| { |
| "layerIndex": 2, |
| "artifactPath": "layers/layer_002.elasticbit", |
| "artifactBytes": 286782490, |
| "artifactSha256": "cf55b102dcc51da7edc80ca826a2ad8ea43c79dec1255df8924ffbb9a8a1ecb2", |
| "artifactType": "elasticbit-block-root-v1", |
| "sourceLayerBytesFP16": 681594880, |
| "compressionSeconds": 5.429288625717163, |
| "threshold": 0.01, |
| "device": "cuda:0", |
| "elasticBitCudaDeviceIndex": 0, |
| "matrixSummary": { |
| "count": 4, |
| "histogram": { |
| "8": 2, |
| "7": 1, |
| "6": 1 |
| }, |
| "effectiveBits": 7.346153846153846, |
| "originalBytes": 681574400, |
| "storageBytes": 313149440, |
| "storageReduction": 0.5405498798076923, |
| "maxMatrixError": 0.00957441795617342 |
| }, |
| "isolatedCalibrationValidation": { |
| "relativeL2": 0.0072191223096902055, |
| "maxAbsDiff": 3.875 |
| }, |
| "independentValidation": { |
| "relativeL2": 0.007312040524336951, |
| "maxAbsDiff": 3.625 |
| }, |
| "artifactRoundTripValidation": { |
| "relativeL2": 0.0, |
| "maxAbsDiff": 0.0 |
| }, |
| "artifactRoundTripLowMem": { |
| "placeholderCount": 4, |
| "avoidedDenseBytesApproxFP16": 681574400, |
| "preloadedStateSlots": 2, |
| "movedNonPersistentBuffers": [], |
| "postloadExactStateSlots": 2, |
| "checkedStateDtypes": 2, |
| "dtypeMismatchCount": 0, |
| "stateValueMismatchCount": 0 |
| }, |
| "artifactRoundTripDevice": "cuda:0", |
| "matrices": [ |
| { |
| "name": "block.self_attn.o_proj", |
| "storageBits": 8, |
| "executionWidth": 8, |
| "error": 0.00939763244241476, |
| "storageBytes": 26234880, |
| "executionBytes": 26234880, |
| "originalBytes": 52428800, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.self_attn.qkv_proj", |
| "storageBits": 7, |
| "executionWidth": 8, |
| "error": 0.00753616401925683, |
| "storageBytes": 34437120, |
| "executionBytes": 39352320, |
| "originalBytes": 78643200, |
| "memoryReduction": 0.562109375 |
| }, |
| { |
| "name": "block.mlp.gate_up_proj", |
| "storageBits": 8, |
| "executionWidth": 8, |
| "error": 0.007165043614804745, |
| "storageBytes": 183644160, |
| "executionBytes": 183644160, |
| "originalBytes": 367001600, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.mlp.down_proj", |
| "storageBits": 6, |
| "executionWidth": 8, |
| "error": 0.00957441795617342, |
| "storageBytes": 68833280, |
| "executionBytes": 91770880, |
| "originalBytes": 183500800, |
| "memoryReduction": 0.6248883928571429 |
| } |
| ] |
| } |
