the last stored byte of the second block carries one live nibble and one past the end of K. The GEMV walks whole bytes, so this is the only shape that exercises its out-of-range activation guard; every other M==1 case has an even K."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 33],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 2, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 33, "N": 5, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q2_gemv_default_zero_m1_k18_n5",
      "provenance": {
        "notes": "The only two-bit M==1 case: four codes share one stored byte, and K=18 leaves two of the last byte's four codes past the end of K. Covers both the two-bit unpack in the GEMV and its out-of-range activation guard at the widest code-per-byte packing."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 18],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 2, 4],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 18, "N": 5, "bits": 2, "block_size": 16 }
    },
    {
      "name": "q4_gemv_default_zero_m1_n13_ncols8",
      "tunables": { "GEMV_N_COLS": 8 },
      "provenance": {
        "notes": "GEMV_N_COLS=8 with N=13: two workgroups, first fully live, second with a partially live first group and one live column in the second."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [13, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [13, 1],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 13], "tolerance": 0.00001 } },
      "attrs": { "K": 32, "N": 13, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_gemv_default_zero_m1_tail_n7_ncols8",
      "tunables": { "GEMV_N_COLS": 8 },
      "provenance": {
        "notes": "GEMV_N_COLS=8 twin of tail_n7: one workgroup, second column group entirely tail-guarded (cols 4-6 active, 7 out of range)."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [7, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [7, 1],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 7], "tolerance": 0.00001 } },
      "attrs": { "K": 32, "N": 7, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_zero_bias_gemv_m1_tail_n5",
      "provenance": {
        "notes": "M=1 q8 GEMV with N=5 (not a multiple of 4), WITH zero_points and bias. Exercises the vec4 tail-column store guards together with the q8 unpack and bias epilogue in matmul-nbits-gemv-q4.wgsl.jinja. Routes to gemv_zero_bias (prio 20)."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [5],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 16, "N": 5, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q8_zero_only_gemv_m1_tail_n5",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "M=1 q8 GEMV with explicit zero points and no bias. This pins the zero-points-only binding layout and epilogue instead of relying on the combined zero-point-plus-bias case."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 16, "N": 5, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q8_bias_only_gemv_m1_tail_n5",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "M=1 q8 GEMV with bias and the schema-default zero point. This pins the bias-only binding layout and epilogue instead of relying on the combined zero-point-plus-bias case."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [5],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 16, "N": 5, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q8_zero_bias_gemv_m1_tail_n5_ncols8",
      "tunables": { "GEMV_N_COLS": 8 },
      "provenance": {
        "notes": "GEMV_N_COLS=8 twin of tail_n5: q8 unpack + zero_points + bias epilogue with only one live column in the second group."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [1, 16],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [5, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [5, 1],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [5],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [1, 5], "tolerance": 0.00001 } },
      "attrs": { "K": 16, "N": 5, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q8_zero_bias_naive_fallback_tailK_m3_n6",
      "provenance": {
        "notes": "K=17,blockSize=16 gives a tail block; N=6 is below the register-tiled floor and the aligned tile paths are ineligible, while M=3 bypasses GEMV. With zero_points+bias present this routes to the priority-0 zero_bias scalar fallback and covers q8 unpack, per-block zero points, bias, and a partial K block."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [3, 17],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.75 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [6, 2, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [6],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [3, 6], "tolerance": 0.001 } },
      "attrs": { "K": 17, "N": 6, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q4_prefill_tiled_reg_tailk_m32_k33_n4096",
      "provenance": {
        "notes": "Compact tail-block lock for the register-tiled prefill path used by the realistic K=2561 benchmark; the final 31 padded weights must not read past A."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 33],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [4096, 2, 16],
          "data": {
            "kind": "cycle",
            "values": { "$ref": "#/fixtureArrays/q4_prefill_tiled_reg_tailk_m32_k33_n4096_input_bT" }
          }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [4096, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 4096], "tolerance": 0.001 } },
      "attrs": { "K": 33, "N": 4096, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_prefill_tiled_reg_aligned_m32_k32_n4096_at_workgroup_floor",
      "provenance": {
        "source": "authored for route coverage",
        "notes": "Aligned-K twin of the tail-K case at the same M and N. A tail-K shape skips the register tile's workgroup floor entirely, so no fixture exercised that floor; this one sits exactly on it at 64 workgroups."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.17, "cosStep": 0.31, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [4096, 1, 16],
          "data": {
            "kind": "cycle",
            "values": { "$ref": "#/fixtureArrays/q4_prefill_tiled_reg_tailk_m32_k33_n4096_input_bT" }
          }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [4096, 1],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 4096], "tolerance": 0.001 } },
      "attrs": { "K": 32, "N": 4096, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_prefill_tiled_reg_tailk_m32_k33_n512_below_wide_n_floor",
      "provenance": {
        "notes": "Tail-block coverage below the ordinary wide register-tile N floor. K=33 exercises the padded final block, while N=512 keeps the case inside the relaxed band that the larger N=4096 case cannot distinguish."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 33],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [512, 2, 16],
          "data": {
            "kind": "cycle",
            "values": { "$ref": "#/fixtureArrays/q4_prefill_tiled_reg_tailk_m32_k33_n4096_input_bT" }
          }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [512, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 512], "tolerance": 0.001 } },
      "attrs": { "K": 33, "N": 512, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_f16_prefill_reg_tailk_m32_k33_n4096",
      "provenance": {
        "notes": "f16 activations through the register-tiled general (non-block-aligned) loader and the M=32 portable geometry, plus the padded tail K."
      },
      "inputs": {
        "aT": {
          "dtype": "float16",
          "shape": [32, 33],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [4096, 2, 16],
          "data": {
            "kind": "cycle",
            "values": { "$ref": "#/fixtureArrays/q4_prefill_tiled_reg_tailk_m32_k33_n4096_input_bT" }
          }
        },
        "scalesT": {
          "dtype": "float16",
          "shape": [4096, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float16", "shape": [32, 4096], "tolerance": 0.0001, "relTolerance": 0.001 } },
      "attrs": { "K": 33, "N": 4096, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_no_zero_prefill_odd_n_fallback",
      "provenance": {
        "notes": "M>1 q8 prefill with N=17 and no zero_points/bias. Odd N excludes subgroup-matrix execution; the portable tile4x4 tail guards handle the final output column used by the odd-column benchmark guardrail."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 64],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [17, 2, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [17, 2],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 17], "tolerance": 0.001 } },
      "attrs": { "K": 64, "N": 17, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_zero_only_prefill_sgmat_m32_storage_floor",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [7.0, 8.0, 6.0, 9.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 32, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_bias_only_prefill_sgmat_m32_storage_floor",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 32, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_zero_only_prefill_sgmat_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_bias_only_prefill_sgmat_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_zero_only_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.0001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_bias_only_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.0001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_zero_only_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_bias_only_prefill_reg_m512_n1024",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 1024, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_zero_only_prefill_tile4x4_small_m8",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_bias_only_prefill_tile4x4_small_m8",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.0001 } },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q8_zero_only_prefill_tile4x4_small_m8",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0, 126.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_bias_only_prefill_tile4x4_small_m8",
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/mixed_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q4_zero_only_prefill_tile4x4_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [6.0, 7.0, 8.0, 9.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.0001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_bias_only_prefill_tile4x4_large",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q4_weight_cycle_b_t" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/quant_scale_cycle_t" } }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.0001 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/matmul_nbits.cc",
        "test": "independent has_zero_points/has_bias epilogues",
        "notes": "zero_points-only and bias-only inputs must stay on the fast kernels instead of falling to the naive per-element path."
      },
      "attrs": { "K": 128, "N": 64, "bits": 4, "block_size": 32 }
    },
    {
      "name": "q4_dp4a_prefill_al4_m64_k256_n64_b32",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [64, 256],
          "data": { "kind": "fillFloat32", "sinStep": 0.013, "cosStep": 0.021 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 8, 16],
          "data": {
            "kind": "cycle",
            "values": [37, 201, 116, 88, 250, 13, 164, 95, 142, 7, 233, 60, 178, 121, 44, 209, 86, 155]
          }
        },
        "scalesT": { "dtype": "float32", "shape": [64, 8], "data": { "kind": "linspace", "start": 0.01, "end": 0.05 } }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [64, 64], "tolerance": 0.03 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/dp4a_matmul_nbits.cc",
        "test": "accuracy_level=4 int8 activation quantization (DP4AMatMulQuantize + DP4AMatMulNBits)",
        "notes": "accuracy_level 4 permits int8 activation math: A is quantized per 128-element block (scale = max_abs / 127, values pack4x8snorm-rounded) and multiplied against the 4-bit weights with packed integer dots. The reference stays float, so tolerances absorb the permitted quantization error."
      },
      "attrs": { "K": 256, "N": 64, "bits": 4, "block_size": 32, "accuracy_level": 4 }
    },
    {
      "name": "q4_dp4a_prefill_al4_m48_k128_n16_b64",
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [48, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.017, "cosStep": 0.027 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [16, 2, 32],
          "data": { "kind": "cycle", "values": [37, 201, 116, 88, 250, 13, 164, 95, 142, 7, 233] }
        },
        "scalesT": { "dtype": "float32", "shape": [16, 2], "data": { "kind": "linspace", "start": 0.02, "end": 0.06 } }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [48, 16], "tolerance": 0.03 } },
      "provenance": {
        "source": "onnxruntime/contrib_ops/webgpu/quantization/dp4a_matmul_nbits.cc",
        "test": "accuracy_level=4 int8 activation quantization (DP4AMatMulQuantize + DP4AMatMulNBits)",
        "notes": "accuracy_level 4 permits int8 activation math: A is quantized per 128-element block (scale = max_abs / 127, values pack4x8snorm-rounded) and multiplied against the 4-bit weights with packed integer dots. The reference stays float, so tolerances absorb the permitted quantization error."
      },
      "attrs": { "K": 128, "N": 16, "bits": 4, "block_size": 64, "accuracy_level": 4 }
    },
    {
      "name": "q2_zero_bias_prefill_tile4x4_small_m8",
      "provenance": {
        "notes": "2-bit weights pack four codes per stored byte, low codes first along K, with a default zero point of 2. Small-M prefill on the 4x4 output tile."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [8, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 8],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [1.0, 2.0, 3.0, 0.0, 2.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [8, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 2, "block_size": 32 }
    },
    {
      "name": "q2_no_zero_prefill_sgmat_large",
      "provenance": {
        "notes": "2-bit weights pack four codes per stored byte, low codes first along K, with a default zero point of 2. Prefill wide enough for the subgroup-matrix tile."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [128, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 8],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [128, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 2, "block_size": 32 }
    },
    {
      "name": "q2_zero_bias_prefill_sgmat_m32_storage_floor",
      "provenance": {
        "notes": "2-bit weights pack four codes per stored byte, low codes first along K, with a default zero point of 2. Subgroup-matrix tile at the M=32 shared-storage floor."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [32, 32],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 1, 8],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 1],
          "data": { "kind": "cycle", "values": [1.0, 2.0, 3.0, 0.0, 2.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [32, 64], "tolerance": 0.001 } },
      "attrs": { "K": 32, "N": 64, "bits": 2, "block_size": 32 }
    },
    {
      "name": "q2_no_zero_prefill_reg_m512_n1024",
      "provenance": {
        "notes": "2-bit weights pack four codes per stored byte, low codes first along K, with a default zero point of 2. Register-tile prefill, whose loader unrolls by K and so needs its own unpack."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 8],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 2, "block_size": 32 }
    },
    {
      "name": "q2_zero_bias_prefill_reg_m512_n1024",
      "provenance": {
        "notes": "2-bit weights pack four codes per stored byte, low codes first along K, with a default zero point of 2. Register-tile prefill with zero points and bias."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [512, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [1024, 4, 8],
          "data": { "kind": "cycle", "values": [27, 180, 75, 226, 33, 150, 201, 108, 57, 246, 129, 66, 195, 84, 237] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [1024, 4],
          "data": { "kind": "cycle", "values": [1.0, 2.0, 3.0, 0.0, 2.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [1024],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [512, 1024], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 1024, "bits": 2, "block_size": 32 }
    },
    {
      "name": "q8_rank3_prefill_tile4x4",
      "provenance": { "notes": "Rank-3 A folding to eight rows, wide enough for the prefill output tile." },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [2, 4, 128],
          "data": { "kind": "fillFloat32", "sinStep": 0.11, "cosStep": 0.23, "scale": 0.35 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [64, 4, 32],
          "data": { "kind": "cycle", "values": [17, 200, 91, 45, 233, 128, 7, 176, 250, 33, 142, 99, 210, 64, 188] }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075, 0.07] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [64, 4],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0, 144.0] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [64],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.1 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [2, 4, 64], "tolerance": 0.001 } },
      "attrs": { "K": 128, "N": 64, "bits": 8, "block_size": 32 }
    },
    {
      "name": "q8_zero_only_naive_fallback_tailk_m3_n6",
      "provenance": {
        "notes": "K=17 leaves a partial final block and M=3 bypasses GEMV. With explicit zero points but no bias, the aligned prefill paths are ineligible and the zero-only scalar fallback handles the tail block."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [3, 17],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.75 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [6, 2, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075] }
        },
        "zeroPointsT": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": { "kind": "cycle", "values": [120.0, 132.0, 128.0, 140.0, 135.0, 118.0] }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [3, 6], "tolerance": 0.001 } },
      "attrs": { "K": 17, "N": 6, "bits": 8, "block_size": 16 }
    },
    {
      "name": "q8_bias_only_naive_fallback_tailk_m3_n6",
      "provenance": {
        "notes": "K=17 leaves a partial final block and M=3 bypasses GEMV. With bias and the schema-default zero point, the aligned prefill paths are ineligible and the bias-only scalar fallback handles the tail block."
      },
      "inputs": {
        "aT": {
          "dtype": "float32",
          "shape": [3, 17],
          "data": { "kind": "fillFloat32", "sinStep": 0.13, "cosStep": 0.29, "scale": 0.75 }
        },
        "bT": {
          "dtype": "uint8",
          "shape": [6, 2, 16],
          "data": { "kind": "cycle", "values": { "$ref": "#/fixtureArrays/q8_zero_bias_gemv_m1_tail_n5_input_bT" } }
        },
        "scalesT": {
          "dtype": "float32",
          "shape": [6, 2],
          "data": { "kind": "cycle", "values": [0.04, 0.055, 0.05, 0.065, 0.06, 0.075] }
        },
        "biasT": {
          "dtype": "float32",
          "shape": [6],
          "data": { "kind": "fillFloat32", "sinStep": 0.07, "cosStep": 0.41, "scale": 0.2 }
        }
      },
      "outputs": { "yT": { "dtype": "float32", "shape": [3, 6], "tolerance": 0.001 } },
      "attrs": { "K": 17, "N": 6, "bits": 8, "block_size": 16 }
    }
  ]
}