{ "hardware": "Apple M3 / 16 GB", "accepted": "Nine native scaled-dot-product attention blocks in the FP32 code-prediction transformer; decoder unchanged.", "defaultEncoderNormFusion": false, "summary": { "baseline": { "wallSeconds": 19.185629937499996, "cpuSeconds": 2.405257999999999, "inferenceSeconds": 16.832070648670197, "encoderSeconds": 5.952198624610901, "decoderSeconds": 10.879868388175964, "peakPhysicalFootprintBytes": 2187224624 }, "norm": { "wallSeconds": 18.507835, "cpuSeconds": 2.1664035000000004, "inferenceSeconds": 16.639364063739777, "encoderSeconds": 5.9536720514297485, "decoderSeconds": 10.685688495635986, "peakPhysicalFootprintBytes": 2202732128, "reductionPercent": { "wallSeconds": 3.5328260771630315, "inferenceSeconds": 1.1448774720158705, "encoderSeconds": -0.024754328808107573, "peakPhysicalFootprintBytes": -0.709003722335555 } }, "attention": { "wallSeconds": 17.0293313335, "cpuSeconds": 1.8463729999999998, "inferenceSeconds": 15.770417392253876, "encoderSeconds": 5.165894567966461, "decoderSeconds": 10.604519486427307, "peakPhysicalFootprintBytes": 2063476272, "reductionPercent": { "wallSeconds": 11.239133721563766, "inferenceSeconds": 6.3073241467186625, "encoderSeconds": 13.210312797581425, "peakPhysicalFootprintBytes": 5.657779756232295 } } }, "measurements": [ { "pair": 0, "variant": "baseline", "wallSeconds": 19.266577124999998, "cpuSeconds": 2.503933, "pngByteParityWithBaseline": true, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 10.474938869476318, "decoderUnits": "gpu", "encoderSeconds": 5.879798054695129, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 16.354741096496582, "loadSeconds": 2.3399150371551514, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedDecoderNorm": true, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2171405872, "pythonRuntime": false, "scale": 2, "seconds": 19.019976019859314, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 0, "thermalStateAtStart": 0 } }, { "pair": 0, "variant": "norm", "wallSeconds": 17.590159416000002, "cpuSeconds": 2.1660500000000003, "pngByteParityWithBaseline": false, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 10.094687938690186, "decoderUnits": "gpu", "encoderSeconds": 5.5853201150894165, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 15.680013179779053, "loadSeconds": 1.3422569036483765, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedDecoderNorm": true, "fusedEncoderNorm": true, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2202240608, "pythonRuntime": false, "scale": 2, "seconds": 17.34317398071289, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 0, "thermalStateAtStart": 0 } }, { "pair": 0, "variant": "attention", "wallSeconds": 18.036012041999996, "cpuSeconds": 2.0608000000000004, "pngByteParityWithBaseline": true, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 10.87925910949707, "decoderUnits": "gpu", "encoderSeconds": 5.322363972663879, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 16.2016259431839, "loadSeconds": 1.3784769773483276, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedCodeAttention": true, "fusedDecoderNorm": true, "fusedEncoderNorm": false, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2077869640, "pythonRuntime": false, "scale": 2, "seconds": 17.82434093952179, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 0, "thermalStateAtStart": 0 } }, { "pair": 1, "variant": "attention", "wallSeconds": 16.022650625000004, "cpuSeconds": 1.6319459999999992, "pngByteParityWithBaseline": true, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 10.329779863357544, "decoderUnits": "gpu", "encoderSeconds": 5.009425163269043, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 15.339208841323853, "loadSeconds": 0.19483697414398193, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedCodeAttention": true, "fusedDecoderNorm": true, "fusedEncoderNorm": false, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2049082904, "pythonRuntime": false, "scale": 2, "seconds": 15.808948993682861, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 1, "thermalStateAtStart": 0 } }, { "pair": 1, "variant": "norm", "wallSeconds": 19.425510583999994, "cpuSeconds": 2.1667570000000005, "pngByteParityWithBaseline": false, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 11.276689052581787, "decoderUnits": "gpu", "encoderSeconds": 6.322023987770081, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 17.5987149477005, "loadSeconds": 1.3740719556808472, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedDecoderNorm": true, "fusedEncoderNorm": true, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2203223648, "pythonRuntime": false, "scale": 2, "seconds": 19.19998598098755, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 1, "thermalStateAtStart": 1 } }, { "pair": 1, "variant": "baseline", "wallSeconds": 19.104682749999995, "cpuSeconds": 2.306582999999998, "pngByteParityWithBaseline": true, "report": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 11.28479790687561, "decoderUnits": "gpu", "encoderSeconds": 6.024599194526672, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 17.30940020084381, "loadSeconds": 1.288451075553894, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedDecoderNorm": true, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2203043376, "pythonRuntime": false, "scale": 2, "seconds": 18.87854504585266, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 1, "thermalStateAtStart": 1 } } ], "graphStats": { "pgtformer": { "total": 2811, "operations": { "const": 1752, "sub": 19, "mul": 24, "conv": 62, "batch_norm": 46, "relu": 27, "max_pool": 1, "add": 92, "avg_pool": 4, "upsample_nearest_neighbor": 3, "sigmoid": 3, "concat": 3, "upsample_bilinear": 2, "reshape": 248, "transpose": 172, "reduce_mean": 36, "square": 17, "sqrt": 17, "real_div": 16, "silu": 15, "pad": 4, "layer_norm": 35, "linear": 96, "slice_by_index": 24, "matmul": 34, "softmax": 17, "gelu": 17, "gather": 17, "reduce_argmax": 1, "greater_equal": 1, "select": 1, "reduce_sum": 1, "tile": 4 } }, "pgt-fusion-norm": { "total": 2676, "operations": { "const": 1707, "sub": 4, "mul": 24, "conv": 62, "batch_norm": 46, "relu": 27, "max_pool": 1, "add": 77, "avg_pool": 4, "upsample_nearest_neighbor": 3, "sigmoid": 3, "concat": 3, "upsample_bilinear": 2, "reshape": 248, "transpose": 172, "layer_norm": 50, "silu": 15, "pad": 4, "linear": 96, "slice_by_index": 24, "matmul": 34, "softmax": 17, "gelu": 17, "gather": 17, "reduce_argmax": 1, "greater_equal": 1, "select": 1, "reduce_sum": 1, "reduce_mean": 6, "square": 2, "sqrt": 2, "tile": 4, "real_div": 1 } }, "pgt-fusion-attention": { "total": 2784, "operations": { "const": 1725, "sub": 19, "mul": 15, "conv": 62, "batch_norm": 46, "relu": 27, "max_pool": 1, "add": 92, "avg_pool": 4, "upsample_nearest_neighbor": 3, "sigmoid": 3, "concat": 3, "upsample_bilinear": 2, "reshape": 275, "transpose": 172, "reduce_mean": 36, "square": 17, "sqrt": 17, "real_div": 16, "silu": 15, "pad": 4, "layer_norm": 35, "linear": 96, "slice_by_index": 24, "matmul": 16, "softmax": 8, "gelu": 17, "gather": 17, "scaled_dot_product_attention": 9, "reduce_argmax": 1, "greater_equal": 1, "select": 1, "reduce_sum": 1, "tile": 4 } } }, "originalNetworkReferenceGeneration": { "model": "pgtformer", "scope": "Actual product video inputs, preprocessing and native graph versus original pretrained network", "passed": true, "frames": [ { "frame": 0, "window": [ 0, 0, 1 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.013160884380340576, "meanAbsoluteError": 8.180399163393304e-05 }, { "frame": 1, "window": [ 0, 1, 2 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.01537981629371643, "meanAbsoluteError": 8.932689524954185e-05 }, { "frame": 2, "window": [ 1, 2, 3 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.013586729764938354, "meanAbsoluteError": 0.00011335056478856131 }, { "frame": 3, "window": [ 2, 3, 4 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.014907270669937134, "meanAbsoluteError": 9.071586100617424e-05 }, { "frame": 4, "window": [ 3, 4, 5 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.012558221817016602, "meanAbsoluteError": 9.717844659462571e-05 }, { "frame": 5, "window": [ 4, 5, 6 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.009725838899612427, "meanAbsoluteError": 8.610969962319359e-05 }, { "frame": 6, "window": [ 5, 6, 7 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.015181079506874084, "meanAbsoluteError": 0.00012718189100269228 }, { "frame": 7, "window": [ 6, 7, 7 ], "inputMaximumError": 0.0, "maximumAbsoluteError": 0.014310456812381744, "meanAbsoluteError": 0.00014974609075579792 } ] }, "encoderNormValidation": { "backend": "native Swift Core ML", "decoderUnits": "gpu", "encoderUnits": "gpu", "loadSeconds": 2.8730820417404175, "maximumAbsoluteError": 0.016909509897232056, "meanAbsoluteError": 0.00015762679654482992, "measurements": [ { "decoderSeconds": 1.48848295211792, "encoderSeconds": 0.8714809417724609, "iteration": 0, "maximumAbsoluteError": 0.012677013874053955, "meanAbsoluteError": 8.39920463995569e-05, "passed": true, "reference": "000000.safetensors" }, { "decoderSeconds": 1.3886889219284058, "encoderSeconds": 0.9401990175247192, "iteration": 0, "maximumAbsoluteError": 0.014373481273651123, "meanAbsoluteError": 9.247053679046771e-05, "passed": true, "reference": "000001.safetensors" }, { "decoderSeconds": 1.565735936164856, "encoderSeconds": 0.9051710367202759, "iteration": 0, "maximumAbsoluteError": 0.01339605450630188, "meanAbsoluteError": 0.0001274039177631183, "passed": true, "reference": "000002.safetensors" }, { "decoderSeconds": 1.5463720560073853, "encoderSeconds": 0.8498330116271973, "iteration": 0, "maximumAbsoluteError": 0.011530846357345581, "meanAbsoluteError": 0.0001002705472701159, "passed": true, "reference": "000003.safetensors" }, { "decoderSeconds": 1.4075289964675903, "encoderSeconds": 0.7971700429916382, "iteration": 0, "maximumAbsoluteError": 0.016909509897232056, "meanAbsoluteError": 0.00010848505156604289, "passed": true, "reference": "000004.safetensors" }, { "decoderSeconds": 1.5271708965301514, "encoderSeconds": 1.4246550798416138, "iteration": 0, "maximumAbsoluteError": 0.014797002077102661, "meanAbsoluteError": 9.283579105300297e-05, "passed": true, "reference": "000005.safetensors" }, { "decoderSeconds": 1.4414219856262207, "encoderSeconds": 0.82452392578125, "iteration": 0, "maximumAbsoluteError": 0.01603156328201294, "meanAbsoluteError": 0.00011911896752330754, "passed": true, "reference": "000006.safetensors" }, { "decoderSeconds": 1.3687210083007812, "encoderSeconds": 0.8771669864654541, "iteration": 0, "maximumAbsoluteError": 0.012937672436237335, "meanAbsoluteError": 0.00015762679654482992, "passed": true, "reference": "000007.safetensors" } ], "passed": true, "peakPhysicalFootprintBytes": 2204124696 }, "fusedAttentionValidation": { "backend": "native Swift Core ML", "decoderUnits": "gpu", "encoderUnits": "gpu", "loadSeconds": 4.362862944602966, "maximumAbsoluteError": 0.01537981629371643, "meanAbsoluteError": 0.00014974610047203637, "measurements": [ { "decoderSeconds": 1.4304579496383667, "encoderSeconds": 0.7125070095062256, "iteration": 0, "maximumAbsoluteError": 0.013160884380340576, "meanAbsoluteError": 8.180398304042859e-05, "passed": true, "reference": "000000.safetensors" }, { "decoderSeconds": 1.3408229351043701, "encoderSeconds": 0.6830980777740479, "iteration": 0, "maximumAbsoluteError": 0.01537981629371643, "meanAbsoluteError": 8.93268901504928e-05, "passed": true, "reference": "000001.safetensors" }, { "decoderSeconds": 1.2520500421524048, "encoderSeconds": 0.6845479011535645, "iteration": 0, "maximumAbsoluteError": 0.013586729764938354, "meanAbsoluteError": 0.00011335056096090178, "passed": true, "reference": "000002.safetensors" }, { "decoderSeconds": 1.3508729934692383, "encoderSeconds": 0.6919339895248413, "iteration": 0, "maximumAbsoluteError": 0.014907270669937134, "meanAbsoluteError": 9.071585954940839e-05, "passed": true, "reference": "000003.safetensors" }, { "decoderSeconds": 1.222872018814087, "encoderSeconds": 0.6258350610733032, "iteration": 0, "maximumAbsoluteError": 0.012558221817016602, "meanAbsoluteError": 9.717844423500631e-05, "passed": true, "reference": "000004.safetensors" }, { "decoderSeconds": 1.2699329853057861, "encoderSeconds": 0.597944974899292, "iteration": 0, "maximumAbsoluteError": 0.009725838899612427, "meanAbsoluteError": 8.610969884778243e-05, "passed": true, "reference": "000005.safetensors" }, { "decoderSeconds": 1.4274159669876099, "encoderSeconds": 0.6805440187454224, "iteration": 0, "maximumAbsoluteError": 0.015181079506874084, "meanAbsoluteError": 0.00012718189698826012, "passed": true, "reference": "000006.safetensors" }, { "decoderSeconds": 1.415895938873291, "encoderSeconds": 0.6880509853363037, "iteration": 0, "maximumAbsoluteError": 0.014310456812381744, "meanAbsoluteError": 0.00014974610047203637, "passed": true, "reference": "000007.safetensors" } ], "passed": true, "peakPhysicalFootprintBytes": 2055177752 }, "acceptance": { "maximumAbsoluteError": 0.02, "meanAbsoluteError": 0.002 }, "torchPreconversionMaximumAbsoluteError": 0.0, "fusedAttentionPNGByteParityWithBaseline": true, "timingContext": "Six serial eight-frame jobs in baseline/norm/attention then attention/norm/baseline order. No conversion, build or reference inference overlapped the timed jobs. Thermal state progressed from nominal to fair; per-job states are retained. Prefer inference and encoder time to wall time because model compilation/loading differs. Small gains should not be treated as hardware-independent guarantees.", "reproduction": "Use convert-pgtformer-split.py with --encoder-only and the existing decoder. Default code attention is fused; --no-fuse-code-attention reproduces baseline. For the normalization-only trial use --fuse-encoder-norm --no-fuse-code-attention. Independent reference tensors can be saved using validate-pgtformer-video.py --reference-dir and checked with waifu-pgtformer --verify-split. Defaults retain original Conv/BatchNorm arithmetic ordering.", "productValidation": { "passed": true, "casesPassed": 1, "casesFailed": 0, "scope": "Product preview, 1x and 2x reconstruction, source timestamps, export with audio, and cancellation using the configured fused package", "scales": { "1": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 2.650662064552307, "decoderUnits": "gpu", "encoderSeconds": 1.414147973060608, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 2, "inferenceSeconds": 4.0648109912872314, "loadSeconds": 1.7018150091171265, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedCodeAttention": true, "fusedDecoderNorm": true, "fusedEncoderNorm": false, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2094253592, "pythonRuntime": false, "scale": 1, "seconds": 5.810317039489746, "sourceDownsampling": 1, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 1 ] ], "thermalStateAtEnd": 1, "thermalStateAtStart": 1 }, "2": { "backend": "native Swift Core ML / Apple GPU", "centerFrame": 1, "compiledModelCache": true, "decoderSeconds": 10.581934928894043, "decoderUnits": "gpu", "encoderSeconds": 5.071166038513184, "encoderUnits": "gpu", "executionPolicy": "gpu-encoder-decoder-v2", "frames": 8, "inferenceSeconds": 15.653102993965149, "loadSeconds": 0.1875300407409668, "memoryBudgetBytes": 6000000000, "model": "pgtformer", "modelInputShape": [ 3, 3, 512, 512 ], "optimizations": { "centerTail": true, "decoderPrecision": "float16", "fusedCodeAttention": true, "fusedDecoderNorm": true, "fusedEncoderNorm": false, "preservesBatchNormOrdering": true }, "peakPhysicalFootprintBytes": 2053948880, "pythonRuntime": false, "scale": 2, "seconds": 15.96576702594757, "sourceDownsampling": 2, "temporalWindows": [ [ 0, 0, 1 ], [ 0, 1, 2 ], [ 1, 2, 3 ], [ 2, 3, 4 ], [ 3, 4, 5 ], [ 4, 5, 6 ], [ 5, 6, 7 ], [ 6, 7, 7 ] ], "thermalStateAtEnd": 1, "thermalStateAtStart": 1 } } }, "validatedPackageSHA256": { "PGTEncoder.mlpackage/Manifest.json": "8bb5d5b7afc3995c464c3d216012147cc87fbfd2c2d39bfc830b49b40b7e75c8", "PGTEncoder.mlpackage/Data/com.apple.CoreML/model.mlmodel": "2848ef8a8f21df7f6ad2a76149fbe0bd284bb71701bf00c405ec7775284739d2", "PGTEncoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "0de003e90039114b9b12f3df3b980ec41135b50e6f5dfeaec53d9963b30fc1b9", "PGTDecoder.mlpackage/Manifest.json": "367b3e06871404f5cfcfae3068df6b1f161ada54955b8ffa90484b7f296dbe59", "PGTDecoder.mlpackage/Data/com.apple.CoreML/model.mlmodel": "aabcb5ec6daadaefebc0c133a8ab6085e82affe0590127233acd28cd305eeaec", "PGTDecoder.mlpackage/Data/com.apple.CoreML/weights/weight.bin": "a551812f9df33db60b07e743b2097bf805c62aac4ad0b6e5bd9317a50be3b3e5" } }