{
"schema": "apr-kernel-registry/v1",
"note": "KREG-001 (aprender#4539). One entry per dispatchable (backend, op, qtype) arm a kernel selector can reach. Shaped by contracts/kernel-registry-v1.yaml; read at runtime by realizar::kernel_registry, which refuses any (backend, arch, isa_features, qtype, layout) with no entry. layout is row_major only: GGUF/APR data is row-major after import (contracts/tensor-layout-v1.yaml, LAYOUT-001/002). tolerance is 'unmeasured' until a parity receipt measures it; it is never typed from memory. ops[] holds the per-forward ops (norm, RoPE, attention, activations): they run on f32 activations whatever the tensor types, so they carry no qtype or ggml_type and admit never selects one (contracts/kernel-registry-v1.yaml, kernel-registry-v1.op).",
"kernels": [
{
"kernel_id": "cpu.matvec.f32",
"op": "matvec",
"qtype": "F32",
"ggml_type": 0,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/matmul_fused.rs",
"source_fn": "fused_matmul_f32",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.f16",
"op": "matvec",
"qtype": "F16",
"ggml_type": 1,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.f16.json",
"source_file": "crates/aprender-serve/src/gguf/inference/matmul_fused.rs",
"source_fn": "float16_matmul",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.bf16",
"op": "matvec",
"qtype": "BF16",
"ggml_type": 30,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.bf16.json",
"source_file": "crates/aprender-serve/src/gguf/inference/matmul_fused.rs",
"source_fn": "float16_matmul",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q4_0",
"op": "matvec",
"qtype": "Q4_0",
"ggml_type": 2,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "i32_f32",
"precision": "q8_0",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.q4_0.json",
"source_file": "crates/aprender-serve/src/quantize/fused_q4_0_q8_0.rs",
"source_fn": "fused_q4_0_q8_0_parallel_matvec",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q8_0",
"op": "matvec",
"qtype": "Q8_0",
"ggml_type": 8,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "i32_f32",
"precision": "q8_0",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.q8_0.json",
"source_file": "crates/aprender-serve/src/quantize/fused_q8_0_q8_0.rs",
"source_fn": "fused_q8_0_q8_0_parallel_matvec",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q4_1",
"op": "matvec",
"qtype": "Q4_1",
"ggml_type": 3,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/dequant.rs",
"source_fn": "dequantize_q4_1",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q5_0",
"op": "matvec",
"qtype": "Q5_0",
"ggml_type": 6,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/dequant.rs",
"source_fn": "dequantize_q5_0",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q5_1",
"op": "matvec",
"qtype": "Q5_1",
"ggml_type": 7,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/dequant.rs",
"source_fn": "dequantize_q5_1",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.apr_q4",
"op": "matvec",
"qtype": "APR_Q4",
"ggml_type": 128,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.apr_q4.json",
"source_file": "crates/aprender-serve/src/apr/dequant.rs",
"source_fn": "dequantize_apr_q4",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.apr_q8",
"op": "matvec",
"qtype": "APR_Q8",
"ggml_type": 129,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.apr_q8.json",
"source_file": "crates/aprender-serve/src/apr/dequant.rs",
"source_fn": "dequantize_apr_q8",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q4_k",
"op": "matvec",
"qtype": "Q4_K",
"ggml_type": 12,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "i32_f32",
"precision": "q8_k",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.q4_k.json",
"source_file": "crates/aprender-serve/src/quantize/parallel_k.rs",
"source_fn": "fused_q4k_parallel_matvec",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q5_k",
"op": "matvec",
"qtype": "Q5_K",
"ggml_type": 13,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.q5_k.json",
"source_file": "crates/aprender-serve/src/quantize/parallel_k.rs",
"source_fn": "fused_q5k_parallel_matvec",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q6_k",
"op": "matvec",
"qtype": "Q6_K",
"ggml_type": 14,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "evidence/kreg/parity/cpu.matvec.q6_k.json",
"source_file": "crates/aprender-serve/src/quantize/q5k_q6k_matvec.rs",
"source_fn": "fused_q6k_parallel_matvec",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q2_k",
"op": "matvec",
"qtype": "Q2_K",
"ggml_type": 10,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/dequant_q4k.rs",
"source_fn": "dequantize_q2_k",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.q3_k",
"op": "matvec",
"qtype": "Q3_K",
"ggml_type": 11,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/dequant_q4k.rs",
"source_fn": "dequantize_q3_k",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq2_xxs",
"op": "matvec",
"qtype": "IQ2_XXS",
"ggml_type": 16,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq3_xxs",
"op": "matvec",
"qtype": "IQ3_XXS",
"ggml_type": 18,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq4_nl",
"op": "matvec",
"qtype": "IQ4_NL",
"ggml_type": 20,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq3_s",
"op": "matvec",
"qtype": "IQ3_S",
"ggml_type": 21,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq2_s",
"op": "matvec",
"qtype": "IQ2_S",
"ggml_type": 22,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cpu.matvec.iq4_xs",
"op": "matvec",
"qtype": "IQ4_XS",
"ggml_type": 23,
"layout": "row_major",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m_any",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/quantize/iq_dispatch.rs",
"source_fn": "iq_parallel_matvec_into",
"selector": "realizar::gguf::OwnedQuantizedModel::fused_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q4_0",
"op": "gemv",
"qtype": "Q4_0",
"ggml_type": 2,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q4_0_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q4_1",
"op": "gemv",
"qtype": "Q4_1",
"ggml_type": 3,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/device.rs",
"source_fn": "q4_1_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q5_0",
"op": "gemv",
"qtype": "Q5_0",
"ggml_type": 6,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q5_0_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q5_1",
"op": "gemv",
"qtype": "Q5_1",
"ggml_type": 7,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q5_1_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q8_0",
"op": "gemv",
"qtype": "Q8_0",
"ggml_type": 8,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q8_0_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q2_k",
"op": "gemv",
"qtype": "Q2_K",
"ggml_type": 10,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q2_k_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q4_k",
"op": "gemv",
"qtype": "Q4_K",
"ggml_type": 12,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "profile_selected",
"precision": "profile_selected",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/q6k_gemv_indexed.rs",
"source_fn": "q4k_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml",
"labels": [
"q4k_gemv@q4_k"
]
},
{
"kernel_id": "cuda.gemv.q5_k",
"op": "gemv",
"qtype": "Q5_K",
"ggml_type": 13,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/device.rs",
"source_fn": "q5k_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.q6_k",
"op": "gemv",
"qtype": "Q6_K",
"ggml_type": 14,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "profile_selected",
"precision": "profile_selected",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "q6k_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.f32",
"op": "gemv",
"qtype": "F32",
"ggml_type": 0,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "f32_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.f16",
"op": "gemv",
"qtype": "F16",
"ggml_type": 1,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "f16_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.bf16",
"op": "gemv",
"qtype": "BF16",
"ggml_type": 30,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "bf16_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq2_xxs",
"op": "gemv",
"qtype": "IQ2_XXS",
"ggml_type": 16,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq2_xxs_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq3_xxs",
"op": "gemv",
"qtype": "IQ3_XXS",
"ggml_type": 18,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq3_xxs_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq4_nl",
"op": "gemv",
"qtype": "IQ4_NL",
"ggml_type": 20,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 32,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq4_nl_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq3_s",
"op": "gemv",
"qtype": "IQ3_S",
"ggml_type": 21,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq3_s_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq2_s",
"op": "gemv",
"qtype": "IQ2_S",
"ggml_type": 22,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq2_s_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "cuda.gemv.iq4_xs",
"op": "gemv",
"qtype": "IQ4_XS",
"ggml_type": 23,
"layout": "row_major",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/weight.rs",
"source_fn": "iq4_xs_gemv_into",
"selector": "realizar::cuda::CudaExecutor::gemv_dispatch",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "wgpu.gemv.q4_k",
"op": "gemv",
"qtype": "Q4_K",
"ggml_type": 12,
"layout": "row_major",
"backend": "wgpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 256,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-compute/src/backends/gpu/device/linalg/wgsl_forward.rs",
"source_fn": "encode_q4k_gemv",
"selector": "realizar::gpu::adapters::wgpu_adapter::raw_q4k_weights",
"contract": "contracts/tensor-layout-v1.yaml"
},
{
"kernel_id": "wgpu.gemv.f32",
"op": "gemv",
"qtype": "F32",
"ggml_type": 0,
"layout": "row_major",
"backend": "wgpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"block_elems": 1,
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-DOT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-compute/src/backends/gpu/device/linalg/wgsl_forward.rs",
"source_fn": "encode_matmul",
"selector": "trueno::backends::gpu::WgslForward::encode_matmul",
"contract": "contracts/tensor-layout-v1.yaml"
}
],
"ops": [
{
"kernel_id": "cpu.rmsnorm.f32",
"op": "rmsnorm",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.rmsnorm.f32.json",
"source_file": "crates/aprender-serve/src/gguf/ops.rs",
"source_fn": "rms_norm_into",
"selector": "realizar::gguf::OwnedQuantizedModel::forward_single_with_scratch",
"contract": "contracts/rmsnorm-kernel-v1.yaml"
},
{
"kernel_id": "cpu.layernorm.f32",
"op": "layernorm",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.layernorm.f32.json",
"source_file": "crates/aprender-serve/src/gguf/ops.rs",
"source_fn": "layer_norm_into",
"selector": "realizar::gguf::OwnedQuantizedModel::forward_single_with_scratch",
"contract": "contracts/layernorm-kernel-v1.yaml"
},
{
"kernel_id": "cpu.rope.f32",
"op": "rope",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ROPE",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.rope.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/rope.rs",
"source_fn": "apply_rope",
"selector": "realizar::gguf::OwnedQuantizedModel::scratch_attention_block",
"contract": "contracts/rope-kernel-v1.yaml"
},
{
"kernel_id": "cpu.attention.f32",
"op": "attention",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ATT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.attention.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/attention_gqa.rs",
"source_fn": "attention_with_cache_gqa_into",
"selector": "realizar::gguf::OwnedQuantizedModel::scratch_attention_block",
"contract": "contracts/attention-kernel-v1.yaml"
},
{
"kernel_id": "cpu.swiglu.f32",
"op": "swiglu",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.swiglu.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/forward/results.rs",
"source_fn": "swiglu_gate_into",
"selector": "realizar::gguf::OwnedQuantizedModel::forward_single_with_scratch",
"contract": "contracts/swiglu-kernel-v1.yaml"
},
{
"kernel_id": "cpu.gelu.f32",
"op": "gelu",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.gelu.f32.json",
"source_file": "crates/aprender-serve/src/gguf/ops.rs",
"source_fn": "gelu",
"selector": "realizar::gguf::OwnedQuantizedModel::scratch_gelu_ffn",
"contract": "contracts/gelu-kernel-v1.yaml"
},
{
"kernel_id": "cpu.embed.f32",
"op": "embed",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.embed.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/matmul_fused.rs",
"source_fn": "embed_into",
"selector": "realizar::gguf::OwnedQuantizedModel::forward_single_with_scratch",
"contract": "contracts/embedding-lookup-v1.yaml"
},
{
"kernel_id": "cpu.kv_write.f32",
"op": "kv_write",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.kv_write.f32.json",
"source_file": "crates/aprender-serve/src/gguf/runtime.rs",
"source_fn": "append",
"selector": "realizar::gguf::OwnedQuantizedModel::scratch_attention_block",
"contract": "contracts/kv-cache-equivalence-v1.yaml"
},
{
"kernel_id": "cpu.residual_add.f32",
"op": "residual_add",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.residual_add.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/forward/results.rs",
"source_fn": "residual_add_into",
"selector": "realizar::gguf::OwnedQuantizedModel::forward_single_with_scratch",
"contract": "contracts/apr-model-graph-v1.yaml"
},
{
"kernel_id": "cpu.argmax.f32",
"op": "argmax",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.argmax.f32.json",
"source_file": "crates/aprender-serve/src/gguf/ops.rs",
"source_fn": "argmax",
"selector": "realizar::gguf::OwnedQuantizedModel::generate_with_cache",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.embed.f32",
"op": "embed",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/inference/matmul_fused.rs",
"source_fn": "embed_into",
"selector": "realizar::gguf::OwnedQuantizedModelCuda::forward_gpu_resident_to_token_id",
"contract": "contracts/embedding-lookup-v1.yaml"
},
{
"kernel_id": "cuda.rmsnorm.f32",
"op": "rmsnorm",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/layer_norm_gpu.rs",
"source_fn": "rmsnorm_into",
"selector": "realizar::cuda::CudaExecutor::workspace_qkv_rope_phase",
"contract": "contracts/rmsnorm-kernel-v1.yaml",
"labels": [
"rmsnorm@f32"
]
},
{
"kernel_id": "cuda.rmsnorm.per_head.f32",
"op": "rmsnorm",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/layer_norm_gpu.rs",
"source_fn": "per_head_rmsnorm_into",
"selector": "realizar::cuda::CudaExecutor::apply_qk_norm",
"contract": "contracts/rmsnorm-kernel-v1.yaml"
},
{
"kernel_id": "cuda.rope.f32",
"op": "rope",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ROPE",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/fused_ffn.rs",
"source_fn": "rope_indirect_into",
"selector": "realizar::cuda::CudaExecutor::apply_rope_to_qk",
"contract": "contracts/rope-kernel-v1.yaml"
},
{
"kernel_id": "cuda.rope.neox.f32",
"op": "rope",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ROPE",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/rope_indirect.rs",
"source_fn": "rope_neox_indirect_into",
"selector": "realizar::cuda::CudaExecutor::apply_rope_to_qk",
"contract": "contracts/rope-kernel-v1.yaml"
},
{
"kernel_id": "cuda.kv_write.f32",
"op": "kv_write",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/kv_scatter.rs",
"source_fn": "scatter_kv_to_cache",
"selector": "realizar::cuda::CudaExecutor::incremental_attention_into_inner",
"contract": "contracts/kv-cache-equivalence-v1.yaml"
},
{
"kernel_id": "cuda.attention.f32",
"op": "attention",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ATT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/kv_scatter.rs",
"source_fn": "launch_attention_kernel",
"selector": "realizar::cuda::CudaExecutor::workspace_attention_residual_phase",
"contract": "contracts/attention-kernel-v1.yaml"
},
{
"kernel_id": "cuda.swiglu.f32",
"op": "swiglu",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/silu.rs",
"source_fn": "fused_swiglu_into",
"selector": "realizar::cuda::CudaExecutor::workspace_ffn_phase",
"contract": "contracts/kernel-fusion-v1.yaml"
},
{
"kernel_id": "cuda.residual_add.f32",
"op": "residual_add",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/residual.rs",
"source_fn": "residual_add_into",
"selector": "realizar::cuda::CudaExecutor::workspace_attention_residual_phase",
"contract": "contracts/apr-model-graph-v1.yaml"
},
{
"kernel_id": "cuda.argmax.f32",
"op": "argmax",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/layers/par-062.rs",
"source_fn": "gpu_argmax",
"selector": "realizar::cuda::CudaExecutor::forward_graphed_replay_to_token_id",
"contract": "contracts/decode-gpu-resident-sampling-v1.yaml"
},
{
"kernel_id": "cuda.rope.partial_neox.f32",
"op": "rope",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ROPE",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_partial_neox_rope_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::attention_layer_inner",
"contract": "contracts/rope-kernel-v1.yaml",
"archs": [
"qwen35",
"qwen35moe",
"qwen3moe"
]
},
{
"kernel_id": "cuda.attention.gdn_decode.f32",
"op": "attention",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ATT",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_decode_attention_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::attention_layer_inner",
"contract": "contracts/attention-kernel-v1.yaml",
"archs": [
"qwen35",
"qwen35moe",
"qwen3moe"
]
},
{
"kernel_id": "cuda.rmsnorm.gated.f32",
"op": "rmsnorm",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_gated_rmsnorm_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::deltanet_layer_inner",
"contract": "contracts/rmsnorm-kernel-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.conv1d.silu.f32",
"op": "conv1d",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_causal_conv1d_silu_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::deltanet_layer_inner",
"contract": "contracts/conv1d-kernel-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.l2norm.per_head.f32",
"op": "l2norm",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_per_head_l2_norm_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::deltanet_layer_inner",
"contract": "contracts/gated-delta-net-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.gdn_gates.f32",
"op": "gdn_gates",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_gates_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::deltanet_layer_inner",
"contract": "contracts/gated-delta-net-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.delta_rule.f32",
"op": "delta_rule",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_delta_rule_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::deltanet_layer_inner",
"contract": "contracts/gated-delta-net-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.sigmoid_gate.f32",
"op": "sigmoid_gate",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_sigmoid_gate_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::attention_layer_inner",
"contract": "contracts/gated-delta-net-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.split.interleaved.f32",
"op": "split",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/gdn_ops.rs",
"source_fn": "gdn_split_interleaved_into",
"selector": "realizar::gguf::cuda::Qwen35CudaModel::attention_layer_inner",
"contract": "contracts/gated-delta-net-v1.yaml",
"archs": [
"qwen35",
"qwen35moe"
]
},
{
"kernel_id": "cuda.elementwise_mul.f32",
"op": "elementwise_mul",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/cuda/executor/silu.rs",
"source_fn": "elementwise_mul_into",
"selector": "realizar::gguf::cuda::Qwen3MoeCudaModel::moe_layer",
"contract": "contracts/moe-expert-dispatch-v1.yaml",
"archs": [
"qwen3moe",
"qwen35moe"
]
},
{
"kernel_id": "cuda.sample.host.f32",
"op": "sample",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/infer/qwen3_moe_generate.rs",
"source_fn": "sample_from_logits",
"selector": "realizar::infer::run_qwen3_moe_generate_gpu",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml",
"archs": [
"qwen3moe",
"qwen35moe"
]
},
{
"kernel_id": "cpu.sample.topk.f32",
"op": "sample",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.sample.topk.f32.json",
"source_file": "crates/aprender-serve/src/sampling.rs",
"source_fn": "draw",
"selector": "realizar::gguf::OwnedQuantizedModel::generate_with_scratch",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.sample.host_topk.f32",
"op": "sample",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/inference/generate_quantized.rs",
"source_fn": "sample_topk",
"selector": "realizar::gguf::cuda::OwnedQuantizedModelCuda::next_token",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.sample.host_seeded.f32",
"op": "sample",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/inference/generate_quantized.rs",
"source_fn": "sample_topk_seeded",
"selector": "realizar::session::choose_token",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.argmax.host.f32",
"op": "argmax",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-RED",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/ops.rs",
"source_fn": "argmax",
"selector": "realizar::session::choose_token",
"contract": "contracts/decode-gpu-resident-sampling-v1.yaml"
},
{
"kernel_id": "cpu.sample.seeded.f32",
"op": "sample",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.sample.seeded.f32.json",
"source_file": "crates/aprender-serve/src/sampling.rs",
"source_fn": "draw",
"selector": "realizar::session::choose_token",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.route_topk.host.f32",
"op": "route_topk",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-SMX",
"determinism": "bounded",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/qwen3_moe_load.rs",
"source_fn": "route_top_k",
"selector": "realizar::gguf::cuda::Qwen3MoeCudaModel::moe_layer",
"contract": "contracts/moe-expert-dispatch-v1.yaml",
"archs": [
"qwen3moe",
"qwen35moe"
]
},
{
"kernel_id": "cpu.repeat_penalty.f32",
"op": "repeat_penalty",
"backend": "cpu",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "evidence/kreg/parity/cpu.repeat_penalty.f32.json",
"source_file": "crates/aprender-serve/src/gguf/inference/generate_quantized.rs",
"source_fn": "apply_repeat_penalty",
"selector": "realizar::session::Session::advance_and_choose",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
},
{
"kernel_id": "cuda.repeat_penalty.host.f32",
"op": "repeat_penalty",
"backend": "cuda",
"arch": "any",
"isa_features": "none",
"requires": "none",
"accumulate": "f32",
"precision": "f32",
"error_model": "EM-ELEM",
"determinism": "bitwise",
"shape_class": "m1",
"tolerance": "unmeasured",
"source_file": "crates/aprender-serve/src/gguf/inference/generate_quantized.rs",
"source_fn": "apply_repeat_penalty",
"selector": "realizar::gguf::cuda::OwnedQuantizedModelCuda::next_token",
"contract": "contracts/apr-run-sampling-plumbing-v1.yaml"
}
]
}