level 3 · 27/63 correct
trajectories
63
finished
63
correct
27
compiled
45
avg speedup (correct)
4.13x
sweep-wide eval metricsaveraged over correct kernels
outcomescorrect 27 · incorrect 18 · compile_fail 18
avg speedup4.13x (over 27 correct)
avg memory ratio (kernel / ref)0.90x (27 samples)
avg fusion ratio (ref launches / kernel launches)32.78x (27 samples)
avg energy ratio (ref / kernel)1.82x (22 samples)
avg SOL score (0–1)0.355 (27 samples)
avg DRAM util %49.7%
avg occupancy %267.7%
avg max abs error6.08e-06
original · 20/50 correct
P1 · gpt-5.5-priority1_MLP.py
compile_fail—
P2 · gpt-5.5-priority2_ShallowWideMLP.py
incorrect—
P3 · gpt-5.5-priority3_DeepNarrowMLP.py
correct1.06x1.7μs
P4 · gpt-5.5-priority4_LeNet5.py
correct1.09x0.9μs
P5 · gpt-5.5-priority5_AlexNet.py
correct1.13x21.5μs
P6 · gpt-5.5-priority6_GoogleNetInceptionModule.py
incorrect—
P7 · gpt-5.5-priority7_GoogleNetInceptionV1.py
correct1.33x1.9μs
P8 · gpt-5.5-priority8_ResNetBasicBlock.py
incorrect—
P9 · gpt-5.5-priority9_ResNet18.py
compile_fail—
P10 · gpt-5.5-priority10_ResNet101.py
correct1.11x8.6μs
P11 · gpt-5.5-priority11_VGG16.py
correct0.61x5.6μs
P12 · gpt-5.5-priority12_VGG19.py
correct1.29x3.0μs
P13 · gpt-5.5-priority13_DenseNet121TransitionLayer.py
incorrect—
P14 · gpt-5.5-priority14_DenseNet121DenseBlock.py
incorrect—
P15 · gpt-5.5-priority15_DenseNet121.py
correct1.03x7.3μs
P16 · gpt-5.5-priority16_DenseNet201.py
compile_fail—
P17 · gpt-5.5-priority17_SqueezeNetFireModule.py
correct1.06x12.6μs
P18 · gpt-5.5-priority18_SqueezeNet.py
correct1.28x22.1μs
P19 · gpt-5.5-priority19_MobileNetV1.py
correct1.03x2.0μs
P20 · gpt-5.5-priority20_MobileNetV2.py
correct0.99x4.7μs
P21 · gpt-5.5-priority21_EfficientNetMBConv.py
compile_fail—
P22 · gpt-5.5-priority22_EfficientNetB0.py
incorrect—
P23 · gpt-5.5-priority23_EfficientNetB1.py
compile_fail—
P24 · gpt-5.5-priority24_EfficientNetB2.py
compile_fail—
P25 · gpt-5.5-priority25_ShuffleNetUnit.py
correct1.05x8.3μs
P26 · gpt-5.5-priority26_ShuffleNet.py
correct1.04x8.4μs
P27 · gpt-5.5-priority27_RegNet.py
correct1.05x2.3μs
P28 · gpt-5.5-priority28_VisionTransformer.py
correct8.57x1.0μs
P29 · gpt-5.5-priority29_SwinMLP.py
compile_fail—
P30 · gpt-5.5-priority30_SwinTransformerV2.py
correct2.06x11.7μs
P31 · gpt-5.5-priority31_VisionAttention.py
incorrect—
P32 · gpt-5.5-priority32_ConvolutionalVisionTransforme
correct1.69x1.4μs
P33 · gpt-5.5-priority33_VanillaRNN.py
correct1.00x6.8μs
P34 · gpt-5.5-priority34_VanillaRNNHidden.py
compile_fail—
P35 · gpt-5.5-priority35_LSTM.py
compile_fail—
P36 · gpt-5.5-priority36_LSTMHn.py
compile_fail—
P37 · gpt-5.5-priority37_LSTMCn.py
compile_fail—
P38 · gpt-5.5-priority38_LSTMBidirectional.py
compile_fail—
P39 · gpt-5.5-priority39_GRU.py
compile_fail—
P40 · gpt-5.5-priority40_GRUHidden.py
incorrect—
P41 · gpt-5.5-priority41_GRUBidirectional.py
compile_fail—
P42 · gpt-5.5-priority42_GRUBidirectionalHidden.py
incorrect—
P43 · gpt-5.5-priority43_MinGPTCausalAttention.py
incorrect—
P44 · gpt-5.5-priority44_MiniGPTBlock.py
incorrect—
P45 · gpt-5.5-priority45_UNetSoftmax.py
incorrect—
P46 · gpt-5.5-priority46_NetVladWithGhostClusters.py
compile_fail—
P47 · gpt-5.5-priority47_NetVladNoGhostClusters.py
correct0.22x6.4μs
P48 · gpt-5.5-priority48_Mamba2ReturnY.py
incorrect—
P49 · gpt-5.5-priority49_Mamba2ReturnFinalState.py
incorrect—
P50 · gpt-5.5-priority50_ReLUSelfAttention.py
incorrect—
popcorn · 7/13 correct
P1 · gpt-5.5-priority1_DilatedConvTower_GenomicBlock.
compile_fail—
P2 · gpt-5.5-priority2_FusedMultiStepDiffusionSolver.
incorrect—
P3 · gpt-5.5-priority3_SSM2D_FusedBidirectionalCrossS
correct20.13x0.0μs
P4 · gpt-5.5-priority4_MambaBlock.py
correct52.27x3.1μs
P5 · gpt-5.5-priority5_FlowMatchingProbabilityFlowODE
compile_fail—
P6 · gpt-5.5-priority6_HyenaOperator.py
correct1.45x3.4μs
P7 · gpt-5.5-priority7_GaussianProcessRegression.py
correct1.06x0.9μs
P8 · gpt-5.5-priority8_EvoformerBlock.py
correct1.33x0.7μs
P9 · gpt-5.5-priority9_BayesianLinearRegression.py
correct2.26x0.1μs
P10 · gpt-5.5-priority10_GaussianMixtureModelEM.py
incorrect—
P11 · gpt-5.5-priority11_pipeline_three_stage_mlp.py
compile_fail—
P12 · gpt-5.5-priority12_NormalizingFlowPlanar.py
incorrect—
P13 · gpt-5.5-priority13_ConditionalVAELoss.py
correct3.22x0.1μs