| { |
| "layerIndex": 31, |
| "artifactPath": "layers/layer_031.elasticbit", |
| "artifactBytes": 325697741, |
| "artifactSha256": "e8294534856975c85168549e3495218005a3d5daea3d2d86665db4fb651c1283", |
| "artifactType": "elasticbit-block-root-v1", |
| "sourceLayerBytesFP16": 681594880, |
| "compressionSeconds": 6.678463459014893, |
| "threshold": 0.01, |
| "device": "cuda:1", |
| "elasticBitCudaDeviceIndex": 1, |
| "matrixSummary": { |
| "count": 4, |
| "histogram": { |
| "8": 3, |
| "9": 1 |
| }, |
| "effectiveBits": 8.26923076923077, |
| "originalBytes": 681574400, |
| "storageBytes": 352471040, |
| "storageReduction": 0.4828575721153846, |
| "maxMatrixError": 0.009832173585891724 |
| }, |
| "isolatedCalibrationValidation": { |
| "relativeL2": 0.002334490643799475, |
| "maxAbsDiff": 0.5 |
| }, |
| "independentValidation": { |
| "relativeL2": 0.00226059629353232, |
| "maxAbsDiff": 0.375 |
| }, |
| "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": 0.009832173585891724, |
| "storageBytes": 26234880, |
| "executionBytes": 26234880, |
| "originalBytes": 52428800, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.self_attn.qkv_proj", |
| "storageBits": 8, |
| "executionWidth": 8, |
| "error": 0.00505587225779891, |
| "storageBytes": 39352320, |
| "executionBytes": 39352320, |
| "originalBytes": 78643200, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.mlp.gate_up_proj", |
| "storageBits": 8, |
| "executionWidth": 8, |
| "error": 0.007346144877374172, |
| "storageBytes": 183644160, |
| "executionBytes": 183644160, |
| "originalBytes": 367001600, |
| "memoryReduction": 0.499609375 |
| }, |
| { |
| "name": "block.mlp.down_proj", |
| "storageBits": 9, |
| "executionWidth": 16, |
| "error": 0.005322652403265238, |
| "storageBytes": 103239680, |
| "executionBytes": 183500800, |
| "originalBytes": 183500800, |
| "memoryReduction": 0.4373883928571428 |
| } |
| ] |
| } |
