ElasticBit layers 004-007 (atomic batch, nextLayer=8)

.gitattributes

@@ -37,3 +37,7 @@ layers/layer_000.elasticbit filter=lfs diff=lfs merge=lfs -text
 layers/layer_001.elasticbit filter=lfs diff=lfs merge=lfs -text
 layers/layer_002.elasticbit filter=lfs diff=lfs merge=lfs -text
 layers/layer_003.elasticbit filter=lfs diff=lfs merge=lfs -text
+layers/layer_004.elasticbit filter=lfs diff=lfs merge=lfs -text
+layers/layer_005.elasticbit filter=lfs diff=lfs merge=lfs -text
+layers/layer_006.elasticbit filter=lfs diff=lfs merge=lfs -text
+layers/layer_007.elasticbit filter=lfs diff=lfs merge=lfs -text

layers/layer_004.elasticbit

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6143d75f8520f0475d4c24e26ce7ca55db80522eef1ac3e6855484a284181058
+size 302328429

layers/layer_005.elasticbit

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2a8e918a6947ffb8f41d4afc47fc7ca19345719a4e0822acb08c3239e5875fa6
+size 325176459

layers/layer_006.elasticbit

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc5be9111fb5f22b44edd501b34fe23524722f6d0c7432e4b3e87ede71b1328d
+size 309010735

layers/layer_007.elasticbit

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:921bc85a5336d45dbbf1183f15b9a80c098044b53ec300254e608d6054e11e41
+size 331483376

progress.json

@@ -221,19 +221,210 @@
         "stateValueMismatchCount": 0
       },
       "artifactRoundTripDevice": "cuda:0"
+    },
+    {
+      "layerIndex": 4,
+      "artifactPath": "layers/layer_004.elasticbit",
+      "artifactBytes": 302328429,
+      "artifactSha256": "6143d75f8520f0475d4c24e26ce7ca55db80522eef1ac3e6855484a284181058",
+      "artifactType": "elasticbit-block-root-v1",
+      "sourceLayerBytesFP16": 681594880,
+      "compressionSeconds": 5.131640434265137,
+      "threshold": 0.01,
+      "device": "cuda:0",
+      "elasticBitCudaDeviceIndex": 0,
+      "matrixSummary": {
+        "count": 4,
+        "histogram": {
+          "8": 2,
+          "7": 2
+        },
+        "effectiveBits": 7.615384615384615,
+        "originalBytes": 681574400,
+        "storageBytes": 324618240,
+        "storageReduction": 0.5237229567307693,
+        "maxMatrixError": 0.009336325339972973
+      },
+      "isolatedCalibrationValidation": {
+        "relativeL2": 0.0029243629728438842,
+        "maxAbsDiff": 3.5
+      },
+      "independentValidation": {
+        "relativeL2": 0.0032304916402272135,
+        "maxAbsDiff": 4.0
+      },
+      "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"
+    },
+    {
+      "layerIndex": 5,
+      "artifactPath": "layers/layer_005.elasticbit",
+      "artifactBytes": 325176459,
+      "artifactSha256": "2a8e918a6947ffb8f41d4afc47fc7ca19345719a4e0822acb08c3239e5875fa6",
+      "artifactType": "elasticbit-block-root-v1",
+      "sourceLayerBytesFP16": 681594880,
+      "compressionSeconds": 6.981692552566528,
+      "threshold": 0.01,
+      "device": "cuda:0",
+      "elasticBitCudaDeviceIndex": 0,
+      "matrixSummary": {
+        "count": 4,
+        "histogram": {
+          "8": 2,
+          "7": 1,
+          "9": 1
+        },
+        "effectiveBits": 8.153846153846153,
+        "originalBytes": 681574400,
+        "storageBytes": 347555840,
+        "storageReduction": 0.49006911057692304,
+        "maxMatrixError": 0.009992124512791634
+      },
+      "isolatedCalibrationValidation": {
+        "relativeL2": 0.0009526971167285513,
+        "maxAbsDiff": 0.125
+      },
+      "independentValidation": {
+        "relativeL2": 0.0009863846305689502,
+        "maxAbsDiff": 0.125
+      },
+      "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"
+    },
+    {
+      "layerIndex": 6,
+      "artifactPath": "layers/layer_006.elasticbit",
+      "artifactBytes": 309010735,
+      "artifactSha256": "cc5be9111fb5f22b44edd501b34fe23524722f6d0c7432e4b3e87ede71b1328d",
+      "artifactType": "elasticbit-block-root-v1",
+      "sourceLayerBytesFP16": 681594880,
+      "compressionSeconds": 5.5142552852630615,
+      "threshold": 0.01,
+      "device": "cuda:0",
+      "elasticBitCudaDeviceIndex": 0,
+      "matrixSummary": {
+        "count": 4,
+        "histogram": {
+          "9": 1,
+          "7": 1,
+          "8": 2
+        },
+        "effectiveBits": 7.961538461538462,
+        "originalBytes": 681574400,
+        "storageBytes": 339363840,
+        "storageReduction": 0.5020883413461539,
+        "maxMatrixError": 0.008528505451977253
+      },
+      "isolatedCalibrationValidation": {
+        "relativeL2": 0.0011323165792623885,
+        "maxAbsDiff": 0.5
+      },
+      "independentValidation": {
+        "relativeL2": 0.0012071607504281967,
+        "maxAbsDiff": 0.5
+      },
+      "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"
+    },
+    {
+      "layerIndex": 7,
+      "artifactPath": "layers/layer_007.elasticbit",
+      "artifactBytes": 331483376,
+      "artifactSha256": "921bc85a5336d45dbbf1183f15b9a80c098044b53ec300254e608d6054e11e41",
+      "artifactType": "elasticbit-block-root-v1",
+      "sourceLayerBytesFP16": 681594880,
+      "compressionSeconds": 7.554826021194458,
+      "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": 0.008778146468102932
+      },
+      "isolatedCalibrationValidation": {
+        "relativeL2": 0.0010294616772501977,
+        "maxAbsDiff": 0.5
+      },
+      "independentValidation": {
+        "relativeL2": 0.0010377453582805136,
+        "maxAbsDiff": 0.044921875
+      },
+      "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"
     }
   ],
-  "nextLayer": 4,
+  "nextLayer": 8,
   "reconstructionSmoke": null,
   "complete": false,
   "aggregate": {
-    "completedLayers": 4,
-    "linearOriginalBytes": 2726297600,
-    "linearElasticBitStorageBytes": 1291919360,
-    "linearStorageReduction": 0.5261268028846153,
-    "parameterWeightedEffectiveBits": 7.576923076923077,
-    "layerArtifactBytes": 1185699138,
+    "completedLayers": 8,
+    "linearOriginalBytes": 5452595200,
+    "linearElasticBitStorageBytes": 2654289920,
+    "linearStorageReduction": 0.5132061298076923,
+    "parameterWeightedEffectiveBits": 7.783653846153846,
+    "layerArtifactBytes": 2453698137,
     "globalComponentBytes": 2055219848,
-    "totalUploadedModelBytesApprox": 3240918986
+    "totalUploadedModelBytesApprox": 4508917985
   }
 }

reports/layer_004.json

@@ -0,0 +1,89 @@
+{
+  "layerIndex": 4,
+  "artifactPath": "layers/layer_004.elasticbit",
+  "artifactBytes": 302328429,
+  "artifactSha256": "6143d75f8520f0475d4c24e26ce7ca55db80522eef1ac3e6855484a284181058",
+  "artifactType": "elasticbit-block-root-v1",
+  "sourceLayerBytesFP16": 681594880,
+  "compressionSeconds": 5.131640434265137,
+  "threshold": 0.01,
+  "device": "cuda:0",
+  "elasticBitCudaDeviceIndex": 0,
+  "matrixSummary": {
+    "count": 4,
+    "histogram": {
+      "8": 2,
+      "7": 2
+    },
+    "effectiveBits": 7.615384615384615,
+    "originalBytes": 681574400,
+    "storageBytes": 324618240,
+    "storageReduction": 0.5237229567307693,
+    "maxMatrixError": 0.009336325339972973
+  },
+  "isolatedCalibrationValidation": {
+    "relativeL2": 0.0029243629728438842,
+    "maxAbsDiff": 3.5
+  },
+  "independentValidation": {
+    "relativeL2": 0.0032304916402272135,
+    "maxAbsDiff": 4.0
+  },
+  "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": 0.009336325339972973,
+      "storageBytes": 26234880,
+      "executionBytes": 26234880,
+      "originalBytes": 52428800,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.self_attn.qkv_proj",
+      "storageBits": 7,
+      "executionWidth": 8,
+      "error": 0.0084691122174263,
+      "storageBytes": 34437120,
+      "executionBytes": 39352320,
+      "originalBytes": 78643200,
+      "memoryReduction": 0.562109375
+    },
+    {
+      "name": "block.mlp.gate_up_proj",
+      "storageBits": 8,
+      "executionWidth": 8,
+      "error": 0.007204472087323666,
+      "storageBytes": 183644160,
+      "executionBytes": 183644160,
+      "originalBytes": 367001600,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.mlp.down_proj",
+      "storageBits": 7,
+      "executionWidth": 8,
+      "error": 0.006671952083706856,
+      "storageBytes": 80302080,
+      "executionBytes": 91770880,
+      "originalBytes": 183500800,
+      "memoryReduction": 0.5623883928571429
+    }
+  ]
+}

reports/layer_005.json

@@ -0,0 +1,90 @@
+{
+  "layerIndex": 5,
+  "artifactPath": "layers/layer_005.elasticbit",
+  "artifactBytes": 325176459,
+  "artifactSha256": "2a8e918a6947ffb8f41d4afc47fc7ca19345719a4e0822acb08c3239e5875fa6",
+  "artifactType": "elasticbit-block-root-v1",
+  "sourceLayerBytesFP16": 681594880,
+  "compressionSeconds": 6.981692552566528,
+  "threshold": 0.01,
+  "device": "cuda:0",
+  "elasticBitCudaDeviceIndex": 0,
+  "matrixSummary": {
+    "count": 4,
+    "histogram": {
+      "8": 2,
+      "7": 1,
+      "9": 1
+    },
+    "effectiveBits": 8.153846153846153,
+    "originalBytes": 681574400,
+    "storageBytes": 347555840,
+    "storageReduction": 0.49006911057692304,
+    "maxMatrixError": 0.009992124512791634
+  },
+  "isolatedCalibrationValidation": {
+    "relativeL2": 0.0009526971167285513,
+    "maxAbsDiff": 0.125
+  },
+  "independentValidation": {
+    "relativeL2": 0.0009863846305689502,
+    "maxAbsDiff": 0.125
+  },
+  "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": 0.009992124512791634,
+      "storageBytes": 26234880,
+      "executionBytes": 26234880,
+      "originalBytes": 52428800,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.self_attn.qkv_proj",
+      "storageBits": 7,
+      "executionWidth": 8,
+      "error": 0.008202957920730114,
+      "storageBytes": 34437120,
+      "executionBytes": 39352320,
+      "originalBytes": 78643200,
+      "memoryReduction": 0.562109375
+    },
+    {
+      "name": "block.mlp.gate_up_proj",
+      "storageBits": 8,
+      "executionWidth": 8,
+      "error": 0.0070442804135382175,
+      "storageBytes": 183644160,
+      "executionBytes": 183644160,
+      "originalBytes": 367001600,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.mlp.down_proj",
+      "storageBits": 9,
+      "executionWidth": 16,
+      "error": 0.005074291490018368,
+      "storageBytes": 103239680,
+      "executionBytes": 183500800,
+      "originalBytes": 183500800,
+      "memoryReduction": 0.4373883928571428
+    }
+  ]
+}

reports/layer_006.json

@@ -0,0 +1,90 @@
+{
+  "layerIndex": 6,
+  "artifactPath": "layers/layer_006.elasticbit",
+  "artifactBytes": 309010735,
+  "artifactSha256": "cc5be9111fb5f22b44edd501b34fe23524722f6d0c7432e4b3e87ede71b1328d",
+  "artifactType": "elasticbit-block-root-v1",
+  "sourceLayerBytesFP16": 681594880,
+  "compressionSeconds": 5.5142552852630615,
+  "threshold": 0.01,
+  "device": "cuda:0",
+  "elasticBitCudaDeviceIndex": 0,
+  "matrixSummary": {
+    "count": 4,
+    "histogram": {
+      "9": 1,
+      "7": 1,
+      "8": 2
+    },
+    "effectiveBits": 7.961538461538462,
+    "originalBytes": 681574400,
+    "storageBytes": 339363840,
+    "storageReduction": 0.5020883413461539,
+    "maxMatrixError": 0.008528505451977253
+  },
+  "isolatedCalibrationValidation": {
+    "relativeL2": 0.0011323165792623885,
+    "maxAbsDiff": 0.5
+  },
+  "independentValidation": {
+    "relativeL2": 0.0012071607504281967,
+    "maxAbsDiff": 0.5
+  },
+  "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": 0.0052634370513260365,
+      "storageBytes": 29511680,
+      "executionBytes": 52428800,
+      "originalBytes": 52428800,
+      "memoryReduction": 0.437109375
+    },
+    {
+      "name": "block.self_attn.qkv_proj",
+      "storageBits": 7,
+      "executionWidth": 8,
+      "error": 0.008528505451977253,
+      "storageBytes": 34437120,
+      "executionBytes": 39352320,
+      "originalBytes": 78643200,
+      "memoryReduction": 0.562109375
+    },
+    {
+      "name": "block.mlp.gate_up_proj",
+      "storageBits": 8,
+      "executionWidth": 8,
+      "error": 0.006829055026173592,
+      "storageBytes": 183644160,
+      "executionBytes": 183644160,
+      "originalBytes": 367001600,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.mlp.down_proj",
+      "storageBits": 8,
+      "executionWidth": 8,
+      "error": 0.007873034104704857,
+      "storageBytes": 91770880,
+      "executionBytes": 91770880,
+      "originalBytes": 183500800,
+      "memoryReduction": 0.4998883928571428
+    }
+  ]
+}

reports/layer_007.json

@@ -0,0 +1,90 @@
+{
+  "layerIndex": 7,
+  "artifactPath": "layers/layer_007.elasticbit",
+  "artifactBytes": 331483376,
+  "artifactSha256": "921bc85a5336d45dbbf1183f15b9a80c098044b53ec300254e608d6054e11e41",
+  "artifactType": "elasticbit-block-root-v1",
+  "sourceLayerBytesFP16": 681594880,
+  "compressionSeconds": 7.554826021194458,
+  "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": 0.008778146468102932
+  },
+  "isolatedCalibrationValidation": {
+    "relativeL2": 0.0010294616772501977,
+    "maxAbsDiff": 0.5
+  },
+  "independentValidation": {
+    "relativeL2": 0.0010377453582805136,
+    "maxAbsDiff": 0.044921875
+  },
+  "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": 0.005046782083809376,
+      "storageBytes": 29511680,
+      "executionBytes": 52428800,
+      "originalBytes": 52428800,
+      "memoryReduction": 0.437109375
+    },
+    {
+      "name": "block.self_attn.qkv_proj",
+      "storageBits": 7,
+      "executionWidth": 8,
+      "error": 0.008778146468102932,
+      "storageBytes": 34437120,
+      "executionBytes": 39352320,
+      "originalBytes": 78643200,
+      "memoryReduction": 0.562109375
+    },
+    {
+      "name": "block.mlp.gate_up_proj",
+      "storageBits": 8,
+      "executionWidth": 8,
+      "error": 0.006849483586847782,
+      "storageBytes": 183644160,
+      "executionBytes": 183644160,
+      "originalBytes": 367001600,
+      "memoryReduction": 0.499609375
+    },
+    {
+      "name": "block.mlp.down_proj",
+      "storageBits": 9,
+      "executionWidth": 16,
+      "error": 0.005052048712968826,
+      "storageBytes": 103239680,
+      "executionBytes": 183500800,
+      "originalBytes": 183500800,
+      "memoryReduction": 0.4373883928571428
+    }
+  ]
+}