phi-4-ElasticBit / reports /layer_008.json
maxzameer's picture
ElasticBit layers 008-011 (atomic batch, nextLayer=12)
484aea2 verified
Raw History Blame Contribute Delete
2.47 kB
{
"layerIndex": 8,
"artifactPath": "layers/layer_008.elasticbit",
"artifactBytes": 331504769,
"artifactSha256": "ddaee56f39e36629e511fea764a367621d5692130db90b17dc410f8084bb8e36",
"artifactType": "elasticbit-block-root-v1",
"sourceLayerBytesFP16": 681594880,
"compressionSeconds": 7.562642574310303,
"threshold": 0.01,
"device": "cuda:0",
"elasticBitCudaDeviceIndex": 0,
"matrixSummary": {
"count": 4,
"histogram": {
"9": 2,
"7": 1,
"8": 1
},
"effectiveBits": 8.23076923076923,
"originalBytes": 681574400,
"storageBytes": 350832640,
"storageReduction": 0.48526141826923075,
"maxMatrixError": 0.008863055147230625
},
"isolatedCalibrationValidation": {
"relativeL2": 0.0011334917249192446,
"maxAbsDiff": 0.046875
},
"independentValidation": {
"relativeL2": 0.0011649601996427913,
"maxAbsDiff": 0.0546875
},
"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": 0.005554458126425743,
"storageBytes": 29511680,
"executionBytes": 52428800,
"originalBytes": 52428800,
"memoryReduction": 0.437109375
},
{
"name": "block.self_attn.qkv_proj",
"storageBits": 7,
"executionWidth": 8,
"error": 0.008863055147230625,
"storageBytes": 34437120,
"executionBytes": 39352320,
"originalBytes": 78643200,
"memoryReduction": 0.562109375
},
{
"name": "block.mlp.gate_up_proj",
"storageBits": 8,
"executionWidth": 8,
"error": 0.006947134621441364,
"storageBytes": 183644160,
"executionBytes": 183644160,
"originalBytes": 367001600,
"memoryReduction": 0.499609375
},
{
"name": "block.mlp.down_proj",
"storageBits": 9,
"executionWidth": 16,
"error": 0.004984528291970491,
"storageBytes": 103239680,
"executionBytes": 183500800,
"originalBytes": 183500800,
"memoryReduction": 0.4373883928571428
}
]
}