| { |
| "layerIndex": 28, |
| "artifactPath": "layers/layer_028.elasticbit", |
| "artifactBytes": 325459304, |
| "artifactSha256": "362edbd06b8dad57918a1022fb46e1d20eb08ecd56beb7a3c5764c2ce06575d2", |
| "artifactType": "elasticbit-block-root-v1", |
| "sourceLayerBytesFP16": 681594880, |
| "compressionSeconds": 6.939639329910278, |
| "threshold": 0.01, |
| "device": "cuda:1", |
| "elasticBitCudaDeviceIndex": 1, |
| "matrixSummary": { |
| "count": 4, |
| "histogram": { |
| "8": 2, |
| "7": 1, |
| "9": 1 |
| }, |
| "effectiveBits": 8.153846153846153, |
| "originalBytes": 681574400, |
| "storageBytes": 347555840, |
| "storageReduction": 0.49006911057692304, |
| "maxMatrixError": 0.009819724597036839 |
| }, |
| "isolatedCalibrationValidation": { |
| "relativeL2": 0.003278049962468948, |
| "maxAbsDiff": 0.5 |
| }, |
| "independentValidation": { |
| "relativeL2": 0.0035124642659136052, |
| "maxAbsDiff": 0.25 |
| }, |
| "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": 0.009595238603651524, |
| "storageBytes": 26234880, |
| "executionBytes": 26234880, |
| "originalBytes": 52428800, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.self_attn.qkv_proj", |
| "storageBits": 7, |
| "executionWidth": 8, |
| "error": 0.009819724597036839, |
| "storageBytes": 34437120, |
| "executionBytes": 39352320, |
| "originalBytes": 78643200, |
| "memoryReduction": 0.562109375 |
| }, |
| { |
| "name": "block.mlp.gate_up_proj", |
| "storageBits": 8, |
| "executionWidth": 8, |
| "error": 0.006956745870411396, |
| "storageBytes": 183644160, |
| "executionBytes": 183644160, |
| "originalBytes": 367001600, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.mlp.down_proj", |
| "storageBits": 9, |
| "executionWidth": 16, |
| "error": 0.0051415106281638145, |
| "storageBytes": 103239680, |
| "executionBytes": 183500800, |
| "originalBytes": 183500800, |
| "memoryReduction": 0.4373883928571428 |
| } |
| ] |
| } |
