{ "fixtureArrays": { "weight48": [129, 126, 131, 138, 139, 131, 133, 130, 131, 122, 143, 148, 133, 129, 132, 144, 138, 148, 132, 136, 137, 129, 115, 135, 131, 130, 168, 150, 144, 139, 135, 130, 132, 138, 140, 133, 130, 129, 132, 144, 152, 161, 149, 170, 212, 128, 149, 160], "weight48v2": [2, -1, 4, 11, 12, 4, 6, 3, 4, -5, 16, 21, 6, 2, 5, 17, 11, 21, 5, 9, 10, 2, -12, 8, 4, 3, 41, 23, 17, 12, 8, 3, 5, 11, 13, 6, 3, 2, 5, 17, 25, 34, 22, 43, 85, 1, 22, 33], "input24": [133, 130, 131, 122, 128, 138, 136, 123, 128, 138, 148, 158, 136, 123, 128, 138, 168, 178, 133, 130, 131, 122, 188, 198], "output16": [11.68952751159668, 2.76993727684021, 7.0500006675720215, 8.350000381469727, 11.690000534057617, 2.7700002193450928, 7.0500006675720215, 8.350000381469727, 14.276561737060547, 5.37415885925293, 9.65000057220459, 10.950000762939453, 14.289074897766113, 5.3702874183654785, 9.65000057220459, 10.950000762939453] }, "cases": [ { "name": "ort_batch1_u8u8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionBatch1", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "uint8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "mask_index": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_batch1_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionBatch1", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "int8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48v2" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "mask_index": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_batch2_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionBatch2", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122, 136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "int8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48v2" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "values", "values": [2, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 2, 4], "data": { "kind": "values", "values": [3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432, 3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_mask_partial_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionMaskPartialSequence", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "int8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48v2" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "mask_index": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [1] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [8.690000534057617, -0.1300000250339508, 4.250000476837158, 5.65000057220459, 8.690000534057617, -0.1300000250339508, 4.250000476837158, 5.65000057220459] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_mask_exceed_u8u8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionMaskExceedSequence", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "uint8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "mask_index": { "dtype": "int32", "shape": [1], "data": { "kind": "values", "values": [3] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_no_mask_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionNoMaskIndex", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [136, 123, 128, 138, 133, 130, 131, 122] } }, "weight": { "dtype": "int8", "shape": [4, 12], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/weight48v2" } } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [3.1495978832244873, 0.1084367036819458, 4.250000476837158, 5.65000057220459, 3.969679594039917, 0.07314367592334747, 4.25, 5.650000095367432] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_unidirectional_u8u8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionUnidirectional_U8U8", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [255, 77, 48, 153, 87, 0, 170, 71] } }, "weight": { "dtype": "uint8", "shape": [4, 12], "data": { "kind": "values", "values": [88, 160, 219, 190, 150, 60, 181, 202, 108, 100, 164, 197, 241, 87, 143, 99, 147, 255, 175, 90, 106, 152, 146, 152, 201, 146, 71, 102, 142, 96, 140, 165, 141, 109, 183, 183, 29, 137, 102, 88, 244, 215, 98, 91, 0, 227, 162, 138] } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.0540979839861393, -0.06444740295410156, 0.03112877532839775, -0.08288222551345825, 0.07840359210968018, 0.039143580943346024, -0.45591455698013306, -0.11876055598258972, 0.3670335114002228, 0.028461361303925514, -0.08913630992174149, 0.28048714995384216] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.006167058367282152] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.007593048270791769] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [107] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [123] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [-0.02771615982055664, 0.09102192521095276, 0.08093852549791336, 0.6191383600234985, 0.3608928322792053, -0.11653690040111542, -0.03012145683169365, 0.4092332720756531] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_unidirectional_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionUnidirectional_U8S8", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [1, 2, 4], "data": { "kind": "values", "values": [255, 77, 48, 153, 87, 0, 170, 71] } }, "weight": { "dtype": "int8", "shape": [4, 12], "data": { "kind": "values", "values": [-39, 33, 92, 63, 23, -66, 53, 75, -19, -27, 37, 70, 114, -40, 15, -28, 19, 127, 48, -37, -21, 25, 19, 25, 74, 18, -56, -25, 15, -31, 13, 38, 14, -18, 55, 56, -98, 10, -25, -39, 117, 88, -29, -36, -127, 99, 35, 11] } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.0540979839861393, -0.06444740295410156, 0.03112877532839775, -0.08288222551345825, 0.07840359210968018, 0.039143580943346024, -0.45591455698013306, -0.11876055598258972, 0.3670335114002228, 0.028461361303925514, -0.08913630992174149, 0.28048714995384216] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.006167058367282152] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.007622942328453064] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [107] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [-4] } } }, "outputs": { "output": { "dtype": "float32", "shape": [1, 2, 4], "data": { "kind": "values", "values": [-0.029270321130752563, 0.08910571783781052, 0.08438176661729813, 0.6204716563224792, 0.3609381914138794, -0.11732138693332672, -0.029981952160596848, 0.4099825322628021] }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_pruned_u8s8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionPrunedModel", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 2, 6], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/input24" } } }, "weight": { "dtype": "int8", "shape": [6, 12], "data": { "kind": "values", "values": [2, -1, 4, 11, 12, 4, 6, 3, 4, -5, 16, 21, 6, 2, 5, 17, 11, 21, 5, 9, 10, 2, -12, 8, 4, 3, 41, 23, 17, 12, 8, 3, 5, 11, 13, 6, 3, 2, 5, 17, 25, 34, 22, 43, 85, 1, 22, 33, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2] } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [1] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 2, 4], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/output16" } }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "ort_pruned_u8u8", "provenance": { "upstream": "onnxruntime/test/contrib_ops/quantize_attention_op_test.cc", "test": "QAttentionTest.QAttentionPrunedModel", "note": "Quantized bytes reconstructed by applying the upstream test's quantizer to its float vectors; expected values are the reference CPU kernel's output on those bytes." }, "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 2, 6], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/input24" } } }, "weight": { "dtype": "uint8", "shape": [6, 12], "data": { "kind": "values", "values": [129, 126, 131, 138, 139, 131, 133, 130, 131, 122, 143, 148, 133, 129, 132, 144, 138, 148, 132, 136, 137, 129, 115, 135, 131, 130, 168, 150, 144, 139, 135, 130, 132, 138, 140, 133, 130, 129, 132, 144, 152, 161, 149, 170, 212, 128, 149, 160, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 140, 139, 138, 137, 136, 135, 134, 133, 132, 131, 130, 129] } }, "bias": { "dtype": "float32", "shape": [12], "data": { "kind": "values", "values": [-0.5, 0.6000000238418579, 1.2000000476837158, 2.0999999046325684, 0.5, 0.699999988079071, 0.20000000298023224, 1.2000000476837158, 0.5, 0.4000000059604645, 0.30000001192092896, 1.2000000476837158] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.10000000149011612] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [128] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 2, 4], "data": { "kind": "values", "values": { "$ref": "#/fixtureArrays/output16" } }, "tolerance": 0.00002, "relTolerance": 0.00002 } } }, { "name": "cover_plain_mask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_mask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_mask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_mask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_nomask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_nomask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_nomask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_plain_nomask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_mask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_mask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_mask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_mask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_nomask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_nomask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_nomask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_present_nomask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 32, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_mask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 37], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_mask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 37], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_mask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 37], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_mask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 37], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_nomask_izp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_nomask_izp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_nomask_noizp_wzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "cover_past_nomask_noizp_nowzp", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "past": { "dtype": "float32", "shape": [2, 2, 4, 5, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 4, 37, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "mask_endpos_u8s8", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [3, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [3], "data": { "kind": "cycle", "values": [4, 2, 3] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [3, 4, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "mask_endstart_u8s8", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [3, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [6], "data": { "kind": "cycle", "values": [4, 3, 4, 3, 4, 3, 0, 1, 0, 1, 0, 1] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [3, 4, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "mask_raw3d_u8u8", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "uint8", "shape": [8, 24], "data": { "kind": "cycle", "values": [10, 245, 130, 60, 199, 88, 171, 27, 222, 105, 149, 4, 240] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 3, 3], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [129] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "percolumn_scale_zero_u8s8", "attrs": { "num_heads": 3 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 36], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.011, 0.007, 0.015, 0.009, 0.013, 0.005, 0.017, 0.008, 0.012, 0.006, 0.014] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [4, 2, 3] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [36], "data": { "kind": "cycle", "values": [3, -7, 1, -2, 5, -4, 0, 6, -1, 2, -6] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 4, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "percolumn_scale_zero_u8u8", "attrs": { "num_heads": 3 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "uint8", "shape": [8, 36], "data": { "kind": "cycle", "values": [10, 245, 130, 60, 199, 88, 171, 27, 222, 105, 149, 4, 240] } }, "bias": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.011, 0.007, 0.015, 0.009, 0.013, 0.005, 0.017, 0.008, 0.012, 0.006, 0.014] } }, "mask_index": { "dtype": "int32", "shape": [2, 4], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "uint8", "shape": [36], "data": { "kind": "cycle", "values": [131, 121, 129, 126, 133, 124, 128, 134, 127, 130, 122] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 4, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "causal_no_past", "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 5, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [5, 3, 4] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 5, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "causal_with_past", "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [7, 5, 6] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 2, 4, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 2, 7, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "decode_single_token", "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 1, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [8, 6, 7] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 2, 7, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 1, 8], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 2, 8, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "head_size_two_padded", "attrs": { "num_heads": 6 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 36], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 4], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 4, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "head_size_two_padded_sgmat", "attrs": { "num_heads": 6 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 36], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "EXACT_MATRIX_MIN_INVOCATIONS": 0 } }, { "name": "head_size_six_padded", "attrs": { "num_heads": 3 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 54], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [54], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [3, 1, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 18], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "explicit_scale", "attrs": { "num_heads": 2, "scale": 0.35 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [3, 1, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "mask_filter_value", "attrs": { "num_heads": 2, "mask_filter_value": -2500 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 4, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 4], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 4, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "contraction_tail_k9", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 9], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [9, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [3, 1, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "contraction_tail_k7_u8u8", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 7], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "uint8", "shape": [7, 24], "data": { "kind": "cycle", "values": [10, 245, 130, 60, 199, 88, 171, 27, 222, 105, 149, 4, 240] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [3, 1, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [129] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "wide_columns_n130", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 390], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [390], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [3, 1, 2] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 130], "tolerance": 0.0001, "relTolerance": 0.0001 } } }, { "name": "qtile_wide_head_unstriped", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 8, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 384], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [384], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 8], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 8, 128], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "ATTEND_MIN_WORKGROUPS": 1 } }, { "name": "qtile_padded_head_striped", "attrs": { "num_heads": 6 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 8, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 36], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [36], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 8], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 8, 12], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "ATTEND_MIN_WORKGROUPS": 1 } }, { "name": "qtile_tail_seq7", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 7, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 96], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [96], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [7, 5, 6] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 7, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "ATTEND_MIN_WORKGROUPS": 1 } }, { "name": "qtile_causal_past", "attrs": { "num_heads": 2, "unidirectional": 1 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 6, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 96], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [96], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [9, 7, 8] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 2, 3, 16], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 6, 32], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 2, 9, 16], "tolerance": 0.00001, "relTolerance": 0.00001 } }, "tunables": { "ATTEND_MIN_WORKGROUPS": 1 } }, { "name": "qtile_raw3d_mask", "attrs": { "num_heads": 2 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 6, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 96], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [96], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 6, 6], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 6, 32], "tolerance": 0.0001, "relTolerance": 0.0001 } }, "tunables": { "ATTEND_MIN_WORKGROUPS": 1 } }, { "name": "share_buffer_zero_with_past", "attrs": { "num_heads": 2, "past_present_share_buffer": 0 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 3, 8], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [8, 24], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [24], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2], "data": { "kind": "cycle", "values": [7, 5, 6] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } }, "past": { "dtype": "float32", "shape": [2, 2, 2, 4, 4], "data": { "kind": "cycle", "values": [0.21, -0.34, 0.52, 0.08, -0.47, 0.16, 0.39, -0.11, 0.28, -0.05, 0.44] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 3, 8], "tolerance": 0.0001, "relTolerance": 0.0001 }, "present": { "dtype": "float32", "shape": [2, 2, 2, 7, 4], "tolerance": 0.00001, "relTolerance": 0.00001 } } }, { "name": "prefill_rows64_default_tunables", "attrs": { "num_heads": 4 }, "inputs": { "input": { "dtype": "uint8", "shape": [2, 32, 32], "data": { "kind": "cycle", "values": [17, 200, 3, 251, 96, 140, 62, 218, 41, 175, 9, 233] } }, "weight": { "dtype": "int8", "shape": [32, 48], "data": { "kind": "cycle", "values": [-118, 37, -5, 92, 61, -74, 23, -101, 13, 88, -47, 5, 111] } }, "bias": { "dtype": "float32", "shape": [48], "data": { "kind": "cycle", "values": [0.31, -0.22, 0.07, 0.44, -0.35, 0.13, -0.08, 0.26, 0.5, -0.17, 0.02, -0.41, 0.19] } }, "input_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.020999999716877937] } }, "weight_scale": { "dtype": "float32", "shape": [1], "data": { "kind": "values", "values": [0.013000000268220901] } }, "mask_index": { "dtype": "int32", "shape": [2, 32], "data": { "kind": "cycle", "values": [1, 1, 0, 1, 0, 1, 1, 1, 0] } }, "input_zero_point": { "dtype": "uint8", "shape": [1], "data": { "kind": "values", "values": [124] } }, "weight_zero_point": { "dtype": "int8", "shape": [1], "data": { "kind": "values", "values": [2] } } }, "outputs": { "output": { "dtype": "float32", "shape": [2, 32, 16], "tolerance": 0.0001, "relTolerance": 0.0001 } } } ] }