phi-4-ElasticBit / reports /layer_019.json
maxzameer's picture
ElasticBit layers 016-019 (atomic batch, nextLayer=20)
f2f63d4 verified
Raw History Blame Contribute Delete
2.47 kB
{
"layerIndex": 19,
"artifactPath": "layers/layer_019.elasticbit",
"artifactBytes": 331409192,
"artifactSha256": "5d42ed457116b5494db163ccf0a84c3e964f45924ab3fbcd1e85985d1e37d9f4",
"artifactType": "elasticbit-block-root-v1",
"sourceLayerBytesFP16": 681594880,
"compressionSeconds": 7.416729211807251,
"threshold": 0.01,
"device": "cuda:1",
"elasticBitCudaDeviceIndex": 1,
"matrixSummary": {
"count": 4,
"histogram": {
"9": 2,
"7": 1,
"8": 1
},
"effectiveBits": 8.23076923076923,
"originalBytes": 681574400,
"storageBytes": 350832640,
"storageReduction": 0.48526141826923075,
"maxMatrixError": 0.008992556482553482
},
"isolatedCalibrationValidation": {
"relativeL2": 0.002284017627333564,
"maxAbsDiff": 0.1484375
},
"independentValidation": {
"relativeL2": 0.0023912012136796195,
"maxAbsDiff": 0.140625
},
"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": 9,
"executionWidth": 16,
"error": 0.006902495864778757,
"storageBytes": 29511680,
"executionBytes": 52428800,
"originalBytes": 52428800,
"memoryReduction": 0.437109375
},
{
"name": "block.self_attn.qkv_proj",
"storageBits": 7,
"executionWidth": 8,
"error": 0.008992556482553482,
"storageBytes": 34437120,
"executionBytes": 39352320,
"originalBytes": 78643200,
"memoryReduction": 0.562109375
},
{
"name": "block.mlp.gate_up_proj",
"storageBits": 8,
"executionWidth": 8,
"error": 0.006906770635396242,
"storageBytes": 183644160,
"executionBytes": 183644160,
"originalBytes": 367001600,
"memoryReduction": 0.499609375
},
{
"name": "block.mlp.down_proj",
"storageBits": 9,
"executionWidth": 16,
"error": 0.005733635742217302,
"storageBytes": 103239680,
"executionBytes": 183500800,
"originalBytes": 183500800,
"memoryReduction": 0.4373883928571428
}
]
}