diff --git a/.gitattributes b/.gitattributes index 11b7704c838d8565ce08b5f3cb2b5af2b94ed814..6bc3707777d9173945467cda0dbb6996c6efd7f9 100644 --- a/.gitattributes +++ b/.gitattributes @@ -5109,3 +5109,4 @@ lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample778-layer4-i lambda0.004/hyperprior-featurecoding-8bit-individual/fc_gsm8k/sample91-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.004/hyperprior-featurecoding-8bit-individual/fc_hellaswag/sample52-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text lambda0.004/hyperprior-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample9-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample90-layer4-item1.zst filter=lfs diff=lfs merge=lfs -text diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample61-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample61-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..de2009e35f11f46227fb15d943139774360386bd --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample61-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fa1065f1d91653c454d278179ee991295181668399d797c09689415fa21c1e5 +size 5825250 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample62-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample62-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..f05d0acc762c3b7354eb775253d7beb7808fbf33 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample62-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6fe6f90dcc6e23c125977740787e9d69778b6156c7e7265149356794606e7620 +size 5345476 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample63-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample63-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..ce363ebbc4958f6b9d13265a180b8c1b202001bd --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample63-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35a36fe9a20f96fe15668372cd19d3945acd4d320e5cd4daf5289aba8aee7ac9 +size 5994081 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample64-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample64-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..d7d22ada689c391022bcdc46827c72fd4eba53c0 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample64-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd7abf3801ffcd7aa36aebba4a00239ee8c8160444f266a7c555156a139d98da +size 5517620 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample65-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample65-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..b86f8bdbfb385af74169f46c41c97d01b4e9a98d --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample65-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:49e042290dd0ad0c9bd0d5ea645ec0f39d9aa6dfb2f2f703fdce2b891d5776be +size 6086896 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample66-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample66-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..ba7710c53e1819b8bff9f8efa8f2500cc086ff8a --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample66-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b64160ea57ffce6ad57471372cadfd7478820a2a8c224b5dc09e2d14f0a9425d +size 5840235 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample67-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample67-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..b7b09c489501873b5a75dfed5aa86f5f44b5c016 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample67-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bc222a7aff278dfe39d93071ccac8deec78f9aee84facb9211dba5325a2cbb9e +size 6161027 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample68-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample68-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..9b4b72516f38daaba9459b3a6de3a184e5242756 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample68-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:870b65564d427a20df77b1e15d7d3fda92ab80f0f88cdcf9b438176afebb6eb1 +size 5040553 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample69-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample69-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..974bd31ae5e0be948e93eb737e70c7a6dd8866b9 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample69-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2fff74eccd75faf9df55fbb63ed004113d790a4f11c52c82f56e4b4b75144f33 +size 5780748 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample7-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample7-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..c8b313029d297239d35ba69787db541b6af665c7 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample7-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69a631c5162e35f90dc9055846a3aa5007efe2d9ef49064a1b3a2473cf4bc838 +size 5844384 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample70-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample70-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..6f0de1d7b733538c0b23786869c8e3a250683a0b --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample70-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c28b4a78f1ecea103b345917c13920be07643e830848e33336bfff79bc8b0b49 +size 5472371 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample71-layer4-item1.zst b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample71-layer4-item1.zst new file mode 100644 index 0000000000000000000000000000000000000000..c9bcc9c7f4e07f8a39154994a7bc87e48c870115 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_hellaswag/sample71-layer4-item1.zst @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b29ad2d02c6346e22650442aa8065e6e0b660265aa978979d36f0f4e07578b3 +size 5614451 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log index a1a1a5642f94db7c060d5e5f8a48e1ecc6337be4..48b039eec2683e958429b1ceff552278b00712ff 100644 --- a/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_elic-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/dtufc_elic-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: elic-featurecoding - handler: qwen - checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar - transform_type: kmeans - transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json - bit_depth: 8 - device: cuda:0 -Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 255 -Loaded elic-featurecoding (1-channel) on cuda:0 -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json -Loaded per-key mappings: model=qwen - Keys: ['layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache', 'layer.4.output'] ----------------- -------------------------------------------------------------------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding -Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Transform type kmeans -Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json -Input ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1 -Output output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1 ----------------- -------------------------------------------------------------------------------------------------------------------- -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 276, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 276, 128) -Output shape: (1, 276, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) -> torch.Size([1, 1, 276, 1024]) - layer.4.output: torch.Size([1, 276, 4096]) -> torch.Size([1, 1, 276, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 5,028B, BPFP=0.1423 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,952B, BPFP=1.6687 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 30,892B, BPFP=0.8744 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,316B, BPFP=1.7922 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 41,868B, BPFP=1.1851 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 63,208B, BPFP=1.7892 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 44,152B, BPFP=1.2498 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 63,044B, BPFP=1.7845 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 30,072B, BPFP=0.8512 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,700B, BPFP=1.8031 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 95,756B, BPFP=0.6776 -⌛️ [2/4] FRONTEND: Frontend time: 2.818s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 276, 128]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.output: torch.Size([1, 276, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.073s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 276, 128]) - layer.0.v_cache: torch.Size([1, 8, 276, 128]) - layer.1.k_cache: torch.Size([1, 8, 276, 128]) - layer.1.v_cache: torch.Size([1, 8, 276, 128]) - layer.2.k_cache: torch.Size([1, 8, 276, 128]) - layer.2.v_cache: torch.Size([1, 8, 276, 128]) - layer.3.k_cache: torch.Size([1, 8, 276, 128]) - layer.3.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.k_cache: torch.Size([1, 8, 276, 128]) - layer.4.v_cache: torch.Size([1, 8, 276, 128]) - layer.4.output: torch.Size([1, 276, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02420060 8.34414894 - layer.0.v_cache 0.00000027 0.00024154 - layer.1.k_cache 0.00304789 1.85440882 - layer.1.v_cache 0.00000080 0.00083202 - layer.2.k_cache 0.00119192 0.62617305 - layer.2.v_cache 0.00000117 0.00123918 - layer.3.k_cache 0.00127914 0.70598348 - layer.3.v_cache 0.00000214 0.00199654 - layer.4.k_cache 0.00374547 1.94008382 - layer.4.v_cache 0.00000302 0.00329724 - layer.4.output 0.00013416 0.07382436 - ------------------------------------------------------------------------------------- - TOTAL 0.00242922 0.98383586 - (elements=3,956,736) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3956736 -Total Bytes 559988 -BPFP 1.1322 bits/point -EBPFP 2.2644 equivalent bits/point -MSE 0.983836 ----------------------- -------------------------------------------------------- -Time: 4.904s Load: 0.013s, Pack+Encode: 2.818s, Decode+Unpack: 2.073s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 276, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 276, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9838 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 283, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.015s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 283, 128) -Output shape: (1, 283, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) -> torch.Size([1, 1, 283, 1024]) - layer.4.output: torch.Size([1, 283, 4096]) -> torch.Size([1, 1, 283, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 5,020B, BPFP=0.1386 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 57,156B, BPFP=1.5778 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 31,612B, BPFP=0.8727 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 61,684B, BPFP=1.7028 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,960B, BPFP=1.0203 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 62,292B, BPFP=1.7196 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 45,592B, BPFP=1.2586 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 61,028B, BPFP=1.6847 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 29,728B, BPFP=0.8207 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,168B, BPFP=1.7438 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 109,264B, BPFP=0.7541 -⌛️ [2/4] FRONTEND: Frontend time: 2.291s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 283, 128]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.output: torch.Size([1, 283, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.977s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 283, 128]) - layer.0.v_cache: torch.Size([1, 8, 283, 128]) - layer.1.k_cache: torch.Size([1, 8, 283, 128]) - layer.1.v_cache: torch.Size([1, 8, 283, 128]) - layer.2.k_cache: torch.Size([1, 8, 283, 128]) - layer.2.v_cache: torch.Size([1, 8, 283, 128]) - layer.3.k_cache: torch.Size([1, 8, 283, 128]) - layer.3.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.k_cache: torch.Size([1, 8, 283, 128]) - layer.4.v_cache: torch.Size([1, 8, 283, 128]) - layer.4.output: torch.Size([1, 283, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02559670 8.43530273 - layer.0.v_cache 0.00000030 0.00026055 - layer.1.k_cache 0.00287003 1.89172385 - layer.1.v_cache 0.00000078 0.00081542 - layer.2.k_cache 0.00118918 0.63965011 - layer.2.v_cache 0.00000127 0.00125248 - layer.3.k_cache 0.00127329 0.70922032 - layer.3.v_cache 0.00000227 0.00206242 - layer.4.k_cache 0.00368869 1.82069920 - layer.4.v_cache 0.00000310 0.00329566 - layer.4.output 0.00017183 0.08491556 - ------------------------------------------------------------------------------------- - TOTAL 0.00252235 0.98885321 - (elements=4,057,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 4057088 -Total Bytes 563504 -BPFP 1.1111 bits/point -EBPFP 2.2223 equivalent bits/point -MSE 0.988853 ----------------------- -------------------------------------------------------- -Time: 4.283s Load: 0.015s, Pack+Encode: 2.291s, Decode+Unpack: 1.977s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 283, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 283, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9889 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample10-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample10-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,016B, BPFP=0.1491 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,480B, BPFP=1.7049 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,804B, BPFP=0.9298 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 36,976B, BPFP=1.8283 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,212B, BPFP=1.2961 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 37,528B, BPFP=1.8556 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,816B, BPFP=1.3754 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 36,480B, BPFP=1.8038 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,824B, BPFP=0.9802 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,556B, BPFP=1.8570 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 73,932B, BPFP=0.9139 -⌛️ [2/4] FRONTEND: Frontend time: 2.199s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.441s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02645765 8.64982122 - layer.0.v_cache 0.00000027 0.00025499 - layer.1.k_cache 0.00311677 1.52423550 - layer.1.v_cache 0.00000085 0.00086764 - layer.2.k_cache 0.00117315 0.65898412 - layer.2.v_cache 0.00000112 0.00132963 - layer.3.k_cache 0.00135825 0.77194088 - layer.3.v_cache 0.00000212 0.00212811 - layer.4.k_cache 0.00352137 1.94804672 - layer.4.v_cache 0.00000299 0.00344926 - layer.4.output 0.00021777 0.10408457 - ------------------------------------------------------------------------------------- - TOTAL 0.00260754 0.99838545 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 352624 -BPFP 1.2454 bits/point -EBPFP 2.4908 equivalent bits/point -MSE 0.998385 ----------------------- -------------------------------------------------------- -Time: 3.648s Load: 0.008s, Pack+Encode: 2.199s, Decode+Unpack: 1.441s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9984 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 154, 128) -Output shape: (1, 154, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.output: torch.Size([1, 154, 4096]) -> torch.Size([1, 1, 154, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,944B, BPFP=0.1494 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,520B, BPFP=1.8019 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,180B, BPFP=0.9223 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,900B, BPFP=1.9734 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,172B, BPFP=1.2263 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,964B, BPFP=1.9767 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,172B, BPFP=1.4799 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,740B, BPFP=1.9146 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,712B, BPFP=0.9493 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,964B, BPFP=1.9767 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,184B, BPFP=0.9028 -⌛️ [2/4] FRONTEND: Frontend time: 1.986s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.494s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02866718 8.66026802 - layer.0.v_cache 0.00000026 0.00025233 - layer.1.k_cache 0.00317794 1.69384171 - layer.1.v_cache 0.00000104 0.00091901 - layer.2.k_cache 0.00114871 0.63062385 - layer.2.v_cache 0.00000119 0.00136465 - layer.3.k_cache 0.00136607 0.77737263 - layer.3.v_cache 0.00000217 0.00220822 - layer.4.k_cache 0.00338641 1.76709975 - layer.4.v_cache 0.00000303 0.00353066 - layer.4.output 0.00020569 0.10372582 - ------------------------------------------------------------------------------------- - TOTAL 0.00275548 0.99659886 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 354452 -BPFP 1.2844 bits/point -EBPFP 2.5688 equivalent bits/point -MSE 0.996599 ----------------------- -------------------------------------------------------- -Time: 3.489s Load: 0.009s, Pack+Encode: 1.986s, Decode+Unpack: 1.494s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9966 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample101-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample101-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 150, 128) -Output shape: (1, 150, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.output: torch.Size([1, 150, 4096]) -> torch.Size([1, 1, 150, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,964B, BPFP=0.1544 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,452B, BPFP=1.8985 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 20,192B, BPFP=1.0517 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,524B, BPFP=2.1627 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,292B, BPFP=1.2652 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 43,808B, BPFP=2.2817 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,364B, BPFP=1.5294 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,368B, BPFP=2.1025 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,848B, BPFP=1.0858 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,324B, BPFP=2.2044 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 92,160B, BPFP=1.2000 -⌛️ [2/4] FRONTEND: Frontend time: 1.908s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.622s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02593268 8.90618815 - layer.0.v_cache 0.00000026 0.00024887 - layer.1.k_cache 0.00305171 1.63631246 - layer.1.v_cache 0.00000092 0.00096403 - layer.2.k_cache 0.00119162 0.63098017 - layer.2.v_cache 0.00000123 0.00145051 - layer.3.k_cache 0.00128981 0.74613098 - layer.3.v_cache 0.00000247 0.00232788 - layer.4.k_cache 0.00345981 1.86082479 - layer.4.v_cache 0.00000333 0.00374501 - layer.4.output 0.00015266 0.09227080 - ------------------------------------------------------------------------------------- - TOTAL 0.00253889 1.01130400 - (elements=2,150,400) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2150400 -Total Bytes 394296 -BPFP 1.4669 bits/point -EBPFP 2.9337 equivalent bits/point -MSE 1.011304 ----------------------- -------------------------------------------------------- -Time: 3.539s Load: 0.009s, Pack+Encode: 1.908s, Decode+Unpack: 1.622s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0113 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample102-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample102-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,152B, BPFP=0.1457 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,552B, BPFP=1.6897 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,164B, BPFP=0.7935 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,896B, BPFP=1.7518 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,032B, BPFP=1.1572 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,500B, BPFP=1.7798 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,088B, BPFP=1.2522 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,516B, BPFP=1.7343 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,016B, BPFP=0.9715 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,504B, BPFP=1.7800 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 64,760B, BPFP=0.7484 -⌛️ [2/4] FRONTEND: Frontend time: 1.935s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.435s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02810763 9.28997983 - layer.0.v_cache 0.00000027 0.00025387 - layer.1.k_cache 0.00304808 1.65358883 - layer.1.v_cache 0.00000079 0.00091538 - layer.2.k_cache 0.00115532 0.66906558 - layer.2.v_cache 0.00000117 0.00136309 - layer.3.k_cache 0.00130126 0.76209345 - layer.3.v_cache 0.00000227 0.00218565 - layer.4.k_cache 0.00349612 1.79958619 - layer.4.v_cache 0.00000313 0.00359875 - layer.4.output 0.00017973 0.09068472 - ------------------------------------------------------------------------------------- - TOTAL 0.00270250 1.03895496 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 347180 -BPFP 1.1464 bits/point -EBPFP 2.2928 equivalent bits/point -MSE 1.038955 ----------------------- -------------------------------------------------------- -Time: 3.379s Load: 0.009s, Pack+Encode: 1.935s, Decode+Unpack: 1.435s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0390 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample103-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample103-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 150, 128) -Output shape: (1, 150, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) -> torch.Size([1, 1, 150, 1024]) - layer.4.output: torch.Size([1, 150, 4096]) -> torch.Size([1, 1, 150, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,996B, BPFP=0.1560 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,556B, BPFP=1.9040 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 20,408B, BPFP=1.0629 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,936B, BPFP=2.1842 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 23,180B, BPFP=1.2073 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,668B, BPFP=2.3265 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,272B, BPFP=1.5246 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,472B, BPFP=2.2642 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,736B, BPFP=1.0800 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 43,908B, BPFP=2.2869 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 69,128B, BPFP=0.9001 -⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.514s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 150, 128]) - layer.0.v_cache: torch.Size([1, 8, 150, 128]) - layer.1.k_cache: torch.Size([1, 8, 150, 128]) - layer.1.v_cache: torch.Size([1, 8, 150, 128]) - layer.2.k_cache: torch.Size([1, 8, 150, 128]) - layer.2.v_cache: torch.Size([1, 8, 150, 128]) - layer.3.k_cache: torch.Size([1, 8, 150, 128]) - layer.3.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.k_cache: torch.Size([1, 8, 150, 128]) - layer.4.v_cache: torch.Size([1, 8, 150, 128]) - layer.4.output: torch.Size([1, 150, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02945780 8.92136882 - layer.0.v_cache 0.00000027 0.00024997 - layer.1.k_cache 0.00309326 1.53333883 - layer.1.v_cache 0.00000086 0.00089856 - layer.2.k_cache 0.00115289 0.65416463 - layer.2.v_cache 0.00000113 0.00131001 - layer.3.k_cache 0.00132321 0.75035746 - layer.3.v_cache 0.00000214 0.00208529 - layer.4.k_cache 0.00338825 1.74421509 - layer.4.v_cache 0.00000324 0.00364681 - layer.4.output 0.00014331 0.09532045 - ------------------------------------------------------------------------------------- - TOTAL 0.00278545 0.99949409 - (elements=2,150,400) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2150400 -Total Bytes 376260 -BPFP 1.3998 bits/point -EBPFP 2.7996 equivalent bits/point -MSE 0.999494 ----------------------- -------------------------------------------------------- -Time: 3.486s Load: 0.009s, Pack+Encode: 1.962s, Decode+Unpack: 1.514s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 150, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 150, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9995 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample105-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample105-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 178, 128) -Output shape: (1, 178, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.output: torch.Size([1, 178, 4096]) -> torch.Size([1, 1, 178, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,264B, BPFP=0.1433 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,524B, BPFP=1.5153 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,068B, BPFP=0.7930 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,676B, BPFP=1.6536 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,480B, BPFP=1.0744 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,640B, BPFP=1.6959 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,288B, BPFP=1.2416 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,512B, BPFP=1.6464 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,096B, BPFP=0.8381 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,680B, BPFP=1.6977 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,808B, BPFP=0.8757 -⌛️ [2/4] FRONTEND: Frontend time: 1.882s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.628s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02793351 8.95456447 - layer.0.v_cache 0.00000028 0.00026188 - layer.1.k_cache 0.00336537 1.66309828 - layer.1.v_cache 0.00000081 0.00084256 - layer.2.k_cache 0.00115814 0.64185998 - layer.2.v_cache 0.00000111 0.00127967 - layer.3.k_cache 0.00131051 0.76508186 - layer.3.v_cache 0.00000213 0.00209271 - layer.4.k_cache 0.00357284 1.93501316 - layer.4.v_cache 0.00000305 0.00336388 - layer.4.output 0.00019607 0.10497372 - ------------------------------------------------------------------------------------- - TOTAL 0.00272372 1.02766809 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 360036 -BPFP 1.1287 bits/point -EBPFP 2.2574 equivalent bits/point -MSE 1.027668 ----------------------- -------------------------------------------------------- -Time: 3.520s Load: 0.009s, Pack+Encode: 1.882s, Decode+Unpack: 1.628s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0277 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample106-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample106-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 217, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 217, 128) -Output shape: (1, 217, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) -> torch.Size([1, 1, 217, 1024]) - layer.4.output: torch.Size([1, 217, 4096]) -> torch.Size([1, 1, 217, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,928B, BPFP=0.1414 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,484B, BPFP=1.7095 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,364B, BPFP=0.9132 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,612B, BPFP=1.8221 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,756B, BPFP=1.2153 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 51,036B, BPFP=1.8374 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,572B, BPFP=1.3167 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 51,084B, BPFP=1.8391 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 27,260B, BPFP=0.9814 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 50,788B, BPFP=1.8285 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 141,140B, BPFP=1.2703 -⌛️ [2/4] FRONTEND: Frontend time: 2.551s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 217, 128]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.output: torch.Size([1, 217, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.903s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 217, 128]) - layer.0.v_cache: torch.Size([1, 8, 217, 128]) - layer.1.k_cache: torch.Size([1, 8, 217, 128]) - layer.1.v_cache: torch.Size([1, 8, 217, 128]) - layer.2.k_cache: torch.Size([1, 8, 217, 128]) - layer.2.v_cache: torch.Size([1, 8, 217, 128]) - layer.3.k_cache: torch.Size([1, 8, 217, 128]) - layer.3.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.k_cache: torch.Size([1, 8, 217, 128]) - layer.4.v_cache: torch.Size([1, 8, 217, 128]) - layer.4.output: torch.Size([1, 217, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02570072 8.09960206 - layer.0.v_cache 0.00000028 0.00024837 - layer.1.k_cache 0.00315590 1.73137781 - layer.1.v_cache 0.00000089 0.00092261 - layer.2.k_cache 0.00124266 0.64704832 - layer.2.v_cache 0.00000119 0.00136371 - layer.3.k_cache 0.00129581 0.73869626 - layer.3.v_cache 0.00000363 0.00239240 - layer.4.k_cache 0.00355111 1.93461925 - layer.4.v_cache 0.00000379 0.00366850 - layer.4.output 0.00017838 0.09621535 - ------------------------------------------------------------------------------------- - TOTAL 0.00254782 0.96748576 - (elements=3,110,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3110912 -Total Bytes 519024 -BPFP 1.3347 bits/point -EBPFP 2.6694 equivalent bits/point -MSE 0.967486 ----------------------- -------------------------------------------------------- -Time: 4.466s Load: 0.012s, Pack+Encode: 2.551s, Decode+Unpack: 1.903s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 217, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 217, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9675 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample11-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample11-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,240B, BPFP=0.1463 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,140B, BPFP=1.5869 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,732B, BPFP=0.8459 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,228B, BPFP=1.7263 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,372B, BPFP=1.1909 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,988B, BPFP=1.7607 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,588B, BPFP=1.2910 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,728B, BPFP=1.7038 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,620B, BPFP=0.8409 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,448B, BPFP=1.7363 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 98,248B, BPFP=1.1092 -⌛️ [2/4] FRONTEND: Frontend time: 1.991s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.477s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02750326 8.21713274 - layer.0.v_cache 0.00000027 0.00026111 - layer.1.k_cache 0.00310050 1.72249043 - layer.1.v_cache 0.00000106 0.00090349 - layer.2.k_cache 0.00119927 0.64151080 - layer.2.v_cache 0.00000121 0.00139140 - layer.3.k_cache 0.00131341 0.76432439 - layer.3.v_cache 0.00000232 0.00222107 - layer.4.k_cache 0.00354461 1.61763803 - layer.4.v_cache 0.00000318 0.00353761 - layer.4.output 0.00022996 0.10585980 - ------------------------------------------------------------------------------------- - TOTAL 0.00268492 0.95677502 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 382332 -BPFP 1.2333 bits/point -EBPFP 2.4665 equivalent bits/point -MSE 0.956775 ----------------------- -------------------------------------------------------- -Time: 3.478s Load: 0.010s, Pack+Encode: 1.991s, Decode+Unpack: 1.477s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9568 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample110-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample110-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 178, 128) -Output shape: (1, 178, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.output: torch.Size([1, 178, 4096]) -> torch.Size([1, 1, 178, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,300B, BPFP=0.1448 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,308B, BPFP=1.5497 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,388B, BPFP=0.8071 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,248B, BPFP=1.6787 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,020B, BPFP=1.1420 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,756B, BPFP=1.7010 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,224B, BPFP=1.2388 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,908B, BPFP=1.6638 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,532B, BPFP=0.8134 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,932B, BPFP=1.7087 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,900B, BPFP=1.1510 -⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.589s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02635046 8.83215469 - layer.0.v_cache 0.00000029 0.00025969 - layer.1.k_cache 0.00306277 1.60615900 - layer.1.v_cache 0.00000083 0.00091535 - layer.2.k_cache 0.00121657 0.65669328 - layer.2.v_cache 0.00000121 0.00132192 - layer.3.k_cache 0.00128658 0.74986456 - layer.3.v_cache 0.00000219 0.00215501 - layer.4.k_cache 0.00351850 1.73522932 - layer.4.v_cache 0.00000322 0.00354779 - layer.4.output 0.00018894 0.10197869 - ------------------------------------------------------------------------------------- - TOTAL 0.00258560 0.99972967 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 388516 -BPFP 1.2180 bits/point -EBPFP 2.4360 equivalent bits/point -MSE 0.999730 ----------------------- -------------------------------------------------------- -Time: 3.507s Load: 0.009s, Pack+Encode: 1.909s, Decode+Unpack: 1.589s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9997 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample111-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample111-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 164, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 164, 128) -Output shape: (1, 164, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) -> torch.Size([1, 1, 164, 1024]) - layer.4.output: torch.Size([1, 164, 4096]) -> torch.Size([1, 1, 164, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,092B, BPFP=0.1473 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,720B, BPFP=1.6540 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,796B, BPFP=0.8478 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,628B, BPFP=1.8401 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,012B, BPFP=1.1439 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,164B, BPFP=1.8180 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,204B, BPFP=1.3436 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,316B, BPFP=1.7776 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,768B, BPFP=0.9893 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,300B, BPFP=1.8245 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 66,312B, BPFP=0.7897 -⌛️ [2/4] FRONTEND: Frontend time: 1.920s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 164, 128]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.output: torch.Size([1, 164, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.533s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 164, 128]) - layer.0.v_cache: torch.Size([1, 8, 164, 128]) - layer.1.k_cache: torch.Size([1, 8, 164, 128]) - layer.1.v_cache: torch.Size([1, 8, 164, 128]) - layer.2.k_cache: torch.Size([1, 8, 164, 128]) - layer.2.v_cache: torch.Size([1, 8, 164, 128]) - layer.3.k_cache: torch.Size([1, 8, 164, 128]) - layer.3.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.k_cache: torch.Size([1, 8, 164, 128]) - layer.4.v_cache: torch.Size([1, 8, 164, 128]) - layer.4.output: torch.Size([1, 164, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02515400 9.07289831 - layer.0.v_cache 0.00000027 0.00025588 - layer.1.k_cache 0.00312787 1.51987290 - layer.1.v_cache 0.00000077 0.00088273 - layer.2.k_cache 0.00115327 0.66164445 - layer.2.v_cache 0.00000115 0.00133927 - layer.3.k_cache 0.00131309 0.78271345 - layer.3.v_cache 0.00000211 0.00217240 - layer.4.k_cache 0.00347769 1.87285595 - layer.4.v_cache 0.00000342 0.00376397 - layer.4.output 0.00018780 0.09579995 - ------------------------------------------------------------------------------------- - TOTAL 0.00249892 1.02154279 - (elements=2,351,104) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2351104 -Total Bytes 347312 -BPFP 1.1818 bits/point -EBPFP 2.3636 equivalent bits/point -MSE 1.021543 ----------------------- -------------------------------------------------------- -Time: 3.462s Load: 0.009s, Pack+Encode: 1.920s, Decode+Unpack: 1.533s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 164, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 164, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0215 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample112-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample112-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 144, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 144, 128) -Output shape: (1, 144, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) -> torch.Size([1, 1, 144, 1024]) - layer.4.output: torch.Size([1, 144, 4096]) -> torch.Size([1, 1, 144, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,852B, BPFP=0.1547 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 39,760B, BPFP=2.1571 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,428B, BPFP=1.0540 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 48,104B, BPFP=2.6098 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 23,448B, BPFP=1.2721 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 41,872B, BPFP=2.2717 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,356B, BPFP=1.5927 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,704B, BPFP=2.3711 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,312B, BPFP=0.9935 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 46,232B, BPFP=2.5082 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 70,584B, BPFP=0.9574 -⌛️ [2/4] FRONTEND: Frontend time: 1.960s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 144, 128]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.output: torch.Size([1, 144, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.426s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 144, 128]) - layer.0.v_cache: torch.Size([1, 8, 144, 128]) - layer.1.k_cache: torch.Size([1, 8, 144, 128]) - layer.1.v_cache: torch.Size([1, 8, 144, 128]) - layer.2.k_cache: torch.Size([1, 8, 144, 128]) - layer.2.v_cache: torch.Size([1, 8, 144, 128]) - layer.3.k_cache: torch.Size([1, 8, 144, 128]) - layer.3.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.k_cache: torch.Size([1, 8, 144, 128]) - layer.4.v_cache: torch.Size([1, 8, 144, 128]) - layer.4.output: torch.Size([1, 144, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02690511 8.90126885 - layer.0.v_cache 0.00000027 0.00026024 - layer.1.k_cache 0.00313749 1.75544294 - layer.1.v_cache 0.00000108 0.00096884 - layer.2.k_cache 0.00118318 0.66132742 - layer.2.v_cache 0.00000114 0.00136441 - layer.3.k_cache 0.00131682 0.75892417 - layer.3.v_cache 0.00000220 0.00226377 - layer.4.k_cache 0.00346269 1.69654761 - layer.4.v_cache 0.00000333 0.00367693 - layer.4.output 0.00016306 0.09957916 - ------------------------------------------------------------------------------------- - TOTAL 0.00261897 1.01288299 - (elements=2,064,384) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2064384 -Total Bytes 383652 -BPFP 1.4867 bits/point -EBPFP 2.9735 equivalent bits/point -MSE 1.012883 ----------------------- -------------------------------------------------------- -Time: 3.394s Load: 0.008s, Pack+Encode: 1.960s, Decode+Unpack: 1.426s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 144, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 144, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0129 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample114-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample114-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,204B, BPFP=0.1464 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,580B, BPFP=1.6255 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,584B, BPFP=0.8490 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,140B, BPFP=1.7882 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,628B, BPFP=1.1709 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,716B, BPFP=1.7688 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,268B, BPFP=1.2915 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,312B, BPFP=1.7047 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,792B, BPFP=0.9956 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,696B, BPFP=1.7679 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 68,920B, BPFP=0.7872 -⌛️ [2/4] FRONTEND: Frontend time: 1.912s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.596s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02584258 8.26071363 - layer.0.v_cache 0.00000026 0.00024404 - layer.1.k_cache 0.00320420 1.53822764 - layer.1.v_cache 0.00000080 0.00091174 - layer.2.k_cache 0.00119855 0.65969684 - layer.2.v_cache 0.00000108 0.00126356 - layer.3.k_cache 0.00129062 0.75439221 - layer.3.v_cache 0.00000212 0.00215915 - layer.4.k_cache 0.00351305 1.92048368 - layer.4.v_cache 0.00000315 0.00361473 - layer.4.output 0.00016406 0.09551655 - ------------------------------------------------------------------------------------- - TOTAL 0.00255090 0.96598382 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 355840 -BPFP 1.1612 bits/point -EBPFP 2.3225 equivalent bits/point -MSE 0.965984 ----------------------- -------------------------------------------------------- -Time: 3.518s Load: 0.010s, Pack+Encode: 1.912s, Decode+Unpack: 1.596s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9660 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample116-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample116-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,216B, BPFP=0.1436 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,928B, BPFP=1.6039 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,516B, BPFP=0.8266 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,032B, BPFP=1.6979 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,984B, BPFP=1.1600 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,084B, BPFP=1.7002 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,076B, BPFP=1.2534 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,588B, BPFP=1.6780 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,180B, BPFP=0.8116 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,264B, BPFP=1.7082 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 73,524B, BPFP=0.8206 -⌛️ [2/4] FRONTEND: Frontend time: 1.926s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.445s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02705980 8.74973493 - layer.0.v_cache 0.00000027 0.00025723 - layer.1.k_cache 0.00311821 1.72464669 - layer.1.v_cache 0.00000085 0.00089160 - layer.2.k_cache 0.00114782 0.66488687 - layer.2.v_cache 0.00000111 0.00129964 - layer.3.k_cache 0.00131325 0.76775391 - layer.3.v_cache 0.00000221 0.00219223 - layer.4.k_cache 0.00342930 1.76234253 - layer.4.v_cache 0.00000305 0.00358577 - layer.4.output 0.00022257 0.10965068 - ------------------------------------------------------------------------------------- - TOTAL 0.00264044 1.00829958 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 355392 -BPFP 1.1333 bits/point -EBPFP 2.2665 equivalent bits/point -MSE 1.008300 ----------------------- -------------------------------------------------------- -Time: 3.380s Load: 0.009s, Pack+Encode: 1.926s, Decode+Unpack: 1.445s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0083 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample117-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample117-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 218, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 218, 128) -Output shape: (1, 218, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) -> torch.Size([1, 1, 218, 1024]) - layer.4.output: torch.Size([1, 218, 4096]) -> torch.Size([1, 1, 218, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,940B, BPFP=0.1412 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,480B, BPFP=1.7015 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,192B, BPFP=0.9028 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,988B, BPFP=1.8273 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,136B, BPFP=1.1875 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 50,864B, BPFP=1.8228 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,712B, BPFP=1.3157 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,840B, BPFP=1.7861 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,732B, BPFP=0.9580 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 51,200B, BPFP=1.8349 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 85,180B, BPFP=0.7632 -⌛️ [2/4] FRONTEND: Frontend time: 2.473s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 218, 128]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.output: torch.Size([1, 218, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.802s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 218, 128]) - layer.0.v_cache: torch.Size([1, 8, 218, 128]) - layer.1.k_cache: torch.Size([1, 8, 218, 128]) - layer.1.v_cache: torch.Size([1, 8, 218, 128]) - layer.2.k_cache: torch.Size([1, 8, 218, 128]) - layer.2.v_cache: torch.Size([1, 8, 218, 128]) - layer.3.k_cache: torch.Size([1, 8, 218, 128]) - layer.3.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.k_cache: torch.Size([1, 8, 218, 128]) - layer.4.v_cache: torch.Size([1, 8, 218, 128]) - layer.4.output: torch.Size([1, 218, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02786819 8.07473615 - layer.0.v_cache 0.00000028 0.00024535 - layer.1.k_cache 0.00304396 1.69908884 - layer.1.v_cache 0.00000081 0.00092191 - layer.2.k_cache 0.00116087 0.64741446 - layer.2.v_cache 0.00000117 0.00131882 - layer.3.k_cache 0.00132673 0.76908854 - layer.3.v_cache 0.00000218 0.00214160 - layer.4.k_cache 0.00351200 1.96618246 - layer.4.v_cache 0.00000322 0.00355305 - layer.4.output 0.00018944 0.09149617 - ------------------------------------------------------------------------------------- - TOTAL 0.00269123 0.96647685 - (elements=3,125,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3125248 -Total Bytes 461264 -BPFP 1.1807 bits/point -EBPFP 2.3615 equivalent bits/point -MSE 0.966477 ----------------------- -------------------------------------------------------- -Time: 4.287s Load: 0.012s, Pack+Encode: 2.473s, Decode+Unpack: 1.802s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 218, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 218, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9665 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample12-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample12-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,164B, BPFP=0.1446 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,524B, BPFP=1.6230 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,204B, BPFP=0.8774 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,388B, BPFP=1.7995 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,852B, BPFP=1.2268 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,500B, BPFP=1.7590 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,048B, BPFP=1.2814 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,624B, BPFP=1.7189 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,868B, BPFP=0.8620 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,240B, BPFP=1.7471 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 65,600B, BPFP=0.7493 -⌛️ [2/4] FRONTEND: Frontend time: 1.958s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.584s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02655128 9.20851251 - layer.0.v_cache 0.00000026 0.00025268 - layer.1.k_cache 0.00304961 1.69513689 - layer.1.v_cache 0.00000080 0.00088642 - layer.2.k_cache 0.00119340 0.67704090 - layer.2.v_cache 0.00000107 0.00126157 - layer.3.k_cache 0.00127930 0.76318350 - layer.3.v_cache 0.00000211 0.00206164 - layer.4.k_cache 0.00352197 1.70310715 - layer.4.v_cache 0.00000299 0.00341608 - layer.4.output 0.00016153 0.08817701 - ------------------------------------------------------------------------------------- - TOTAL 0.00258921 1.02911196 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 351012 -BPFP 1.1455 bits/point -EBPFP 2.2910 equivalent bits/point -MSE 1.029112 ----------------------- -------------------------------------------------------- -Time: 3.551s Load: 0.009s, Pack+Encode: 1.958s, Decode+Unpack: 1.584s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0291 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample120-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample120-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 190, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 190, 128) -Output shape: (1, 190, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) -> torch.Size([1, 1, 190, 1024]) - layer.4.output: torch.Size([1, 190, 4096]) -> torch.Size([1, 1, 190, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,552B, BPFP=0.1461 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 33,664B, BPFP=1.3842 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,784B, BPFP=0.7312 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 36,800B, BPFP=1.5132 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,980B, BPFP=1.0683 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 37,784B, BPFP=1.5536 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,188B, BPFP=1.1590 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 36,884B, BPFP=1.5166 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,608B, BPFP=0.7651 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,060B, BPFP=1.5650 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,252B, BPFP=0.8044 -⌛️ [2/4] FRONTEND: Frontend time: 1.881s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 190, 128]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.output: torch.Size([1, 190, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.509s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 190, 128]) - layer.0.v_cache: torch.Size([1, 8, 190, 128]) - layer.1.k_cache: torch.Size([1, 8, 190, 128]) - layer.1.v_cache: torch.Size([1, 8, 190, 128]) - layer.2.k_cache: torch.Size([1, 8, 190, 128]) - layer.2.v_cache: torch.Size([1, 8, 190, 128]) - layer.3.k_cache: torch.Size([1, 8, 190, 128]) - layer.3.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.k_cache: torch.Size([1, 8, 190, 128]) - layer.4.v_cache: torch.Size([1, 8, 190, 128]) - layer.4.output: torch.Size([1, 190, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02682430 9.10809776 - layer.0.v_cache 0.00000028 0.00027233 - layer.1.k_cache 0.00318839 1.84535362 - layer.1.v_cache 0.00000076 0.00081040 - layer.2.k_cache 0.00117212 0.69659576 - layer.2.v_cache 0.00000110 0.00118052 - layer.3.k_cache 0.00129629 0.81928743 - layer.3.v_cache 0.00000212 0.00203685 - layer.4.k_cache 0.00366745 1.98752474 - layer.4.v_cache 0.00000296 0.00308916 - layer.4.output 0.00018717 0.10432664 - ------------------------------------------------------------------------------------- - TOTAL 0.00263603 1.06296822 - (elements=2,723,840) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2723840 -Total Bytes 355556 -BPFP 1.0443 bits/point -EBPFP 2.0886 equivalent bits/point -MSE 1.062968 ----------------------- -------------------------------------------------------- -Time: 3.401s Load: 0.011s, Pack+Encode: 1.881s, Decode+Unpack: 1.509s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 190, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 190, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0630 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample122-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample122-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,152B, BPFP=0.1432 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,348B, BPFP=1.6056 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,888B, BPFP=0.8125 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,556B, BPFP=1.7059 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,864B, BPFP=1.1294 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,252B, BPFP=1.7375 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,948B, BPFP=1.2694 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,152B, BPFP=1.6875 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,468B, BPFP=0.8843 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,156B, BPFP=1.7331 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 74,532B, BPFP=0.8463 -⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.438s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02742122 9.40959416 - layer.0.v_cache 0.00000027 0.00025589 - layer.1.k_cache 0.00311153 1.66462956 - layer.1.v_cache 0.00000078 0.00086935 - layer.2.k_cache 0.00118723 0.64620572 - layer.2.v_cache 0.00000107 0.00127928 - layer.3.k_cache 0.00131758 0.75495733 - layer.3.v_cache 0.00000212 0.00210827 - layer.4.k_cache 0.00355508 1.78747488 - layer.4.v_cache 0.00000305 0.00348713 - layer.4.output 0.00016643 0.09833274 - ------------------------------------------------------------------------------------- - TOTAL 0.00266183 1.04744232 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 354316 -BPFP 1.1495 bits/point -EBPFP 2.2991 equivalent bits/point -MSE 1.047442 ----------------------- -------------------------------------------------------- -Time: 3.417s Load: 0.011s, Pack+Encode: 1.968s, Decode+Unpack: 1.438s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0474 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample125-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample125-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,116B, BPFP=0.1475 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,636B, BPFP=1.6873 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,452B, BPFP=0.8737 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,760B, BPFP=1.8352 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,328B, BPFP=1.2466 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,520B, BPFP=1.8239 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,636B, BPFP=1.3559 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,476B, BPFP=1.8691 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,192B, BPFP=0.9561 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,384B, BPFP=1.8174 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,212B, BPFP=1.1981 -⌛️ [2/4] FRONTEND: Frontend time: 1.942s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.563s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02563144 10.68736757 - layer.0.v_cache 0.00000028 0.00026715 - layer.1.k_cache 0.00322103 1.68883538 - layer.1.v_cache 0.00000086 0.00090780 - layer.2.k_cache 0.00118144 0.68146852 - layer.2.v_cache 0.00000111 0.00132324 - layer.3.k_cache 0.00130833 0.78673614 - layer.3.v_cache 0.00000230 0.00228011 - layer.4.k_cache 0.00357125 1.85639611 - layer.4.v_cache 0.00000311 0.00355979 - layer.4.output 0.00020081 0.10360762 - ------------------------------------------------------------------------------------- - TOTAL 0.00255174 1.15168373 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 388712 -BPFP 1.3146 bits/point -EBPFP 2.6293 equivalent bits/point -MSE 1.151684 ----------------------- -------------------------------------------------------- -Time: 3.514s Load: 0.009s, Pack+Encode: 1.942s, Decode+Unpack: 1.563s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1517 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample126-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample126-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 174, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 174, 128) -Output shape: (1, 174, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) -> torch.Size([1, 1, 174, 1024]) - layer.4.output: torch.Size([1, 174, 4096]) -> torch.Size([1, 1, 174, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,252B, BPFP=0.1460 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,280B, BPFP=1.5841 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,564B, BPFP=0.7886 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,088B, BPFP=1.6652 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 23,036B, BPFP=1.0343 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 37,728B, BPFP=1.6940 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,776B, BPFP=1.2022 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 36,652B, BPFP=1.6457 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,452B, BPFP=0.8285 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,544B, BPFP=1.6857 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 69,028B, BPFP=0.7748 -⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 174, 128]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.output: torch.Size([1, 174, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.509s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 174, 128]) - layer.0.v_cache: torch.Size([1, 8, 174, 128]) - layer.1.k_cache: torch.Size([1, 8, 174, 128]) - layer.1.v_cache: torch.Size([1, 8, 174, 128]) - layer.2.k_cache: torch.Size([1, 8, 174, 128]) - layer.2.v_cache: torch.Size([1, 8, 174, 128]) - layer.3.k_cache: torch.Size([1, 8, 174, 128]) - layer.3.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.k_cache: torch.Size([1, 8, 174, 128]) - layer.4.v_cache: torch.Size([1, 8, 174, 128]) - layer.4.output: torch.Size([1, 174, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02884222 8.87635400 - layer.0.v_cache 0.00000026 0.00025361 - layer.1.k_cache 0.00305335 1.64736518 - layer.1.v_cache 0.00000075 0.00085916 - layer.2.k_cache 0.00114914 0.64673334 - layer.2.v_cache 0.00000112 0.00125548 - layer.3.k_cache 0.00130410 0.73880496 - layer.3.v_cache 0.00000204 0.00206857 - layer.4.k_cache 0.00362435 1.73972750 - layer.4.v_cache 0.00000301 0.00338014 - layer.4.output 0.00017507 0.08917772 - ------------------------------------------------------------------------------------- - TOTAL 0.00276290 1.00096520 - (elements=2,494,464) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2494464 -Total Bytes 342400 -BPFP 1.0981 bits/point -EBPFP 2.1962 equivalent bits/point -MSE 1.000965 ----------------------- -------------------------------------------------------- -Time: 3.405s Load: 0.009s, Pack+Encode: 1.887s, Decode+Unpack: 1.509s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 174, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 174, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0010 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample129-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample129-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 152, 128) -Output shape: (1, 152, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.output: torch.Size([1, 152, 4096]) -> torch.Size([1, 1, 152, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,988B, BPFP=0.1536 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,684B, BPFP=1.8855 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 20,296B, BPFP=1.0432 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,368B, BPFP=2.1262 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,704B, BPFP=1.2697 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 42,748B, BPFP=2.1972 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,476B, BPFP=1.4636 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,452B, BPFP=2.0278 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,016B, BPFP=0.9260 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 41,144B, BPFP=2.1147 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,264B, BPFP=1.0185 -⌛️ [2/4] FRONTEND: Frontend time: 1.977s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.425s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02732493 8.67468904 - layer.0.v_cache 0.00000027 0.00025756 - layer.1.k_cache 0.00311055 1.61858860 - layer.1.v_cache 0.00000090 0.00093887 - layer.2.k_cache 0.00122034 0.63591044 - layer.2.v_cache 0.00000121 0.00139375 - layer.3.k_cache 0.00130561 0.74594046 - layer.3.v_cache 0.00000231 0.00226791 - layer.4.k_cache 0.00351005 1.75563973 - layer.4.v_cache 0.00000314 0.00358285 - layer.4.output 0.00016281 0.08860906 - ------------------------------------------------------------------------------------- - TOTAL 0.00265218 0.98526039 - (elements=2,179,072) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2179072 -Total Bytes 375140 -BPFP 1.3772 bits/point -EBPFP 2.7545 equivalent bits/point -MSE 0.985260 ----------------------- -------------------------------------------------------- -Time: 3.411s Load: 0.009s, Pack+Encode: 1.977s, Decode+Unpack: 1.425s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9853 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample130-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample130-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 154, 128) -Output shape: (1, 154, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.output: torch.Size([1, 154, 4096]) -> torch.Size([1, 1, 154, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,948B, BPFP=0.1496 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,260B, BPFP=1.7888 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,356B, BPFP=0.9312 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,984B, BPFP=1.9777 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 22,384B, BPFP=1.1356 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,032B, BPFP=1.9294 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,580B, BPFP=1.4499 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,252B, BPFP=1.9405 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,268B, BPFP=0.9775 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,008B, BPFP=1.9789 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,740B, BPFP=0.9099 -⌛️ [2/4] FRONTEND: Frontend time: 1.932s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.608s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02760743 8.57062560 - layer.0.v_cache 0.00000026 0.00026196 - layer.1.k_cache 0.00317114 1.70662739 - layer.1.v_cache 0.00000086 0.00088974 - layer.2.k_cache 0.00117006 0.65249624 - layer.2.v_cache 0.00000109 0.00130466 - layer.3.k_cache 0.00134568 0.77972615 - layer.3.v_cache 0.00000205 0.00213617 - layer.4.k_cache 0.00345514 1.71997031 - layer.4.v_cache 0.00000304 0.00354792 - layer.4.output 0.00019092 0.09935538 - ------------------------------------------------------------------------------------- - TOTAL 0.00268003 0.98821483 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 352812 -BPFP 1.2785 bits/point -EBPFP 2.5569 equivalent bits/point -MSE 0.988215 ----------------------- -------------------------------------------------------- -Time: 3.549s Load: 0.009s, Pack+Encode: 1.932s, Decode+Unpack: 1.608s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9882 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample144-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample144-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,100B, BPFP=0.1468 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,228B, BPFP=1.6206 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,836B, BPFP=0.8445 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,208B, BPFP=1.7617 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,636B, BPFP=1.3085 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,172B, BPFP=1.8074 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,868B, BPFP=1.3195 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 36,660B, BPFP=1.7358 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,260B, BPFP=0.9119 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,768B, BPFP=1.7883 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 68,796B, BPFP=0.8143 -⌛️ [2/4] FRONTEND: Frontend time: 1.862s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.561s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02663295 9.34988089 - layer.0.v_cache 0.00000027 0.00025455 - layer.1.k_cache 0.00314363 1.56395893 - layer.1.v_cache 0.00000082 0.00091366 - layer.2.k_cache 0.00118021 0.68869499 - layer.2.v_cache 0.00000116 0.00136282 - layer.3.k_cache 0.00128344 0.77193969 - layer.3.v_cache 0.00000214 0.00216618 - layer.4.k_cache 0.00361995 1.71780987 - layer.4.v_cache 0.00000298 0.00351997 - layer.4.output 0.00016477 0.09895671 - ------------------------------------------------------------------------------------- - TOTAL 0.00260904 1.03545203 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 348532 -BPFP 1.1787 bits/point -EBPFP 2.3575 equivalent bits/point -MSE 1.035452 ----------------------- -------------------------------------------------------- -Time: 3.432s Load: 0.009s, Pack+Encode: 1.862s, Decode+Unpack: 1.561s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0355 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample147-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample147-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 209, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 209, 128) -Output shape: (1, 209, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) -> torch.Size([1, 1, 209, 1024]) - layer.4.output: torch.Size([1, 209, 4096]) -> torch.Size([1, 1, 209, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,788B, BPFP=0.1416 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,364B, BPFP=1.7705 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,964B, BPFP=0.9332 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 49,556B, BPFP=1.8524 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,252B, BPFP=1.2056 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,672B, BPFP=1.8568 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,936B, BPFP=1.3807 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 48,552B, BPFP=1.8149 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,084B, BPFP=0.9750 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 49,996B, BPFP=1.8689 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 80,340B, BPFP=0.7508 -⌛️ [2/4] FRONTEND: Frontend time: 2.375s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 209, 128]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.output: torch.Size([1, 209, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.815s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 209, 128]) - layer.0.v_cache: torch.Size([1, 8, 209, 128]) - layer.1.k_cache: torch.Size([1, 8, 209, 128]) - layer.1.v_cache: torch.Size([1, 8, 209, 128]) - layer.2.k_cache: torch.Size([1, 8, 209, 128]) - layer.2.v_cache: torch.Size([1, 8, 209, 128]) - layer.3.k_cache: torch.Size([1, 8, 209, 128]) - layer.3.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.k_cache: torch.Size([1, 8, 209, 128]) - layer.4.v_cache: torch.Size([1, 8, 209, 128]) - layer.4.output: torch.Size([1, 209, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02572924 8.74925414 - layer.0.v_cache 0.00000027 0.00025647 - layer.1.k_cache 0.00299828 1.72624228 - layer.1.v_cache 0.00000076 0.00084511 - layer.2.k_cache 0.00126581 0.62887727 - layer.2.v_cache 0.00000114 0.00126165 - layer.3.k_cache 0.00131754 0.76238707 - layer.3.v_cache 0.00000208 0.00208528 - layer.4.k_cache 0.00354781 1.99364957 - layer.4.v_cache 0.00000302 0.00340955 - layer.4.output 0.00016605 0.09639503 - ------------------------------------------------------------------------------------- - TOTAL 0.00253787 1.01813204 - (elements=2,996,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2996224 -Total Bytes 449504 -BPFP 1.2002 bits/point -EBPFP 2.4004 equivalent bits/point -MSE 1.018132 ----------------------- -------------------------------------------------------- -Time: 4.202s Load: 0.012s, Pack+Encode: 2.375s, Decode+Unpack: 1.815s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 209, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 209, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0181 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample15-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample15-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 171, 128) -Output shape: (1, 171, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) -> torch.Size([1, 1, 171, 1024]) - layer.4.output: torch.Size([1, 171, 4096]) -> torch.Size([1, 1, 171, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,156B, BPFP=0.1442 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,108B, BPFP=1.6040 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,348B, BPFP=0.8840 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,408B, BPFP=1.7548 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,860B, BPFP=1.2272 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,976B, BPFP=1.7807 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,456B, BPFP=1.3001 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,532B, BPFP=1.7604 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,844B, BPFP=0.8609 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,420B, BPFP=1.7553 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 87,844B, BPFP=1.0033 -⌛️ [2/4] FRONTEND: Frontend time: 2.024s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.472s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 171, 128]) - layer.0.v_cache: torch.Size([1, 8, 171, 128]) - layer.1.k_cache: torch.Size([1, 8, 171, 128]) - layer.1.v_cache: torch.Size([1, 8, 171, 128]) - layer.2.k_cache: torch.Size([1, 8, 171, 128]) - layer.2.v_cache: torch.Size([1, 8, 171, 128]) - layer.3.k_cache: torch.Size([1, 8, 171, 128]) - layer.3.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.k_cache: torch.Size([1, 8, 171, 128]) - layer.4.v_cache: torch.Size([1, 8, 171, 128]) - layer.4.output: torch.Size([1, 171, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710774 9.45482399 - layer.0.v_cache 0.00000028 0.00026916 - layer.1.k_cache 0.00313176 1.73725686 - layer.1.v_cache 0.00000087 0.00094501 - layer.2.k_cache 0.00119500 0.67604159 - layer.2.v_cache 0.00000129 0.00136487 - layer.3.k_cache 0.00128502 0.77720062 - layer.3.v_cache 0.00000223 0.00222287 - layer.4.k_cache 0.00358421 1.66091705 - layer.4.v_cache 0.00000308 0.00355177 - layer.4.output 0.00020665 0.09794719 - ------------------------------------------------------------------------------------- - TOTAL 0.00265272 1.05045590 - (elements=2,451,456) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2451456 -Total Bytes 373952 -BPFP 1.2203 bits/point -EBPFP 2.4407 equivalent bits/point -MSE 1.050456 ----------------------- -------------------------------------------------------- -Time: 3.506s Load: 0.010s, Pack+Encode: 2.024s, Decode+Unpack: 1.472s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 171, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 171, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0505 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample152-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample152-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,456B, BPFP=0.1459 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,768B, BPFP=1.5105 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,784B, BPFP=0.7932 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,692B, BPFP=1.6340 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,156B, BPFP=1.1046 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,352B, BPFP=1.6618 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,096B, BPFP=1.1865 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,380B, BPFP=1.6208 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,420B, BPFP=0.8201 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,088B, BPFP=1.6507 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 95,424B, BPFP=1.0074 -⌛️ [2/4] FRONTEND: Frontend time: 1.911s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.684s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02889625 8.92241607 - layer.0.v_cache 0.00000026 0.00025365 - layer.1.k_cache 0.00305597 1.81500805 - layer.1.v_cache 0.00000081 0.00092012 - layer.2.k_cache 0.00119284 0.63463102 - layer.2.v_cache 0.00000119 0.00140246 - layer.3.k_cache 0.00128722 0.76163107 - layer.3.v_cache 0.00000243 0.00239728 - layer.4.k_cache 0.00400834 1.87332170 - layer.4.v_cache 0.00000313 0.00361332 - layer.4.output 0.00018370 0.09523816 - ------------------------------------------------------------------------------------- - TOTAL 0.00279880 1.02832481 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 382616 -BPFP 1.1541 bits/point -EBPFP 2.3083 equivalent bits/point -MSE 1.028325 ----------------------- -------------------------------------------------------- -Time: 3.607s Load: 0.012s, Pack+Encode: 1.911s, Decode+Unpack: 1.684s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0283 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample16-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample16-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 139, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 139, 128) -Output shape: (1, 139, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) -> torch.Size([1, 1, 139, 1024]) - layer.4.output: torch.Size([1, 139, 4096]) -> torch.Size([1, 1, 139, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,732B, BPFP=0.1536 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,060B, BPFP=2.0268 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,516B, BPFP=1.0407 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,644B, BPFP=2.5092 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 19,580B, BPFP=1.1005 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,356B, BPFP=2.7741 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,384B, BPFP=1.4829 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 45,376B, BPFP=2.5504 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,924B, BPFP=1.0074 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 44,188B, BPFP=2.4836 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 76,672B, BPFP=1.0773 -⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 139, 128]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.output: torch.Size([1, 139, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.417s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 139, 128]) - layer.0.v_cache: torch.Size([1, 8, 139, 128]) - layer.1.k_cache: torch.Size([1, 8, 139, 128]) - layer.1.v_cache: torch.Size([1, 8, 139, 128]) - layer.2.k_cache: torch.Size([1, 8, 139, 128]) - layer.2.v_cache: torch.Size([1, 8, 139, 128]) - layer.3.k_cache: torch.Size([1, 8, 139, 128]) - layer.3.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.k_cache: torch.Size([1, 8, 139, 128]) - layer.4.v_cache: torch.Size([1, 8, 139, 128]) - layer.4.output: torch.Size([1, 139, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03086843 9.50145343 - layer.0.v_cache 0.00000027 0.00026360 - layer.1.k_cache 0.00318250 1.73043077 - layer.1.v_cache 0.00000093 0.00093194 - layer.2.k_cache 0.00113687 0.66798977 - layer.2.v_cache 0.00000113 0.00137488 - layer.3.k_cache 0.00135365 0.76098216 - layer.3.v_cache 0.00000214 0.00221235 - layer.4.k_cache 0.00345519 1.73227983 - layer.4.v_cache 0.00000310 0.00362968 - layer.4.output 0.00019842 0.10667342 - ------------------------------------------------------------------------------------- - TOTAL 0.00291414 1.05916015 - (elements=1,992,704) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1992704 -Total Bytes 381432 -BPFP 1.5313 bits/point -EBPFP 3.0626 equivalent bits/point -MSE 1.059160 ----------------------- -------------------------------------------------------- -Time: 3.334s Load: 0.008s, Pack+Encode: 1.909s, Decode+Unpack: 1.417s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 139, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 139, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0592 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample162-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample162-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,152B, BPFP=0.1432 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,304B, BPFP=1.6036 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,108B, BPFP=0.8679 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,940B, BPFP=1.7233 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,268B, BPFP=1.1931 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,180B, BPFP=1.7342 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,812B, BPFP=1.2633 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,344B, BPFP=1.6962 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,092B, BPFP=0.7763 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,176B, BPFP=1.7340 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 75,972B, BPFP=0.8627 -⌛️ [2/4] FRONTEND: Frontend time: 1.983s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.503s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02678066 9.01464773 - layer.0.v_cache 0.00000028 0.00026084 - layer.1.k_cache 0.00311261 1.65765647 - layer.1.v_cache 0.00000082 0.00090009 - layer.2.k_cache 0.00118278 0.70122989 - layer.2.v_cache 0.00000114 0.00132202 - layer.3.k_cache 0.00135865 0.79522758 - layer.3.v_cache 0.00000220 0.00223682 - layer.4.k_cache 0.00351030 1.64475818 - layer.4.v_cache 0.00000304 0.00351977 - layer.4.output 0.00023867 0.10648786 - ------------------------------------------------------------------------------------- - TOTAL 0.00263622 1.01769363 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 356348 -BPFP 1.1561 bits/point -EBPFP 2.3123 equivalent bits/point -MSE 1.017694 ----------------------- -------------------------------------------------------- -Time: 3.497s Load: 0.012s, Pack+Encode: 1.983s, Decode+Unpack: 1.503s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0177 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample166-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample166-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 213, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 213, 128) -Output shape: (1, 213, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) -> torch.Size([1, 1, 213, 1024]) - layer.4.output: torch.Size([1, 213, 4096]) -> torch.Size([1, 1, 213, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,888B, BPFP=0.1426 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 43,076B, BPFP=1.5800 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,748B, BPFP=0.9077 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 47,152B, BPFP=1.7295 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 29,316B, BPFP=1.0753 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 48,932B, BPFP=1.7947 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,608B, BPFP=1.3427 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 48,340B, BPFP=1.7730 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,100B, BPFP=0.8839 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 48,880B, BPFP=1.7928 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 80,348B, BPFP=0.7368 -⌛️ [2/4] FRONTEND: Frontend time: 2.256s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 213, 128]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.output: torch.Size([1, 213, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.139s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 213, 128]) - layer.0.v_cache: torch.Size([1, 8, 213, 128]) - layer.1.k_cache: torch.Size([1, 8, 213, 128]) - layer.1.v_cache: torch.Size([1, 8, 213, 128]) - layer.2.k_cache: torch.Size([1, 8, 213, 128]) - layer.2.v_cache: torch.Size([1, 8, 213, 128]) - layer.3.k_cache: torch.Size([1, 8, 213, 128]) - layer.3.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.k_cache: torch.Size([1, 8, 213, 128]) - layer.4.v_cache: torch.Size([1, 8, 213, 128]) - layer.4.output: torch.Size([1, 213, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02775377 8.49116280 - layer.0.v_cache 0.00000028 0.00027450 - layer.1.k_cache 0.00299314 1.64893552 - layer.1.v_cache 0.00000076 0.00080843 - layer.2.k_cache 0.00116592 0.64473026 - layer.2.v_cache 0.00000113 0.00121062 - layer.3.k_cache 0.00131338 0.76348103 - layer.3.v_cache 0.00000208 0.00206680 - layer.4.k_cache 0.00376460 1.86375048 - layer.4.v_cache 0.00000302 0.00322482 - layer.4.output 0.00016925 0.09878862 - ------------------------------------------------------------------------------------- - TOTAL 0.00269108 0.98677141 - (elements=3,053,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3053568 -Total Bytes 435388 -BPFP 1.1407 bits/point -EBPFP 2.2813 equivalent bits/point -MSE 0.986771 ----------------------- -------------------------------------------------------- -Time: 4.407s Load: 0.012s, Pack+Encode: 2.256s, Decode+Unpack: 2.139s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 213, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 213, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9868 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample175-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample175-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 179, 128) -Output shape: (1, 179, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.output: torch.Size([1, 179, 4096]) -> torch.Size([1, 1, 179, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,400B, BPFP=0.1484 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,640B, BPFP=1.5992 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,224B, BPFP=0.7954 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,592B, BPFP=1.6844 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,192B, BPFP=1.0995 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,184B, BPFP=1.7102 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,164B, BPFP=1.1856 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,204B, BPFP=1.6674 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,740B, BPFP=0.8179 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,504B, BPFP=1.7242 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 98,348B, BPFP=1.0731 -⌛️ [2/4] FRONTEND: Frontend time: 1.880s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.473s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02728191 9.27698981 - layer.0.v_cache 0.00000029 0.00025386 - layer.1.k_cache 0.00306154 1.63687935 - layer.1.v_cache 0.00000084 0.00093479 - layer.2.k_cache 0.00116357 0.64837519 - layer.2.v_cache 0.00000127 0.00136230 - layer.3.k_cache 0.00133958 0.74734420 - layer.3.v_cache 0.00000230 0.00225161 - layer.4.k_cache 0.00349468 1.78227225 - layer.4.v_cache 0.00000328 0.00376893 - layer.4.output 0.00017175 0.09105601 - ------------------------------------------------------------------------------------- - TOTAL 0.00264545 1.03318974 - (elements=2,566,144) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2566144 -Total Bytes 383192 -BPFP 1.1946 bits/point -EBPFP 2.3892 equivalent bits/point -MSE 1.033190 ----------------------- -------------------------------------------------------- -Time: 3.362s Load: 0.009s, Pack+Encode: 1.880s, Decode+Unpack: 1.473s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0332 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample18-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample18-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 180, 128) -Output shape: (1, 180, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.output: torch.Size([1, 180, 4096]) -> torch.Size([1, 1, 180, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,352B, BPFP=0.1455 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,000B, BPFP=1.5625 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,800B, BPFP=0.7726 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,260B, BPFP=1.6606 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,072B, BPFP=1.0882 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,796B, BPFP=1.6839 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,408B, BPFP=1.1896 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,716B, BPFP=1.6370 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,176B, BPFP=0.7889 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,368B, BPFP=1.6653 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 72,444B, BPFP=0.7861 -⌛️ [2/4] FRONTEND: Frontend time: 1.946s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.458s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03239844 8.80190091 - layer.0.v_cache 0.00000026 0.00024805 - layer.1.k_cache 0.00307858 1.57316250 - layer.1.v_cache 0.00000079 0.00088517 - layer.2.k_cache 0.00118466 0.63751899 - layer.2.v_cache 0.00000116 0.00134640 - layer.3.k_cache 0.00130112 0.73069492 - layer.3.v_cache 0.00000222 0.00223383 - layer.4.k_cache 0.00350099 1.75610453 - layer.4.v_cache 0.00000308 0.00353109 - layer.4.output 0.00016632 0.08830441 - ------------------------------------------------------------------------------------- - TOTAL 0.00300976 0.99006029 - (elements=2,580,480) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2580480 -Total Bytes 353392 -BPFP 1.0956 bits/point -EBPFP 2.1912 equivalent bits/point -MSE 0.990060 ----------------------- -------------------------------------------------------- -Time: 3.415s Load: 0.011s, Pack+Encode: 1.946s, Decode+Unpack: 1.458s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9901 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample19-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample19-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 145, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 145, 128) -Output shape: (1, 145, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) -> torch.Size([1, 1, 145, 1024]) - layer.4.output: torch.Size([1, 145, 4096]) -> torch.Size([1, 1, 145, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,816B, BPFP=0.1517 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,272B, BPFP=1.9543 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,212B, BPFP=1.0351 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 44,456B, BPFP=2.3953 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,896B, BPFP=1.3414 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,484B, BPFP=2.3968 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,848B, BPFP=1.5543 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 43,324B, BPFP=2.3343 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,236B, BPFP=1.1442 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 45,604B, BPFP=2.4571 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 64,236B, BPFP=0.8652 -⌛️ [2/4] FRONTEND: Frontend time: 1.893s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 145, 128]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.output: torch.Size([1, 145, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.628s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 145, 128]) - layer.0.v_cache: torch.Size([1, 8, 145, 128]) - layer.1.k_cache: torch.Size([1, 8, 145, 128]) - layer.1.v_cache: torch.Size([1, 8, 145, 128]) - layer.2.k_cache: torch.Size([1, 8, 145, 128]) - layer.2.v_cache: torch.Size([1, 8, 145, 128]) - layer.3.k_cache: torch.Size([1, 8, 145, 128]) - layer.3.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.k_cache: torch.Size([1, 8, 145, 128]) - layer.4.v_cache: torch.Size([1, 8, 145, 128]) - layer.4.output: torch.Size([1, 145, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02766713 9.28187466 - layer.0.v_cache 0.00000028 0.00025531 - layer.1.k_cache 0.00309559 1.54769761 - layer.1.v_cache 0.00000087 0.00093153 - layer.2.k_cache 0.00119714 0.65775804 - layer.2.v_cache 0.00000118 0.00135423 - layer.3.k_cache 0.00133867 0.76159647 - layer.3.v_cache 0.00000208 0.00213879 - layer.4.k_cache 0.00347171 1.96718476 - layer.4.v_cache 0.00000305 0.00339794 - layer.4.output 0.00016146 0.09673381 - ------------------------------------------------------------------------------------- - TOTAL 0.00267311 1.04365176 - (elements=2,078,720) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2078720 -Total Bytes 375384 -BPFP 1.4447 bits/point -EBPFP 2.8893 equivalent bits/point -MSE 1.043652 ----------------------- -------------------------------------------------------- -Time: 3.529s Load: 0.008s, Pack+Encode: 1.893s, Decode+Unpack: 1.628s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 145, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 145, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0437 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample190-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample190-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 179, 128) -Output shape: (1, 179, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) -> torch.Size([1, 1, 179, 1024]) - layer.4.output: torch.Size([1, 179, 4096]) -> torch.Size([1, 1, 179, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,332B, BPFP=0.1454 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,808B, BPFP=1.5628 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,036B, BPFP=0.7872 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,828B, BPFP=1.6510 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,028B, BPFP=1.1360 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,488B, BPFP=1.6798 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,608B, BPFP=1.2050 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,216B, BPFP=1.6243 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,868B, BPFP=0.7799 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,032B, BPFP=1.6599 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 75,760B, BPFP=0.8266 -⌛️ [2/4] FRONTEND: Frontend time: 1.884s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.503s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 179, 128]) - layer.0.v_cache: torch.Size([1, 8, 179, 128]) - layer.1.k_cache: torch.Size([1, 8, 179, 128]) - layer.1.v_cache: torch.Size([1, 8, 179, 128]) - layer.2.k_cache: torch.Size([1, 8, 179, 128]) - layer.2.v_cache: torch.Size([1, 8, 179, 128]) - layer.3.k_cache: torch.Size([1, 8, 179, 128]) - layer.3.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.k_cache: torch.Size([1, 8, 179, 128]) - layer.4.v_cache: torch.Size([1, 8, 179, 128]) - layer.4.output: torch.Size([1, 179, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02963084 8.23724195 - layer.0.v_cache 0.00000026 0.00025238 - layer.1.k_cache 0.00308074 1.51712036 - layer.1.v_cache 0.00000080 0.00087941 - layer.2.k_cache 0.00119595 0.67064965 - layer.2.v_cache 0.00000115 0.00134839 - layer.3.k_cache 0.00135258 0.77666895 - layer.3.v_cache 0.00000217 0.00218595 - layer.4.k_cache 0.00351189 1.70917072 - layer.4.v_cache 0.00000298 0.00340998 - layer.4.output 0.00021521 0.10939544 - ------------------------------------------------------------------------------------- - TOTAL 0.00283144 0.95403639 - (elements=2,566,144) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2566144 -Total Bytes 356004 -BPFP 1.1098 bits/point -EBPFP 2.2197 equivalent bits/point -MSE 0.954036 ----------------------- -------------------------------------------------------- -Time: 3.397s Load: 0.010s, Pack+Encode: 1.884s, Decode+Unpack: 1.503s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 179, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 179, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9540 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample20-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample20-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 157, 128) -Output shape: (1, 157, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.output: torch.Size([1, 157, 4096]) -> torch.Size([1, 1, 157, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,060B, BPFP=0.1523 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,172B, BPFP=1.8000 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,380B, BPFP=0.9644 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,564B, BPFP=1.9688 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,720B, BPFP=1.2799 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,132B, BPFP=1.9473 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 29,092B, BPFP=1.4477 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,840B, BPFP=1.8830 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,812B, BPFP=0.9361 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 40,444B, BPFP=2.0125 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,524B, BPFP=0.9769 -⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.458s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02651285 8.70360084 - layer.0.v_cache 0.00000026 0.00024464 - layer.1.k_cache 0.00308002 1.68196524 - layer.1.v_cache 0.00000091 0.00095079 - layer.2.k_cache 0.00121857 0.67731869 - layer.2.v_cache 0.00000134 0.00140161 - layer.3.k_cache 0.00131702 0.76178353 - layer.3.v_cache 0.00000230 0.00223270 - layer.4.k_cache 0.00350090 1.82050532 - layer.4.v_cache 0.00000345 0.00371529 - layer.4.output 0.00019599 0.09879016 - ------------------------------------------------------------------------------------- - TOTAL 0.00260154 1.00349138 - (elements=2,250,752) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2250752 -Total Bytes 367740 -BPFP 1.3071 bits/point -EBPFP 2.6142 equivalent bits/point -MSE 1.003491 ----------------------- -------------------------------------------------------- -Time: 3.412s Load: 0.010s, Pack+Encode: 1.944s, Decode+Unpack: 1.458s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0035 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample21-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample21-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,296B, BPFP=0.1471 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,280B, BPFP=1.6196 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,596B, BPFP=0.8302 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,400B, BPFP=1.7143 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,224B, BPFP=1.0814 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,836B, BPFP=1.7337 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,548B, BPFP=1.2298 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,924B, BPFP=1.6930 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,828B, BPFP=0.8852 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,848B, BPFP=1.7343 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 85,712B, BPFP=0.9566 -⌛️ [2/4] FRONTEND: Frontend time: 1.905s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.635s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02748643 9.58889160 - layer.0.v_cache 0.00000027 0.00025390 - layer.1.k_cache 0.00298274 1.46080811 - layer.1.v_cache 0.00000083 0.00091940 - layer.2.k_cache 0.00117838 0.62908713 - layer.2.v_cache 0.00000115 0.00136298 - layer.3.k_cache 0.00127749 0.73970372 - layer.3.v_cache 0.00000229 0.00227748 - layer.4.k_cache 0.00388258 1.87304879 - layer.4.v_cache 0.00000318 0.00363722 - layer.4.output 0.00017171 0.09515113 - ------------------------------------------------------------------------------------- - TOTAL 0.00267873 1.04861392 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 369492 -BPFP 1.1782 bits/point -EBPFP 2.3565 equivalent bits/point -MSE 1.048614 ----------------------- -------------------------------------------------------- -Time: 3.550s Load: 0.010s, Pack+Encode: 1.905s, Decode+Unpack: 1.635s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0486 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample22-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample22-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,188B, BPFP=0.1448 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,132B, BPFP=1.6412 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,796B, BPFP=0.8537 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,060B, BPFP=1.7742 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,608B, BPFP=1.2086 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,384B, BPFP=1.7889 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,700B, BPFP=1.3036 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,936B, BPFP=1.7231 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,228B, BPFP=0.8734 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,740B, BPFP=1.7596 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 94,660B, BPFP=1.0749 -⌛️ [2/4] FRONTEND: Frontend time: 1.885s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.515s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02961682 9.35204085 - layer.0.v_cache 0.00000027 0.00025853 - layer.1.k_cache 0.00302783 1.78692201 - layer.1.v_cache 0.00000102 0.00097182 - layer.2.k_cache 0.00123476 0.68330765 - layer.2.v_cache 0.00000123 0.00139523 - layer.3.k_cache 0.00126690 0.76005182 - layer.3.v_cache 0.00000235 0.00228390 - layer.4.k_cache 0.00342846 1.67550003 - layer.4.v_cache 0.00000335 0.00376856 - layer.4.output 0.00019350 0.09035336 - ------------------------------------------------------------------------------------- - TOTAL 0.00281121 1.04485099 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 382432 -BPFP 1.2408 bits/point -EBPFP 2.4815 equivalent bits/point -MSE 1.044851 ----------------------- -------------------------------------------------------- -Time: 3.409s Load: 0.009s, Pack+Encode: 1.885s, Decode+Unpack: 1.515s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0449 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample23-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample23-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 178, 128) -Output shape: (1, 178, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.output: torch.Size([1, 178, 4096]) -> torch.Size([1, 1, 178, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,300B, BPFP=0.1448 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,924B, BPFP=1.5328 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,652B, BPFP=0.8186 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,064B, BPFP=1.6706 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,272B, BPFP=1.0653 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,692B, BPFP=1.6982 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,016B, BPFP=1.2296 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,444B, BPFP=1.6434 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,320B, BPFP=0.8919 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,748B, BPFP=1.7007 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,176B, BPFP=0.7810 -⌛️ [2/4] FRONTEND: Frontend time: 2.010s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.428s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02741563 8.26078539 - layer.0.v_cache 0.00000026 0.00025293 - layer.1.k_cache 0.00307030 1.61682300 - layer.1.v_cache 0.00000077 0.00087595 - layer.2.k_cache 0.00116729 0.63695873 - layer.2.v_cache 0.00000113 0.00127455 - layer.3.k_cache 0.00134025 0.75667041 - layer.3.v_cache 0.00000215 0.00210038 - layer.4.k_cache 0.00345504 1.92663334 - layer.4.v_cache 0.00000307 0.00348268 - layer.4.output 0.00018601 0.09589163 - ------------------------------------------------------------------------------------- - TOTAL 0.00265714 0.97067313 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 353608 -BPFP 1.1086 bits/point -EBPFP 2.2171 equivalent bits/point -MSE 0.970673 ----------------------- -------------------------------------------------------- -Time: 3.449s Load: 0.012s, Pack+Encode: 2.010s, Decode+Unpack: 1.428s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9707 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample24-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample24-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,480B, BPFP=0.1438 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,300B, BPFP=1.5005 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,828B, BPFP=0.7783 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,200B, BPFP=1.5790 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,420B, BPFP=1.1334 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,696B, BPFP=1.5995 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,496B, BPFP=1.1366 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,700B, BPFP=1.5584 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,424B, BPFP=0.8029 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,660B, BPFP=1.5980 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 83,720B, BPFP=0.8652 -⌛️ [2/4] FRONTEND: Frontend time: 1.916s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.628s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02569163 9.23740544 - layer.0.v_cache 0.00000027 0.00025225 - layer.1.k_cache 0.00297363 1.75619862 - layer.1.v_cache 0.00000081 0.00090595 - layer.2.k_cache 0.00119805 0.69051567 - layer.2.v_cache 0.00000116 0.00131317 - layer.3.k_cache 0.00128923 0.78776090 - layer.3.v_cache 0.00000232 0.00224132 - layer.4.k_cache 0.00350442 2.01950364 - layer.4.v_cache 0.00000328 0.00357553 - layer.4.output 0.00017922 0.09862976 - ------------------------------------------------------------------------------------- - TOTAL 0.00252727 1.06387083 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 369924 -BPFP 1.0922 bits/point -EBPFP 2.1845 equivalent bits/point -MSE 1.063871 ----------------------- -------------------------------------------------------- -Time: 3.555s Load: 0.010s, Pack+Encode: 1.916s, Decode+Unpack: 1.628s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0639 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample25-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample25-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 178, 128) -Output shape: (1, 178, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) -> torch.Size([1, 1, 178, 1024]) - layer.4.output: torch.Size([1, 178, 4096]) -> torch.Size([1, 1, 178, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,356B, BPFP=0.1473 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,836B, BPFP=1.6167 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,828B, BPFP=0.8264 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,392B, BPFP=1.6850 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,684B, BPFP=1.1712 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,472B, BPFP=1.6886 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,844B, BPFP=1.2221 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,424B, BPFP=1.6426 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,136B, BPFP=0.8838 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,764B, BPFP=1.7014 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 77,196B, BPFP=0.8470 -⌛️ [2/4] FRONTEND: Frontend time: 1.888s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.514s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 178, 128]) - layer.0.v_cache: torch.Size([1, 8, 178, 128]) - layer.1.k_cache: torch.Size([1, 8, 178, 128]) - layer.1.v_cache: torch.Size([1, 8, 178, 128]) - layer.2.k_cache: torch.Size([1, 8, 178, 128]) - layer.2.v_cache: torch.Size([1, 8, 178, 128]) - layer.3.k_cache: torch.Size([1, 8, 178, 128]) - layer.3.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.k_cache: torch.Size([1, 8, 178, 128]) - layer.4.v_cache: torch.Size([1, 8, 178, 128]) - layer.4.output: torch.Size([1, 178, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02863186 9.08592104 - layer.0.v_cache 0.00000027 0.00025356 - layer.1.k_cache 0.00302134 1.98448147 - layer.1.v_cache 0.00000082 0.00089230 - layer.2.k_cache 0.00115360 0.65538989 - layer.2.v_cache 0.00000119 0.00129343 - layer.3.k_cache 0.00131482 0.76807704 - layer.3.v_cache 0.00000222 0.00214382 - layer.4.k_cache 0.00345682 1.96845871 - layer.4.v_cache 0.00000324 0.00357220 - layer.4.output 0.00020381 0.09619245 - ------------------------------------------------------------------------------------- - TOTAL 0.00274296 1.06108952 - (elements=2,551,808) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2551808 -Total Bytes 363932 -BPFP 1.1409 bits/point -EBPFP 2.2819 equivalent bits/point -MSE 1.061090 ----------------------- -------------------------------------------------------- -Time: 3.412s Load: 0.009s, Pack+Encode: 1.888s, Decode+Unpack: 1.514s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 178, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 178, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0611 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample26-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample26-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 182, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 182, 128) -Output shape: (1, 182, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) -> torch.Size([1, 1, 182, 1024]) - layer.4.output: torch.Size([1, 182, 4096]) -> torch.Size([1, 1, 182, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,416B, BPFP=0.1466 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,372B, BPFP=1.5613 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,280B, BPFP=0.8276 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,064B, BPFP=1.6339 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,088B, BPFP=1.0769 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,724B, BPFP=1.6623 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,780B, BPFP=1.1925 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,596B, BPFP=1.6138 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,292B, BPFP=0.8281 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,708B, BPFP=1.6616 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 77,548B, BPFP=0.8322 -⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 182, 128]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.output: torch.Size([1, 182, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.473s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 182, 128]) - layer.0.v_cache: torch.Size([1, 8, 182, 128]) - layer.1.k_cache: torch.Size([1, 8, 182, 128]) - layer.1.v_cache: torch.Size([1, 8, 182, 128]) - layer.2.k_cache: torch.Size([1, 8, 182, 128]) - layer.2.v_cache: torch.Size([1, 8, 182, 128]) - layer.3.k_cache: torch.Size([1, 8, 182, 128]) - layer.3.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.k_cache: torch.Size([1, 8, 182, 128]) - layer.4.v_cache: torch.Size([1, 8, 182, 128]) - layer.4.output: torch.Size([1, 182, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02794281 9.15686639 - layer.0.v_cache 0.00000027 0.00025764 - layer.1.k_cache 0.00307012 1.77050614 - layer.1.v_cache 0.00000079 0.00089167 - layer.2.k_cache 0.00119021 0.63311348 - layer.2.v_cache 0.00000113 0.00130417 - layer.3.k_cache 0.00130892 0.75774970 - layer.3.v_cache 0.00000221 0.00216800 - layer.4.k_cache 0.00347814 1.87066835 - layer.4.v_cache 0.00000308 0.00347675 - layer.4.output 0.00018397 0.09378833 - ------------------------------------------------------------------------------------- - TOTAL 0.00269526 1.04086826 - (elements=2,609,152) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2609152 -Total Bytes 361868 -BPFP 1.1095 bits/point -EBPFP 2.2191 equivalent bits/point -MSE 1.040868 ----------------------- -------------------------------------------------------- -Time: 3.454s Load: 0.010s, Pack+Encode: 1.971s, Decode+Unpack: 1.473s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 182, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 182, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0409 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample27-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample27-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 271, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.014s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 271, 128) -Output shape: (1, 271, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) -> torch.Size([1, 1, 271, 1024]) - layer.4.output: torch.Size([1, 271, 4096]) -> torch.Size([1, 1, 271, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,860B, BPFP=0.1401 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 58,284B, BPFP=1.6802 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 30,684B, BPFP=0.8846 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 63,036B, BPFP=1.8172 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 40,096B, BPFP=1.1559 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 62,896B, BPFP=1.8132 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 44,904B, BPFP=1.2945 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 62,128B, BPFP=1.7911 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 30,596B, BPFP=0.8820 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 63,704B, BPFP=1.8365 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 96,260B, BPFP=0.6938 -⌛️ [2/4] FRONTEND: Frontend time: 2.336s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 271, 128]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.output: torch.Size([1, 271, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.951s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 271, 128]) - layer.0.v_cache: torch.Size([1, 8, 271, 128]) - layer.1.k_cache: torch.Size([1, 8, 271, 128]) - layer.1.v_cache: torch.Size([1, 8, 271, 128]) - layer.2.k_cache: torch.Size([1, 8, 271, 128]) - layer.2.v_cache: torch.Size([1, 8, 271, 128]) - layer.3.k_cache: torch.Size([1, 8, 271, 128]) - layer.3.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.k_cache: torch.Size([1, 8, 271, 128]) - layer.4.v_cache: torch.Size([1, 8, 271, 128]) - layer.4.output: torch.Size([1, 271, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02401601 8.28231305 - layer.0.v_cache 0.00000027 0.00024649 - layer.1.k_cache 0.00298786 1.80833970 - layer.1.v_cache 0.00000080 0.00083555 - layer.2.k_cache 0.00117700 0.63343321 - layer.2.v_cache 0.00000117 0.00122774 - layer.3.k_cache 0.00127449 0.73341848 - layer.3.v_cache 0.00000214 0.00200445 - layer.4.k_cache 0.00368668 1.97311413 - layer.4.v_cache 0.00000306 0.00332858 - layer.4.output 0.00014564 0.08475356 - ------------------------------------------------------------------------------------- - TOTAL 0.00240943 0.98409112 - (elements=3,885,056) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3885056 -Total Bytes 557448 -BPFP 1.1479 bits/point -EBPFP 2.2958 equivalent bits/point -MSE 0.984091 ----------------------- -------------------------------------------------------- -Time: 4.300s Load: 0.014s, Pack+Encode: 2.336s, Decode+Unpack: 1.951s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 271, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 271, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9841 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample3-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample3-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 184, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 184, 128) -Output shape: (1, 184, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) -> torch.Size([1, 1, 184, 1024]) - layer.4.output: torch.Size([1, 184, 4096]) -> torch.Size([1, 1, 184, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,464B, BPFP=0.1471 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,908B, BPFP=1.5246 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,456B, BPFP=0.7836 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,312B, BPFP=1.6267 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,036B, BPFP=1.1055 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,028B, BPFP=1.6571 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,444B, BPFP=1.2077 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,028B, BPFP=1.6146 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,196B, BPFP=0.7726 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,960B, BPFP=1.6542 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 112,104B, BPFP=1.1900 -⌛️ [2/4] FRONTEND: Frontend time: 1.915s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 184, 128]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.output: torch.Size([1, 184, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.674s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 184, 128]) - layer.0.v_cache: torch.Size([1, 8, 184, 128]) - layer.1.k_cache: torch.Size([1, 8, 184, 128]) - layer.1.v_cache: torch.Size([1, 8, 184, 128]) - layer.2.k_cache: torch.Size([1, 8, 184, 128]) - layer.2.v_cache: torch.Size([1, 8, 184, 128]) - layer.3.k_cache: torch.Size([1, 8, 184, 128]) - layer.3.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.k_cache: torch.Size([1, 8, 184, 128]) - layer.4.v_cache: torch.Size([1, 8, 184, 128]) - layer.4.output: torch.Size([1, 184, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02872292 9.07619111 - layer.0.v_cache 0.00000026 0.00025398 - layer.1.k_cache 0.00314513 1.61386457 - layer.1.v_cache 0.00000084 0.00091726 - layer.2.k_cache 0.00123314 0.65226157 - layer.2.v_cache 0.00000120 0.00138065 - layer.3.k_cache 0.00131532 0.77432881 - layer.3.v_cache 0.00000224 0.00224841 - layer.4.k_cache 0.00360741 1.74902078 - layer.4.v_cache 0.00000320 0.00357696 - layer.4.output 0.00023634 0.11201829 - ------------------------------------------------------------------------------------- - TOTAL 0.00278407 1.02300838 - (elements=2,637,824) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2637824 -Total Bytes 396936 -BPFP 1.2038 bits/point -EBPFP 2.4077 equivalent bits/point -MSE 1.023008 ----------------------- -------------------------------------------------------- -Time: 3.598s Load: 0.009s, Pack+Encode: 1.915s, Decode+Unpack: 1.674s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 184, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 184, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0230 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample30-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample30-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,208B, BPFP=0.1474 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,660B, BPFP=1.6847 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,924B, BPFP=0.9156 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,272B, BPFP=1.9426 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,720B, BPFP=1.1820 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,192B, BPFP=1.8011 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,384B, BPFP=1.3044 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,140B, BPFP=1.7987 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,380B, BPFP=0.9825 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,660B, BPFP=1.8226 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 92,708B, BPFP=1.0651 -⌛️ [2/4] FRONTEND: Frontend time: 1.951s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.429s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02895958 9.42058464 - layer.0.v_cache 0.00000027 0.00025631 - layer.1.k_cache 0.00314466 1.76976247 - layer.1.v_cache 0.00000087 0.00094303 - layer.2.k_cache 0.00118284 0.67765144 - layer.2.v_cache 0.00000133 0.00136919 - layer.3.k_cache 0.00130601 0.76757507 - layer.3.v_cache 0.00000238 0.00227843 - layer.4.k_cache 0.00330381 1.85122914 - layer.4.v_cache 0.00000327 0.00364414 - layer.4.output 0.00019227 0.09817790 - ------------------------------------------------------------------------------------- - TOTAL 0.00276244 1.06342896 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 388248 -BPFP 1.2744 bits/point -EBPFP 2.5489 equivalent bits/point -MSE 1.063429 ----------------------- -------------------------------------------------------- -Time: 3.389s Load: 0.010s, Pack+Encode: 1.951s, Decode+Unpack: 1.429s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0634 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample31-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample31-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 206, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 206, 128) -Output shape: (1, 206, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) -> torch.Size([1, 1, 206, 1024]) - layer.4.output: torch.Size([1, 206, 4096]) -> torch.Size([1, 1, 206, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,776B, BPFP=0.1432 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 46,112B, BPFP=1.7488 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,572B, BPFP=0.9319 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 49,712B, BPFP=1.8853 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,784B, BPFP=1.0916 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,640B, BPFP=1.8826 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,704B, BPFP=1.3920 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,452B, BPFP=1.8755 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,900B, BPFP=0.8685 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 50,608B, BPFP=1.9193 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 145,968B, BPFP=1.3840 -⌛️ [2/4] FRONTEND: Frontend time: 2.451s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 206, 128]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.output: torch.Size([1, 206, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.834s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 206, 128]) - layer.0.v_cache: torch.Size([1, 8, 206, 128]) - layer.1.k_cache: torch.Size([1, 8, 206, 128]) - layer.1.v_cache: torch.Size([1, 8, 206, 128]) - layer.2.k_cache: torch.Size([1, 8, 206, 128]) - layer.2.v_cache: torch.Size([1, 8, 206, 128]) - layer.3.k_cache: torch.Size([1, 8, 206, 128]) - layer.3.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.k_cache: torch.Size([1, 8, 206, 128]) - layer.4.v_cache: torch.Size([1, 8, 206, 128]) - layer.4.output: torch.Size([1, 206, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02717657 9.13499377 - layer.0.v_cache 0.00000026 0.00025117 - layer.1.k_cache 0.00302264 1.66995476 - layer.1.v_cache 0.00000082 0.00089043 - layer.2.k_cache 0.00113561 0.67655115 - layer.2.v_cache 0.00000128 0.00127402 - layer.3.k_cache 0.00127405 0.75468963 - layer.3.v_cache 0.00000234 0.00221793 - layer.4.k_cache 0.00358677 1.81542761 - layer.4.v_cache 0.00000306 0.00329745 - layer.4.output 0.00018722 0.10286529 - ------------------------------------------------------------------------------------- - TOTAL 0.00263945 1.03364351 - (elements=2,953,216) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2953216 -Total Bytes 508228 -BPFP 1.3767 bits/point -EBPFP 2.7535 equivalent bits/point -MSE 1.033644 ----------------------- -------------------------------------------------------- -Time: 4.296s Load: 0.011s, Pack+Encode: 2.451s, Decode+Unpack: 1.834s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 206, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 206, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0336 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample32-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample32-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 191, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 191, 128) -Output shape: (1, 191, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) -> torch.Size([1, 1, 191, 1024]) - layer.4.output: torch.Size([1, 191, 4096]) -> torch.Size([1, 1, 191, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,360B, BPFP=0.1374 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,116B, BPFP=1.4364 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 16,816B, BPFP=0.6878 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,884B, BPFP=1.5496 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,904B, BPFP=1.1005 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,452B, BPFP=1.5728 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,432B, BPFP=1.0812 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,688B, BPFP=1.5416 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,640B, BPFP=0.7215 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,824B, BPFP=1.5880 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 82,940B, BPFP=0.8481 -⌛️ [2/4] FRONTEND: Frontend time: 1.939s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 191, 128]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.output: torch.Size([1, 191, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.584s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 191, 128]) - layer.0.v_cache: torch.Size([1, 8, 191, 128]) - layer.1.k_cache: torch.Size([1, 8, 191, 128]) - layer.1.v_cache: torch.Size([1, 8, 191, 128]) - layer.2.k_cache: torch.Size([1, 8, 191, 128]) - layer.2.v_cache: torch.Size([1, 8, 191, 128]) - layer.3.k_cache: torch.Size([1, 8, 191, 128]) - layer.3.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.k_cache: torch.Size([1, 8, 191, 128]) - layer.4.v_cache: torch.Size([1, 8, 191, 128]) - layer.4.output: torch.Size([1, 191, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702075 8.94761586 - layer.0.v_cache 0.00000026 0.00023852 - layer.1.k_cache 0.00310674 1.74768610 - layer.1.v_cache 0.00000076 0.00080417 - layer.2.k_cache 0.00119798 0.67187364 - layer.2.v_cache 0.00000111 0.00119511 - layer.3.k_cache 0.00132973 0.75172640 - layer.3.v_cache 0.00000216 0.00202505 - layer.4.k_cache 0.00342601 2.09448338 - layer.4.v_cache 0.00000312 0.00333445 - layer.4.output 0.00019316 0.09176560 - ------------------------------------------------------------------------------------- - TOTAL 0.00263295 1.04200322 - (elements=2,738,176) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2738176 -Total Bytes 362056 -BPFP 1.0578 bits/point -EBPFP 2.1156 equivalent bits/point -MSE 1.042003 ----------------------- -------------------------------------------------------- -Time: 3.534s Load: 0.010s, Pack+Encode: 1.939s, Decode+Unpack: 1.584s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 191, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 191, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0420 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample33-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample33-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,276B, BPFP=0.1479 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,484B, BPFP=1.6476 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,016B, BPFP=0.8587 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,380B, BPFP=1.7784 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,388B, BPFP=1.1013 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,036B, BPFP=1.7628 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,996B, BPFP=1.2191 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,356B, BPFP=1.7321 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,704B, BPFP=0.8898 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,760B, BPFP=1.7504 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,876B, BPFP=0.8905 -⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.560s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02805312 9.16135296 - layer.0.v_cache 0.00000027 0.00024797 - layer.1.k_cache 0.00341978 1.77088443 - layer.1.v_cache 0.00000082 0.00092445 - layer.2.k_cache 0.00114483 0.62437682 - layer.2.v_cache 0.00000117 0.00132734 - layer.3.k_cache 0.00131871 0.75173033 - layer.3.v_cache 0.00000219 0.00217969 - layer.4.k_cache 0.00345933 1.99177560 - layer.4.v_cache 0.00000316 0.00359276 - layer.4.output 0.00017521 0.10267263 - ------------------------------------------------------------------------------------- - TOTAL 0.00272173 1.05136306 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 364272 -BPFP 1.1750 bits/point -EBPFP 2.3500 equivalent bits/point -MSE 1.051363 ----------------------- -------------------------------------------------------- -Time: 3.456s Load: 0.009s, Pack+Encode: 1.887s, Decode+Unpack: 1.560s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0514 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample35-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample35-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,152B, BPFP=0.1457 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,868B, BPFP=1.6581 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,112B, BPFP=0.8835 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,716B, BPFP=1.7898 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,672B, BPFP=1.2792 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,120B, BPFP=1.7622 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,672B, BPFP=1.3254 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,548B, BPFP=1.7358 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,752B, BPFP=0.9131 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,548B, BPFP=1.7820 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 81,968B, BPFP=0.9473 -⌛️ [2/4] FRONTEND: Frontend time: 1.969s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.437s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02802299 9.12907744 - layer.0.v_cache 0.00000027 0.00025290 - layer.1.k_cache 0.00309164 1.67810456 - layer.1.v_cache 0.00000083 0.00093972 - layer.2.k_cache 0.00116255 0.66628568 - layer.2.v_cache 0.00000115 0.00132864 - layer.3.k_cache 0.00130840 0.79944146 - layer.3.v_cache 0.00000220 0.00223636 - layer.4.k_cache 0.00347752 1.78893537 - layer.4.v_cache 0.00000302 0.00352220 - layer.4.output 0.00021363 0.10793614 - ------------------------------------------------------------------------------------- - TOTAL 0.00270893 1.03584778 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 369128 -BPFP 1.2189 bits/point -EBPFP 2.4377 equivalent bits/point -MSE 1.035848 ----------------------- -------------------------------------------------------- -Time: 3.415s Load: 0.009s, Pack+Encode: 1.969s, Decode+Unpack: 1.437s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0358 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,500B, BPFP=0.1447 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,540B, BPFP=1.4691 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,000B, BPFP=0.7854 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,572B, BPFP=1.5531 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,568B, BPFP=1.0569 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,236B, BPFP=1.5805 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,560B, BPFP=1.1392 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,172B, BPFP=1.5365 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,584B, BPFP=0.8095 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,208B, BPFP=1.5794 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 94,492B, BPFP=0.9765 -⌛️ [2/4] FRONTEND: Frontend time: 1.938s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.600s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02754041 9.59657247 - layer.0.v_cache 0.00000027 0.00026249 - layer.1.k_cache 0.00311113 1.95622116 - layer.1.v_cache 0.00000084 0.00091247 - layer.2.k_cache 0.00124070 0.66249625 - layer.2.v_cache 0.00000144 0.00139469 - layer.3.k_cache 0.00134127 0.81280203 - layer.3.v_cache 0.00000226 0.00227443 - layer.4.k_cache 0.00350879 1.92226446 - layer.4.v_cache 0.00000320 0.00369376 - layer.4.output 0.00022464 0.12374474 - ------------------------------------------------------------------------------------- - TOTAL 0.00268921 1.10384809 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 376432 -BPFP 1.1114 bits/point -EBPFP 2.2229 equivalent bits/point -MSE 1.103848 ----------------------- -------------------------------------------------------- -Time: 3.549s Load: 0.011s, Pack+Encode: 1.938s, Decode+Unpack: 1.600s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1038 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample37-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample37-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,296B, BPFP=0.1455 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,096B, BPFP=1.5491 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,092B, BPFP=0.7986 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,756B, BPFP=1.6665 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,472B, BPFP=1.1243 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,312B, BPFP=1.6910 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,964B, BPFP=1.2343 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,152B, BPFP=1.6398 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,348B, BPFP=0.8099 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,488B, BPFP=1.6988 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,692B, BPFP=0.8794 -⌛️ [2/4] FRONTEND: Frontend time: 1.867s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.606s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02865209 9.16175489 - layer.0.v_cache 0.00000026 0.00024831 - layer.1.k_cache 0.00315540 1.72573042 - layer.1.v_cache 0.00000078 0.00088771 - layer.2.k_cache 0.00119797 0.67400201 - layer.2.v_cache 0.00000110 0.00128343 - layer.3.k_cache 0.00133161 0.75403832 - layer.3.v_cache 0.00000204 0.00205564 - layer.4.k_cache 0.00358407 1.79185667 - layer.4.v_cache 0.00000307 0.00342850 - layer.4.output 0.00019353 0.09655066 - ------------------------------------------------------------------------------------- - TOTAL 0.00276446 1.03582061 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 359668 -BPFP 1.1339 bits/point -EBPFP 2.2679 equivalent bits/point -MSE 1.035821 ----------------------- -------------------------------------------------------- -Time: 3.482s Load: 0.009s, Pack+Encode: 1.867s, Decode+Unpack: 1.606s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0358 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample38-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample38-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,252B, BPFP=0.1469 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,372B, BPFP=1.6425 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,420B, BPFP=0.8770 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 40,312B, BPFP=1.8204 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,092B, BPFP=1.1331 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,112B, BPFP=1.7663 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,972B, BPFP=1.2632 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,656B, BPFP=1.7457 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,456B, BPFP=0.8786 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,768B, BPFP=1.7507 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 98,428B, BPFP=1.1112 -⌛️ [2/4] FRONTEND: Frontend time: 1.963s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.439s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02612204 9.26535546 - layer.0.v_cache 0.00000027 0.00025513 - layer.1.k_cache 0.00302657 1.63265409 - layer.1.v_cache 0.00000086 0.00091501 - layer.2.k_cache 0.00117788 0.63341685 - layer.2.v_cache 0.00000116 0.00134726 - layer.3.k_cache 0.00128963 0.74453056 - layer.3.v_cache 0.00000233 0.00221092 - layer.4.k_cache 0.00354125 1.87740577 - layer.4.v_cache 0.00000310 0.00348187 - layer.4.output 0.00019853 0.10746117 - ------------------------------------------------------------------------------------- - TOTAL 0.00256851 1.04224411 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 386840 -BPFP 1.2478 bits/point -EBPFP 2.4956 equivalent bits/point -MSE 1.042244 ----------------------- -------------------------------------------------------- -Time: 3.412s Load: 0.010s, Pack+Encode: 1.963s, Decode+Unpack: 1.439s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0422 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample39-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample39-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 243, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 243, 128) -Output shape: (1, 243, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) -> torch.Size([1, 1, 243, 1024]) - layer.4.output: torch.Size([1, 243, 4096]) -> torch.Size([1, 1, 243, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,280B, BPFP=0.1376 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,436B, BPFP=1.5251 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,484B, BPFP=0.7872 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 51,024B, BPFP=1.6404 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 36,036B, BPFP=1.1586 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 51,680B, BPFP=1.6615 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 35,592B, BPFP=1.1443 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 50,016B, BPFP=1.6080 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,440B, BPFP=0.7858 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 51,224B, BPFP=1.6469 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 116,108B, BPFP=0.9332 -⌛️ [2/4] FRONTEND: Frontend time: 2.326s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 243, 128]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.output: torch.Size([1, 243, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.001s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 243, 128]) - layer.0.v_cache: torch.Size([1, 8, 243, 128]) - layer.1.k_cache: torch.Size([1, 8, 243, 128]) - layer.1.v_cache: torch.Size([1, 8, 243, 128]) - layer.2.k_cache: torch.Size([1, 8, 243, 128]) - layer.2.v_cache: torch.Size([1, 8, 243, 128]) - layer.3.k_cache: torch.Size([1, 8, 243, 128]) - layer.3.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.k_cache: torch.Size([1, 8, 243, 128]) - layer.4.v_cache: torch.Size([1, 8, 243, 128]) - layer.4.output: torch.Size([1, 243, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02455453 7.95266485 - layer.0.v_cache 0.00000027 0.00024798 - layer.1.k_cache 0.00295349 1.62457363 - layer.1.v_cache 0.00000082 0.00092298 - layer.2.k_cache 0.00119501 0.66077257 - layer.2.v_cache 0.00000127 0.00138735 - layer.3.k_cache 0.00131209 0.72251226 - layer.3.v_cache 0.00000246 0.00230157 - layer.4.k_cache 0.00353116 2.01401121 - layer.4.v_cache 0.00000321 0.00353263 - layer.4.output 0.00017298 0.09291335 - ------------------------------------------------------------------------------------- - TOTAL 0.00244616 0.95389860 - (elements=3,483,648) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3483648 -Total Bytes 492320 -BPFP 1.1306 bits/point -EBPFP 2.2612 equivalent bits/point -MSE 0.953899 ----------------------- -------------------------------------------------------- -Time: 4.340s Load: 0.013s, Pack+Encode: 2.326s, Decode+Unpack: 2.001s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 243, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 243, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9539 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample4-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample4-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 152, 128) -Output shape: (1, 152, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) -> torch.Size([1, 1, 152, 1024]) - layer.4.output: torch.Size([1, 152, 4096]) -> torch.Size([1, 1, 152, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,912B, BPFP=0.1497 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,996B, BPFP=1.9015 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,088B, BPFP=0.9811 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,240B, BPFP=2.1711 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,736B, BPFP=1.2714 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 44,640B, BPFP=2.2944 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,900B, BPFP=1.4854 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,008B, BPFP=2.1591 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,944B, BPFP=1.0251 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 42,096B, BPFP=2.1637 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 86,732B, BPFP=1.1145 -⌛️ [2/4] FRONTEND: Frontend time: 1.903s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.609s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 152, 128]) - layer.0.v_cache: torch.Size([1, 8, 152, 128]) - layer.1.k_cache: torch.Size([1, 8, 152, 128]) - layer.1.v_cache: torch.Size([1, 8, 152, 128]) - layer.2.k_cache: torch.Size([1, 8, 152, 128]) - layer.2.v_cache: torch.Size([1, 8, 152, 128]) - layer.3.k_cache: torch.Size([1, 8, 152, 128]) - layer.3.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.k_cache: torch.Size([1, 8, 152, 128]) - layer.4.v_cache: torch.Size([1, 8, 152, 128]) - layer.4.output: torch.Size([1, 152, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02793655 9.51049162 - layer.0.v_cache 0.00000027 0.00026804 - layer.1.k_cache 0.00308216 1.51824038 - layer.1.v_cache 0.00000096 0.00092070 - layer.2.k_cache 0.00119397 0.65045106 - layer.2.v_cache 0.00000115 0.00137198 - layer.3.k_cache 0.00132015 0.74450909 - layer.3.v_cache 0.00000229 0.00234673 - layer.4.k_cache 0.00355994 1.90181812 - layer.4.v_cache 0.00000322 0.00374423 - layer.4.output 0.00016979 0.09797885 - ------------------------------------------------------------------------------------- - TOTAL 0.00269856 1.05186267 - (elements=2,179,072) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2179072 -Total Bytes 390292 -BPFP 1.4329 bits/point -EBPFP 2.8657 equivalent bits/point -MSE 1.051863 ----------------------- -------------------------------------------------------- -Time: 3.521s Load: 0.009s, Pack+Encode: 1.903s, Decode+Unpack: 1.609s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 152, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 152, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0519 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,452B, BPFP=0.1458 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,964B, BPFP=1.5188 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,688B, BPFP=0.8314 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,552B, BPFP=1.6280 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,796B, BPFP=1.0894 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,916B, BPFP=1.6434 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,752B, BPFP=1.1720 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,960B, BPFP=1.6030 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,372B, BPFP=0.7758 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,096B, BPFP=1.6510 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 77,788B, BPFP=0.8212 -⌛️ [2/4] FRONTEND: Frontend time: 1.914s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.429s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02732194 8.81897501 - layer.0.v_cache 0.00000026 0.00025938 - layer.1.k_cache 0.00316422 1.63736605 - layer.1.v_cache 0.00000081 0.00092900 - layer.2.k_cache 0.00119231 0.66132775 - layer.2.v_cache 0.00000116 0.00133504 - layer.3.k_cache 0.00128139 0.74792827 - layer.3.v_cache 0.00000230 0.00228321 - layer.4.k_cache 0.00354606 1.78219093 - layer.4.v_cache 0.00000329 0.00357424 - layer.4.output 0.00016551 0.08757306 - ------------------------------------------------------------------------------------- - TOTAL 0.00265541 1.00046151 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 363336 -BPFP 1.0960 bits/point -EBPFP 2.1919 equivalent bits/point -MSE 1.000462 ----------------------- -------------------------------------------------------- -Time: 3.352s Load: 0.009s, Pack+Encode: 1.914s, Decode+Unpack: 1.429s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0005 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample41-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample41-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,120B, BPFP=0.1442 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,600B, BPFP=1.6919 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,056B, BPFP=0.8347 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,912B, BPFP=1.7988 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,008B, BPFP=1.2485 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,436B, BPFP=1.7768 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,876B, BPFP=1.3349 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,588B, BPFP=1.7376 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,316B, BPFP=0.9392 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,364B, BPFP=1.7735 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,752B, BPFP=0.9217 -⌛️ [2/4] FRONTEND: Frontend time: 1.961s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.539s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02650603 8.98175302 - layer.0.v_cache 0.00000028 0.00026137 - layer.1.k_cache 0.00313577 1.87723338 - layer.1.v_cache 0.00000090 0.00092658 - layer.2.k_cache 0.00117709 0.66966053 - layer.2.v_cache 0.00000116 0.00135744 - layer.3.k_cache 0.00129027 0.77544814 - layer.3.v_cache 0.00000225 0.00227563 - layer.4.k_cache 0.00358651 1.86877080 - layer.4.v_cache 0.00000317 0.00364844 - layer.4.output 0.00019489 0.10011039 - ------------------------------------------------------------------------------------- - TOTAL 0.00260593 1.04155549 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 367028 -BPFP 1.2119 bits/point -EBPFP 2.4238 equivalent bits/point -MSE 1.041555 ----------------------- -------------------------------------------------------- -Time: 3.509s Load: 0.009s, Pack+Encode: 1.961s, Decode+Unpack: 1.539s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0416 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample42-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample42-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,216B, BPFP=0.1478 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,464B, BPFP=1.6298 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,844B, BPFP=0.8660 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 40,332B, BPFP=1.8535 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,648B, BPFP=1.2246 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,056B, BPFP=1.7949 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,412B, BPFP=1.2597 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,752B, BPFP=1.7349 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,716B, BPFP=0.9520 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,288B, BPFP=1.7596 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 64,968B, BPFP=0.7464 -⌛️ [2/4] FRONTEND: Frontend time: 1.890s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.583s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02682906 9.18835377 - layer.0.v_cache 0.00000026 0.00024883 - layer.1.k_cache 0.00324658 1.69055445 - layer.1.v_cache 0.00000081 0.00090961 - layer.2.k_cache 0.00118449 0.67555739 - layer.2.v_cache 0.00000107 0.00126221 - layer.3.k_cache 0.00127207 0.75909397 - layer.3.v_cache 0.00000215 0.00212749 - layer.4.k_cache 0.00355165 1.93346486 - layer.4.v_cache 0.00000309 0.00355587 - layer.4.output 0.00015524 0.09534495 - ------------------------------------------------------------------------------------- - TOTAL 0.00262230 1.04546487 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 352696 -BPFP 1.1577 bits/point -EBPFP 2.3155 equivalent bits/point -MSE 1.045465 ----------------------- -------------------------------------------------------- -Time: 3.482s Load: 0.009s, Pack+Encode: 1.890s, Decode+Unpack: 1.583s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0455 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample43-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample43-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 194, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 194, 128) -Output shape: (1, 194, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) -> torch.Size([1, 1, 194, 1024]) - layer.4.output: torch.Size([1, 194, 4096]) -> torch.Size([1, 1, 194, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,600B, BPFP=0.1450 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 46,440B, BPFP=1.8702 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,468B, BPFP=1.0256 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,756B, BPFP=2.0440 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,560B, BPFP=1.1501 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,748B, BPFP=2.0034 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,028B, BPFP=1.4509 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,248B, BPFP=1.9832 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,444B, BPFP=1.0246 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 49,252B, BPFP=1.9834 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 89,480B, BPFP=0.9009 -⌛️ [2/4] FRONTEND: Frontend time: 2.278s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 194, 128]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.output: torch.Size([1, 194, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.913s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 194, 128]) - layer.0.v_cache: torch.Size([1, 8, 194, 128]) - layer.1.k_cache: torch.Size([1, 8, 194, 128]) - layer.1.v_cache: torch.Size([1, 8, 194, 128]) - layer.2.k_cache: torch.Size([1, 8, 194, 128]) - layer.2.v_cache: torch.Size([1, 8, 194, 128]) - layer.3.k_cache: torch.Size([1, 8, 194, 128]) - layer.3.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.k_cache: torch.Size([1, 8, 194, 128]) - layer.4.v_cache: torch.Size([1, 8, 194, 128]) - layer.4.output: torch.Size([1, 194, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710510 9.10116797 - layer.0.v_cache 0.00000027 0.00025635 - layer.1.k_cache 0.00315558 1.73084865 - layer.1.v_cache 0.00000085 0.00090679 - layer.2.k_cache 0.00115955 0.63541436 - layer.2.v_cache 0.00000114 0.00131354 - layer.3.k_cache 0.00127895 0.76197319 - layer.3.v_cache 0.00000244 0.00222614 - layer.4.k_cache 0.00355671 1.83023480 - layer.4.v_cache 0.00000315 0.00356732 - layer.4.output 0.00014532 0.09460480 - ------------------------------------------------------------------------------------- - TOTAL 0.00263179 1.03188059 - (elements=2,781,184) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2781184 -Total Bytes 454024 -BPFP 1.3060 bits/point -EBPFP 2.6120 equivalent bits/point -MSE 1.031881 ----------------------- -------------------------------------------------------- -Time: 4.201s Load: 0.010s, Pack+Encode: 2.278s, Decode+Unpack: 1.913s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 194, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 194, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0319 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,148B, BPFP=0.1447 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,544B, BPFP=1.6335 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,216B, BPFP=0.8371 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,260B, BPFP=1.8042 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,036B, BPFP=1.1506 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,908B, BPFP=1.7881 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,628B, BPFP=1.2697 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,360B, BPFP=1.7629 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 16,724B, BPFP=0.7686 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,332B, BPFP=1.7616 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 87,452B, BPFP=1.0047 -⌛️ [2/4] FRONTEND: Frontend time: 1.981s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.435s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02613039 9.24408892 - layer.0.v_cache 0.00000027 0.00026445 - layer.1.k_cache 0.00316806 1.69518289 - layer.1.v_cache 0.00000083 0.00092296 - layer.2.k_cache 0.00117426 0.66932701 - layer.2.v_cache 0.00000117 0.00132034 - layer.3.k_cache 0.00130856 0.78285280 - layer.3.v_cache 0.00000223 0.00221447 - layer.4.k_cache 0.00356532 1.68983442 - layer.4.v_cache 0.00000295 0.00333981 - layer.4.output 0.00018112 0.09941420 - ------------------------------------------------------------------------------------- - TOTAL 0.00257704 1.03478606 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 368608 -BPFP 1.2100 bits/point -EBPFP 2.4200 equivalent bits/point -MSE 1.034786 ----------------------- -------------------------------------------------------- -Time: 3.426s Load: 0.009s, Pack+Encode: 1.981s, Decode+Unpack: 1.435s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0348 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample45-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample45-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 198, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 198, 128) -Output shape: (1, 198, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) -> torch.Size([1, 1, 198, 1024]) - layer.4.output: torch.Size([1, 198, 4096]) -> torch.Size([1, 1, 198, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,616B, BPFP=0.1427 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 44,836B, BPFP=1.7691 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,920B, BPFP=1.0227 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 49,240B, BPFP=1.9429 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,620B, BPFP=1.2082 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,900B, BPFP=1.9689 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,048B, BPFP=1.4618 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 48,984B, BPFP=1.9328 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 26,456B, BPFP=1.0439 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 49,800B, BPFP=1.9650 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 74,104B, BPFP=0.7310 -⌛️ [2/4] FRONTEND: Frontend time: 2.309s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 198, 128]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.output: torch.Size([1, 198, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.970s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 198, 128]) - layer.0.v_cache: torch.Size([1, 8, 198, 128]) - layer.1.k_cache: torch.Size([1, 8, 198, 128]) - layer.1.v_cache: torch.Size([1, 8, 198, 128]) - layer.2.k_cache: torch.Size([1, 8, 198, 128]) - layer.2.v_cache: torch.Size([1, 8, 198, 128]) - layer.3.k_cache: torch.Size([1, 8, 198, 128]) - layer.3.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.k_cache: torch.Size([1, 8, 198, 128]) - layer.4.v_cache: torch.Size([1, 8, 198, 128]) - layer.4.output: torch.Size([1, 198, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02819830 8.54346196 - layer.0.v_cache 0.00000029 0.00026534 - layer.1.k_cache 0.00303339 1.77475870 - layer.1.v_cache 0.00000075 0.00084239 - layer.2.k_cache 0.00117903 0.66737851 - layer.2.v_cache 0.00000119 0.00128088 - layer.3.k_cache 0.00132068 0.76923679 - layer.3.v_cache 0.00000222 0.00217796 - layer.4.k_cache 0.00360013 1.93949705 - layer.4.v_cache 0.00000329 0.00347936 - layer.4.output 0.00021245 0.10174848 - ------------------------------------------------------------------------------------- - TOTAL 0.00272779 1.00781235 - (elements=2,838,528) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2838528 -Total Bytes 440524 -BPFP 1.2416 bits/point -EBPFP 2.4831 equivalent bits/point -MSE 1.007812 ----------------------- -------------------------------------------------------- -Time: 4.290s Load: 0.011s, Pack+Encode: 2.309s, Decode+Unpack: 1.970s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 198, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 198, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0078 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample46-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample46-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,252B, BPFP=0.1469 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,652B, BPFP=1.6100 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,464B, BPFP=0.8338 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,836B, BPFP=1.7990 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,012B, BPFP=1.1747 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,072B, BPFP=1.7645 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,560B, BPFP=1.2897 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,096B, BPFP=1.7204 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,228B, BPFP=0.9586 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,016B, BPFP=1.7619 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,824B, BPFP=0.9012 -⌛️ [2/4] FRONTEND: Frontend time: 1.894s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.598s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02631039 8.77336271 - layer.0.v_cache 0.00000026 0.00025013 - layer.1.k_cache 0.00315605 1.60819928 - layer.1.v_cache 0.00000083 0.00091762 - layer.2.k_cache 0.00121195 0.66939346 - layer.2.v_cache 0.00000119 0.00137020 - layer.3.k_cache 0.00131129 0.76680548 - layer.3.v_cache 0.00000228 0.00225030 - layer.4.k_cache 0.00352345 2.01135624 - layer.4.v_cache 0.00000333 0.00376774 - layer.4.output 0.00017266 0.10740735 - ------------------------------------------------------------------------------------- - TOTAL 0.00258655 1.01909304 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 369012 -BPFP 1.1903 bits/point -EBPFP 2.3806 equivalent bits/point -MSE 1.019093 ----------------------- -------------------------------------------------------- -Time: 3.503s Load: 0.010s, Pack+Encode: 1.894s, Decode+Unpack: 1.598s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0191 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 175, 128) -Output shape: (1, 175, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) -> torch.Size([1, 1, 175, 1024]) - layer.4.output: torch.Size([1, 175, 4096]) -> torch.Size([1, 1, 175, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,260B, BPFP=0.1455 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,076B, BPFP=1.6105 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,096B, BPFP=0.8525 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,040B, BPFP=1.6982 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,200B, BPFP=1.1250 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,568B, BPFP=1.7218 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,488B, BPFP=1.2718 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,564B, BPFP=1.6770 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,752B, BPFP=0.8371 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,948B, BPFP=1.6941 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 80,404B, BPFP=0.8974 -⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.462s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 175, 128]) - layer.0.v_cache: torch.Size([1, 8, 175, 128]) - layer.1.k_cache: torch.Size([1, 8, 175, 128]) - layer.1.v_cache: torch.Size([1, 8, 175, 128]) - layer.2.k_cache: torch.Size([1, 8, 175, 128]) - layer.2.v_cache: torch.Size([1, 8, 175, 128]) - layer.3.k_cache: torch.Size([1, 8, 175, 128]) - layer.3.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.k_cache: torch.Size([1, 8, 175, 128]) - layer.4.v_cache: torch.Size([1, 8, 175, 128]) - layer.4.output: torch.Size([1, 175, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03100830 8.73972866 - layer.0.v_cache 0.00000027 0.00026152 - layer.1.k_cache 0.00306070 1.55416818 - layer.1.v_cache 0.00000080 0.00090075 - layer.2.k_cache 0.00119082 0.65014862 - layer.2.v_cache 0.00000118 0.00138553 - layer.3.k_cache 0.00134761 0.76605861 - layer.3.v_cache 0.00000229 0.00229267 - layer.4.k_cache 0.00348653 1.73818447 - layer.4.v_cache 0.00000313 0.00348828 - layer.4.output 0.00023428 0.10562609 - ------------------------------------------------------------------------------------- - TOTAL 0.00293134 0.99136583 - (elements=2,508,800) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2508800 -Total Bytes 363396 -BPFP 1.1588 bits/point -EBPFP 2.3176 equivalent bits/point -MSE 0.991366 ----------------------- -------------------------------------------------------- -Time: 3.416s Load: 0.011s, Pack+Encode: 1.944s, Decode+Unpack: 1.462s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 175, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 175, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9914 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample49-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample49-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 240, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.014s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 240, 128) -Output shape: (1, 240, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) -> torch.Size([1, 1, 240, 1024]) - layer.4.output: torch.Size([1, 240, 4096]) -> torch.Size([1, 1, 240, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,352B, BPFP=0.1417 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 46,992B, BPFP=1.5297 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,496B, BPFP=0.7974 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,880B, BPFP=1.6562 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 34,652B, BPFP=1.1280 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 51,360B, BPFP=1.6719 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,020B, BPFP=1.2051 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 50,000B, BPFP=1.6276 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 24,716B, BPFP=0.8046 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 51,364B, BPFP=1.6720 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 121,620B, BPFP=0.9897 -⌛️ [2/4] FRONTEND: Frontend time: 2.421s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 240, 128]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.output: torch.Size([1, 240, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.842s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 240, 128]) - layer.0.v_cache: torch.Size([1, 8, 240, 128]) - layer.1.k_cache: torch.Size([1, 8, 240, 128]) - layer.1.v_cache: torch.Size([1, 8, 240, 128]) - layer.2.k_cache: torch.Size([1, 8, 240, 128]) - layer.2.v_cache: torch.Size([1, 8, 240, 128]) - layer.3.k_cache: torch.Size([1, 8, 240, 128]) - layer.3.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.k_cache: torch.Size([1, 8, 240, 128]) - layer.4.v_cache: torch.Size([1, 8, 240, 128]) - layer.4.output: torch.Size([1, 240, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02660088 8.56784871 - layer.0.v_cache 0.00000028 0.00026016 - layer.1.k_cache 0.00295962 1.59051107 - layer.1.v_cache 0.00000087 0.00091636 - layer.2.k_cache 0.00121828 0.64100291 - layer.2.v_cache 0.00000120 0.00131165 - layer.3.k_cache 0.00127180 0.73601907 - layer.3.v_cache 0.00000260 0.00225467 - layer.4.k_cache 0.00365680 1.73777212 - layer.4.v_cache 0.00000312 0.00347457 - layer.4.output 0.00015244 0.08784887 - ------------------------------------------------------------------------------------- - TOTAL 0.00259466 0.97376905 - (elements=3,440,640) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3440640 -Total Bytes 497452 -BPFP 1.1566 bits/point -EBPFP 2.3133 equivalent bits/point -MSE 0.973769 ----------------------- -------------------------------------------------------- -Time: 4.277s Load: 0.014s, Pack+Encode: 2.421s, Decode+Unpack: 1.842s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 240, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 240, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9738 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample5-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample5-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,580B, BPFP=0.1480 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,784B, BPFP=1.4792 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,492B, BPFP=0.8057 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,440B, BPFP=1.5890 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,084B, BPFP=1.1195 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,628B, BPFP=1.5967 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,124B, BPFP=1.1212 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,544B, BPFP=1.5519 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,736B, BPFP=0.7745 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,772B, BPFP=1.6027 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 96,464B, BPFP=0.9969 -⌛️ [2/4] FRONTEND: Frontend time: 1.910s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.670s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02646759 8.89445504 - layer.0.v_cache 0.00000027 0.00025682 - layer.1.k_cache 0.00304090 1.94681464 - layer.1.v_cache 0.00000085 0.00094374 - layer.2.k_cache 0.00115999 0.69540841 - layer.2.v_cache 0.00000120 0.00136873 - layer.3.k_cache 0.00125869 0.77425502 - layer.3.v_cache 0.00000235 0.00225415 - layer.4.k_cache 0.00349331 1.92496373 - layer.4.v_cache 0.00000329 0.00361780 - layer.4.output 0.00016341 0.09527834 - ------------------------------------------------------------------------------------- - TOTAL 0.00257729 1.04467510 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 381648 -BPFP 1.1268 bits/point -EBPFP 2.2537 equivalent bits/point -MSE 1.044675 ----------------------- -------------------------------------------------------- -Time: 3.590s Load: 0.010s, Pack+Encode: 1.910s, Decode+Unpack: 1.670s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0447 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 170, 128) -Output shape: (1, 170, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) -> torch.Size([1, 1, 170, 1024]) - layer.4.output: torch.Size([1, 170, 4096]) -> torch.Size([1, 1, 170, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,244B, BPFP=0.1491 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,552B, BPFP=1.6338 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,200B, BPFP=0.8364 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,040B, BPFP=1.7482 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,664B, BPFP=1.1794 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,300B, BPFP=1.7601 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,616B, BPFP=1.3151 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,628B, BPFP=1.7292 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,440B, BPFP=0.9393 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,788B, BPFP=1.7825 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 79,368B, BPFP=0.9119 -⌛️ [2/4] FRONTEND: Frontend time: 1.878s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.528s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 170, 128]) - layer.0.v_cache: torch.Size([1, 8, 170, 128]) - layer.1.k_cache: torch.Size([1, 8, 170, 128]) - layer.1.v_cache: torch.Size([1, 8, 170, 128]) - layer.2.k_cache: torch.Size([1, 8, 170, 128]) - layer.2.v_cache: torch.Size([1, 8, 170, 128]) - layer.3.k_cache: torch.Size([1, 8, 170, 128]) - layer.3.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.k_cache: torch.Size([1, 8, 170, 128]) - layer.4.v_cache: torch.Size([1, 8, 170, 128]) - layer.4.output: torch.Size([1, 170, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03233885 10.25079992 - layer.0.v_cache 0.00000026 0.00025012 - layer.1.k_cache 0.00307676 1.68396732 - layer.1.v_cache 0.00000080 0.00089128 - layer.2.k_cache 0.00113683 0.64509968 - layer.2.v_cache 0.00000124 0.00132683 - layer.3.k_cache 0.00130099 0.75817494 - layer.3.v_cache 0.00000223 0.00219458 - layer.4.k_cache 0.00352303 1.95528277 - layer.4.v_cache 0.00000334 0.00362567 - layer.4.output 0.00016991 0.10112504 - ------------------------------------------------------------------------------------- - TOTAL 0.00300457 1.12186524 - (elements=2,437,120) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2437120 -Total Bytes 363840 -BPFP 1.1943 bits/point -EBPFP 2.3887 equivalent bits/point -MSE 1.121865 ----------------------- -------------------------------------------------------- -Time: 3.415s Load: 0.009s, Pack+Encode: 1.878s, Decode+Unpack: 1.528s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 170, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 170, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1219 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample51-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample51-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,008B, BPFP=0.1487 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,164B, BPFP=1.7387 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,700B, BPFP=0.9246 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,884B, BPFP=1.9721 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,908B, BPFP=1.2316 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,276B, BPFP=1.8926 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,016B, BPFP=1.3853 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,132B, BPFP=1.8360 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,616B, BPFP=1.0194 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,584B, BPFP=1.9573 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,460B, BPFP=0.9699 -⌛️ [2/4] FRONTEND: Frontend time: 1.949s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.433s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02846188 8.65889923 - layer.0.v_cache 0.00000028 0.00026210 - layer.1.k_cache 0.00321952 1.98062597 - layer.1.v_cache 0.00000089 0.00093488 - layer.2.k_cache 0.00116433 0.64450295 - layer.2.v_cache 0.00000127 0.00138371 - layer.3.k_cache 0.00135431 0.78872719 - layer.3.v_cache 0.00000232 0.00227678 - layer.4.k_cache 0.00338441 1.89481363 - layer.4.v_cache 0.00000314 0.00359349 - layer.4.output 0.00020823 0.09494488 - ------------------------------------------------------------------------------------- - TOTAL 0.00274466 1.02541425 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 363748 -BPFP 1.2847 bits/point -EBPFP 2.5694 equivalent bits/point -MSE 1.025414 ----------------------- -------------------------------------------------------- -Time: 3.390s Load: 0.009s, Pack+Encode: 1.949s, Decode+Unpack: 1.433s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0254 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample52-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample52-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 185, 128) -Output shape: (1, 185, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) -> torch.Size([1, 1, 185, 1024]) - layer.4.output: torch.Size([1, 185, 4096]) -> torch.Size([1, 1, 185, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,464B, BPFP=0.1463 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,048B, BPFP=1.5223 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,816B, BPFP=0.8368 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,052B, BPFP=1.6069 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,872B, BPFP=1.1348 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,764B, BPFP=1.6370 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,056B, BPFP=1.1848 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,832B, BPFP=1.5976 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,380B, BPFP=0.8184 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,068B, BPFP=1.6498 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 77,516B, BPFP=0.8184 -⌛️ [2/4] FRONTEND: Frontend time: 1.907s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.687s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 185, 128]) - layer.0.v_cache: torch.Size([1, 8, 185, 128]) - layer.1.k_cache: torch.Size([1, 8, 185, 128]) - layer.1.v_cache: torch.Size([1, 8, 185, 128]) - layer.2.k_cache: torch.Size([1, 8, 185, 128]) - layer.2.v_cache: torch.Size([1, 8, 185, 128]) - layer.3.k_cache: torch.Size([1, 8, 185, 128]) - layer.3.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.k_cache: torch.Size([1, 8, 185, 128]) - layer.4.v_cache: torch.Size([1, 8, 185, 128]) - layer.4.output: torch.Size([1, 185, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02684415 8.59376518 - layer.0.v_cache 0.00000026 0.00025680 - layer.1.k_cache 0.00308337 1.67585614 - layer.1.v_cache 0.00000086 0.00090611 - layer.2.k_cache 0.00119263 0.66472465 - layer.2.v_cache 0.00000115 0.00134094 - layer.3.k_cache 0.00133167 0.76854660 - layer.3.v_cache 0.00000224 0.00230201 - layer.4.k_cache 0.00346133 1.78905128 - layer.4.v_cache 0.00000321 0.00376674 - layer.4.output 0.00019684 0.09083755 - ------------------------------------------------------------------------------------- - TOTAL 0.00262202 0.99027619 - (elements=2,652,160) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2652160 -Total Bytes 364868 -BPFP 1.1006 bits/point -EBPFP 2.2012 equivalent bits/point -MSE 0.990276 ----------------------- -------------------------------------------------------- -Time: 3.605s Load: 0.011s, Pack+Encode: 1.907s, Decode+Unpack: 1.687s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 185, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 185, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9903 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 165, 128) -Output shape: (1, 165, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) -> torch.Size([1, 1, 165, 1024]) - layer.4.output: torch.Size([1, 165, 4096]) -> torch.Size([1, 1, 165, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,152B, BPFP=0.1492 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,536B, BPFP=1.6826 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,932B, BPFP=0.8964 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,420B, BPFP=1.8191 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,372B, BPFP=1.2487 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,080B, BPFP=1.8504 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,104B, BPFP=1.3307 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,540B, BPFP=1.8248 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 22,164B, BPFP=1.0494 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,280B, BPFP=1.8125 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,508B, BPFP=0.9293 -⌛️ [2/4] FRONTEND: Frontend time: 1.874s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.523s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 165, 128]) - layer.0.v_cache: torch.Size([1, 8, 165, 128]) - layer.1.k_cache: torch.Size([1, 8, 165, 128]) - layer.1.v_cache: torch.Size([1, 8, 165, 128]) - layer.2.k_cache: torch.Size([1, 8, 165, 128]) - layer.2.v_cache: torch.Size([1, 8, 165, 128]) - layer.3.k_cache: torch.Size([1, 8, 165, 128]) - layer.3.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.k_cache: torch.Size([1, 8, 165, 128]) - layer.4.v_cache: torch.Size([1, 8, 165, 128]) - layer.4.output: torch.Size([1, 165, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02722176 9.18663885 - layer.0.v_cache 0.00000026 0.00025128 - layer.1.k_cache 0.00312822 1.73330245 - layer.1.v_cache 0.00000086 0.00091610 - layer.2.k_cache 0.00119859 0.70275255 - layer.2.v_cache 0.00000118 0.00134966 - layer.3.k_cache 0.00130010 0.77126160 - layer.3.v_cache 0.00000225 0.00217383 - layer.4.k_cache 0.00349983 1.94479407 - layer.4.v_cache 0.00000307 0.00352348 - layer.4.output 0.00018894 0.09522231 - ------------------------------------------------------------------------------------- - TOTAL 0.00265085 1.05198951 - (elements=2,365,440) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2365440 -Total Bytes 367088 -BPFP 1.2415 bits/point -EBPFP 2.4830 equivalent bits/point -MSE 1.051990 ----------------------- -------------------------------------------------------- -Time: 3.405s Load: 0.009s, Pack+Encode: 1.874s, Decode+Unpack: 1.523s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 165, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 165, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0520 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample54-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample54-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 210, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.013s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 210, 128) -Output shape: (1, 210, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) -> torch.Size([1, 1, 210, 1024]) - layer.4.output: torch.Size([1, 210, 4096]) -> torch.Size([1, 1, 210, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,760B, BPFP=0.1399 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 44,736B, BPFP=1.6643 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,764B, BPFP=0.8841 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 48,276B, BPFP=1.7960 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 30,000B, BPFP=1.1161 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,592B, BPFP=1.8449 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,632B, BPFP=1.3628 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 48,616B, BPFP=1.8086 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,748B, BPFP=0.9579 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 49,692B, BPFP=1.8487 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 80,256B, BPFP=0.7464 -⌛️ [2/4] FRONTEND: Frontend time: 2.403s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 210, 128]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.output: torch.Size([1, 210, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.820s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 210, 128]) - layer.0.v_cache: torch.Size([1, 8, 210, 128]) - layer.1.k_cache: torch.Size([1, 8, 210, 128]) - layer.1.v_cache: torch.Size([1, 8, 210, 128]) - layer.2.k_cache: torch.Size([1, 8, 210, 128]) - layer.2.v_cache: torch.Size([1, 8, 210, 128]) - layer.3.k_cache: torch.Size([1, 8, 210, 128]) - layer.3.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.k_cache: torch.Size([1, 8, 210, 128]) - layer.4.v_cache: torch.Size([1, 8, 210, 128]) - layer.4.output: torch.Size([1, 210, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702481 8.41819894 - layer.0.v_cache 0.00000029 0.00025388 - layer.1.k_cache 0.00301925 1.65261303 - layer.1.v_cache 0.00000076 0.00081302 - layer.2.k_cache 0.00118472 0.64992654 - layer.2.v_cache 0.00000128 0.00120310 - layer.3.k_cache 0.00128382 0.74264788 - layer.3.v_cache 0.00000210 0.00202007 - layer.4.k_cache 0.00361221 1.95860625 - layer.4.v_cache 0.00000301 0.00324083 - layer.4.output 0.00015169 0.08593610 - ------------------------------------------------------------------------------------- - TOTAL 0.00262422 0.98380485 - (elements=3,010,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3010560 -Total Bytes 441072 -BPFP 1.1721 bits/point -EBPFP 2.3441 equivalent bits/point -MSE 0.983805 ----------------------- -------------------------------------------------------- -Time: 4.236s Load: 0.013s, Pack+Encode: 2.403s, Decode+Unpack: 1.820s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 210, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 210, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9838 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample56-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample56-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 146, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 146, 128) -Output shape: (1, 146, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) -> torch.Size([1, 1, 146, 1024]) - layer.4.output: torch.Size([1, 146, 4096]) -> torch.Size([1, 1, 146, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,880B, BPFP=0.1541 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,952B, BPFP=1.9238 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,868B, BPFP=1.0096 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,216B, BPFP=2.2590 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 22,132B, BPFP=1.1843 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 41,396B, BPFP=2.2151 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,160B, BPFP=1.5068 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 42,620B, BPFP=2.2806 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,612B, BPFP=1.0494 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 41,940B, BPFP=2.2442 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 86,092B, BPFP=1.1517 -⌛️ [2/4] FRONTEND: Frontend time: 1.942s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 146, 128]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.output: torch.Size([1, 146, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.535s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 146, 128]) - layer.0.v_cache: torch.Size([1, 8, 146, 128]) - layer.1.k_cache: torch.Size([1, 8, 146, 128]) - layer.1.v_cache: torch.Size([1, 8, 146, 128]) - layer.2.k_cache: torch.Size([1, 8, 146, 128]) - layer.2.v_cache: torch.Size([1, 8, 146, 128]) - layer.3.k_cache: torch.Size([1, 8, 146, 128]) - layer.3.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.k_cache: torch.Size([1, 8, 146, 128]) - layer.4.v_cache: torch.Size([1, 8, 146, 128]) - layer.4.output: torch.Size([1, 146, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02740652 9.24501017 - layer.0.v_cache 0.00000027 0.00025496 - layer.1.k_cache 0.00310827 1.56208686 - layer.1.v_cache 0.00000089 0.00094531 - layer.2.k_cache 0.00118257 0.65062003 - layer.2.v_cache 0.00000116 0.00136181 - layer.3.k_cache 0.00130374 0.75590170 - layer.3.v_cache 0.00000225 0.00224526 - layer.4.k_cache 0.00337717 1.80190497 - layer.4.v_cache 0.00000316 0.00365712 - layer.4.output 0.00015578 0.09651232 - ------------------------------------------------------------------------------------- - TOTAL 0.00264351 1.02928839 - (elements=2,093,056) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2093056 -Total Bytes 381868 -BPFP 1.4596 bits/point -EBPFP 2.9191 equivalent bits/point -MSE 1.029288 ----------------------- -------------------------------------------------------- -Time: 3.486s Load: 0.009s, Pack+Encode: 1.942s, Decode+Unpack: 1.535s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 146, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 146, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0293 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample57-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample57-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 223, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 223, 128) -Output shape: (1, 223, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) -> torch.Size([1, 1, 223, 1024]) - layer.4.output: torch.Size([1, 223, 4096]) -> torch.Size([1, 1, 223, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 4,012B, BPFP=0.1406 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,500B, BPFP=1.6641 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,704B, BPFP=0.8655 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,648B, BPFP=1.7744 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,792B, BPFP=1.1839 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 50,996B, BPFP=1.7866 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,248B, BPFP=1.2699 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,856B, BPFP=1.7466 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,168B, BPFP=0.8817 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 51,360B, BPFP=1.7993 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 113,148B, BPFP=0.9910 -⌛️ [2/4] FRONTEND: Frontend time: 2.282s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 223, 128]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.output: torch.Size([1, 223, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.078s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 223, 128]) - layer.0.v_cache: torch.Size([1, 8, 223, 128]) - layer.1.k_cache: torch.Size([1, 8, 223, 128]) - layer.1.v_cache: torch.Size([1, 8, 223, 128]) - layer.2.k_cache: torch.Size([1, 8, 223, 128]) - layer.2.v_cache: torch.Size([1, 8, 223, 128]) - layer.3.k_cache: torch.Size([1, 8, 223, 128]) - layer.3.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.k_cache: torch.Size([1, 8, 223, 128]) - layer.4.v_cache: torch.Size([1, 8, 223, 128]) - layer.4.output: torch.Size([1, 223, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02710374 9.07727051 - layer.0.v_cache 0.00000027 0.00025831 - layer.1.k_cache 0.00293174 1.62861312 - layer.1.v_cache 0.00000088 0.00092944 - layer.2.k_cache 0.00135432 0.66590587 - layer.2.v_cache 0.00000122 0.00133455 - layer.3.k_cache 0.00128520 0.75242375 - layer.3.v_cache 0.00000265 0.00232433 - layer.4.k_cache 0.00356170 1.82552019 - layer.4.v_cache 0.00000337 0.00354296 - layer.4.output 0.00016975 0.08817101 - ------------------------------------------------------------------------------------- - TOTAL 0.00263743 1.02220051 - (elements=3,196,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3196928 -Total Bytes 487432 -BPFP 1.2198 bits/point -EBPFP 2.4395 equivalent bits/point -MSE 1.022201 ----------------------- -------------------------------------------------------- -Time: 4.371s Load: 0.011s, Pack+Encode: 2.282s, Decode+Unpack: 2.078s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 223, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 223, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0222 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample6-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample6-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 167, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 167, 128) -Output shape: (1, 167, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) -> torch.Size([1, 1, 167, 1024]) - layer.4.output: torch.Size([1, 167, 4096]) -> torch.Size([1, 1, 167, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,120B, BPFP=0.1460 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,608B, BPFP=1.6658 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,988B, BPFP=0.8883 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,792B, BPFP=1.8147 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,292B, BPFP=1.2300 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,536B, BPFP=1.8028 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,592B, BPFP=1.2908 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,120B, BPFP=1.7833 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,852B, BPFP=0.9287 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,224B, BPFP=1.7882 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 76,452B, BPFP=0.8941 -⌛️ [2/4] FRONTEND: Frontend time: 1.916s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 167, 128]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.output: torch.Size([1, 167, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.441s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 167, 128]) - layer.0.v_cache: torch.Size([1, 8, 167, 128]) - layer.1.k_cache: torch.Size([1, 8, 167, 128]) - layer.1.v_cache: torch.Size([1, 8, 167, 128]) - layer.2.k_cache: torch.Size([1, 8, 167, 128]) - layer.2.v_cache: torch.Size([1, 8, 167, 128]) - layer.3.k_cache: torch.Size([1, 8, 167, 128]) - layer.3.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.k_cache: torch.Size([1, 8, 167, 128]) - layer.4.v_cache: torch.Size([1, 8, 167, 128]) - layer.4.output: torch.Size([1, 167, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02821272 8.53342168 - layer.0.v_cache 0.00000027 0.00025617 - layer.1.k_cache 0.00309555 1.70050213 - layer.1.v_cache 0.00000088 0.00093664 - layer.2.k_cache 0.00116749 0.65849249 - layer.2.v_cache 0.00000113 0.00133348 - layer.3.k_cache 0.00133544 0.80374337 - layer.3.v_cache 0.00000222 0.00226055 - layer.4.k_cache 0.00347765 1.88166599 - layer.4.v_cache 0.00000297 0.00351439 - layer.4.output 0.00020541 0.09589151 - ------------------------------------------------------------------------------------- - TOTAL 0.00272271 0.99783521 - (elements=2,394,112) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2394112 -Total Bytes 361576 -BPFP 1.2082 bits/point -EBPFP 2.4164 equivalent bits/point -MSE 0.997835 ----------------------- -------------------------------------------------------- -Time: 3.367s Load: 0.010s, Pack+Encode: 1.916s, Decode+Unpack: 1.441s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 167, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 167, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9978 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,076B, BPFP=0.1483 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,264B, BPFP=1.7488 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,852B, BPFP=0.8609 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 40,892B, BPFP=1.9720 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 22,784B, BPFP=1.0988 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 40,152B, BPFP=1.9363 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,184B, BPFP=1.3592 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,520B, BPFP=1.9059 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,332B, BPFP=0.8841 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,824B, BPFP=1.8723 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 83,220B, BPFP=1.0033 -⌛️ [2/4] FRONTEND: Frontend time: 1.980s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.506s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02804119 8.27078737 - layer.0.v_cache 0.00000028 0.00025512 - layer.1.k_cache 0.00315224 1.65045147 - layer.1.v_cache 0.00000083 0.00094284 - layer.2.k_cache 0.00116199 0.65503688 - layer.2.v_cache 0.00000117 0.00133910 - layer.3.k_cache 0.00129529 0.75843152 - layer.3.v_cache 0.00000233 0.00227555 - layer.4.k_cache 0.00343653 1.72757354 - layer.4.v_cache 0.00000316 0.00366536 - layer.4.output 0.00017336 0.09489525 - ------------------------------------------------------------------------------------- - TOTAL 0.00269917 0.96073855 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 369100 -BPFP 1.2714 bits/point -EBPFP 2.5429 equivalent bits/point -MSE 0.960739 ----------------------- -------------------------------------------------------- -Time: 3.494s Load: 0.009s, Pack+Encode: 1.980s, Decode+Unpack: 1.506s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9607 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample61-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample61-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,304B, BPFP=0.1467 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,816B, BPFP=1.6342 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,032B, BPFP=0.8004 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,200B, BPFP=1.7401 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,748B, BPFP=1.1873 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,036B, BPFP=1.7328 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,508B, BPFP=1.2211 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,280B, BPFP=1.7436 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,136B, BPFP=0.8494 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,960B, BPFP=1.7294 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 97,696B, BPFP=1.0842 -⌛️ [2/4] FRONTEND: Frontend time: 1.874s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.600s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02849088 8.92194505 - layer.0.v_cache 0.00000028 0.00025554 - layer.1.k_cache 0.00305327 1.66850818 - layer.1.v_cache 0.00000091 0.00092563 - layer.2.k_cache 0.00120035 0.65487775 - layer.2.v_cache 0.00000116 0.00134562 - layer.3.k_cache 0.00128172 0.73449984 - layer.3.v_cache 0.00000269 0.00225421 - layer.4.k_cache 0.00351652 1.70554820 - layer.4.v_cache 0.00000314 0.00358230 - layer.4.output 0.00018003 0.09306320 - ------------------------------------------------------------------------------------- - TOTAL 0.00273365 1.00471394 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 385716 -BPFP 1.2230 bits/point -EBPFP 2.4459 equivalent bits/point -MSE 1.004714 ----------------------- -------------------------------------------------------- -Time: 3.485s Load: 0.011s, Pack+Encode: 1.874s, Decode+Unpack: 1.600s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0047 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample62-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample62-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 168, 128) -Output shape: (1, 168, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.output: torch.Size([1, 168, 4096]) -> torch.Size([1, 1, 168, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,112B, BPFP=0.1447 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 38,392B, BPFP=1.7853 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,172B, BPFP=0.8916 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 43,268B, BPFP=2.0121 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,228B, BPFP=1.2197 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,804B, BPFP=1.8045 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,316B, BPFP=1.3168 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,868B, BPFP=1.8540 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,584B, BPFP=0.9572 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,568B, BPFP=1.8400 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 69,648B, BPFP=0.8097 -⌛️ [2/4] FRONTEND: Frontend time: 1.896s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.433s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02847742 9.42324466 - layer.0.v_cache 0.00000027 0.00026507 - layer.1.k_cache 0.00308642 1.77965037 - layer.1.v_cache 0.00000083 0.00092562 - layer.2.k_cache 0.00117363 0.66810871 - layer.2.v_cache 0.00000114 0.00130903 - layer.3.k_cache 0.00131494 0.76743071 - layer.3.v_cache 0.00000212 0.00213117 - layer.4.k_cache 0.00375810 2.04833676 - layer.4.v_cache 0.00000313 0.00354084 - layer.4.output 0.00017449 0.09429989 - ------------------------------------------------------------------------------------- - TOTAL 0.00275114 1.07658161 - (elements=2,408,448) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2408448 -Total Bytes 366960 -BPFP 1.2189 bits/point -EBPFP 2.4378 equivalent bits/point -MSE 1.076582 ----------------------- -------------------------------------------------------- -Time: 3.339s Load: 0.011s, Pack+Encode: 1.896s, Decode+Unpack: 1.433s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0766 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample64-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample64-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 160, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 160, 128) -Output shape: (1, 160, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) -> torch.Size([1, 1, 160, 1024]) - layer.4.output: torch.Size([1, 160, 4096]) -> torch.Size([1, 1, 160, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,964B, BPFP=0.1447 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,320B, BPFP=1.7734 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,428B, BPFP=0.9486 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 40,348B, BPFP=1.9701 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,688B, BPFP=1.3031 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,516B, BPFP=1.8807 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,992B, BPFP=1.3668 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 39,260B, BPFP=1.9170 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,356B, BPFP=0.8475 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,252B, BPFP=1.9166 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 68,496B, BPFP=0.8361 -⌛️ [2/4] FRONTEND: Frontend time: 2.002s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 160, 128]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.output: torch.Size([1, 160, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.463s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 160, 128]) - layer.0.v_cache: torch.Size([1, 8, 160, 128]) - layer.1.k_cache: torch.Size([1, 8, 160, 128]) - layer.1.v_cache: torch.Size([1, 8, 160, 128]) - layer.2.k_cache: torch.Size([1, 8, 160, 128]) - layer.2.v_cache: torch.Size([1, 8, 160, 128]) - layer.3.k_cache: torch.Size([1, 8, 160, 128]) - layer.3.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.k_cache: torch.Size([1, 8, 160, 128]) - layer.4.v_cache: torch.Size([1, 8, 160, 128]) - layer.4.output: torch.Size([1, 160, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02762303 9.08161316 - layer.0.v_cache 0.00000027 0.00025851 - layer.1.k_cache 0.00317047 1.67671566 - layer.1.v_cache 0.00000088 0.00093379 - layer.2.k_cache 0.00116413 0.66368732 - layer.2.v_cache 0.00000117 0.00138374 - layer.3.k_cache 0.00133371 0.78741503 - layer.3.v_cache 0.00000211 0.00219350 - layer.4.k_cache 0.00351672 1.74339199 - layer.4.v_cache 0.00000299 0.00354639 - layer.4.output 0.00019762 0.10425466 - ------------------------------------------------------------------------------------- - TOTAL 0.00268614 1.02701127 - (elements=2,293,760) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2293760 -Total Bytes 356620 -BPFP 1.2438 bits/point -EBPFP 2.4876 equivalent bits/point -MSE 1.027011 ----------------------- -------------------------------------------------------- -Time: 3.474s Load: 0.009s, Pack+Encode: 2.002s, Decode+Unpack: 1.463s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 160, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 160, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0270 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 168, 128) -Output shape: (1, 168, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) -> torch.Size([1, 1, 168, 1024]) - layer.4.output: torch.Size([1, 168, 4096]) -> torch.Size([1, 1, 168, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,132B, BPFP=0.1456 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,668B, BPFP=1.6587 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,700B, BPFP=0.8696 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,904B, BPFP=1.8557 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,124B, BPFP=1.2148 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,904B, BPFP=1.8557 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,400B, BPFP=1.3207 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,564B, BPFP=1.7933 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 21,468B, BPFP=0.9983 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,360B, BPFP=1.8304 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 75,384B, BPFP=0.8764 -⌛️ [2/4] FRONTEND: Frontend time: 1.870s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.586s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 168, 128]) - layer.0.v_cache: torch.Size([1, 8, 168, 128]) - layer.1.k_cache: torch.Size([1, 8, 168, 128]) - layer.1.v_cache: torch.Size([1, 8, 168, 128]) - layer.2.k_cache: torch.Size([1, 8, 168, 128]) - layer.2.v_cache: torch.Size([1, 8, 168, 128]) - layer.3.k_cache: torch.Size([1, 8, 168, 128]) - layer.3.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.k_cache: torch.Size([1, 8, 168, 128]) - layer.4.v_cache: torch.Size([1, 8, 168, 128]) - layer.4.output: torch.Size([1, 168, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03000965 10.06079538 - layer.0.v_cache 0.00000027 0.00025734 - layer.1.k_cache 0.00306367 1.66285270 - layer.1.v_cache 0.00000085 0.00092401 - layer.2.k_cache 0.00121342 0.65721952 - layer.2.v_cache 0.00000122 0.00135134 - layer.3.k_cache 0.00130310 0.75064223 - layer.3.v_cache 0.00000229 0.00222890 - layer.4.k_cache 0.00355794 1.85718754 - layer.4.v_cache 0.00000315 0.00364385 - layer.4.output 0.00017029 0.10015015 - ------------------------------------------------------------------------------------- - TOTAL 0.00284548 1.09983596 - (elements=2,408,448) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2408448 -Total Bytes 366608 -BPFP 1.2177 bits/point -EBPFP 2.4355 equivalent bits/point -MSE 1.099836 ----------------------- -------------------------------------------------------- -Time: 3.466s Load: 0.009s, Pack+Encode: 1.870s, Decode+Unpack: 1.586s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 168, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 168, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0998 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample67-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample67-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,300B, BPFP=0.1465 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,964B, BPFP=1.5964 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,940B, BPFP=0.7963 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,696B, BPFP=1.7177 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,032B, BPFP=1.1112 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,248B, BPFP=1.6978 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,796B, BPFP=1.2338 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,660B, BPFP=1.6717 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,876B, BPFP=0.8823 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,960B, BPFP=1.7294 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 72,464B, BPFP=0.8042 -⌛️ [2/4] FRONTEND: Frontend time: 1.905s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.438s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02702593 8.78051758 - layer.0.v_cache 0.00000026 0.00024985 - layer.1.k_cache 0.00320345 1.62787611 - layer.1.v_cache 0.00000081 0.00086802 - layer.2.k_cache 0.00117342 0.66054448 - layer.2.v_cache 0.00000112 0.00128899 - layer.3.k_cache 0.00133382 0.75014981 - layer.3.v_cache 0.00000216 0.00216478 - layer.4.k_cache 0.00342137 1.93675371 - layer.4.v_cache 0.00000324 0.00367361 - layer.4.output 0.00018686 0.09748113 - ------------------------------------------------------------------------------------- - TOTAL 0.00263664 1.01100082 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 355936 -BPFP 1.1286 bits/point -EBPFP 2.2571 equivalent bits/point -MSE 1.011001 ----------------------- -------------------------------------------------------- -Time: 3.353s Load: 0.010s, Pack+Encode: 1.905s, Decode+Unpack: 1.438s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0110 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 157, 128) -Output shape: (1, 157, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) -> torch.Size([1, 1, 157, 1024]) - layer.4.output: torch.Size([1, 157, 4096]) -> torch.Size([1, 1, 157, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,060B, BPFP=0.1523 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,832B, BPFP=1.7830 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,924B, BPFP=0.9417 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,160B, BPFP=1.8989 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,412B, BPFP=1.2645 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,392B, BPFP=1.9104 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,324B, BPFP=1.4094 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,348B, BPFP=1.8585 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,936B, BPFP=0.9423 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,564B, BPFP=1.9688 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,664B, BPFP=0.8915 -⌛️ [2/4] FRONTEND: Frontend time: 1.981s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.486s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 157, 128]) - layer.0.v_cache: torch.Size([1, 8, 157, 128]) - layer.1.k_cache: torch.Size([1, 8, 157, 128]) - layer.1.v_cache: torch.Size([1, 8, 157, 128]) - layer.2.k_cache: torch.Size([1, 8, 157, 128]) - layer.2.v_cache: torch.Size([1, 8, 157, 128]) - layer.3.k_cache: torch.Size([1, 8, 157, 128]) - layer.3.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.k_cache: torch.Size([1, 8, 157, 128]) - layer.4.v_cache: torch.Size([1, 8, 157, 128]) - layer.4.output: torch.Size([1, 157, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02778443 9.34870668 - layer.0.v_cache 0.00000027 0.00026014 - layer.1.k_cache 0.00313050 1.60853373 - layer.1.v_cache 0.00000087 0.00094667 - layer.2.k_cache 0.00120658 0.66680680 - layer.2.v_cache 0.00000114 0.00135035 - layer.3.k_cache 0.00130588 0.75212379 - layer.3.v_cache 0.00000220 0.00223506 - layer.4.k_cache 0.00347338 1.87405532 - layer.4.v_cache 0.00000322 0.00381467 - layer.4.output 0.00013960 0.09624249 - ------------------------------------------------------------------------------------- - TOTAL 0.00267620 1.04598594 - (elements=2,250,752) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2250752 -Total Bytes 355616 -BPFP 1.2640 bits/point -EBPFP 2.5280 equivalent bits/point -MSE 1.045986 ----------------------- -------------------------------------------------------- -Time: 3.477s Load: 0.010s, Pack+Encode: 1.981s, Decode+Unpack: 1.486s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 157, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 157, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0460 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample69-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample69-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 221, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 221, 128) -Output shape: (1, 221, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) -> torch.Size([1, 1, 221, 1024]) - layer.4.output: torch.Size([1, 221, 4096]) -> torch.Size([1, 1, 221, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,944B, BPFP=0.1394 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 45,840B, BPFP=1.6205 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 24,476B, BPFP=0.8652 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 49,316B, BPFP=1.7434 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 33,372B, BPFP=1.1797 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 49,984B, BPFP=1.7670 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 36,324B, BPFP=1.2841 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,168B, BPFP=1.7381 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 23,204B, BPFP=0.8203 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 50,080B, BPFP=1.7704 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 86,596B, BPFP=0.7653 -⌛️ [2/4] FRONTEND: Frontend time: 2.302s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 221, 128]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.output: torch.Size([1, 221, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.054s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 221, 128]) - layer.0.v_cache: torch.Size([1, 8, 221, 128]) - layer.1.k_cache: torch.Size([1, 8, 221, 128]) - layer.1.v_cache: torch.Size([1, 8, 221, 128]) - layer.2.k_cache: torch.Size([1, 8, 221, 128]) - layer.2.v_cache: torch.Size([1, 8, 221, 128]) - layer.3.k_cache: torch.Size([1, 8, 221, 128]) - layer.3.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.k_cache: torch.Size([1, 8, 221, 128]) - layer.4.v_cache: torch.Size([1, 8, 221, 128]) - layer.4.output: torch.Size([1, 221, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02443308 8.73692688 - layer.0.v_cache 0.00000027 0.00025134 - layer.1.k_cache 0.00299519 1.74182571 - layer.1.v_cache 0.00000096 0.00086449 - layer.2.k_cache 0.00117758 0.64294585 - layer.2.v_cache 0.00000111 0.00126094 - layer.3.k_cache 0.00134364 0.76168644 - layer.3.v_cache 0.00000231 0.00220066 - layer.4.k_cache 0.00352016 1.85869243 - layer.4.v_cache 0.00000295 0.00337685 - layer.4.output 0.00021856 0.10090100 - ------------------------------------------------------------------------------------- - TOTAL 0.00245368 1.01097397 - (elements=3,168,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3168256 -Total Bytes 452304 -BPFP 1.1421 bits/point -EBPFP 2.2842 equivalent bits/point -MSE 1.010974 ----------------------- -------------------------------------------------------- -Time: 4.368s Load: 0.012s, Pack+Encode: 2.302s, Decode+Unpack: 2.054s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 221, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 221, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0110 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,188B, BPFP=0.1448 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,156B, BPFP=1.6423 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,784B, BPFP=0.8078 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 40,024B, BPFP=1.8180 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,724B, BPFP=1.1684 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,928B, BPFP=1.7682 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,472B, BPFP=1.2932 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,368B, BPFP=1.7427 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,032B, BPFP=0.9099 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 39,264B, BPFP=1.7834 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,112B, BPFP=0.8870 -⌛️ [2/4] FRONTEND: Frontend time: 1.872s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.517s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02652868 8.98763612 - layer.0.v_cache 0.00000028 0.00025455 - layer.1.k_cache 0.00309189 1.65384887 - layer.1.v_cache 0.00000082 0.00091305 - layer.2.k_cache 0.00116742 0.66647326 - layer.2.v_cache 0.00000113 0.00132168 - layer.3.k_cache 0.00134765 0.76193645 - layer.3.v_cache 0.00000211 0.00216543 - layer.4.k_cache 0.00352837 1.76730666 - layer.4.v_cache 0.00000329 0.00373569 - layer.4.output 0.00017786 0.09497695 - ------------------------------------------------------------------------------------- - TOTAL 0.00259879 1.01610711 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 366052 -BPFP 1.1876 bits/point -EBPFP 2.3752 equivalent bits/point -MSE 1.016107 ----------------------- -------------------------------------------------------- -Time: 3.399s Load: 0.010s, Pack+Encode: 1.872s, Decode+Unpack: 1.517s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0161 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample70-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample70-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 161, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 161, 128) -Output shape: (1, 161, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) -> torch.Size([1, 1, 161, 1024]) - layer.4.output: torch.Size([1, 161, 4096]) -> torch.Size([1, 1, 161, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,020B, BPFP=0.1465 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,260B, BPFP=1.7595 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,896B, BPFP=0.8684 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,128B, BPFP=1.8987 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,472B, BPFP=1.1875 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 40,708B, BPFP=1.9753 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,324B, BPFP=1.3744 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,268B, BPFP=1.8569 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,248B, BPFP=0.9340 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,488B, BPFP=1.8676 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 67,392B, BPFP=0.8175 -⌛️ [2/4] FRONTEND: Frontend time: 2.003s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 161, 128]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.output: torch.Size([1, 161, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.429s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 161, 128]) - layer.0.v_cache: torch.Size([1, 8, 161, 128]) - layer.1.k_cache: torch.Size([1, 8, 161, 128]) - layer.1.v_cache: torch.Size([1, 8, 161, 128]) - layer.2.k_cache: torch.Size([1, 8, 161, 128]) - layer.2.v_cache: torch.Size([1, 8, 161, 128]) - layer.3.k_cache: torch.Size([1, 8, 161, 128]) - layer.3.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.k_cache: torch.Size([1, 8, 161, 128]) - layer.4.v_cache: torch.Size([1, 8, 161, 128]) - layer.4.output: torch.Size([1, 161, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02852683 8.73013893 - layer.0.v_cache 0.00000027 0.00025409 - layer.1.k_cache 0.00307027 1.54033893 - layer.1.v_cache 0.00000080 0.00090648 - layer.2.k_cache 0.00118547 0.64772290 - layer.2.v_cache 0.00000112 0.00132865 - layer.3.k_cache 0.00132345 0.76484003 - layer.3.v_cache 0.00000221 0.00217907 - layer.4.k_cache 0.00350348 1.86964824 - layer.4.v_cache 0.00000302 0.00350984 - layer.4.output 0.00018782 0.09399451 - ------------------------------------------------------------------------------------- - TOTAL 0.00274058 0.99548894 - (elements=2,308,096) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2308096 -Total Bytes 353204 -BPFP 1.2242 bits/point -EBPFP 2.4485 equivalent bits/point -MSE 0.995489 ----------------------- -------------------------------------------------------- -Time: 3.441s Load: 0.009s, Pack+Encode: 2.003s, Decode+Unpack: 1.429s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 161, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 161, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9955 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample72-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample72-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 176, 128) -Output shape: (1, 176, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) -> torch.Size([1, 1, 176, 1024]) - layer.4.output: torch.Size([1, 176, 4096]) -> torch.Size([1, 1, 176, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,228B, BPFP=0.1433 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,800B, BPFP=1.5891 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,816B, BPFP=0.8352 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,832B, BPFP=1.7237 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,104B, BPFP=1.1587 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,772B, BPFP=1.7211 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,140B, BPFP=1.2491 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,272B, BPFP=1.6989 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,252B, BPFP=0.8546 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,884B, BPFP=1.7260 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 89,476B, BPFP=0.9929 -⌛️ [2/4] FRONTEND: Frontend time: 1.941s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.613s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 176, 128]) - layer.0.v_cache: torch.Size([1, 8, 176, 128]) - layer.1.k_cache: torch.Size([1, 8, 176, 128]) - layer.1.v_cache: torch.Size([1, 8, 176, 128]) - layer.2.k_cache: torch.Size([1, 8, 176, 128]) - layer.2.v_cache: torch.Size([1, 8, 176, 128]) - layer.3.k_cache: torch.Size([1, 8, 176, 128]) - layer.3.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.k_cache: torch.Size([1, 8, 176, 128]) - layer.4.v_cache: torch.Size([1, 8, 176, 128]) - layer.4.output: torch.Size([1, 176, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02614397 8.90414013 - layer.0.v_cache 0.00000027 0.00025971 - layer.1.k_cache 0.00303479 1.51707320 - layer.1.v_cache 0.00000100 0.00097012 - layer.2.k_cache 0.00119315 0.65852161 - layer.2.v_cache 0.00000111 0.00130083 - layer.3.k_cache 0.00127363 0.74481227 - layer.3.v_cache 0.00000250 0.00238763 - layer.4.k_cache 0.00361948 1.71284173 - layer.4.v_cache 0.00000340 0.00374451 - layer.4.output 0.00017471 0.09022209 - ------------------------------------------------------------------------------------- - TOTAL 0.00256944 0.99335286 - (elements=2,523,136) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2523136 -Total Bytes 375576 -BPFP 1.1908 bits/point -EBPFP 2.3816 equivalent bits/point -MSE 0.993353 ----------------------- -------------------------------------------------------- -Time: 3.563s Load: 0.009s, Pack+Encode: 1.941s, Decode+Unpack: 1.613s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 176, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 176, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9934 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample73-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample73-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 153, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 153, 128) -Output shape: (1, 153, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) -> torch.Size([1, 1, 153, 1024]) - layer.4.output: torch.Size([1, 153, 4096]) -> torch.Size([1, 1, 153, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,040B, BPFP=0.1552 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,040B, BPFP=1.8403 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,904B, BPFP=1.0163 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 42,608B, BPFP=2.1757 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 22,676B, BPFP=1.1579 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 46,316B, BPFP=2.3650 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,720B, BPFP=1.4665 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,696B, BPFP=2.0780 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,100B, BPFP=0.9242 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 40,188B, BPFP=2.0521 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 89,424B, BPFP=1.1415 -⌛️ [2/4] FRONTEND: Frontend time: 1.893s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 153, 128]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.output: torch.Size([1, 153, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.504s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 153, 128]) - layer.0.v_cache: torch.Size([1, 8, 153, 128]) - layer.1.k_cache: torch.Size([1, 8, 153, 128]) - layer.1.v_cache: torch.Size([1, 8, 153, 128]) - layer.2.k_cache: torch.Size([1, 8, 153, 128]) - layer.2.v_cache: torch.Size([1, 8, 153, 128]) - layer.3.k_cache: torch.Size([1, 8, 153, 128]) - layer.3.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.k_cache: torch.Size([1, 8, 153, 128]) - layer.4.v_cache: torch.Size([1, 8, 153, 128]) - layer.4.output: torch.Size([1, 153, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02796515 8.66897643 - layer.0.v_cache 0.00000026 0.00025370 - layer.1.k_cache 0.00317772 1.61205426 - layer.1.v_cache 0.00000088 0.00093266 - layer.2.k_cache 0.00119895 0.62010248 - layer.2.v_cache 0.00000122 0.00139042 - layer.3.k_cache 0.00130332 0.74556408 - layer.3.v_cache 0.00000226 0.00226508 - layer.4.k_cache 0.00344657 1.64479304 - layer.4.v_cache 0.00000314 0.00358925 - layer.4.output 0.00015128 0.10648737 - ------------------------------------------------------------------------------------- - TOTAL 0.00269319 0.98041935 - (elements=2,193,408) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2193408 -Total Bytes 387712 -BPFP 1.4141 bits/point -EBPFP 2.8282 equivalent bits/point -MSE 0.980419 ----------------------- -------------------------------------------------------- -Time: 3.407s Load: 0.009s, Pack+Encode: 1.893s, Decode+Unpack: 1.504s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 153, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 153, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9804 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 154, 128) -Output shape: (1, 154, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) -> torch.Size([1, 1, 154, 1024]) - layer.4.output: torch.Size([1, 154, 4096]) -> torch.Size([1, 1, 154, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,016B, BPFP=0.1530 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,696B, BPFP=1.8616 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,260B, BPFP=0.9771 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,372B, BPFP=1.9974 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,940B, BPFP=1.2652 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,864B, BPFP=1.9716 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,968B, BPFP=1.4696 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,896B, BPFP=1.9732 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,944B, BPFP=1.0625 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 40,440B, BPFP=2.0515 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 78,728B, BPFP=0.9985 -⌛️ [2/4] FRONTEND: Frontend time: 1.974s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.426s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 154, 128]) - layer.0.v_cache: torch.Size([1, 8, 154, 128]) - layer.1.k_cache: torch.Size([1, 8, 154, 128]) - layer.1.v_cache: torch.Size([1, 8, 154, 128]) - layer.2.k_cache: torch.Size([1, 8, 154, 128]) - layer.2.v_cache: torch.Size([1, 8, 154, 128]) - layer.3.k_cache: torch.Size([1, 8, 154, 128]) - layer.3.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.k_cache: torch.Size([1, 8, 154, 128]) - layer.4.v_cache: torch.Size([1, 8, 154, 128]) - layer.4.output: torch.Size([1, 154, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02770888 8.69350047 - layer.0.v_cache 0.00000027 0.00025567 - layer.1.k_cache 0.00299767 1.67498185 - layer.1.v_cache 0.00000089 0.00095116 - layer.2.k_cache 0.00118955 0.64797141 - layer.2.v_cache 0.00000124 0.00138600 - layer.3.k_cache 0.00128217 0.75414306 - layer.3.v_cache 0.00000233 0.00231164 - layer.4.k_cache 0.00329405 1.78905378 - layer.4.v_cache 0.00000340 0.00386980 - layer.4.output 0.00014754 0.09844382 - ------------------------------------------------------------------------------------- - TOTAL 0.00264790 0.99730001 - (elements=2,207,744) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2207744 -Total Bytes 370124 -BPFP 1.3412 bits/point -EBPFP 2.6824 equivalent bits/point -MSE 0.997300 ----------------------- -------------------------------------------------------- -Time: 3.409s Load: 0.008s, Pack+Encode: 1.974s, Decode+Unpack: 1.426s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 154, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 154, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9973 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample76-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample76-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,172B, BPFP=0.1466 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,568B, BPFP=1.6442 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,108B, BPFP=0.8371 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,740B, BPFP=1.7909 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,580B, BPFP=1.2287 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,576B, BPFP=1.7833 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,012B, BPFP=1.2949 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,120B, BPFP=1.7622 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,972B, BPFP=0.8770 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,600B, BPFP=1.7844 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 112,092B, BPFP=1.2954 -⌛️ [2/4] FRONTEND: Frontend time: 1.924s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.642s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02814460 9.01806930 - layer.0.v_cache 0.00000027 0.00024829 - layer.1.k_cache 0.00305883 1.71277650 - layer.1.v_cache 0.00000085 0.00093593 - layer.2.k_cache 0.00120620 0.65759485 - layer.2.v_cache 0.00000155 0.00141260 - layer.3.k_cache 0.00127672 0.78149748 - layer.3.v_cache 0.00000231 0.00226924 - layer.4.k_cache 0.00421318 1.79047787 - layer.4.v_cache 0.00000320 0.00358753 - layer.4.output 0.00020059 0.11473976 - ------------------------------------------------------------------------------------- - TOTAL 0.00276501 1.03055919 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 396540 -BPFP 1.3094 bits/point -EBPFP 2.6187 equivalent bits/point -MSE 1.030559 ----------------------- -------------------------------------------------------- -Time: 3.575s Load: 0.009s, Pack+Encode: 1.924s, Decode+Unpack: 1.642s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0306 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample78-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample78-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 180, 128) -Output shape: (1, 180, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) -> torch.Size([1, 1, 180, 1024]) - layer.4.output: torch.Size([1, 180, 4096]) -> torch.Size([1, 1, 180, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,388B, BPFP=0.1470 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 37,632B, BPFP=1.6333 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,656B, BPFP=0.7663 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,832B, BPFP=1.6854 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,088B, BPFP=1.1757 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,368B, BPFP=1.7087 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,552B, BPFP=1.1958 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,852B, BPFP=1.6863 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,276B, BPFP=0.7932 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,972B, BPFP=1.6915 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 90,136B, BPFP=0.9780 -⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.475s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 180, 128]) - layer.0.v_cache: torch.Size([1, 8, 180, 128]) - layer.1.k_cache: torch.Size([1, 8, 180, 128]) - layer.1.v_cache: torch.Size([1, 8, 180, 128]) - layer.2.k_cache: torch.Size([1, 8, 180, 128]) - layer.2.v_cache: torch.Size([1, 8, 180, 128]) - layer.3.k_cache: torch.Size([1, 8, 180, 128]) - layer.3.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.k_cache: torch.Size([1, 8, 180, 128]) - layer.4.v_cache: torch.Size([1, 8, 180, 128]) - layer.4.output: torch.Size([1, 180, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02888792 9.24086507 - layer.0.v_cache 0.00000029 0.00026047 - layer.1.k_cache 0.00301893 1.61964705 - layer.1.v_cache 0.00000086 0.00090131 - layer.2.k_cache 0.00125228 0.65756048 - layer.2.v_cache 0.00000118 0.00136594 - layer.3.k_cache 0.00126354 0.73008957 - layer.3.v_cache 0.00000246 0.00238822 - layer.4.k_cache 0.00353636 1.86514553 - layer.4.v_cache 0.00000320 0.00358883 - layer.4.output 0.00017167 0.09700774 - ------------------------------------------------------------------------------------- - TOTAL 0.00276098 1.03641739 - (elements=2,580,480) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2580480 -Total Bytes 377752 -BPFP 1.1711 bits/point -EBPFP 2.3422 equivalent bits/point -MSE 1.036417 ----------------------- -------------------------------------------------------- -Time: 3.373s Load: 0.011s, Pack+Encode: 1.887s, Decode+Unpack: 1.475s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 180, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 180, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0364 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample79-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample79-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 212, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 212, 128) -Output shape: (1, 212, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) -> torch.Size([1, 1, 212, 1024]) - layer.4.output: torch.Size([1, 212, 4096]) -> torch.Size([1, 1, 212, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,904B, BPFP=0.1439 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,512B, BPFP=1.7509 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 23,780B, BPFP=0.8763 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 50,024B, BPFP=1.8435 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 32,320B, BPFP=1.1910 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 50,568B, BPFP=1.8635 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 37,808B, BPFP=1.3933 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 50,084B, BPFP=1.8457 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,196B, BPFP=0.9285 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 50,348B, BPFP=1.8554 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 104,132B, BPFP=0.9594 -⌛️ [2/4] FRONTEND: Frontend time: 2.407s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 212, 128]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.output: torch.Size([1, 212, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.836s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 212, 128]) - layer.0.v_cache: torch.Size([1, 8, 212, 128]) - layer.1.k_cache: torch.Size([1, 8, 212, 128]) - layer.1.v_cache: torch.Size([1, 8, 212, 128]) - layer.2.k_cache: torch.Size([1, 8, 212, 128]) - layer.2.v_cache: torch.Size([1, 8, 212, 128]) - layer.3.k_cache: torch.Size([1, 8, 212, 128]) - layer.3.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.k_cache: torch.Size([1, 8, 212, 128]) - layer.4.v_cache: torch.Size([1, 8, 212, 128]) - layer.4.output: torch.Size([1, 212, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02614697 8.26912934 - layer.0.v_cache 0.00000027 0.00024847 - layer.1.k_cache 0.00297715 1.75622271 - layer.1.v_cache 0.00000079 0.00087410 - layer.2.k_cache 0.00116049 0.65837464 - layer.2.v_cache 0.00000116 0.00134104 - layer.3.k_cache 0.00132944 0.74351206 - layer.3.v_cache 0.00000234 0.00220882 - layer.4.k_cache 0.00407070 1.97483869 - layer.4.v_cache 0.00000311 0.00350486 - layer.4.output 0.00021146 0.10341920 - ------------------------------------------------------------------------------------- - TOTAL 0.00260988 0.98742368 - (elements=3,039,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 3039232 -Total Bytes 475676 -BPFP 1.2521 bits/point -EBPFP 2.5042 equivalent bits/point -MSE 0.987424 ----------------------- -------------------------------------------------------- -Time: 4.254s Load: 0.011s, Pack+Encode: 2.407s, Decode+Unpack: 1.836s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 212, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 212, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9874 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample8-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample8-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 172, 128) -Output shape: (1, 172, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) -> torch.Size([1, 1, 172, 1024]) - layer.4.output: torch.Size([1, 172, 4096]) -> torch.Size([1, 1, 172, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,220B, BPFP=0.1463 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,172B, BPFP=1.6430 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,568B, BPFP=0.8434 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,804B, BPFP=1.7625 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,508B, BPFP=1.2495 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,848B, BPFP=1.7645 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,952B, BPFP=1.2696 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,228B, BPFP=1.7364 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,120B, BPFP=0.9139 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,452B, BPFP=1.7465 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 72,064B, BPFP=0.8183 -⌛️ [2/4] FRONTEND: Frontend time: 1.966s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.556s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 172, 128]) - layer.0.v_cache: torch.Size([1, 8, 172, 128]) - layer.1.k_cache: torch.Size([1, 8, 172, 128]) - layer.1.v_cache: torch.Size([1, 8, 172, 128]) - layer.2.k_cache: torch.Size([1, 8, 172, 128]) - layer.2.v_cache: torch.Size([1, 8, 172, 128]) - layer.3.k_cache: torch.Size([1, 8, 172, 128]) - layer.3.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.k_cache: torch.Size([1, 8, 172, 128]) - layer.4.v_cache: torch.Size([1, 8, 172, 128]) - layer.4.output: torch.Size([1, 172, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02862963 9.52565180 - layer.0.v_cache 0.00000027 0.00025119 - layer.1.k_cache 0.00312686 1.77519226 - layer.1.v_cache 0.00000083 0.00092874 - layer.2.k_cache 0.00115513 0.67455602 - layer.2.v_cache 0.00000117 0.00135128 - layer.3.k_cache 0.00132852 0.74997117 - layer.3.v_cache 0.00000218 0.00215702 - layer.4.k_cache 0.00345250 1.87761919 - layer.4.v_cache 0.00000308 0.00354338 - layer.4.output 0.00017775 0.09257735 - ------------------------------------------------------------------------------------- - TOTAL 0.00274366 1.07010939 - (elements=2,465,792) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2465792 -Total Bytes 359936 -BPFP 1.1678 bits/point -EBPFP 2.3355 equivalent bits/point -MSE 1.070109 ----------------------- -------------------------------------------------------- -Time: 3.533s Load: 0.011s, Pack+Encode: 1.966s, Decode+Unpack: 1.556s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 172, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 172, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0701 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample80-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample80-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,056B, BPFP=0.1511 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,752B, BPFP=1.8172 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,940B, BPFP=0.9365 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 41,212B, BPFP=2.0378 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,392B, BPFP=1.2061 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 39,924B, BPFP=1.9741 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,828B, BPFP=1.3760 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 40,740B, BPFP=2.0144 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,668B, BPFP=0.9231 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 40,660B, BPFP=2.0105 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 65,912B, BPFP=0.8148 -⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.629s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02915672 9.07772711 - layer.0.v_cache 0.00000028 0.00025912 - layer.1.k_cache 0.00318231 1.64446056 - layer.1.v_cache 0.00000085 0.00091326 - layer.2.k_cache 0.00121414 0.64603086 - layer.2.v_cache 0.00000113 0.00130338 - layer.3.k_cache 0.00130024 0.76229250 - layer.3.v_cache 0.00000226 0.00219798 - layer.4.k_cache 0.00343194 1.67467586 - layer.4.v_cache 0.00000318 0.00362232 - layer.4.output 0.00014314 0.09240306 - ------------------------------------------------------------------------------------- - TOTAL 0.00277611 1.01307823 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 358084 -BPFP 1.2647 bits/point -EBPFP 2.5294 equivalent bits/point -MSE 1.013078 ----------------------- -------------------------------------------------------- -Time: 3.497s Load: 0.010s, Pack+Encode: 1.858s, Decode+Unpack: 1.629s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0131 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample81-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample81-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,128B, BPFP=0.1446 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,996B, BPFP=1.6178 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,468B, BPFP=0.8537 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,536B, BPFP=1.7352 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,420B, BPFP=1.1751 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 37,932B, BPFP=1.7535 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,488B, BPFP=1.3169 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,044B, BPFP=1.7125 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,668B, BPFP=0.9554 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,604B, BPFP=1.7384 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,848B, BPFP=0.8303 -⌛️ [2/4] FRONTEND: Frontend time: 1.925s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.438s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02741317 8.88799507 - layer.0.v_cache 0.00000027 0.00025400 - layer.1.k_cache 0.00313704 1.71729093 - layer.1.v_cache 0.00000078 0.00088498 - layer.2.k_cache 0.00120918 0.67587578 - layer.2.v_cache 0.00000116 0.00132642 - layer.3.k_cache 0.00136700 0.80025062 - layer.3.v_cache 0.00000220 0.00218476 - layer.4.k_cache 0.00352943 1.90565121 - layer.4.v_cache 0.00000294 0.00344774 - layer.4.output 0.00020829 0.10943362 - ------------------------------------------------------------------------------------- - TOTAL 0.00267831 1.03092114 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 353132 -BPFP 1.1660 bits/point -EBPFP 2.3321 equivalent bits/point -MSE 1.030921 ----------------------- -------------------------------------------------------- -Time: 3.373s Load: 0.010s, Pack+Encode: 1.925s, Decode+Unpack: 1.438s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0309 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample88-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample88-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,016B, BPFP=0.1454 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,228B, BPFP=1.6989 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,704B, BPFP=0.8538 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,196B, BPFP=1.8420 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,064B, BPFP=1.2087 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,856B, BPFP=1.8738 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,804B, BPFP=1.3409 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,316B, BPFP=1.7996 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 15,924B, BPFP=0.7679 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,796B, BPFP=1.8227 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 70,364B, BPFP=0.8483 -⌛️ [2/4] FRONTEND: Frontend time: 1.995s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.560s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02771055 8.63840513 - layer.0.v_cache 0.00000026 0.00026347 - layer.1.k_cache 0.00323549 1.60547252 - layer.1.v_cache 0.00000081 0.00091990 - layer.2.k_cache 0.00117871 0.67174558 - layer.2.v_cache 0.00000115 0.00132410 - layer.3.k_cache 0.00135828 0.81577066 - layer.3.v_cache 0.00000214 0.00218930 - layer.4.k_cache 0.00384025 1.67053581 - layer.4.v_cache 0.00000294 0.00340779 - layer.4.output 0.00023325 0.10587456 - ------------------------------------------------------------------------------------- - TOTAL 0.00273311 0.98810946 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 347268 -BPFP 1.1962 bits/point -EBPFP 2.3924 equivalent bits/point -MSE 0.988109 ----------------------- -------------------------------------------------------- -Time: 3.566s Load: 0.011s, Pack+Encode: 1.995s, Decode+Unpack: 1.560s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9881 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample89-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample89-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 193, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 193, 128) -Output shape: (1, 193, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) -> torch.Size([1, 1, 193, 1024]) - layer.4.output: torch.Size([1, 193, 4096]) -> torch.Size([1, 1, 193, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,624B, BPFP=0.1467 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 47,480B, BPFP=1.9220 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 25,640B, BPFP=1.0379 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 53,932B, BPFP=2.1831 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 28,968B, BPFP=1.1726 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 51,068B, BPFP=2.0672 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 35,420B, BPFP=1.4338 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 49,824B, BPFP=2.0168 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 25,200B, BPFP=1.0201 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 50,092B, BPFP=2.0277 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 101,052B, BPFP=1.0226 -⌛️ [2/4] FRONTEND: Frontend time: 2.243s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 193, 128]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.output: torch.Size([1, 193, 4096]) -⌛️ [3/4] BACKEND: Backend time: 2.085s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 193, 128]) - layer.0.v_cache: torch.Size([1, 8, 193, 128]) - layer.1.k_cache: torch.Size([1, 8, 193, 128]) - layer.1.v_cache: torch.Size([1, 8, 193, 128]) - layer.2.k_cache: torch.Size([1, 8, 193, 128]) - layer.2.v_cache: torch.Size([1, 8, 193, 128]) - layer.3.k_cache: torch.Size([1, 8, 193, 128]) - layer.3.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.k_cache: torch.Size([1, 8, 193, 128]) - layer.4.v_cache: torch.Size([1, 8, 193, 128]) - layer.4.output: torch.Size([1, 193, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02858143 8.28117537 - layer.0.v_cache 0.00000027 0.00025203 - layer.1.k_cache 0.00303266 1.59836179 - layer.1.v_cache 0.00000097 0.00097852 - layer.2.k_cache 0.00124314 0.64274830 - layer.2.v_cache 0.00000124 0.00141848 - layer.3.k_cache 0.00125873 0.75097087 - layer.3.v_cache 0.00000241 0.00230922 - layer.4.k_cache 0.00343145 1.69064821 - layer.4.v_cache 0.00000350 0.00388540 - layer.4.output 0.00017011 0.12277587 - ------------------------------------------------------------------------------------- - TOTAL 0.00273116 0.96170369 - (elements=2,766,848) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2766848 -Total Bytes 472300 -BPFP 1.3656 bits/point -EBPFP 2.7312 equivalent bits/point -MSE 0.961704 ----------------------- -------------------------------------------------------- -Time: 4.339s Load: 0.010s, Pack+Encode: 2.243s, Decode+Unpack: 2.085s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 193, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 193, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9617 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample9-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample9-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,568B, BPFP=0.1475 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 37,140B, BPFP=1.5352 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,844B, BPFP=0.7789 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,320B, BPFP=1.5840 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,320B, BPFP=1.1293 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,520B, BPFP=1.5923 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,928B, BPFP=1.1131 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,628B, BPFP=1.5554 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,656B, BPFP=0.8125 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,768B, BPFP=1.6025 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 89,972B, BPFP=0.9298 -⌛️ [2/4] FRONTEND: Frontend time: 1.893s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.400s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02726509 9.01242921 - layer.0.v_cache 0.00000027 0.00025783 - layer.1.k_cache 0.00303300 1.84624017 - layer.1.v_cache 0.00000085 0.00094138 - layer.2.k_cache 0.00116837 0.68670896 - layer.2.v_cache 0.00000118 0.00132657 - layer.3.k_cache 0.00127025 0.77274117 - layer.3.v_cache 0.00000234 0.00228646 - layer.4.k_cache 0.00356333 1.92405903 - layer.4.v_cache 0.00000334 0.00363352 - layer.4.output 0.00017162 0.09507475 - ------------------------------------------------------------------------------------- - TOTAL 0.00264246 1.04506595 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 376664 -BPFP 1.1121 bits/point -EBPFP 2.2243 equivalent bits/point -MSE 1.045066 ----------------------- -------------------------------------------------------- -Time: 3.304s Load: 0.011s, Pack+Encode: 1.893s, Decode+Unpack: 1.400s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0451 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample90-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample90-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.012s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 189, 128) -Output shape: (1, 189, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) -> torch.Size([1, 1, 189, 1024]) - layer.4.output: torch.Size([1, 189, 4096]) -> torch.Size([1, 1, 189, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,568B, BPFP=0.1475 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,724B, BPFP=1.4767 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 19,592B, BPFP=0.8099 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,200B, BPFP=1.5790 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 27,224B, BPFP=1.1253 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,528B, BPFP=1.5926 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,348B, BPFP=1.1305 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,388B, BPFP=1.5455 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,232B, BPFP=0.7950 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,512B, BPFP=1.5919 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 90,292B, BPFP=0.9331 -⌛️ [2/4] FRONTEND: Frontend time: 1.878s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.468s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 189, 128]) - layer.0.v_cache: torch.Size([1, 8, 189, 128]) - layer.1.k_cache: torch.Size([1, 8, 189, 128]) - layer.1.v_cache: torch.Size([1, 8, 189, 128]) - layer.2.k_cache: torch.Size([1, 8, 189, 128]) - layer.2.v_cache: torch.Size([1, 8, 189, 128]) - layer.3.k_cache: torch.Size([1, 8, 189, 128]) - layer.3.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.k_cache: torch.Size([1, 8, 189, 128]) - layer.4.v_cache: torch.Size([1, 8, 189, 128]) - layer.4.output: torch.Size([1, 189, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02698849 9.07180060 - layer.0.v_cache 0.00000027 0.00026268 - layer.1.k_cache 0.00302237 1.85477992 - layer.1.v_cache 0.00000087 0.00095722 - layer.2.k_cache 0.00117769 0.68950802 - layer.2.v_cache 0.00000119 0.00134062 - layer.3.k_cache 0.00126503 0.78376278 - layer.3.v_cache 0.00000236 0.00230510 - layer.4.k_cache 0.00350408 1.90314883 - layer.4.v_cache 0.00000327 0.00367145 - layer.4.output 0.00016083 0.09480397 - ------------------------------------------------------------------------------------- - TOTAL 0.00261492 1.04933951 - (elements=2,709,504) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2709504 -Total Bytes 375608 -BPFP 1.1090 bits/point -EBPFP 2.2180 equivalent bits/point -MSE 1.049340 ----------------------- -------------------------------------------------------- -Time: 3.358s Load: 0.012s, Pack+Encode: 1.878s, Decode+Unpack: 1.468s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 189, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 189, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0493 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample91-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample91-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 158, 128) -Output shape: (1, 158, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) -> torch.Size([1, 1, 158, 1024]) - layer.4.output: torch.Size([1, 158, 4096]) -> torch.Size([1, 1, 158, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,996B, BPFP=0.1481 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,244B, BPFP=1.7427 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,788B, BPFP=0.8795 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,316B, BPFP=1.8451 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 26,268B, BPFP=1.2989 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,172B, BPFP=1.8875 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,280B, BPFP=1.3489 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,548B, BPFP=1.8566 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,764B, BPFP=0.8784 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,220B, BPFP=1.8898 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 69,408B, BPFP=0.8580 -⌛️ [2/4] FRONTEND: Frontend time: 1.961s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.547s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 158, 128]) - layer.0.v_cache: torch.Size([1, 8, 158, 128]) - layer.1.k_cache: torch.Size([1, 8, 158, 128]) - layer.1.v_cache: torch.Size([1, 8, 158, 128]) - layer.2.k_cache: torch.Size([1, 8, 158, 128]) - layer.2.v_cache: torch.Size([1, 8, 158, 128]) - layer.3.k_cache: torch.Size([1, 8, 158, 128]) - layer.3.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.k_cache: torch.Size([1, 8, 158, 128]) - layer.4.v_cache: torch.Size([1, 8, 158, 128]) - layer.4.output: torch.Size([1, 158, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02813752 9.24814577 - layer.0.v_cache 0.00000027 0.00026065 - layer.1.k_cache 0.00306488 1.55726739 - layer.1.v_cache 0.00000089 0.00092507 - layer.2.k_cache 0.00118156 0.66326030 - layer.2.v_cache 0.00000118 0.00137221 - layer.3.k_cache 0.00130882 0.76125886 - layer.3.v_cache 0.00000213 0.00217215 - layer.4.k_cache 0.00347855 1.72627683 - layer.4.v_cache 0.00000307 0.00358126 - layer.4.output 0.00015371 0.09488247 - ------------------------------------------------------------------------------------- - TOTAL 0.00269955 1.02457503 - (elements=2,265,088) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2265088 -Total Bytes 348004 -BPFP 1.2291 bits/point -EBPFP 2.4582 equivalent bits/point -MSE 1.024575 ----------------------- -------------------------------------------------------- -Time: 3.517s Load: 0.009s, Pack+Encode: 1.961s, Decode+Unpack: 1.547s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 158, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 158, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0246 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample93-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample93-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,292B, BPFP=0.1453 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,912B, BPFP=1.5851 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,616B, BPFP=0.8217 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,824B, BPFP=1.7136 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,984B, BPFP=1.1028 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,668B, BPFP=1.7067 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,676B, BPFP=1.2216 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,900B, BPFP=1.6728 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 19,304B, BPFP=0.8520 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,452B, BPFP=1.6972 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 80,612B, BPFP=0.8895 -⌛️ [2/4] FRONTEND: Frontend time: 1.925s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.604s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02696669 8.64561609 - layer.0.v_cache 0.00000028 0.00024791 - layer.1.k_cache 0.00303349 1.62178272 - layer.1.v_cache 0.00000089 0.00090356 - layer.2.k_cache 0.00118683 0.63473200 - layer.2.v_cache 0.00000112 0.00133180 - layer.3.k_cache 0.00129356 0.74059153 - layer.3.v_cache 0.00000212 0.00210135 - layer.4.k_cache 0.00349999 2.02050436 - layer.4.v_cache 0.00000320 0.00357485 - layer.4.output 0.00017955 0.09496909 - ------------------------------------------------------------------------------------- - TOTAL 0.00262188 1.00366161 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 364240 -BPFP 1.1484 bits/point -EBPFP 2.2967 equivalent bits/point -MSE 1.003662 ----------------------- -------------------------------------------------------- -Time: 3.541s Load: 0.011s, Pack+Encode: 1.925s, Decode+Unpack: 1.604s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0037 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample94-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample94-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 169, 128) -Output shape: (1, 169, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) -> torch.Size([1, 1, 169, 1024]) - layer.4.output: torch.Size([1, 169, 4096]) -> torch.Size([1, 1, 169, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,140B, BPFP=0.1452 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,028B, BPFP=1.5730 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,440B, BPFP=0.8062 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,160B, BPFP=1.7178 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,884B, BPFP=1.1966 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 37,992B, BPFP=1.7563 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 26,960B, BPFP=1.2463 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 36,464B, BPFP=1.6857 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,248B, BPFP=0.8436 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 37,512B, BPFP=1.7341 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 65,972B, BPFP=0.7624 -⌛️ [2/4] FRONTEND: Frontend time: 1.922s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.417s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 169, 128]) - layer.0.v_cache: torch.Size([1, 8, 169, 128]) - layer.1.k_cache: torch.Size([1, 8, 169, 128]) - layer.1.v_cache: torch.Size([1, 8, 169, 128]) - layer.2.k_cache: torch.Size([1, 8, 169, 128]) - layer.2.v_cache: torch.Size([1, 8, 169, 128]) - layer.3.k_cache: torch.Size([1, 8, 169, 128]) - layer.3.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.k_cache: torch.Size([1, 8, 169, 128]) - layer.4.v_cache: torch.Size([1, 8, 169, 128]) - layer.4.output: torch.Size([1, 169, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02729142 9.17361071 - layer.0.v_cache 0.00000026 0.00025053 - layer.1.k_cache 0.00319685 1.48964985 - layer.1.v_cache 0.00000078 0.00088474 - layer.2.k_cache 0.00117458 0.67527311 - layer.2.v_cache 0.00000113 0.00129267 - layer.3.k_cache 0.00133789 0.78139789 - layer.3.v_cache 0.00000210 0.00218068 - layer.4.k_cache 0.00349792 1.69551709 - layer.4.v_cache 0.00000290 0.00346466 - layer.4.output 0.00017352 0.09110332 - ------------------------------------------------------------------------------------- - TOTAL 0.00265714 1.01342394 - (elements=2,422,784) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2422784 -Total Bytes 340800 -BPFP 1.1253 bits/point -EBPFP 2.2506 equivalent bits/point -MSE 1.013424 ----------------------- -------------------------------------------------------- -Time: 3.348s Load: 0.009s, Pack+Encode: 1.922s, Decode+Unpack: 1.417s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 169, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 169, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0134 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample95-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample95-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 162, 128) -Output shape: (1, 162, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) -> torch.Size([1, 1, 162, 1024]) - layer.4.output: torch.Size([1, 162, 4096]) -> torch.Size([1, 1, 162, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,052B, BPFP=0.1472 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 34,820B, BPFP=1.6792 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 17,652B, BPFP=0.8513 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 38,568B, BPFP=1.8600 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,336B, BPFP=1.2218 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 40,464B, BPFP=1.9514 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,248B, BPFP=1.3623 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,924B, BPFP=1.8289 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 17,904B, BPFP=0.8634 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,820B, BPFP=1.8721 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,260B, BPFP=0.8591 -⌛️ [2/4] FRONTEND: Frontend time: 2.003s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.495s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 162, 128]) - layer.0.v_cache: torch.Size([1, 8, 162, 128]) - layer.1.k_cache: torch.Size([1, 8, 162, 128]) - layer.1.v_cache: torch.Size([1, 8, 162, 128]) - layer.2.k_cache: torch.Size([1, 8, 162, 128]) - layer.2.v_cache: torch.Size([1, 8, 162, 128]) - layer.3.k_cache: torch.Size([1, 8, 162, 128]) - layer.3.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.k_cache: torch.Size([1, 8, 162, 128]) - layer.4.v_cache: torch.Size([1, 8, 162, 128]) - layer.4.output: torch.Size([1, 162, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02640657 8.78472373 - layer.0.v_cache 0.00000026 0.00025705 - layer.1.k_cache 0.00309680 1.55617307 - layer.1.v_cache 0.00000084 0.00092471 - layer.2.k_cache 0.00116892 0.67770075 - layer.2.v_cache 0.00000121 0.00137346 - layer.3.k_cache 0.00130413 0.77193333 - layer.3.v_cache 0.00000229 0.00223563 - layer.4.k_cache 0.00341342 1.74462740 - layer.4.v_cache 0.00000318 0.00367696 - layer.4.output 0.00016848 0.09630943 - ------------------------------------------------------------------------------------- - TOTAL 0.00257654 0.99491884 - (elements=2,322,432) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2322432 -Total Bytes 354048 -BPFP 1.2196 bits/point -EBPFP 2.4392 equivalent bits/point -MSE 0.994919 ----------------------- -------------------------------------------------------- -Time: 3.507s Load: 0.009s, Pack+Encode: 2.003s, Decode+Unpack: 1.495s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 162, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 162, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9949 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample96-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample96-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 173, 128) -Output shape: (1, 173, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) -> torch.Size([1, 1, 173, 1024]) - layer.4.output: torch.Size([1, 173, 4096]) -> torch.Size([1, 1, 173, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,224B, BPFP=0.1456 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 36,496B, BPFP=1.6481 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,924B, BPFP=0.8546 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 39,320B, BPFP=1.7757 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 24,000B, BPFP=1.0838 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,932B, BPFP=1.7581 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 28,096B, BPFP=1.2688 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 38,060B, BPFP=1.7188 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 18,876B, BPFP=0.8524 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,780B, BPFP=1.7513 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 76,040B, BPFP=0.8585 -⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.674s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 173, 128]) - layer.0.v_cache: torch.Size([1, 8, 173, 128]) - layer.1.k_cache: torch.Size([1, 8, 173, 128]) - layer.1.v_cache: torch.Size([1, 8, 173, 128]) - layer.2.k_cache: torch.Size([1, 8, 173, 128]) - layer.2.v_cache: torch.Size([1, 8, 173, 128]) - layer.3.k_cache: torch.Size([1, 8, 173, 128]) - layer.3.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.k_cache: torch.Size([1, 8, 173, 128]) - layer.4.v_cache: torch.Size([1, 8, 173, 128]) - layer.4.output: torch.Size([1, 173, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02785747 8.74270894 - layer.0.v_cache 0.00000027 0.00025571 - layer.1.k_cache 0.00305214 1.70974996 - layer.1.v_cache 0.00000084 0.00093145 - layer.2.k_cache 0.00114171 0.63173499 - layer.2.v_cache 0.00000110 0.00128353 - layer.3.k_cache 0.00128082 0.75662999 - layer.3.v_cache 0.00000218 0.00215347 - layer.4.k_cache 0.00367794 1.85570372 - layer.4.v_cache 0.00000299 0.00345070 - layer.4.output 0.00016652 0.09040855 - ------------------------------------------------------------------------------------- - TOTAL 0.00269168 1.00473119 - (elements=2,480,128) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2480128 -Total Bytes 360748 -BPFP 1.1636 bits/point -EBPFP 2.3273 equivalent bits/point -MSE 1.004731 ----------------------- -------------------------------------------------------- -Time: 3.592s Load: 0.009s, Pack+Encode: 1.909s, Decode+Unpack: 1.674s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 173, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 173, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0047 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample98-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample98-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.010s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 177, 128) -Output shape: (1, 177, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) -> torch.Size([1, 1, 177, 1024]) - layer.4.output: torch.Size([1, 177, 4096]) -> torch.Size([1, 1, 177, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 3,312B, BPFP=0.1462 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 35,980B, BPFP=1.5881 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 18,144B, BPFP=0.8008 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 37,812B, BPFP=1.6690 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 25,052B, BPFP=1.1058 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 38,304B, BPFP=1.6907 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 27,320B, BPFP=1.2059 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 37,376B, BPFP=1.6497 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 20,184B, BPFP=0.8909 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 38,600B, BPFP=1.7037 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 71,948B, BPFP=0.7939 -⌛️ [2/4] FRONTEND: Frontend time: 1.936s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.512s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 177, 128]) - layer.0.v_cache: torch.Size([1, 8, 177, 128]) - layer.1.k_cache: torch.Size([1, 8, 177, 128]) - layer.1.v_cache: torch.Size([1, 8, 177, 128]) - layer.2.k_cache: torch.Size([1, 8, 177, 128]) - layer.2.v_cache: torch.Size([1, 8, 177, 128]) - layer.3.k_cache: torch.Size([1, 8, 177, 128]) - layer.3.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.k_cache: torch.Size([1, 8, 177, 128]) - layer.4.v_cache: torch.Size([1, 8, 177, 128]) - layer.4.output: torch.Size([1, 177, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02698118 8.75613386 - layer.0.v_cache 0.00000026 0.00025200 - layer.1.k_cache 0.00304540 1.78189621 - layer.1.v_cache 0.00000079 0.00087805 - layer.2.k_cache 0.00119876 0.66109350 - layer.2.v_cache 0.00000114 0.00129302 - layer.3.k_cache 0.00131734 0.75223856 - layer.3.v_cache 0.00000220 0.00214750 - layer.4.k_cache 0.00342134 1.82741861 - layer.4.v_cache 0.00000318 0.00361668 - layer.4.output 0.00016969 0.09642704 - ------------------------------------------------------------------------------------- - TOTAL 0.00261788 1.01233401 - (elements=2,537,472) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 2537472 -Total Bytes 354032 -BPFP 1.1162 bits/point -EBPFP 2.2323 equivalent bits/point -MSE 1.012334 ----------------------- -------------------------------------------------------- -Time: 3.458s Load: 0.010s, Pack+Encode: 1.936s, Decode+Unpack: 1.512s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 177, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 177, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0123 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_truthfulqa_mc1/sample99-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_truthfulqa_mc1/sample99-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 1.2129 bits/point -Avg EBPFP 2.4257 equivalent bits/point -Avg MSE 1.019808 -Avg Time 3.613s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:6958cc1b0c93ea89ca248c6724015776245e10e8bfeb704fc64e34cde93cdd9a +size 1125653 diff --git a/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log index f68b920475cc8c73d172b0a691fa88d8fe3b85fa..fd02a5c047ce985a29b25d23cc82ca1741a5b7c4 100644 --- a/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log +++ b/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log @@ -1,16958 +1,3 @@ -Experiment: dtufc_elic-featurecoding_qwen_individual -Log file: output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/dtufc_elic-featurecoding_qwen_individual.log -DTUFCCodecConfig: - arch: elic-featurecoding - handler: qwen - checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar - transform_type: kmeans - transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json - bit_depth: 8 - device: cuda:0 -Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Checkpoint epoch: 255 -Loaded elic-featurecoding (1-channel) on cuda:0 -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_0_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_1_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_2_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_3_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_k.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_layer_4_v.json -Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json: torch.Size([256]) -Loaded per-key quantization points for key 'layer.4.output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/qwen3/arc_fewshot-8bit_feature.json -Loaded per-key mappings: model=qwen - Keys: ['layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache', 'layer.4.output'] ----------------- -------------------------------------------------------------------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding -Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar -Transform type kmeans -Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json -Input ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande -Output output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande ----------------- -------------------------------------------------------------------------------------------------------------------- -Files found: 100 ----------------------------------------------------------------------- - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-layer4-item1.zst (1/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 100, 128) -Output shape: (1, 100, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.output: torch.Size([1, 100, 4096]) -> torch.Size([1, 1, 100, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,120B, BPFP=0.1656 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,484B, BPFP=1.8347 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,068B, BPFP=1.0209 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,440B, BPFP=1.9875 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,988B, BPFP=1.3272 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,780B, BPFP=2.0141 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,004B, BPFP=1.4847 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,008B, BPFP=1.9538 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,572B, BPFP=0.9822 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,908B, BPFP=2.0241 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 55,704B, BPFP=1.0880 -⌛️ [2/4] FRONTEND: Frontend time: 2.197s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.356s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03019863 10.89289185 - layer.0.v_cache 0.00000027 0.00026709 - layer.1.k_cache 0.00347471 1.72186676 - layer.1.v_cache 0.00000091 0.00094480 - layer.2.k_cache 0.00113692 0.66980049 - layer.2.v_cache 0.00000108 0.00136132 - layer.3.k_cache 0.00132974 0.77265884 - layer.3.v_cache 0.00000211 0.00233634 - layer.4.k_cache 0.00328898 1.56643875 - layer.4.v_cache 0.00000306 0.00376313 - layer.4.output 0.00017021 0.10375260 - ------------------------------------------------------------------------------------- - TOTAL 0.00286552 1.14623856 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 245076 -BPFP 1.3676 bits/point -EBPFP 2.7352 equivalent bits/point -MSE 1.146239 ----------------------- -------------------------------------------------------- -Time: 3.559s Load: 0.006s, Pack+Encode: 2.197s, Decode+Unpack: 1.356s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1462 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample0-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample0-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-layer4-item1.zst (2/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 98, 128) -Output shape: (1, 98, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.output: torch.Size([1, 98, 4096]) -> torch.Size([1, 1, 98, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,040B, BPFP=0.1626 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,248B, BPFP=1.8533 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,160B, BPFP=1.0491 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,460B, BPFP=2.0297 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,548B, BPFP=1.2395 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,744B, BPFP=2.0523 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,396B, BPFP=1.5462 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,184B, BPFP=2.0077 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,680B, BPFP=1.0108 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 26,036B, BPFP=2.0756 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,688B, BPFP=0.9305 -⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.325s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03033472 9.32006836 - layer.0.v_cache 0.00000026 0.00026281 - layer.1.k_cache 0.00340990 1.62362484 - layer.1.v_cache 0.00000094 0.00092408 - layer.2.k_cache 0.00111132 0.66824761 - layer.2.v_cache 0.00000109 0.00136808 - layer.3.k_cache 0.00132423 0.79570506 - layer.3.v_cache 0.00000205 0.00225329 - layer.4.k_cache 0.00330658 1.53245794 - layer.4.v_cache 0.00000307 0.00380063 - layer.4.output 0.00020590 0.11606418 - ------------------------------------------------------------------------------------- - TOTAL 0.00287984 1.02949782 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 235184 -BPFP 1.3392 bits/point -EBPFP 2.6784 equivalent bits/point -MSE 1.029498 ----------------------- -------------------------------------------------------- -Time: 3.037s Load: 0.008s, Pack+Encode: 1.704s, Decode+Unpack: 1.325s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0295 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample1-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample1-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-layer4-item1.zst (3/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,956B, BPFP=0.1626 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,880B, BPFP=1.9016 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,316B, BPFP=1.0236 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,200B, BPFP=2.0944 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,144B, BPFP=1.3418 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,536B, BPFP=2.1223 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,656B, BPFP=1.5505 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,000B, BPFP=2.0778 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,212B, BPFP=1.0150 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,696B, BPFP=2.1356 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 55,752B, BPFP=1.1584 -⌛️ [2/4] FRONTEND: Frontend time: 1.818s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.264s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03043811 10.01205899 - layer.0.v_cache 0.00000028 0.00027166 - layer.1.k_cache 0.00344812 1.57522502 - layer.1.v_cache 0.00000082 0.00094978 - layer.2.k_cache 0.00115027 0.70304725 - layer.2.v_cache 0.00000109 0.00139925 - layer.3.k_cache 0.00130609 0.78437123 - layer.3.v_cache 0.00000216 0.00235694 - layer.4.k_cache 0.00321842 1.49187275 - layer.4.v_cache 0.00000315 0.00383012 - layer.4.output 0.00017058 0.11902417 - ------------------------------------------------------------------------------------- - TOTAL 0.00287506 1.07510569 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 241348 -BPFP 1.4328 bits/point -EBPFP 2.8655 equivalent bits/point -MSE 1.075106 ----------------------- -------------------------------------------------------- -Time: 3.090s Load: 0.008s, Pack+Encode: 1.818s, Decode+Unpack: 1.264s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0751 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample10-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample10-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst (4/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,860B, BPFP=0.1651 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,172B, BPFP=2.0572 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,920B, BPFP=1.0582 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,296B, BPFP=2.2457 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,728B, BPFP=1.3075 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,596B, BPFP=2.2724 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,968B, BPFP=1.5952 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,920B, BPFP=2.2124 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,380B, BPFP=0.8327 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,592B, BPFP=2.2720 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,856B, BPFP=0.9290 -⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.442s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03339012 10.01735340 - layer.0.v_cache 0.00000028 0.00027341 - layer.1.k_cache 0.00350260 1.36581490 - layer.1.v_cache 0.00000079 0.00094882 - layer.2.k_cache 0.00114403 0.66585328 - layer.2.v_cache 0.00000109 0.00141699 - layer.3.k_cache 0.00131544 0.75915233 - layer.3.v_cache 0.00000207 0.00235015 - layer.4.k_cache 0.00326494 1.53470924 - layer.4.v_cache 0.00000300 0.00390586 - layer.4.output 0.00017308 0.11854799 - ------------------------------------------------------------------------------------- - TOTAL 0.00309405 1.05899788 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 222288 -BPFP 1.4096 bits/point -EBPFP 2.8192 equivalent bits/point -MSE 1.058998 ----------------------- -------------------------------------------------------- -Time: 3.212s Load: 0.008s, Pack+Encode: 1.763s, Decode+Unpack: 1.442s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0590 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample100-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample100-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-layer4-item1.zst (5/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,956B, BPFP=0.1661 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,164B, BPFP=1.9671 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,456B, BPFP=1.0577 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,280B, BPFP=2.1467 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,264B, BPFP=1.2962 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,772B, BPFP=2.1885 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,456B, BPFP=1.5673 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,864B, BPFP=2.1114 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,520B, BPFP=0.8933 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,636B, BPFP=2.1770 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,912B, BPFP=1.0172 -⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.235s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03121759 9.43754379 - layer.0.v_cache 0.00000027 0.00026816 - layer.1.k_cache 0.00337054 1.55959420 - layer.1.v_cache 0.00000081 0.00093505 - layer.2.k_cache 0.00113856 0.66111183 - layer.2.v_cache 0.00000109 0.00140807 - layer.3.k_cache 0.00130416 0.76818217 - layer.3.v_cache 0.00000204 0.00232982 - layer.4.k_cache 0.00326587 1.49687775 - layer.4.v_cache 0.00000304 0.00387892 - layer.4.output 0.00017424 0.11672042 - ------------------------------------------------------------------------------------- - TOTAL 0.00292864 1.02850082 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 231280 -BPFP 1.4029 bits/point -EBPFP 2.8057 equivalent bits/point -MSE 1.028501 ----------------------- -------------------------------------------------------- -Time: 2.945s Load: 0.006s, Pack+Encode: 1.704s, Decode+Unpack: 1.235s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0285 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample103-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample103-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst (6/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,916B, BPFP=0.1645 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,880B, BPFP=1.9643 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,612B, BPFP=1.0828 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,856B, BPFP=2.1339 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,428B, BPFP=1.1528 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,440B, BPFP=2.1841 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,748B, BPFP=1.6095 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,516B, BPFP=2.1047 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,104B, BPFP=0.8674 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,316B, BPFP=2.1734 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,356B, BPFP=0.9091 -⌛️ [2/4] FRONTEND: Frontend time: 1.817s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.287s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03135960 10.21605506 - layer.0.v_cache 0.00000027 0.00027367 - layer.1.k_cache 0.00345943 1.59767637 - layer.1.v_cache 0.00000081 0.00092782 - layer.2.k_cache 0.00114090 0.65941096 - layer.2.v_cache 0.00000105 0.00135052 - layer.3.k_cache 0.00133425 0.77927365 - layer.3.v_cache 0.00000201 0.00228186 - layer.4.k_cache 0.00327424 1.49772644 - layer.4.v_cache 0.00000301 0.00368060 - layer.4.output 0.00017460 0.11731650 - ------------------------------------------------------------------------------------- - TOTAL 0.00294814 1.08770878 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 222172 -BPFP 1.3624 bits/point -EBPFP 2.7248 equivalent bits/point -MSE 1.087709 ----------------------- -------------------------------------------------------- -Time: 3.112s Load: 0.007s, Pack+Encode: 1.817s, Decode+Unpack: 1.287s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0877 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample105-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample105-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-layer4-item1.zst (7/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,928B, BPFP=0.1674 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,216B, BPFP=2.0153 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,724B, BPFP=1.1045 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,208B, BPFP=2.1882 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,140B, BPFP=1.1406 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,476B, BPFP=2.2115 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,196B, BPFP=1.6663 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,804B, BPFP=2.1531 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,096B, BPFP=0.8764 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,620B, BPFP=2.2240 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,628B, BPFP=0.8817 -⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.421s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03118788 9.64392361 - layer.0.v_cache 0.00000027 0.00027646 - layer.1.k_cache 0.00346696 1.53097144 - layer.1.v_cache 0.00000079 0.00093165 - layer.2.k_cache 0.00116280 0.67655106 - layer.2.v_cache 0.00000105 0.00136548 - layer.3.k_cache 0.00130692 0.79504971 - layer.3.v_cache 0.00000207 0.00225255 - layer.4.k_cache 0.00320606 1.57675815 - layer.4.v_cache 0.00000305 0.00375973 - layer.4.output 0.00017695 0.11594163 - ------------------------------------------------------------------------------------- - TOTAL 0.00293183 1.04968617 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 222036 -BPFP 1.3767 bits/point -EBPFP 2.7534 equivalent bits/point -MSE 1.049686 ----------------------- -------------------------------------------------------- -Time: 3.137s Load: 0.006s, Pack+Encode: 1.710s, Decode+Unpack: 1.421s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0497 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample107-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample107-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-layer4-item1.zst (8/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,940B, BPFP=0.1647 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,960B, BPFP=1.9497 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,064B, BPFP=1.1094 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,300B, BPFP=2.1484 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,908B, BPFP=1.4358 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,568B, BPFP=2.1712 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,964B, BPFP=1.6104 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,880B, BPFP=2.1128 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,452B, BPFP=0.8026 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,468B, BPFP=2.1627 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,492B, BPFP=0.9233 -⌛️ [2/4] FRONTEND: Frontend time: 1.777s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.259s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03125828 9.99308578 - layer.0.v_cache 0.00000028 0.00027061 - layer.1.k_cache 0.00363872 1.45870292 - layer.1.v_cache 0.00000085 0.00095212 - layer.2.k_cache 0.00115386 0.67171594 - layer.2.v_cache 0.00000106 0.00137936 - layer.3.k_cache 0.00130545 0.79346068 - layer.3.v_cache 0.00000205 0.00231543 - layer.4.k_cache 0.00323476 1.53337910 - layer.4.v_cache 0.00000303 0.00381035 - layer.4.output 0.00019770 0.12560225 - ------------------------------------------------------------------------------------- - TOTAL 0.00295637 1.06867723 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 227996 -BPFP 1.3829 bits/point -EBPFP 2.7659 equivalent bits/point -MSE 1.068677 ----------------------- -------------------------------------------------------- -Time: 3.044s Load: 0.008s, Pack+Encode: 1.777s, Decode+Unpack: 1.259s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0687 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample111-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample111-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-layer4-item1.zst (9/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1685 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,264B, BPFP=2.0891 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,768B, BPFP=0.9670 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,632B, BPFP=2.3017 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,624B, BPFP=1.1336 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,656B, BPFP=2.3039 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,540B, BPFP=1.5751 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,100B, BPFP=2.2540 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,400B, BPFP=0.8441 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,840B, BPFP=2.3204 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,084B, BPFP=0.9223 -⌛️ [2/4] FRONTEND: Frontend time: 1.759s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.440s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03173990 9.67585000 - layer.0.v_cache 0.00000027 0.00026860 - layer.1.k_cache 0.00353239 1.49238955 - layer.1.v_cache 0.00000080 0.00091813 - layer.2.k_cache 0.00117039 0.67564699 - layer.2.v_cache 0.00000104 0.00132603 - layer.3.k_cache 0.00134488 0.77379968 - layer.3.v_cache 0.00000203 0.00224405 - layer.4.k_cache 0.00324251 1.45432571 - layer.4.v_cache 0.00000295 0.00368540 - layer.4.output 0.00019562 0.11342575 - ------------------------------------------------------------------------------------- - TOTAL 0.00298712 1.03815408 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 218784 -BPFP 1.4033 bits/point -EBPFP 2.8067 equivalent bits/point -MSE 1.038154 ----------------------- -------------------------------------------------------- -Time: 3.205s Load: 0.006s, Pack+Encode: 1.759s, Decode+Unpack: 1.440s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0382 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample112-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample112-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-layer4-item1.zst (10/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,864B, BPFP=0.1674 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,332B, BPFP=2.0952 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,196B, BPFP=1.0952 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,344B, BPFP=2.2759 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,252B, BPFP=1.1900 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,672B, BPFP=2.3053 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,200B, BPFP=1.5445 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,920B, BPFP=2.2378 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,404B, BPFP=0.9343 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,544B, BPFP=2.2938 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,824B, BPFP=0.8716 -⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.314s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03084559 9.28257313 - layer.0.v_cache 0.00000028 0.00027429 - layer.1.k_cache 0.00352810 1.31872348 - layer.1.v_cache 0.00000080 0.00093107 - layer.2.k_cache 0.00113435 0.65245183 - layer.2.v_cache 0.00000106 0.00136939 - layer.3.k_cache 0.00133457 0.76111384 - layer.3.v_cache 0.00000203 0.00228880 - layer.4.k_cache 0.00324423 1.59170690 - layer.4.v_cache 0.00000296 0.00370504 - layer.4.output 0.00017424 0.11773398 - ------------------------------------------------------------------------------------- - TOTAL 0.00291364 1.00614812 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 218552 -BPFP 1.4018 bits/point -EBPFP 2.8037 equivalent bits/point -MSE 1.006148 ----------------------- -------------------------------------------------------- -Time: 3.043s Load: 0.006s, Pack+Encode: 1.722s, Decode+Unpack: 1.314s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0061 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample113-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample113-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-layer4-item1.zst (11/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,856B, BPFP=0.1706 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,908B, BPFP=2.1055 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,108B, BPFP=1.2967 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,416B, BPFP=2.3360 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,388B, BPFP=1.1386 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,864B, BPFP=2.3772 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,008B, BPFP=1.6551 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,964B, BPFP=2.2945 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,256B, BPFP=0.9426 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,756B, BPFP=2.3673 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,664B, BPFP=0.9344 -⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.273s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03172885 10.12065430 - layer.0.v_cache 0.00000027 0.00027355 - layer.1.k_cache 0.00352961 1.61031081 - layer.1.v_cache 0.00000080 0.00094955 - layer.2.k_cache 0.00112280 0.68887957 - layer.2.v_cache 0.00000110 0.00141593 - layer.3.k_cache 0.00133082 0.75507220 - layer.3.v_cache 0.00000207 0.00229175 - layer.4.k_cache 0.00329730 1.48480593 - layer.4.v_cache 0.00000296 0.00383208 - layer.4.output 0.00017366 0.10956180 - ------------------------------------------------------------------------------------- - TOTAL 0.00297937 1.07905235 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 222188 -BPFP 1.4587 bits/point -EBPFP 2.9174 equivalent bits/point -MSE 1.079052 ----------------------- -------------------------------------------------------- -Time: 3.108s Load: 0.007s, Pack+Encode: 1.828s, Decode+Unpack: 1.273s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0791 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample117-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample117-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-layer4-item1.zst (12/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1633 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,536B, BPFP=1.8931 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,952B, BPFP=1.0040 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,048B, BPFP=2.1042 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,724B, BPFP=1.2369 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,360B, BPFP=2.1304 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,204B, BPFP=1.5292 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,620B, BPFP=2.0682 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,316B, BPFP=1.1186 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,304B, BPFP=2.1257 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,868B, BPFP=0.8373 -⌛️ [2/4] FRONTEND: Frontend time: 1.693s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.393s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03136649 9.87706076 - layer.0.v_cache 0.00000028 0.00026482 - layer.1.k_cache 0.00359034 1.45845212 - layer.1.v_cache 0.00000079 0.00094022 - layer.2.k_cache 0.00112294 0.69932737 - layer.2.v_cache 0.00000105 0.00138801 - layer.3.k_cache 0.00134461 0.78078526 - layer.3.v_cache 0.00000211 0.00231027 - layer.4.k_cache 0.00331549 1.57872584 - layer.4.v_cache 0.00000311 0.00388127 - layer.4.output 0.00017890 0.10913132 - ------------------------------------------------------------------------------------- - TOTAL 0.00296163 1.05997580 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 222876 -BPFP 1.3373 bits/point -EBPFP 2.6747 equivalent bits/point -MSE 1.059976 ----------------------- -------------------------------------------------------- -Time: 3.092s Load: 0.006s, Pack+Encode: 1.693s, Decode+Unpack: 1.393s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0600 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample12-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample12-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-layer4-item1.zst (13/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 84, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 84, 128) -Output shape: (1, 84, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) -> torch.Size([1, 1, 84, 1024]) - layer.4.output: torch.Size([1, 84, 4096]) -> torch.Size([1, 1, 84, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,780B, BPFP=0.1656 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,372B, BPFP=2.1737 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,612B, BPFP=1.1730 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,016B, BPFP=2.3266 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,212B, BPFP=1.0428 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,568B, BPFP=2.3780 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,800B, BPFP=1.7485 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,904B, BPFP=2.3162 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,928B, BPFP=0.8304 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,596B, BPFP=2.3806 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 37,064B, BPFP=0.8618 -⌛️ [2/4] FRONTEND: Frontend time: 1.770s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.output: torch.Size([1, 84, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.237s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 84, 128]) - layer.0.v_cache: torch.Size([1, 8, 84, 128]) - layer.1.k_cache: torch.Size([1, 8, 84, 128]) - layer.1.v_cache: torch.Size([1, 8, 84, 128]) - layer.2.k_cache: torch.Size([1, 8, 84, 128]) - layer.2.v_cache: torch.Size([1, 8, 84, 128]) - layer.3.k_cache: torch.Size([1, 8, 84, 128]) - layer.3.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.k_cache: torch.Size([1, 8, 84, 128]) - layer.4.v_cache: torch.Size([1, 8, 84, 128]) - layer.4.output: torch.Size([1, 84, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03306588 9.77406020 - layer.0.v_cache 0.00000027 0.00026643 - layer.1.k_cache 0.00337037 1.34102594 - layer.1.v_cache 0.00000088 0.00091886 - layer.2.k_cache 0.00113180 0.71181079 - layer.2.v_cache 0.00000103 0.00135187 - layer.3.k_cache 0.00132828 0.77602804 - layer.3.v_cache 0.00000204 0.00224522 - layer.4.k_cache 0.00320824 1.50588644 - layer.4.v_cache 0.00000300 0.00380668 - layer.4.output 0.00017823 0.12128035 - ------------------------------------------------------------------------------------- - TOTAL 0.00305891 1.04303728 - (elements=1,204,224) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1204224 -Total Bytes 214852 -BPFP 1.4273 bits/point -EBPFP 2.8546 equivalent bits/point -MSE 1.043037 ----------------------- -------------------------------------------------------- -Time: 3.012s Load: 0.005s, Pack+Encode: 1.770s, Decode+Unpack: 1.237s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 84, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 84, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0430 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample121-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample121-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-layer4-item1.zst (14/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,924B, BPFP=0.1670 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,100B, BPFP=2.0052 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,380B, BPFP=1.0747 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,208B, BPFP=2.1882 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,516B, BPFP=1.1733 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,596B, BPFP=2.2219 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,956B, BPFP=1.6455 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,924B, BPFP=2.1635 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,300B, BPFP=0.8941 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,760B, BPFP=2.2361 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,872B, BPFP=0.9304 -⌛️ [2/4] FRONTEND: Frontend time: 1.792s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.426s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03166741 9.82503526 - layer.0.v_cache 0.00000027 0.00027082 - layer.1.k_cache 0.00338869 1.41006995 - layer.1.v_cache 0.00000080 0.00094059 - layer.2.k_cache 0.00112306 0.66113972 - layer.2.v_cache 0.00000109 0.00137482 - layer.3.k_cache 0.00129226 0.78247087 - layer.3.v_cache 0.00000208 0.00223893 - layer.4.k_cache 0.00325011 1.46071438 - layer.4.v_cache 0.00000309 0.00383288 - layer.4.output 0.00016397 0.10613736 - ------------------------------------------------------------------------------------- - TOTAL 0.00295605 1.04090269 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 224536 -BPFP 1.3922 bits/point -EBPFP 2.7844 equivalent bits/point -MSE 1.040903 ----------------------- -------------------------------------------------------- -Time: 3.224s Load: 0.007s, Pack+Encode: 1.792s, Decode+Unpack: 1.426s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0409 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample123-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample123-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst (15/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.011s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,832B, BPFP=0.1626 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,448B, BPFP=2.0817 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,492B, BPFP=1.1090 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,464B, BPFP=2.2607 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,200B, BPFP=1.1719 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,716B, BPFP=2.2830 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,648B, BPFP=1.6555 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,072B, BPFP=2.2259 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,896B, BPFP=0.7898 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,876B, BPFP=2.2972 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,432B, BPFP=0.9640 -⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.304s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03171545 10.57628701 - layer.0.v_cache 0.00000027 0.00026918 - layer.1.k_cache 0.00364625 1.37997090 - layer.1.v_cache 0.00000082 0.00093682 - layer.2.k_cache 0.00114710 0.67890774 - layer.2.v_cache 0.00000106 0.00137552 - layer.3.k_cache 0.00132811 0.77730326 - layer.3.v_cache 0.00000203 0.00228808 - layer.4.k_cache 0.00332194 1.53912440 - layer.4.v_cache 0.00000300 0.00376658 - layer.4.output 0.00016865 0.12704874 - ------------------------------------------------------------------------------------- - TOTAL 0.00298862 1.10488746 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 224076 -BPFP 1.4209 bits/point -EBPFP 2.8419 equivalent bits/point -MSE 1.104887 ----------------------- -------------------------------------------------------- -Time: 3.032s Load: 0.011s, Pack+Encode: 1.717s, Decode+Unpack: 1.304s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1049 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample124-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample124-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-layer4-item1.zst (16/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,896B, BPFP=0.1664 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,136B, BPFP=2.0309 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,328B, BPFP=1.0822 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,144B, BPFP=2.2072 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,196B, BPFP=1.3339 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,592B, BPFP=2.2465 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,788B, BPFP=1.6492 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,896B, BPFP=2.1854 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,472B, BPFP=1.0070 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,496B, BPFP=2.2381 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,112B, BPFP=0.9022 -⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.252s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03218093 10.57047983 - layer.0.v_cache 0.00000027 0.00026785 - layer.1.k_cache 0.00364370 1.55779584 - layer.1.v_cache 0.00000081 0.00096199 - layer.2.k_cache 0.00115129 0.68152743 - layer.2.v_cache 0.00000108 0.00141514 - layer.3.k_cache 0.00133425 0.78206789 - layer.3.v_cache 0.00000206 0.00234812 - layer.4.k_cache 0.00326870 1.57811986 - layer.4.v_cache 0.00000299 0.00381103 - layer.4.output 0.00018569 0.12203220 - ------------------------------------------------------------------------------------- - TOTAL 0.00302349 1.11906598 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 225056 -BPFP 1.4111 bits/point -EBPFP 2.8222 equivalent bits/point -MSE 1.119066 ----------------------- -------------------------------------------------------- -Time: 3.096s Load: 0.006s, Pack+Encode: 1.838s, Decode+Unpack: 1.252s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1191 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample126-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample126-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-layer4-item1.zst (17/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1650 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,452B, BPFP=1.8861 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,100B, BPFP=1.0165 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,992B, BPFP=2.0995 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,364B, BPFP=1.2907 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,328B, BPFP=2.1277 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,416B, BPFP=1.5470 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,636B, BPFP=2.0696 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,912B, BPFP=1.0007 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,316B, BPFP=2.1267 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,656B, BPFP=0.8328 -⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.436s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03082583 9.54080561 - layer.0.v_cache 0.00000028 0.00026480 - layer.1.k_cache 0.00340863 1.50794474 - layer.1.v_cache 0.00000080 0.00094255 - layer.2.k_cache 0.00115869 0.68924967 - layer.2.v_cache 0.00000105 0.00138447 - layer.3.k_cache 0.00132054 0.78447846 - layer.3.v_cache 0.00000208 0.00230683 - layer.4.k_cache 0.00325612 1.57230189 - layer.4.v_cache 0.00000315 0.00387481 - layer.4.output 0.00020065 0.11156053 - ------------------------------------------------------------------------------------- - TOTAL 0.00291284 1.03927114 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 222136 -BPFP 1.3329 bits/point -EBPFP 2.6658 equivalent bits/point -MSE 1.039271 ----------------------- -------------------------------------------------------- -Time: 3.153s Load: 0.007s, Pack+Encode: 1.710s, Decode+Unpack: 1.436s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0393 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample13-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample13-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-layer4-item1.zst (18/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,892B, BPFP=0.1642 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,024B, BPFP=1.9986 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,144B, BPFP=0.9674 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,296B, BPFP=2.1958 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,956B, BPFP=1.1247 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,460B, BPFP=2.2101 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,736B, BPFP=1.5396 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,780B, BPFP=2.1510 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,208B, BPFP=0.8861 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,564B, BPFP=2.2191 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,700B, BPFP=0.8398 -⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.235s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03096869 9.89753215 - layer.0.v_cache 0.00000027 0.00027381 - layer.1.k_cache 0.00338461 1.44569668 - layer.1.v_cache 0.00000079 0.00092632 - layer.2.k_cache 0.00111556 0.65783619 - layer.2.v_cache 0.00000105 0.00135375 - layer.3.k_cache 0.00129967 0.77993842 - layer.3.v_cache 0.00000204 0.00223708 - layer.4.k_cache 0.00324687 1.49737362 - layer.4.v_cache 0.00000301 0.00364955 - layer.4.output 0.00020513 0.11891696 - ------------------------------------------------------------------------------------- - TOTAL 0.00291736 1.05446324 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 216760 -BPFP 1.3440 bits/point -EBPFP 2.6880 equivalent bits/point -MSE 1.054463 ----------------------- -------------------------------------------------------- -Time: 2.976s Load: 0.007s, Pack+Encode: 1.735s, Decode+Unpack: 1.235s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0545 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample132-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample132-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-layer4-item1.zst (19/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,892B, BPFP=0.1661 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,056B, BPFP=2.0239 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,120B, BPFP=1.1517 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,216B, BPFP=2.2135 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,312B, BPFP=1.1685 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,680B, BPFP=2.2542 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,392B, BPFP=1.6145 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,008B, BPFP=2.1952 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,528B, BPFP=0.7486 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,564B, BPFP=2.2440 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,320B, BPFP=0.8848 -⌛️ [2/4] FRONTEND: Frontend time: 1.808s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.328s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03072730 9.55709187 - layer.0.v_cache 0.00000028 0.00027005 - layer.1.k_cache 0.00346408 1.57619622 - layer.1.v_cache 0.00000080 0.00094569 - layer.2.k_cache 0.00114708 0.68092741 - layer.2.v_cache 0.00000106 0.00139337 - layer.3.k_cache 0.00132643 0.76085757 - layer.3.v_cache 0.00000205 0.00235146 - layer.4.k_cache 0.00323342 1.44504521 - layer.4.v_cache 0.00000308 0.00385076 - layer.4.output 0.00016383 0.11232437 - ------------------------------------------------------------------------------------- - TOTAL 0.00289721 1.03415908 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 220088 -BPFP 1.3800 bits/point -EBPFP 2.7599 equivalent bits/point -MSE 1.034159 ----------------------- -------------------------------------------------------- -Time: 3.142s Load: 0.006s, Pack+Encode: 1.808s, Decode+Unpack: 1.328s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0342 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample134-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample134-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-layer4-item1.zst (20/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,932B, BPFP=0.1641 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,020B, BPFP=1.9548 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,664B, BPFP=1.0754 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,944B, BPFP=2.1182 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,820B, BPFP=1.1736 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,456B, BPFP=2.1617 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,564B, BPFP=1.5764 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,580B, BPFP=2.0873 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,644B, BPFP=0.9888 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,396B, BPFP=2.1566 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,292B, BPFP=0.8554 -⌛️ [2/4] FRONTEND: Frontend time: 1.696s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.363s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03070580 9.97173342 - layer.0.v_cache 0.00000027 0.00026925 - layer.1.k_cache 0.00347185 1.54611505 - layer.1.v_cache 0.00000080 0.00093127 - layer.2.k_cache 0.00112340 0.64934896 - layer.2.v_cache 0.00000106 0.00139570 - layer.3.k_cache 0.00133882 0.76243666 - layer.3.v_cache 0.00000200 0.00226039 - layer.4.k_cache 0.00333409 1.60862334 - layer.4.v_cache 0.00000307 0.00380796 - layer.4.output 0.00016236 0.11208983 - ------------------------------------------------------------------------------------- - TOTAL 0.00290219 1.07109152 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 222312 -BPFP 1.3485 bits/point -EBPFP 2.6969 equivalent bits/point -MSE 1.071092 ----------------------- -------------------------------------------------------- -Time: 3.065s Load: 0.006s, Pack+Encode: 1.696s, Decode+Unpack: 1.363s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0711 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample135-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample135-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-layer4-item1.zst (21/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,856B, BPFP=0.1667 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,472B, BPFP=2.1078 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,344B, BPFP=1.1085 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,628B, BPFP=2.3014 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,576B, BPFP=1.1293 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,688B, BPFP=2.3068 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,788B, BPFP=1.6871 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,056B, BPFP=2.2500 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,532B, BPFP=0.9458 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,740B, BPFP=2.3114 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,836B, BPFP=0.8719 -⌛️ [2/4] FRONTEND: Frontend time: 1.808s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.257s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03132415 9.48566726 - layer.0.v_cache 0.00000028 0.00026897 - layer.1.k_cache 0.00348164 1.45307852 - layer.1.v_cache 0.00000079 0.00090418 - layer.2.k_cache 0.00113511 0.66111098 - layer.2.v_cache 0.00000106 0.00132377 - layer.3.k_cache 0.00131762 0.76663050 - layer.3.v_cache 0.00000207 0.00225537 - layer.4.k_cache 0.00320374 1.46070914 - layer.4.v_cache 0.00000303 0.00376276 - layer.4.output 0.00016514 0.11330682 - ------------------------------------------------------------------------------------- - TOTAL 0.00293786 1.02063848 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 220516 -BPFP 1.4144 bits/point -EBPFP 2.8289 equivalent bits/point -MSE 1.020638 ----------------------- -------------------------------------------------------- -Time: 3.071s Load: 0.006s, Pack+Encode: 1.808s, Decode+Unpack: 1.257s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0206 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample136-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample136-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst (22/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,852B, BPFP=0.1644 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,072B, BPFP=2.0483 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,004B, BPFP=0.9769 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,100B, BPFP=2.2283 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,424B, BPFP=1.1918 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,440B, BPFP=2.2585 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,400B, BPFP=1.6335 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,720B, BPFP=2.1946 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,244B, BPFP=0.8207 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,576B, BPFP=2.2706 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,860B, BPFP=0.8847 -⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.425s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03137242 9.60769168 - layer.0.v_cache 0.00000027 0.00027140 - layer.1.k_cache 0.00337550 1.42557543 - layer.1.v_cache 0.00000078 0.00093797 - layer.2.k_cache 0.00112765 0.65684188 - layer.2.v_cache 0.00000106 0.00139041 - layer.3.k_cache 0.00130259 0.76746620 - layer.3.v_cache 0.00000201 0.00228693 - layer.4.k_cache 0.00311145 1.46465562 - layer.4.v_cache 0.00000304 0.00388196 - layer.4.output 0.00020755 0.12537449 - ------------------------------------------------------------------------------------- - TOTAL 0.00293764 1.03089267 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 217692 -BPFP 1.3805 bits/point -EBPFP 2.7609 equivalent bits/point -MSE 1.030893 ----------------------- -------------------------------------------------------- -Time: 3.166s Load: 0.006s, Pack+Encode: 1.735s, Decode+Unpack: 1.425s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0309 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample137-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample137-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-layer4-item1.zst (23/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,492B, BPFP=1.8693 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,464B, BPFP=1.0359 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,252B, BPFP=2.0987 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,448B, BPFP=1.2839 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,468B, BPFP=2.1167 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,788B, BPFP=1.5615 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,680B, BPFP=2.0512 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,492B, BPFP=1.0382 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,612B, BPFP=2.1287 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,320B, BPFP=0.8170 -⌛️ [2/4] FRONTEND: Frontend time: 1.700s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.242s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03043337 9.11985227 - layer.0.v_cache 0.00000027 0.00027050 - layer.1.k_cache 0.00337587 1.54495889 - layer.1.v_cache 0.00000080 0.00094687 - layer.2.k_cache 0.00115990 0.70619072 - layer.2.v_cache 0.00000105 0.00137327 - layer.3.k_cache 0.00132055 0.78394236 - layer.3.v_cache 0.00000202 0.00227380 - layer.4.k_cache 0.00326599 1.51200688 - layer.4.v_cache 0.00000301 0.00377434 - layer.4.output 0.00017344 0.11311491 - ------------------------------------------------------------------------------------- - TOTAL 0.00287547 1.00914640 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 223980 -BPFP 1.3297 bits/point -EBPFP 2.6593 equivalent bits/point -MSE 1.009146 ----------------------- -------------------------------------------------------- -Time: 2.947s Load: 0.006s, Pack+Encode: 1.700s, Decode+Unpack: 1.242s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0091 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample138-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample138-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst (24/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,840B, BPFP=0.1634 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,184B, BPFP=2.0582 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,800B, BPFP=1.0476 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,360B, BPFP=2.2514 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,732B, BPFP=1.2191 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,632B, BPFP=2.2756 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 16,684B, BPFP=1.4812 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,996B, BPFP=2.2191 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,908B, BPFP=0.8796 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,652B, BPFP=2.2773 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,708B, BPFP=0.9035 -⌛️ [2/4] FRONTEND: Frontend time: 1.815s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.317s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03171148 10.36345395 - layer.0.v_cache 0.00000028 0.00027006 - layer.1.k_cache 0.00359844 1.46455470 - layer.1.v_cache 0.00000084 0.00094081 - layer.2.k_cache 0.00114211 0.65597196 - layer.2.v_cache 0.00000108 0.00139735 - layer.3.k_cache 0.00132976 0.75953232 - layer.3.v_cache 0.00000213 0.00234696 - layer.4.k_cache 0.00324865 1.50587290 - layer.4.v_cache 0.00000293 0.00375674 - layer.4.output 0.00016674 0.11224034 - ------------------------------------------------------------------------------------- - TOTAL 0.00297890 1.08621851 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 219496 -BPFP 1.3919 bits/point -EBPFP 2.7838 equivalent bits/point -MSE 1.086219 ----------------------- -------------------------------------------------------- -Time: 3.138s Load: 0.006s, Pack+Encode: 1.815s, Decode+Unpack: 1.317s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0862 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample139-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample139-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-layer4-item1.zst (25/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,888B, BPFP=0.1657 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,112B, BPFP=2.0288 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,856B, BPFP=1.0407 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,112B, BPFP=2.2044 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,832B, BPFP=1.1264 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,456B, BPFP=2.2346 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 20,032B, BPFP=1.7584 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,736B, BPFP=2.1713 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,532B, BPFP=0.8367 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,496B, BPFP=2.2381 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,200B, BPFP=0.9700 -⌛️ [2/4] FRONTEND: Frontend time: 1.700s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.345s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03094440 9.99708129 - layer.0.v_cache 0.00000027 0.00027173 - layer.1.k_cache 0.00351605 1.46904386 - layer.1.v_cache 0.00000081 0.00093475 - layer.2.k_cache 0.00112935 0.67920256 - layer.2.v_cache 0.00000105 0.00135446 - layer.3.k_cache 0.00132761 0.79861673 - layer.3.v_cache 0.00000216 0.00233332 - layer.4.k_cache 0.00319280 1.50770689 - layer.4.v_cache 0.00000300 0.00377903 - layer.4.output 0.00020683 0.12153718 - ------------------------------------------------------------------------------------- - TOTAL 0.00292463 1.06760524 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 224252 -BPFP 1.4061 bits/point -EBPFP 2.8121 equivalent bits/point -MSE 1.067605 ----------------------- -------------------------------------------------------- -Time: 3.052s Load: 0.007s, Pack+Encode: 1.700s, Decode+Unpack: 1.345s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0676 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample140-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample140-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-layer4-item1.zst (26/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,980B, BPFP=0.1681 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,156B, BPFP=1.9664 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,136B, BPFP=1.0306 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,216B, BPFP=2.1413 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,296B, BPFP=1.2989 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,532B, BPFP=2.1681 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,156B, BPFP=1.5418 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,832B, BPFP=2.1087 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,916B, BPFP=1.0968 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,772B, BPFP=2.1885 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,628B, BPFP=0.9262 -⌛️ [2/4] FRONTEND: Frontend time: 1.787s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.248s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03142291 10.33124774 - layer.0.v_cache 0.00000027 0.00026977 - layer.1.k_cache 0.00332372 1.56226034 - layer.1.v_cache 0.00000083 0.00092819 - layer.2.k_cache 0.00117076 0.65906135 - layer.2.v_cache 0.00000109 0.00140620 - layer.3.k_cache 0.00132340 0.76654683 - layer.3.v_cache 0.00000203 0.00224144 - layer.4.k_cache 0.00333891 1.50647304 - layer.4.v_cache 0.00000309 0.00386665 - layer.4.output 0.00017306 0.11921089 - ------------------------------------------------------------------------------------- - TOTAL 0.00294852 1.09365322 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 228620 -BPFP 1.3867 bits/point -EBPFP 2.7734 equivalent bits/point -MSE 1.093653 ----------------------- -------------------------------------------------------- -Time: 3.042s Load: 0.006s, Pack+Encode: 1.787s, Decode+Unpack: 1.248s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0937 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample141-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample141-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-layer4-item1.zst (27/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1633 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,864B, BPFP=1.9207 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,368B, BPFP=1.1230 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,252B, BPFP=2.1213 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,848B, BPFP=1.3313 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,640B, BPFP=2.1539 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,320B, BPFP=1.5390 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,864B, BPFP=2.0887 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,508B, BPFP=1.0507 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,476B, BPFP=2.1401 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 45,344B, BPFP=0.9523 -⌛️ [2/4] FRONTEND: Frontend time: 1.750s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.459s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03274377 9.97115597 - layer.0.v_cache 0.00000028 0.00027036 - layer.1.k_cache 0.00350561 1.63069924 - layer.1.v_cache 0.00000082 0.00096909 - layer.2.k_cache 0.00114290 0.68004198 - layer.2.v_cache 0.00000109 0.00143141 - layer.3.k_cache 0.00132787 0.78582739 - layer.3.v_cache 0.00000209 0.00240939 - layer.4.k_cache 0.00329855 1.55353439 - layer.4.v_cache 0.00000319 0.00390237 - layer.4.output 0.00017541 0.10872663 - ------------------------------------------------------------------------------------- - TOTAL 0.00305199 1.07608201 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 231428 -BPFP 1.3887 bits/point -EBPFP 2.7773 equivalent bits/point -MSE 1.076082 ----------------------- -------------------------------------------------------- -Time: 3.216s Load: 0.007s, Pack+Encode: 1.750s, Decode+Unpack: 1.459s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0761 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample142-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample142-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-layer4-item1.zst (28/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 82, 128) -Output shape: (1, 82, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) -> torch.Size([1, 1, 82, 1024]) - layer.4.output: torch.Size([1, 82, 4096]) -> torch.Size([1, 1, 82, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,804B, BPFP=0.1719 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,364B, BPFP=2.2260 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 9,684B, BPFP=0.9226 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,312B, BPFP=2.4116 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 11,584B, BPFP=1.1037 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,856B, BPFP=2.4634 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 16,012B, BPFP=1.5255 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,180B, BPFP=2.3990 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,292B, BPFP=0.7900 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,840B, BPFP=2.4619 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,744B, BPFP=0.9466 -⌛️ [2/4] FRONTEND: Frontend time: 1.689s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.263s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 82, 128]) - layer.0.v_cache: torch.Size([1, 8, 82, 128]) - layer.1.k_cache: torch.Size([1, 8, 82, 128]) - layer.1.v_cache: torch.Size([1, 8, 82, 128]) - layer.2.k_cache: torch.Size([1, 8, 82, 128]) - layer.2.v_cache: torch.Size([1, 8, 82, 128]) - layer.3.k_cache: torch.Size([1, 8, 82, 128]) - layer.3.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.k_cache: torch.Size([1, 8, 82, 128]) - layer.4.v_cache: torch.Size([1, 8, 82, 128]) - layer.4.output: torch.Size([1, 82, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03240362 10.01397705 - layer.0.v_cache 0.00000028 0.00027083 - layer.1.k_cache 0.00358562 1.61043065 - layer.1.v_cache 0.00000082 0.00092233 - layer.2.k_cache 0.00116773 0.69988307 - layer.2.v_cache 0.00000108 0.00137847 - layer.3.k_cache 0.00133166 0.77561969 - layer.3.v_cache 0.00000211 0.00229257 - layer.4.k_cache 0.00321141 1.46422847 - layer.4.v_cache 0.00000303 0.00372957 - layer.4.output 0.00019620 0.12769113 - ------------------------------------------------------------------------------------- - TOTAL 0.00303515 1.07739266 - (elements=1,175,552) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1175552 -Total Bytes 212672 -BPFP 1.4473 bits/point -EBPFP 2.8946 equivalent bits/point -MSE 1.077393 ----------------------- -------------------------------------------------------- -Time: 2.957s Load: 0.005s, Pack+Encode: 1.689s, Decode+Unpack: 1.263s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 82, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 82, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0774 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample145-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample145-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-layer4-item1.zst (29/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1685 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,552B, BPFP=2.1149 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,808B, BPFP=1.2399 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,640B, BPFP=2.3024 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,216B, BPFP=1.0970 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,712B, BPFP=2.3089 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,308B, BPFP=1.5542 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,204B, BPFP=2.2633 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,748B, BPFP=1.0550 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,732B, BPFP=2.3107 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,148B, BPFP=1.0360 -⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.289s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03155136 9.95576495 - layer.0.v_cache 0.00000028 0.00027333 - layer.1.k_cache 0.00336733 1.49374846 - layer.1.v_cache 0.00000080 0.00093712 - layer.2.k_cache 0.00113514 0.66065983 - layer.2.v_cache 0.00000108 0.00137221 - layer.3.k_cache 0.00133108 0.75384451 - layer.3.v_cache 0.00000214 0.00229580 - layer.4.k_cache 0.00321453 1.49940315 - layer.4.v_cache 0.00000314 0.00376390 - layer.4.output 0.00017293 0.10914861 - ------------------------------------------------------------------------------------- - TOTAL 0.00294990 1.05776126 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 228944 -BPFP 1.4685 bits/point -EBPFP 2.9370 equivalent bits/point -MSE 1.057761 ----------------------- -------------------------------------------------------- -Time: 3.120s Load: 0.006s, Pack+Encode: 1.825s, Decode+Unpack: 1.289s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0578 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample147-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample147-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-layer4-item1.zst (30/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,940B, BPFP=0.1647 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,980B, BPFP=1.9514 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,372B, BPFP=1.0506 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,132B, BPFP=2.1342 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,880B, BPFP=1.2636 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,552B, BPFP=2.1698 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,424B, BPFP=1.6495 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,816B, BPFP=2.1073 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,852B, BPFP=1.0914 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,476B, BPFP=2.1634 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,468B, BPFP=0.8803 -⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.440s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03082907 9.66592938 - layer.0.v_cache 0.00000028 0.00027073 - layer.1.k_cache 0.00336151 1.54810582 - layer.1.v_cache 0.00000078 0.00092742 - layer.2.k_cache 0.00113870 0.66007017 - layer.2.v_cache 0.00000105 0.00136980 - layer.3.k_cache 0.00130133 0.77515561 - layer.3.v_cache 0.00000208 0.00231117 - layer.4.k_cache 0.00328680 1.53245909 - layer.4.v_cache 0.00000308 0.00384117 - layer.4.output 0.00016380 0.10902934 - ------------------------------------------------------------------------------------- - TOTAL 0.00289856 1.04475412 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 226892 -BPFP 1.3762 bits/point -EBPFP 2.7525 equivalent bits/point -MSE 1.044754 ----------------------- -------------------------------------------------------- -Time: 3.150s Load: 0.007s, Pack+Encode: 1.704s, Decode+Unpack: 1.440s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0448 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample149-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample149-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-layer4-item1.zst (31/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,900B, BPFP=0.1649 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,820B, BPFP=1.9809 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,192B, BPFP=0.9715 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,260B, BPFP=2.1927 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,132B, BPFP=1.1399 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,636B, BPFP=2.2253 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,760B, BPFP=1.6285 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,900B, BPFP=2.1615 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,120B, BPFP=0.8785 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,548B, BPFP=2.2177 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,740B, BPFP=0.9058 -⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.249s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03195141 9.96253798 - layer.0.v_cache 0.00000027 0.00027336 - layer.1.k_cache 0.00349494 1.55829197 - layer.1.v_cache 0.00000082 0.00093083 - layer.2.k_cache 0.00113744 0.67257300 - layer.2.v_cache 0.00000112 0.00141192 - layer.3.k_cache 0.00129971 0.77178641 - layer.3.v_cache 0.00000221 0.00233429 - layer.4.k_cache 0.00331958 1.53178067 - layer.4.v_cache 0.00000310 0.00381295 - layer.4.output 0.00017227 0.10755233 - ------------------------------------------------------------------------------------- - TOTAL 0.00299284 1.06685305 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 221008 -BPFP 1.3703 bits/point -EBPFP 2.7407 equivalent bits/point -MSE 1.066853 ----------------------- -------------------------------------------------------- -Time: 3.018s Load: 0.006s, Pack+Encode: 1.763s, Decode+Unpack: 1.249s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0669 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample15-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample15-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-layer4-item1.zst (32/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,568B, BPFP=1.8757 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,192B, BPFP=1.0133 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,240B, BPFP=2.0977 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,296B, BPFP=1.2713 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,420B, BPFP=2.1127 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,616B, BPFP=1.5472 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,628B, BPFP=2.0469 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,092B, BPFP=1.0050 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,548B, BPFP=2.1233 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,148B, BPFP=0.8134 -⌛️ [2/4] FRONTEND: Frontend time: 1.780s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.344s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03049540 9.09146118 - layer.0.v_cache 0.00000028 0.00027115 - layer.1.k_cache 0.00333180 1.50487648 - layer.1.v_cache 0.00000080 0.00094374 - layer.2.k_cache 0.00116407 0.71755770 - layer.2.v_cache 0.00000105 0.00137027 - layer.3.k_cache 0.00132929 0.77843654 - layer.3.v_cache 0.00000200 0.00227509 - layer.4.k_cache 0.00323017 1.52398260 - layer.4.v_cache 0.00000303 0.00378163 - layer.4.output 0.00017160 0.11472025 - ------------------------------------------------------------------------------------- - TOTAL 0.00287459 1.00598839 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 222712 -BPFP 1.3221 bits/point -EBPFP 2.6443 equivalent bits/point -MSE 1.005988 ----------------------- -------------------------------------------------------- -Time: 3.131s Load: 0.007s, Pack+Encode: 1.780s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0060 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample153-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample153-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-layer4-item1.zst (33/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,852B, BPFP=0.1702 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,444B, BPFP=2.1548 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,340B, BPFP=1.3180 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,368B, BPFP=2.3316 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,220B, BPFP=1.1232 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,836B, BPFP=2.3746 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,484B, BPFP=1.7908 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,084B, BPFP=2.3055 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,152B, BPFP=0.8412 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,796B, BPFP=2.3710 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,020B, BPFP=0.8736 -⌛️ [2/4] FRONTEND: Frontend time: 1.692s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.309s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03223823 10.97735668 - layer.0.v_cache 0.00000028 0.00026909 - layer.1.k_cache 0.00365975 1.45888564 - layer.1.v_cache 0.00000080 0.00093237 - layer.2.k_cache 0.00113736 0.67850719 - layer.2.v_cache 0.00000107 0.00140295 - layer.3.k_cache 0.00130376 0.75756603 - layer.3.v_cache 0.00000210 0.00226788 - layer.4.k_cache 0.00330933 1.45835607 - layer.4.v_cache 0.00000299 0.00378442 - layer.4.output 0.00016397 0.11482234 - ------------------------------------------------------------------------------------- - TOTAL 0.00302225 1.12847269 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 220596 -BPFP 1.4482 bits/point -EBPFP 2.8965 equivalent bits/point -MSE 1.128473 ----------------------- -------------------------------------------------------- -Time: 3.007s Load: 0.006s, Pack+Encode: 1.692s, Decode+Unpack: 1.309s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1285 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample154-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample154-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-layer4-item1.zst (34/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,868B, BPFP=0.1677 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,500B, BPFP=2.1103 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,788B, BPFP=0.9688 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,632B, BPFP=2.3017 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,828B, BPFP=1.1519 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,836B, BPFP=2.3200 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,940B, BPFP=1.6110 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,136B, BPFP=2.2572 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,092B, BPFP=0.8165 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,800B, BPFP=2.3168 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,720B, BPFP=0.9142 -⌛️ [2/4] FRONTEND: Frontend time: 1.810s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.241s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03126283 9.93361760 - layer.0.v_cache 0.00000028 0.00027110 - layer.1.k_cache 0.00362225 1.56604495 - layer.1.v_cache 0.00000080 0.00094347 - layer.2.k_cache 0.00113278 0.66045152 - layer.2.v_cache 0.00000108 0.00140008 - layer.3.k_cache 0.00132193 0.75753468 - layer.3.v_cache 0.00000209 0.00231197 - layer.4.k_cache 0.00322631 1.45778121 - layer.4.v_cache 0.00000306 0.00375971 - layer.4.output 0.00017090 0.10613041 - ------------------------------------------------------------------------------------- - TOTAL 0.00294693 1.05775985 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 219140 -BPFP 1.4056 bits/point -EBPFP 2.8112 equivalent bits/point -MSE 1.057760 ----------------------- -------------------------------------------------------- -Time: 3.058s Load: 0.007s, Pack+Encode: 1.810s, Decode+Unpack: 1.241s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0578 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample155-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample155-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst (35/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,848B, BPFP=0.1641 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,040B, BPFP=2.0455 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,544B, BPFP=0.9361 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,144B, BPFP=2.2322 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,172B, BPFP=1.1694 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,680B, BPFP=2.2798 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,944B, BPFP=1.6818 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,000B, BPFP=2.2195 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,192B, BPFP=0.9048 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,764B, BPFP=2.2873 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,892B, BPFP=0.9298 -⌛️ [2/4] FRONTEND: Frontend time: 1.706s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.432s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03162268 10.00297477 - layer.0.v_cache 0.00000028 0.00027037 - layer.1.k_cache 0.00348150 1.34194756 - layer.1.v_cache 0.00000080 0.00095026 - layer.2.k_cache 0.00116795 0.66052172 - layer.2.v_cache 0.00000108 0.00142404 - layer.3.k_cache 0.00133447 0.78158483 - layer.3.v_cache 0.00000211 0.00239099 - layer.4.k_cache 0.00331899 1.51253596 - layer.4.v_cache 0.00000308 0.00385921 - layer.4.output 0.00016829 0.11435692 - ------------------------------------------------------------------------------------- - TOTAL 0.00297187 1.05470624 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 221220 -BPFP 1.4028 bits/point -EBPFP 2.8057 equivalent bits/point -MSE 1.054706 ----------------------- -------------------------------------------------------- -Time: 3.145s Load: 0.007s, Pack+Encode: 1.706s, Decode+Unpack: 1.432s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0547 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample156-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample156-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-layer4-item1.zst (36/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,912B, BPFP=0.1660 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,240B, BPFP=2.0174 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,772B, BPFP=1.1087 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,100B, BPFP=2.1788 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,752B, BPFP=1.1069 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,348B, BPFP=2.2003 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,952B, BPFP=1.7319 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,696B, BPFP=2.1437 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,428B, BPFP=0.8184 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,496B, BPFP=2.2132 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,300B, BPFP=0.8529 -⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.242s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03130752 9.79668918 - layer.0.v_cache 0.00000027 0.00027281 - layer.1.k_cache 0.00339397 1.50617998 - layer.1.v_cache 0.00000079 0.00092763 - layer.2.k_cache 0.00115273 0.66510434 - layer.2.v_cache 0.00000105 0.00135948 - layer.3.k_cache 0.00133725 0.80423821 - layer.3.v_cache 0.00000203 0.00225951 - layer.4.k_cache 0.00329241 1.52314317 - layer.4.v_cache 0.00000297 0.00370133 - layer.4.output 0.00017286 0.11673074 - ------------------------------------------------------------------------------------- - TOTAL 0.00294160 1.05505704 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 219996 -BPFP 1.3641 bits/point -EBPFP 2.7281 equivalent bits/point -MSE 1.055057 ----------------------- -------------------------------------------------------- -Time: 3.019s Load: 0.007s, Pack+Encode: 1.771s, Decode+Unpack: 1.242s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0551 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample157-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample157-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-layer4-item1.zst (37/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,940B, BPFP=0.1684 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,948B, BPFP=1.9920 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,728B, BPFP=1.1049 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,128B, BPFP=2.1812 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,112B, BPFP=1.0514 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,384B, BPFP=2.2035 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,640B, BPFP=1.6181 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,720B, BPFP=2.1458 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,928B, BPFP=0.8618 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,424B, BPFP=2.2069 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,800B, BPFP=0.8854 -⌛️ [2/4] FRONTEND: Frontend time: 1.806s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.378s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03118540 9.92364434 - layer.0.v_cache 0.00000028 0.00027582 - layer.1.k_cache 0.00362682 1.54584978 - layer.1.v_cache 0.00000082 0.00094879 - layer.2.k_cache 0.00111630 0.68193351 - layer.2.v_cache 0.00000105 0.00137737 - layer.3.k_cache 0.00134410 0.81154056 - layer.3.v_cache 0.00000205 0.00228589 - layer.4.k_cache 0.00329490 1.54266934 - layer.4.v_cache 0.00000297 0.00378068 - layer.4.output 0.00020247 0.11848047 - ------------------------------------------------------------------------------------- - TOTAL 0.00295604 1.07058771 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 219752 -BPFP 1.3625 bits/point -EBPFP 2.7251 equivalent bits/point -MSE 1.070588 ----------------------- -------------------------------------------------------- -Time: 3.190s Load: 0.006s, Pack+Encode: 1.806s, Decode+Unpack: 1.378s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0706 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample16-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample16-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-layer4-item1.zst (38/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1644 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,864B, BPFP=1.9416 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,780B, BPFP=1.1702 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,012B, BPFP=2.1240 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,848B, BPFP=1.2609 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,512B, BPFP=2.1664 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,212B, BPFP=1.5465 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,804B, BPFP=2.1063 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,488B, BPFP=0.9755 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,592B, BPFP=2.1732 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,840B, BPFP=0.9307 -⌛️ [2/4] FRONTEND: Frontend time: 1.684s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.342s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03142573 9.75133349 - layer.0.v_cache 0.00000027 0.00026725 - layer.1.k_cache 0.00360029 1.60239046 - layer.1.v_cache 0.00000081 0.00095319 - layer.2.k_cache 0.00114206 0.64874894 - layer.2.v_cache 0.00000107 0.00139553 - layer.3.k_cache 0.00131314 0.78419677 - layer.3.v_cache 0.00000205 0.00233150 - layer.4.k_cache 0.00327277 1.59333138 - layer.4.v_cache 0.00000301 0.00385309 - layer.4.output 0.00017705 0.11412423 - ------------------------------------------------------------------------------------- - TOTAL 0.00296210 1.06037846 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 227888 -BPFP 1.3823 bits/point -EBPFP 2.7646 equivalent bits/point -MSE 1.060378 ----------------------- -------------------------------------------------------- -Time: 3.033s Load: 0.006s, Pack+Encode: 1.684s, Decode+Unpack: 1.342s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0604 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample17-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample17-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-layer4-item1.zst (39/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,924B, BPFP=0.1634 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,972B, BPFP=1.9507 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,660B, BPFP=0.9901 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,104B, BPFP=2.1318 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,724B, BPFP=1.2503 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,380B, BPFP=2.1552 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,160B, BPFP=1.6270 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,760B, BPFP=2.1026 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,656B, BPFP=0.9049 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,424B, BPFP=2.1590 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,912B, BPFP=0.8898 -⌛️ [2/4] FRONTEND: Frontend time: 1.824s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.248s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03108079 9.83050935 - layer.0.v_cache 0.00000027 0.00026337 - layer.1.k_cache 0.00341615 1.54387781 - layer.1.v_cache 0.00000080 0.00092781 - layer.2.k_cache 0.00113779 0.66655387 - layer.2.v_cache 0.00000105 0.00135953 - layer.3.k_cache 0.00132123 0.79094033 - layer.3.v_cache 0.00000201 0.00226276 - layer.4.k_cache 0.00323701 1.51263859 - layer.4.v_cache 0.00000300 0.00378256 - layer.4.output 0.00016915 0.12707944 - ------------------------------------------------------------------------------------- - TOTAL 0.00291977 1.06153098 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 223676 -BPFP 1.3567 bits/point -EBPFP 2.7135 equivalent bits/point -MSE 1.061531 ----------------------- -------------------------------------------------------- -Time: 3.079s Load: 0.007s, Pack+Encode: 1.824s, Decode+Unpack: 1.248s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0615 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample18-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample18-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-layer4-item1.zst (40/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,916B, BPFP=0.1663 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,004B, BPFP=1.9969 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,728B, BPFP=1.1049 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,200B, BPFP=2.1875 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,636B, BPFP=1.0969 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,380B, BPFP=2.2031 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,300B, BPFP=1.5885 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,720B, BPFP=2.1458 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,144B, BPFP=0.7937 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,444B, BPFP=2.2087 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,712B, BPFP=0.8835 -⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.449s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03108226 9.54052531 - layer.0.v_cache 0.00000028 0.00027725 - layer.1.k_cache 0.00355800 1.49863112 - layer.1.v_cache 0.00000082 0.00095289 - layer.2.k_cache 0.00111789 0.64641575 - layer.2.v_cache 0.00000106 0.00137944 - layer.3.k_cache 0.00134609 0.78596446 - layer.3.v_cache 0.00000204 0.00230316 - layer.4.k_cache 0.00329523 1.51446838 - layer.4.v_cache 0.00000295 0.00377231 - layer.4.output 0.00029330 0.11592556 - ------------------------------------------------------------------------------------- - TOTAL 0.00296998 1.03274231 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 219184 -BPFP 1.3590 bits/point -EBPFP 2.7181 equivalent bits/point -MSE 1.032742 ----------------------- -------------------------------------------------------- -Time: 3.200s Load: 0.007s, Pack+Encode: 1.744s, Decode+Unpack: 1.449s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0327 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample19-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample19-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-layer4-item1.zst (41/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 98, 128) -Output shape: (1, 98, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.output: torch.Size([1, 98, 4096]) -> torch.Size([1, 1, 98, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,080B, BPFP=0.1658 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,296B, BPFP=1.8571 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,232B, BPFP=1.0548 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,500B, BPFP=2.0328 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,712B, BPFP=1.1728 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,752B, BPFP=2.0529 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,952B, BPFP=1.5108 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,072B, BPFP=1.9987 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,888B, BPFP=1.0274 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 26,144B, BPFP=2.0842 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,896B, BPFP=0.9944 -⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.243s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03050847 9.50578028 - layer.0.v_cache 0.00000027 0.00026969 - layer.1.k_cache 0.00347784 1.63184372 - layer.1.v_cache 0.00000091 0.00092912 - layer.2.k_cache 0.00115394 0.67259699 - layer.2.v_cache 0.00000107 0.00136183 - layer.3.k_cache 0.00132167 0.79251363 - layer.3.v_cache 0.00000207 0.00229623 - layer.4.k_cache 0.00336398 1.49748526 - layer.4.v_cache 0.00000299 0.00371097 - layer.4.output 0.00016871 0.11032356 - ------------------------------------------------------------------------------------- - TOTAL 0.00289343 1.03929157 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 237524 -BPFP 1.3525 bits/point -EBPFP 2.7050 equivalent bits/point -MSE 1.039292 ----------------------- -------------------------------------------------------- -Time: 2.966s Load: 0.008s, Pack+Encode: 1.715s, Decode+Unpack: 1.243s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0393 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample2-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample2-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-layer4-item1.zst (42/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1633 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,924B, BPFP=1.9257 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,472B, BPFP=0.9637 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,088B, BPFP=2.1075 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,080B, BPFP=1.1828 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,608B, BPFP=2.1512 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,852B, BPFP=1.4997 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,816B, BPFP=2.0847 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,924B, BPFP=0.9177 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,504B, BPFP=2.1425 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,404B, BPFP=0.8905 -⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.349s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03217121 9.86404255 - layer.0.v_cache 0.00000028 0.00026953 - layer.1.k_cache 0.00347451 1.52520932 - layer.1.v_cache 0.00000083 0.00093756 - layer.2.k_cache 0.00115003 0.68662176 - layer.2.v_cache 0.00000108 0.00139887 - layer.3.k_cache 0.00134333 0.76527610 - layer.3.v_cache 0.00000208 0.00234579 - layer.4.k_cache 0.00325638 1.45820880 - layer.4.v_cache 0.00000311 0.00384508 - layer.4.output 0.00016214 0.10638611 - ------------------------------------------------------------------------------------- - TOTAL 0.00300367 1.05240713 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 222616 -BPFP 1.3358 bits/point -EBPFP 2.6716 equivalent bits/point -MSE 1.052407 ----------------------- -------------------------------------------------------- -Time: 3.181s Load: 0.007s, Pack+Encode: 1.825s, Decode+Unpack: 1.349s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0524 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample20-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample20-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-layer4-item1.zst (43/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,952B, BPFP=0.1622 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,596B, BPFP=1.8780 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,100B, BPFP=1.0057 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,712B, BPFP=2.0539 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,768B, BPFP=1.2274 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,096B, BPFP=2.0858 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,328B, BPFP=1.5233 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,412B, BPFP=2.0289 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,720B, BPFP=0.9741 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,160B, BPFP=2.0911 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 37,232B, BPFP=0.7736 -⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.399s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03064084 9.60624792 - layer.0.v_cache 0.00000027 0.00026772 - layer.1.k_cache 0.00361138 1.48072604 - layer.1.v_cache 0.00000084 0.00091443 - layer.2.k_cache 0.00114431 0.68144356 - layer.2.v_cache 0.00000106 0.00135821 - layer.3.k_cache 0.00136507 0.78136931 - layer.3.v_cache 0.00000197 0.00220764 - layer.4.k_cache 0.00326917 1.57136730 - layer.4.v_cache 0.00000291 0.00364935 - layer.4.output 0.00019732 0.11615443 - ------------------------------------------------------------------------------------- - TOTAL 0.00291622 1.04244066 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 218076 -BPFP 1.2946 bits/point -EBPFP 2.5892 equivalent bits/point -MSE 1.042441 ----------------------- -------------------------------------------------------- -Time: 3.108s Load: 0.007s, Pack+Encode: 1.702s, Decode+Unpack: 1.399s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0424 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample21-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample21-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-layer4-item1.zst (44/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,016B, BPFP=0.1641 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,060B, BPFP=1.8766 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,600B, BPFP=1.1068 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,996B, BPFP=2.0342 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,996B, BPFP=1.0576 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,360B, BPFP=2.0638 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,588B, BPFP=1.5127 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,856B, BPFP=2.0228 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,876B, BPFP=1.1292 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,408B, BPFP=2.0677 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 45,892B, BPFP=0.9337 -⌛️ [2/4] FRONTEND: Frontend time: 1.782s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.245s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03181871 9.54093170 - layer.0.v_cache 0.00000028 0.00027421 - layer.1.k_cache 0.00341635 1.72049872 - layer.1.v_cache 0.00000081 0.00095024 - layer.2.k_cache 0.00112215 0.66797241 - layer.2.v_cache 0.00000111 0.00141210 - layer.3.k_cache 0.00131706 0.79089967 - layer.3.v_cache 0.00000210 0.00233178 - layer.4.k_cache 0.00330853 1.54953035 - layer.4.v_cache 0.00000300 0.00379954 - layer.4.output 0.00017506 0.11644580 - ------------------------------------------------------------------------------------- - TOTAL 0.00297788 1.05317028 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 230648 -BPFP 1.3407 bits/point -EBPFP 2.6815 equivalent bits/point -MSE 1.053170 ----------------------- -------------------------------------------------------- -Time: 3.035s Load: 0.007s, Pack+Encode: 1.782s, Decode+Unpack: 1.245s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0532 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample23-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample23-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-layer4-item1.zst (45/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,928B, BPFP=0.1637 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,916B, BPFP=1.9460 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,016B, BPFP=1.0204 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,012B, BPFP=2.1240 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,504B, BPFP=1.1467 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,396B, BPFP=2.1566 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,068B, BPFP=1.6192 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,736B, BPFP=2.1005 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,992B, BPFP=0.8485 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,436B, BPFP=2.1600 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,064B, BPFP=0.8930 -⌛️ [2/4] FRONTEND: Frontend time: 1.756s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.428s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03098850 9.70387467 - layer.0.v_cache 0.00000027 0.00026342 - layer.1.k_cache 0.00332311 1.27289233 - layer.1.v_cache 0.00000079 0.00092437 - layer.2.k_cache 0.00114598 0.67789260 - layer.2.v_cache 0.00000105 0.00136150 - layer.3.k_cache 0.00130880 0.79316454 - layer.3.v_cache 0.00000204 0.00227669 - layer.4.k_cache 0.00318646 1.48876489 - layer.4.v_cache 0.00000304 0.00382170 - layer.4.output 0.00017235 0.12685492 - ------------------------------------------------------------------------------------- - TOTAL 0.00290353 1.03233260 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 222068 -BPFP 1.3470 bits/point -EBPFP 2.6940 equivalent bits/point -MSE 1.032333 ----------------------- -------------------------------------------------------- -Time: 3.191s Load: 0.007s, Pack+Encode: 1.756s, Decode+Unpack: 1.428s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0323 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample24-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample24-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-layer4-item1.zst (46/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,988B, BPFP=0.1635 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,096B, BPFP=1.8993 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,940B, BPFP=0.9819 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,816B, BPFP=2.0408 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,460B, BPFP=1.2714 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,204B, BPFP=2.0727 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,432B, BPFP=1.5158 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,584B, BPFP=2.0217 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,280B, BPFP=1.0921 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,600B, BPFP=2.1053 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,636B, BPFP=0.8971 -⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.302s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03093621 10.13966129 - layer.0.v_cache 0.00000027 0.00027637 - layer.1.k_cache 0.00341424 1.61175055 - layer.1.v_cache 0.00000077 0.00090150 - layer.2.k_cache 0.00114464 0.68159862 - layer.2.v_cache 0.00000103 0.00133559 - layer.3.k_cache 0.00132643 0.78017482 - layer.3.v_cache 0.00000197 0.00222467 - layer.4.k_cache 0.00330159 1.55188085 - layer.4.v_cache 0.00000302 0.00384615 - layer.4.output 0.00016139 0.10749722 - ------------------------------------------------------------------------------------- - TOTAL 0.00291255 1.08597423 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 228036 -BPFP 1.3395 bits/point -EBPFP 2.6790 equivalent bits/point -MSE 1.085974 ----------------------- -------------------------------------------------------- -Time: 3.021s Load: 0.007s, Pack+Encode: 1.712s, Decode+Unpack: 1.302s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0860 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample27-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample27-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-layer4-item1.zst (47/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,984B, BPFP=0.1632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,048B, BPFP=1.8954 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,500B, BPFP=1.0280 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,864B, BPFP=2.0447 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,312B, BPFP=1.2592 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,212B, BPFP=2.0734 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,408B, BPFP=1.5138 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,624B, BPFP=2.0250 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,768B, BPFP=1.0500 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,628B, BPFP=2.1076 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,412B, BPFP=0.8925 -⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.288s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03089984 9.74789525 - layer.0.v_cache 0.00000027 0.00027590 - layer.1.k_cache 0.00345231 1.64437497 - layer.1.v_cache 0.00000077 0.00090065 - layer.2.k_cache 0.00115012 0.68694675 - layer.2.v_cache 0.00000103 0.00132970 - layer.3.k_cache 0.00131461 0.77274507 - layer.3.v_cache 0.00000196 0.00221285 - layer.4.k_cache 0.00336511 1.57278185 - layer.4.v_cache 0.00000303 0.00384004 - layer.4.output 0.00016181 0.10869256 - ------------------------------------------------------------------------------------- - TOTAL 0.00291688 1.06200523 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 227760 -BPFP 1.3379 bits/point -EBPFP 2.6758 equivalent bits/point -MSE 1.062005 ----------------------- -------------------------------------------------------- -Time: 3.135s Load: 0.007s, Pack+Encode: 1.840s, Decode+Unpack: 1.288s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0620 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample28-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample28-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-layer4-item1.zst (48/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1651 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,088B, BPFP=1.9606 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,956B, BPFP=1.1002 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,240B, BPFP=2.1433 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,932B, BPFP=1.0982 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,440B, BPFP=2.1603 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,576B, BPFP=1.5774 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,748B, BPFP=2.1016 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,764B, BPFP=0.8291 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,392B, BPFP=2.1562 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,528B, BPFP=0.8392 -⌛️ [2/4] FRONTEND: Frontend time: 1.697s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.372s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03058780 10.12583260 - layer.0.v_cache 0.00000028 0.00026909 - layer.1.k_cache 0.00355631 1.51119580 - layer.1.v_cache 0.00000081 0.00095715 - layer.2.k_cache 0.00116497 0.66852731 - layer.2.v_cache 0.00000107 0.00139078 - layer.3.k_cache 0.00134281 0.77374375 - layer.3.v_cache 0.00000206 0.00230326 - layer.4.k_cache 0.00330740 1.50322657 - layer.4.v_cache 0.00000300 0.00374669 - layer.4.output 0.00016528 0.11605250 - ------------------------------------------------------------------------------------- - TOTAL 0.00290197 1.07538593 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 219608 -BPFP 1.3321 bits/point -EBPFP 2.6641 equivalent bits/point -MSE 1.075386 ----------------------- -------------------------------------------------------- -Time: 3.076s Load: 0.006s, Pack+Encode: 1.697s, Decode+Unpack: 1.372s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0754 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample29-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample29-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-layer4-item1.zst (49/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 98, 128) -Output shape: (1, 98, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) -> torch.Size([1, 1, 98, 1024]) - layer.4.output: torch.Size([1, 98, 4096]) -> torch.Size([1, 1, 98, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,052B, BPFP=0.1636 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,080B, BPFP=1.8399 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,264B, BPFP=1.0574 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,232B, BPFP=2.0115 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,068B, BPFP=1.2809 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,496B, BPFP=2.0325 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,276B, BPFP=1.5367 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,740B, BPFP=1.9723 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 14,232B, BPFP=1.1346 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,800B, BPFP=2.0568 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,784B, BPFP=0.8726 -⌛️ [2/4] FRONTEND: Frontend time: 1.762s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.240s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 98, 128]) - layer.0.v_cache: torch.Size([1, 8, 98, 128]) - layer.1.k_cache: torch.Size([1, 8, 98, 128]) - layer.1.v_cache: torch.Size([1, 8, 98, 128]) - layer.2.k_cache: torch.Size([1, 8, 98, 128]) - layer.2.v_cache: torch.Size([1, 8, 98, 128]) - layer.3.k_cache: torch.Size([1, 8, 98, 128]) - layer.3.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.k_cache: torch.Size([1, 8, 98, 128]) - layer.4.v_cache: torch.Size([1, 8, 98, 128]) - layer.4.output: torch.Size([1, 98, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02931073 9.39926583 - layer.0.v_cache 0.00000028 0.00028002 - layer.1.k_cache 0.00346088 1.64799998 - layer.1.v_cache 0.00000093 0.00090468 - layer.2.k_cache 0.00114617 0.69278997 - layer.2.v_cache 0.00000108 0.00134989 - layer.3.k_cache 0.00130775 0.78250558 - layer.3.v_cache 0.00000199 0.00216241 - layer.4.k_cache 0.00329454 1.55608010 - layer.4.v_cache 0.00000306 0.00368035 - layer.4.output 0.00016599 0.11069830 - ------------------------------------------------------------------------------------- - TOTAL 0.00279938 1.03784372 - (elements=1,404,928) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1404928 -Total Bytes 233024 -BPFP 1.3269 bits/point -EBPFP 2.6538 equivalent bits/point -MSE 1.037844 ----------------------- -------------------------------------------------------- -Time: 3.009s Load: 0.006s, Pack+Encode: 1.762s, Decode+Unpack: 1.240s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 98, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 98, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0378 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample30-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample30-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-layer4-item1.zst (50/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,992B, BPFP=0.1621 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,780B, BPFP=1.8538 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,952B, BPFP=1.0540 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,680B, BPFP=2.0085 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,080B, BPFP=1.0645 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,216B, BPFP=2.0521 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,528B, BPFP=1.5078 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,576B, BPFP=2.0000 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,368B, BPFP=0.9251 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,300B, BPFP=2.0589 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,624B, BPFP=0.8265 -⌛️ [2/4] FRONTEND: Frontend time: 1.782s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.384s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03011865 9.73773702 - layer.0.v_cache 0.00000028 0.00026787 - layer.1.k_cache 0.00341880 1.74337387 - layer.1.v_cache 0.00000077 0.00090879 - layer.2.k_cache 0.00113755 0.67878326 - layer.2.v_cache 0.00000103 0.00134979 - layer.3.k_cache 0.00137011 0.78426663 - layer.3.v_cache 0.00000200 0.00227451 - layer.4.k_cache 0.00324136 1.62510395 - layer.4.v_cache 0.00000292 0.00382895 - layer.4.output 0.00022292 0.11872918 - ------------------------------------------------------------------------------------- - TOTAL 0.00287037 1.07520081 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 221096 -BPFP 1.2852 bits/point -EBPFP 2.5704 equivalent bits/point -MSE 1.075201 ----------------------- -------------------------------------------------------- -Time: 3.174s Load: 0.007s, Pack+Encode: 1.782s, Decode+Unpack: 1.384s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0752 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample31-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample31-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-layer4-item1.zst (51/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,952B, BPFP=0.1622 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,804B, BPFP=1.8953 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,168B, BPFP=1.0944 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,996B, BPFP=2.0775 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,736B, BPFP=1.3910 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,280B, BPFP=2.1011 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,576B, BPFP=1.5439 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,624B, BPFP=2.0465 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,244B, BPFP=0.8514 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,488B, BPFP=2.1184 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,256B, BPFP=0.8157 -⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.282s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03206433 9.07020212 - layer.0.v_cache 0.00000027 0.00027073 - layer.1.k_cache 0.00350816 1.63307320 - layer.1.v_cache 0.00000080 0.00094043 - layer.2.k_cache 0.00113886 0.66112218 - layer.2.v_cache 0.00000104 0.00133426 - layer.3.k_cache 0.00133101 0.78054891 - layer.3.v_cache 0.00000202 0.00226375 - layer.4.k_cache 0.00327566 1.43884699 - layer.4.v_cache 0.00000298 0.00376643 - layer.4.output 0.00018563 0.11372944 - ------------------------------------------------------------------------------------- - TOTAL 0.00300483 1.00337762 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 223124 -BPFP 1.3246 bits/point -EBPFP 2.6492 equivalent bits/point -MSE 1.003378 ----------------------- -------------------------------------------------------- -Time: 3.000s Load: 0.007s, Pack+Encode: 1.711s, Decode+Unpack: 1.282s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0034 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample32-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample32-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-layer4-item1.zst (52/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,048B, BPFP=0.1649 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,760B, BPFP=1.8331 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,060B, BPFP=0.9713 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,056B, BPFP=2.0180 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,668B, BPFP=1.1814 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,388B, BPFP=2.0448 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,468B, BPFP=1.4874 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,616B, BPFP=1.9826 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,884B, BPFP=0.8766 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,416B, BPFP=2.0470 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,124B, BPFP=0.8885 -⌛️ [2/4] FRONTEND: Frontend time: 1.808s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.258s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.02997011 9.65613548 - layer.0.v_cache 0.00000027 0.00026070 - layer.1.k_cache 0.00344926 1.49261663 - layer.1.v_cache 0.00000085 0.00093213 - layer.2.k_cache 0.00114066 0.66021469 - layer.2.v_cache 0.00000106 0.00138597 - layer.3.k_cache 0.00133063 0.79561646 - layer.3.v_cache 0.00000201 0.00225557 - layer.4.k_cache 0.00322564 1.47907800 - layer.4.v_cache 0.00000314 0.00387705 - layer.4.output 0.00016478 0.11417204 - ------------------------------------------------------------------------------------- - TOTAL 0.00284162 1.03921863 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 225488 -BPFP 1.2972 bits/point -EBPFP 2.5944 equivalent bits/point -MSE 1.039219 ----------------------- -------------------------------------------------------- -Time: 3.073s Load: 0.007s, Pack+Encode: 1.808s, Decode+Unpack: 1.258s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0392 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample33-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample33-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-layer4-item1.zst (53/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,984B, BPFP=0.1632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,196B, BPFP=1.9076 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,540B, BPFP=1.1135 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,164B, BPFP=2.0694 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,540B, BPFP=1.3602 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,576B, BPFP=2.1033 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,432B, BPFP=1.5158 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,812B, BPFP=2.0405 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,536B, BPFP=0.9487 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,688B, BPFP=2.1125 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,968B, BPFP=0.8217 -⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.403s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03212489 9.48308812 - layer.0.v_cache 0.00000028 0.00027575 - layer.1.k_cache 0.00341497 1.42942633 - layer.1.v_cache 0.00000078 0.00091852 - layer.2.k_cache 0.00114913 0.67001616 - layer.2.v_cache 0.00000106 0.00137250 - layer.3.k_cache 0.00131997 0.78575215 - layer.3.v_cache 0.00000202 0.00219687 - layer.4.k_cache 0.00329406 1.46884830 - layer.4.v_cache 0.00000310 0.00382709 - layer.4.output 0.00015936 0.10605878 - ------------------------------------------------------------------------------------- - TOTAL 0.00299627 1.01928264 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 226436 -BPFP 1.3301 bits/point -EBPFP 2.6602 equivalent bits/point -MSE 1.019283 ----------------------- -------------------------------------------------------- -Time: 3.127s Load: 0.008s, Pack+Encode: 1.716s, Decode+Unpack: 1.403s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0193 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample35-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample35-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-layer4-item1.zst (54/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1644 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,596B, BPFP=1.9188 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,304B, BPFP=1.0448 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,716B, BPFP=2.0988 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,272B, BPFP=1.2120 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,252B, BPFP=2.1444 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,804B, BPFP=1.5968 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,568B, BPFP=2.0863 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,716B, BPFP=0.9949 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,048B, BPFP=2.1270 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,072B, BPFP=0.8932 -⌛️ [2/4] FRONTEND: Frontend time: 1.742s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.250s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03130991 9.85292982 - layer.0.v_cache 0.00000027 0.00027009 - layer.1.k_cache 0.00358715 1.52138685 - layer.1.v_cache 0.00000080 0.00094902 - layer.2.k_cache 0.00113844 0.66814551 - layer.2.v_cache 0.00000105 0.00140708 - layer.3.k_cache 0.00133437 0.77137624 - layer.3.v_cache 0.00000205 0.00235453 - layer.4.k_cache 0.00328807 1.58546945 - layer.4.v_cache 0.00000299 0.00380443 - layer.4.output 0.00019541 0.11499599 - ------------------------------------------------------------------------------------- - TOTAL 0.00296048 1.06200550 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 223284 -BPFP 1.3544 bits/point -EBPFP 2.7087 equivalent bits/point -MSE 1.062005 ----------------------- -------------------------------------------------------- -Time: 2.998s Load: 0.007s, Pack+Encode: 1.742s, Decode+Unpack: 1.250s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0620 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample36-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample36-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-layer4-item1.zst (55/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,980B, BPFP=0.1628 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,020B, BPFP=1.8931 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,132B, BPFP=1.0799 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,240B, BPFP=2.0757 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,524B, BPFP=1.2766 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,524B, BPFP=2.0990 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,752B, BPFP=1.4599 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,892B, BPFP=2.0470 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,840B, BPFP=1.0559 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,656B, BPFP=2.1099 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,744B, BPFP=0.8377 -⌛️ [2/4] FRONTEND: Frontend time: 1.797s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.357s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03212384 9.34819336 - layer.0.v_cache 0.00000028 0.00027105 - layer.1.k_cache 0.00332589 1.60385517 - layer.1.v_cache 0.00000080 0.00093394 - layer.2.k_cache 0.00116734 0.70142999 - layer.2.v_cache 0.00000106 0.00135696 - layer.3.k_cache 0.00135603 0.78853744 - layer.3.v_cache 0.00000201 0.00224294 - layer.4.k_cache 0.00321713 1.54221336 - layer.4.v_cache 0.00000315 0.00386336 - layer.4.output 0.00019059 0.12168273 - ------------------------------------------------------------------------------------- - TOTAL 0.00299713 1.03425918 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 226304 -BPFP 1.3293 bits/point -EBPFP 2.6586 equivalent bits/point -MSE 1.034259 ----------------------- -------------------------------------------------------- -Time: 3.161s Load: 0.006s, Pack+Encode: 1.797s, Decode+Unpack: 1.357s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0343 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample37-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample37-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-layer4-item1.zst (56/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1626 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,908B, BPFP=1.9244 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,688B, BPFP=1.0659 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,004B, BPFP=2.1005 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,892B, BPFP=1.3350 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,428B, BPFP=2.1361 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,276B, BPFP=1.5353 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,624B, BPFP=2.0685 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,024B, BPFP=1.0101 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,396B, BPFP=2.1334 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,108B, BPFP=0.9053 -⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.353s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03063477 9.97495458 - layer.0.v_cache 0.00000027 0.00026624 - layer.1.k_cache 0.00346550 1.42312753 - layer.1.v_cache 0.00000078 0.00093661 - layer.2.k_cache 0.00113041 0.67244536 - layer.2.v_cache 0.00000112 0.00142818 - layer.3.k_cache 0.00132028 0.78566463 - layer.3.v_cache 0.00000203 0.00235903 - layer.4.k_cache 0.00322248 1.58413992 - layer.4.v_cache 0.00000299 0.00387905 - layer.4.output 0.00017253 0.11533827 - ------------------------------------------------------------------------------------- - TOTAL 0.00289077 1.06503959 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 227284 -BPFP 1.3638 bits/point -EBPFP 2.7276 equivalent bits/point -MSE 1.065040 ----------------------- -------------------------------------------------------- -Time: 3.070s Load: 0.007s, Pack+Encode: 1.710s, Decode+Unpack: 1.353s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0650 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample39-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample39-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-layer4-item1.zst (57/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 100, 128) -Output shape: (1, 100, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) -> torch.Size([1, 1, 100, 1024]) - layer.4.output: torch.Size([1, 100, 4096]) -> torch.Size([1, 1, 100, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,124B, BPFP=0.1659 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,544B, BPFP=1.8394 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,100B, BPFP=1.0234 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,492B, BPFP=1.9916 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 17,016B, BPFP=1.3294 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,796B, BPFP=2.0153 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,396B, BPFP=1.4372 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,012B, BPFP=1.9541 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,596B, BPFP=1.0622 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 26,100B, BPFP=2.0391 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 59,108B, BPFP=1.1545 -⌛️ [2/4] FRONTEND: Frontend time: 1.820s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.238s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 100, 128]) - layer.0.v_cache: torch.Size([1, 8, 100, 128]) - layer.1.k_cache: torch.Size([1, 8, 100, 128]) - layer.1.v_cache: torch.Size([1, 8, 100, 128]) - layer.2.k_cache: torch.Size([1, 8, 100, 128]) - layer.2.v_cache: torch.Size([1, 8, 100, 128]) - layer.3.k_cache: torch.Size([1, 8, 100, 128]) - layer.3.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.k_cache: torch.Size([1, 8, 100, 128]) - layer.4.v_cache: torch.Size([1, 8, 100, 128]) - layer.4.output: torch.Size([1, 100, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03019520 10.12390137 - layer.0.v_cache 0.00000027 0.00026755 - layer.1.k_cache 0.00345483 1.67382965 - layer.1.v_cache 0.00000092 0.00093599 - layer.2.k_cache 0.00115188 0.70528824 - layer.2.v_cache 0.00000108 0.00134976 - layer.3.k_cache 0.00131981 0.78557358 - layer.3.v_cache 0.00000202 0.00227813 - layer.4.k_cache 0.00326509 1.59807159 - layer.4.v_cache 0.00000302 0.00379545 - layer.4.output 0.00016666 0.12323204 - ------------------------------------------------------------------------------------- - TOTAL 0.00286148 1.09915853 - (elements=1,433,600) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1433600 -Total Bytes 249284 -BPFP 1.3911 bits/point -EBPFP 2.7822 equivalent bits/point -MSE 1.099159 ----------------------- -------------------------------------------------------- -Time: 3.066s Load: 0.008s, Pack+Encode: 1.820s, Decode+Unpack: 1.238s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 100, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 100, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0992 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample4-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample4-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-layer4-item1.zst (58/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,904B, BPFP=0.1671 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,208B, BPFP=2.0372 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,616B, BPFP=1.1074 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,404B, BPFP=2.2300 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,776B, BPFP=1.1215 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 26,136B, BPFP=2.2942 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,148B, BPFP=1.6808 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,120B, BPFP=2.2051 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,612B, BPFP=0.8438 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,800B, BPFP=2.2647 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,104B, BPFP=1.0118 -⌛️ [2/4] FRONTEND: Frontend time: 1.723s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.410s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03166650 9.59141969 - layer.0.v_cache 0.00000027 0.00027059 - layer.1.k_cache 0.00343321 1.41842266 - layer.1.v_cache 0.00000094 0.00097791 - layer.2.k_cache 0.00113193 0.66153186 - layer.2.v_cache 0.00000110 0.00142024 - layer.3.k_cache 0.00132786 0.78751768 - layer.3.v_cache 0.00000214 0.00239975 - layer.4.k_cache 0.00323286 1.41961567 - layer.4.v_cache 0.00000312 0.00386265 - layer.4.output 0.00018081 0.11567072 - ------------------------------------------------------------------------------------- - TOTAL 0.00296594 1.02500868 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 227828 -BPFP 1.4285 bits/point -EBPFP 2.8570 equivalent bits/point -MSE 1.025009 ----------------------- -------------------------------------------------------- -Time: 3.140s Load: 0.006s, Pack+Encode: 1.723s, Decode+Unpack: 1.410s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0250 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample40-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample40-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-layer4-item1.zst (59/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1650 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,868B, BPFP=1.9210 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,588B, BPFP=1.1415 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,076B, BPFP=2.1065 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,244B, BPFP=1.2806 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,588B, BPFP=2.1495 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,732B, BPFP=1.5736 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,672B, BPFP=2.0726 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 14,088B, BPFP=1.1835 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,576B, BPFP=2.1485 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 47,628B, BPFP=1.0003 -⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.222s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03068307 10.00834016 - layer.0.v_cache 0.00000028 0.00027355 - layer.1.k_cache 0.00346304 1.55821737 - layer.1.v_cache 0.00000079 0.00095303 - layer.2.k_cache 0.00114171 0.68508665 - layer.2.v_cache 0.00000107 0.00139755 - layer.3.k_cache 0.00131636 0.77997950 - layer.3.v_cache 0.00000209 0.00234444 - layer.4.k_cache 0.00323274 1.60663662 - layer.4.v_cache 0.00000311 0.00382373 - layer.4.output 0.00016873 0.10408480 - ------------------------------------------------------------------------------------- - TOTAL 0.00289423 1.07595656 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 235024 -BPFP 1.4102 bits/point -EBPFP 2.8205 equivalent bits/point -MSE 1.075957 ----------------------- -------------------------------------------------------- -Time: 2.959s Load: 0.009s, Pack+Encode: 1.729s, Decode+Unpack: 1.222s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0760 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample41-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample41-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-layer4-item1.zst (60/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,912B, BPFP=0.1660 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,172B, BPFP=2.0115 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,460B, BPFP=1.0816 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,036B, BPFP=2.1733 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,064B, BPFP=1.1340 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,416B, BPFP=2.2062 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,452B, BPFP=1.6017 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,892B, BPFP=2.1608 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,440B, BPFP=0.8194 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,688B, BPFP=2.2299 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,164B, BPFP=0.9150 -⌛️ [2/4] FRONTEND: Frontend time: 1.821s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.319s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03151998 9.52444255 - layer.0.v_cache 0.00000027 0.00026755 - layer.1.k_cache 0.00334378 1.55381012 - layer.1.v_cache 0.00000080 0.00091366 - layer.2.k_cache 0.00114309 0.68054589 - layer.2.v_cache 0.00000106 0.00134751 - layer.3.k_cache 0.00133556 0.77607142 - layer.3.v_cache 0.00000205 0.00224464 - layer.4.k_cache 0.00317137 1.47868669 - layer.4.v_cache 0.00000297 0.00370081 - layer.4.output 0.00023356 0.12403313 - ------------------------------------------------------------------------------------- - TOTAL 0.00296108 1.03701167 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 221696 -BPFP 1.3746 bits/point -EBPFP 2.7492 equivalent bits/point -MSE 1.037012 ----------------------- -------------------------------------------------------- -Time: 3.147s Load: 0.007s, Pack+Encode: 1.821s, Decode+Unpack: 1.319s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0370 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample43-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample43-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-layer4-item1.zst (61/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,880B, BPFP=0.1688 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,368B, BPFP=2.0984 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,880B, BPFP=1.1566 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,512B, BPFP=2.2909 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,420B, BPFP=1.1153 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,772B, BPFP=2.3143 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,092B, BPFP=1.6246 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,136B, BPFP=2.2572 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,404B, BPFP=0.9343 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,968B, BPFP=2.3319 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,664B, BPFP=0.9353 -⌛️ [2/4] FRONTEND: Frontend time: 1.679s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.396s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03168404 9.76568954 - layer.0.v_cache 0.00000027 0.00027275 - layer.1.k_cache 0.00359645 1.43613004 - layer.1.v_cache 0.00000085 0.00091958 - layer.2.k_cache 0.00115081 0.66974184 - layer.2.v_cache 0.00000106 0.00135715 - layer.3.k_cache 0.00133256 0.76614713 - layer.3.v_cache 0.00000205 0.00226531 - layer.4.k_cache 0.00340755 1.49383194 - layer.4.v_cache 0.00000299 0.00373143 - layer.4.output 0.00018399 0.12336845 - ------------------------------------------------------------------------------------- - TOTAL 0.00299390 1.04525432 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 223096 -BPFP 1.4310 bits/point -EBPFP 2.8620 equivalent bits/point -MSE 1.045254 ----------------------- -------------------------------------------------------- -Time: 3.082s Load: 0.007s, Pack+Encode: 1.679s, Decode+Unpack: 1.396s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0453 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample44-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample44-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-layer4-item1.zst (62/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,972B, BPFP=0.1605 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,748B, BPFP=1.8512 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,184B, BPFP=1.1543 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,644B, BPFP=2.0055 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,168B, BPFP=1.0716 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 24,956B, BPFP=2.0309 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,744B, BPFP=1.5254 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,492B, BPFP=1.9932 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,624B, BPFP=1.1087 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,252B, BPFP=2.0550 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,016B, BPFP=0.8141 -⌛️ [2/4] FRONTEND: Frontend time: 1.800s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.251s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03074032 9.19006666 - layer.0.v_cache 0.00000027 0.00026665 - layer.1.k_cache 0.00337797 1.72067070 - layer.1.v_cache 0.00000078 0.00091703 - layer.2.k_cache 0.00113475 0.69610755 - layer.2.v_cache 0.00000103 0.00134071 - layer.3.k_cache 0.00138182 0.79157329 - layer.3.v_cache 0.00000203 0.00227979 - layer.4.k_cache 0.00323877 1.56461112 - layer.4.v_cache 0.00000295 0.00380294 - layer.4.output 0.00017303 0.11127399 - ------------------------------------------------------------------------------------- - TOTAL 0.00289806 1.02976660 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 223800 -BPFP 1.3009 bits/point -EBPFP 2.6018 equivalent bits/point -MSE 1.029767 ----------------------- -------------------------------------------------------- -Time: 3.059s Load: 0.007s, Pack+Encode: 1.800s, Decode+Unpack: 1.251s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0298 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample46-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample46-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-layer4-item1.zst (63/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,884B, BPFP=0.1692 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,368B, BPFP=2.0984 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,332B, BPFP=1.1074 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,536B, BPFP=2.2931 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,684B, BPFP=1.2288 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,692B, BPFP=2.3071 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,536B, BPFP=1.7543 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,164B, BPFP=2.2597 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,296B, BPFP=0.9246 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,912B, BPFP=2.3269 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,452B, BPFP=0.9081 -⌛️ [2/4] FRONTEND: Frontend time: 1.752s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.391s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03178630 9.85369522 - layer.0.v_cache 0.00000027 0.00027150 - layer.1.k_cache 0.00361903 1.40903298 - layer.1.v_cache 0.00000083 0.00091778 - layer.2.k_cache 0.00113732 0.66247094 - layer.2.v_cache 0.00000106 0.00136578 - layer.3.k_cache 0.00132704 0.77332543 - layer.3.v_cache 0.00000206 0.00226647 - layer.4.k_cache 0.00332258 1.52206053 - layer.4.v_cache 0.00000303 0.00372006 - layer.4.output 0.00018372 0.12441591 - ------------------------------------------------------------------------------------- - TOTAL 0.00299531 1.05191360 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 223856 -BPFP 1.4359 bits/point -EBPFP 2.8717 equivalent bits/point -MSE 1.051914 ----------------------- -------------------------------------------------------- -Time: 3.150s Load: 0.007s, Pack+Encode: 1.752s, Decode+Unpack: 1.391s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0519 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample47-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample47-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-layer4-item1.zst (64/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,056B, BPFP=0.1656 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,284B, BPFP=1.8753 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,344B, BPFP=1.0747 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,416B, BPFP=2.0470 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,368B, BPFP=1.1572 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,568B, BPFP=2.0593 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,572B, BPFP=1.5764 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,056B, BPFP=2.0180 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,492B, BPFP=0.9256 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,936B, BPFP=2.0889 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 49,996B, BPFP=1.0067 -⌛️ [2/4] FRONTEND: Frontend time: 1.703s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.307s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03078698 9.87107110 - layer.0.v_cache 0.00000027 0.00026550 - layer.1.k_cache 0.00338495 1.75952746 - layer.1.v_cache 0.00000090 0.00092033 - layer.2.k_cache 0.00114245 0.67345877 - layer.2.v_cache 0.00000106 0.00136834 - layer.3.k_cache 0.00133426 0.78826763 - layer.3.v_cache 0.00000206 0.00229270 - layer.4.k_cache 0.00327800 1.52244804 - layer.4.v_cache 0.00000314 0.00377537 - layer.4.output 0.00016719 0.10389063 - ------------------------------------------------------------------------------------- - TOTAL 0.00290020 1.07421127 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 236088 -BPFP 1.3582 bits/point -EBPFP 2.7164 equivalent bits/point -MSE 1.074211 ----------------------- -------------------------------------------------------- -Time: 3.017s Load: 0.007s, Pack+Encode: 1.703s, Decode+Unpack: 1.307s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0742 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample48-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample48-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-layer4-item1.zst (65/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,932B, BPFP=0.1641 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,912B, BPFP=1.9457 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,444B, BPFP=0.9718 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,032B, BPFP=2.1257 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,768B, BPFP=1.1692 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,356B, BPFP=2.1532 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,076B, BPFP=1.6199 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,496B, BPFP=2.0802 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,184B, BPFP=0.8648 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,284B, BPFP=2.1471 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,032B, BPFP=0.8286 -⌛️ [2/4] FRONTEND: Frontend time: 1.811s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.297s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03091334 9.86517268 - layer.0.v_cache 0.00000027 0.00026671 - layer.1.k_cache 0.00334237 1.55482085 - layer.1.v_cache 0.00000084 0.00092562 - layer.2.k_cache 0.00114123 0.67269139 - layer.2.v_cache 0.00000105 0.00136042 - layer.3.k_cache 0.00132733 0.79158775 - layer.3.v_cache 0.00000206 0.00224553 - layer.4.k_cache 0.00332026 1.45748503 - layer.4.v_cache 0.00000301 0.00384277 - layer.4.output 0.00018747 0.11645059 - ------------------------------------------------------------------------------------- - TOTAL 0.00291440 1.05830008 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 218516 -BPFP 1.3254 bits/point -EBPFP 2.6509 equivalent bits/point -MSE 1.058300 ----------------------- -------------------------------------------------------- -Time: 3.114s Load: 0.006s, Pack+Encode: 1.811s, Decode+Unpack: 1.297s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0583 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample50-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample50-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-layer4-item1.zst (66/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,976B, BPFP=0.1642 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,964B, BPFP=1.9086 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,444B, BPFP=1.0342 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,108B, BPFP=2.0868 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,596B, BPFP=1.3793 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,444B, BPFP=2.1147 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,504B, BPFP=1.5379 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,760B, BPFP=2.0578 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,660B, BPFP=0.8860 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,596B, BPFP=2.1273 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,776B, BPFP=0.8680 -⌛️ [2/4] FRONTEND: Frontend time: 1.697s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.423s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03107318 9.23251538 - layer.0.v_cache 0.00000027 0.00026814 - layer.1.k_cache 0.00348481 1.54036339 - layer.1.v_cache 0.00000081 0.00092785 - layer.2.k_cache 0.00116411 0.67188283 - layer.2.v_cache 0.00000106 0.00136597 - layer.3.k_cache 0.00133018 0.77901248 - layer.3.v_cache 0.00000204 0.00227673 - layer.4.k_cache 0.00337179 1.49631711 - layer.4.v_cache 0.00000293 0.00368418 - layer.4.output 0.00016660 0.10652131 - ------------------------------------------------------------------------------------- - TOTAL 0.00293554 1.01104995 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 225828 -BPFP 1.3406 bits/point -EBPFP 2.6813 equivalent bits/point -MSE 1.011050 ----------------------- -------------------------------------------------------- -Time: 3.126s Load: 0.006s, Pack+Encode: 1.697s, Decode+Unpack: 1.423s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0110 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample51-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample51-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-layer4-item1.zst (67/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.009s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,864B, BPFP=0.1674 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,256B, BPFP=2.0884 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,924B, BPFP=1.0708 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,384B, BPFP=2.2795 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,192B, BPFP=1.1846 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,528B, BPFP=2.2924 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 16,832B, BPFP=1.5115 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,768B, BPFP=2.2241 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,348B, BPFP=0.8394 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,744B, BPFP=2.3118 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 37,680B, BPFP=0.8459 -⌛️ [2/4] FRONTEND: Frontend time: 1.786s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.240s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03151683 9.81529359 - layer.0.v_cache 0.00000028 0.00027458 - layer.1.k_cache 0.00361716 1.41268114 - layer.1.v_cache 0.00000078 0.00090768 - layer.2.k_cache 0.00116662 0.66704103 - layer.2.v_cache 0.00000103 0.00133616 - layer.3.k_cache 0.00135197 0.74582453 - layer.3.v_cache 0.00000203 0.00219970 - layer.4.k_cache 0.00327724 1.55076862 - layer.4.v_cache 0.00000296 0.00368747 - layer.4.output 0.00017041 0.11496564 - ------------------------------------------------------------------------------------- - TOTAL 0.00297275 1.04713407 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 215520 -BPFP 1.3824 bits/point -EBPFP 2.7648 equivalent bits/point -MSE 1.047134 ----------------------- -------------------------------------------------------- -Time: 3.035s Load: 0.009s, Pack+Encode: 1.786s, Decode+Unpack: 1.240s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0471 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample53-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample53-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-layer4-item1.zst (68/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 95, 128) -Output shape: (1, 95, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) -> torch.Size([1, 1, 95, 1024]) - layer.4.output: torch.Size([1, 95, 4096]) -> torch.Size([1, 1, 95, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,016B, BPFP=0.1658 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,732B, BPFP=1.8694 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,968B, BPFP=1.0664 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,980B, BPFP=2.0543 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,132B, BPFP=1.1622 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,332B, BPFP=2.0832 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,092B, BPFP=1.4878 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,512B, BPFP=2.0158 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,284B, BPFP=0.8457 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,716B, BPFP=2.1148 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,592B, BPFP=0.8345 -⌛️ [2/4] FRONTEND: Frontend time: 1.768s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.434s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 95, 128]) - layer.0.v_cache: torch.Size([1, 8, 95, 128]) - layer.1.k_cache: torch.Size([1, 8, 95, 128]) - layer.1.v_cache: torch.Size([1, 8, 95, 128]) - layer.2.k_cache: torch.Size([1, 8, 95, 128]) - layer.2.v_cache: torch.Size([1, 8, 95, 128]) - layer.3.k_cache: torch.Size([1, 8, 95, 128]) - layer.3.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.k_cache: torch.Size([1, 8, 95, 128]) - layer.4.v_cache: torch.Size([1, 8, 95, 128]) - layer.4.output: torch.Size([1, 95, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03080931 8.85226023 - layer.0.v_cache 0.00000028 0.00027228 - layer.1.k_cache 0.00333058 1.59407236 - layer.1.v_cache 0.00000087 0.00092978 - layer.2.k_cache 0.00113656 0.65631200 - layer.2.v_cache 0.00000104 0.00137799 - layer.3.k_cache 0.00135899 0.77732600 - layer.3.v_cache 0.00000196 0.00221321 - layer.4.k_cache 0.00332312 1.49257347 - layer.4.v_cache 0.00000307 0.00389912 - layer.4.output 0.00019584 0.11010411 - ------------------------------------------------------------------------------------- - TOTAL 0.00291065 0.98726092 - (elements=1,361,920) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1361920 -Total Bytes 221356 -BPFP 1.3003 bits/point -EBPFP 2.6005 equivalent bits/point -MSE 0.987261 ----------------------- -------------------------------------------------------- -Time: 3.208s Load: 0.006s, Pack+Encode: 1.768s, Decode+Unpack: 1.434s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 95, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 95, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 0.9873 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample54-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample54-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-layer4-item1.zst (69/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1685 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,460B, BPFP=2.1067 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 10,480B, BPFP=0.9411 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,644B, BPFP=2.3028 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,308B, BPFP=1.1052 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,656B, BPFP=2.3039 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 16,448B, BPFP=1.4770 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,044B, BPFP=2.2489 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,872B, BPFP=0.8865 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,600B, BPFP=2.2989 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,832B, BPFP=0.9840 -⌛️ [2/4] FRONTEND: Frontend time: 1.701s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.301s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03144652 9.67377972 - layer.0.v_cache 0.00000028 0.00027274 - layer.1.k_cache 0.00348072 1.32463837 - layer.1.v_cache 0.00000083 0.00093919 - layer.2.k_cache 0.00116226 0.64237266 - layer.2.v_cache 0.00000106 0.00138700 - layer.3.k_cache 0.00133044 0.76438588 - layer.3.v_cache 0.00000208 0.00225975 - layer.4.k_cache 0.00328583 1.51188291 - layer.4.v_cache 0.00000293 0.00366486 - layer.4.output 0.00019032 0.12692036 - ------------------------------------------------------------------------------------- - TOTAL 0.00296244 1.03094747 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 220220 -BPFP 1.4125 bits/point -EBPFP 2.8251 equivalent bits/point -MSE 1.030947 ----------------------- -------------------------------------------------------- -Time: 3.008s Load: 0.006s, Pack+Encode: 1.701s, Decode+Unpack: 1.301s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0309 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample55-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample55-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst (70/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,932B, BPFP=0.1659 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,940B, BPFP=1.9694 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,884B, BPFP=1.0203 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,192B, BPFP=2.1628 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,292B, BPFP=1.1411 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,592B, BPFP=2.1971 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,268B, BPFP=1.5683 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,732B, BPFP=2.1233 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,776B, BPFP=0.8393 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,528B, BPFP=2.1916 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,588B, BPFP=0.8711 -⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.251s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03082078 9.55348826 - layer.0.v_cache 0.00000027 0.00027470 - layer.1.k_cache 0.00345050 1.39929627 - layer.1.v_cache 0.00000081 0.00093940 - layer.2.k_cache 0.00112876 0.70279886 - layer.2.v_cache 0.00000108 0.00137353 - layer.3.k_cache 0.00131171 0.77794085 - layer.3.v_cache 0.00000203 0.00227283 - layer.4.k_cache 0.00324214 1.45126594 - layer.4.v_cache 0.00000302 0.00372103 - layer.4.output 0.00017376 0.11164103 - ------------------------------------------------------------------------------------- - TOTAL 0.00290401 1.02428113 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 219724 -BPFP 1.3474 bits/point -EBPFP 2.6948 equivalent bits/point -MSE 1.024281 ----------------------- -------------------------------------------------------- -Time: 3.090s Load: 0.006s, Pack+Encode: 1.833s, Decode+Unpack: 1.251s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0243 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample59-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample59-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-layer4-item1.zst (71/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,872B, BPFP=0.1681 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,188B, BPFP=2.0823 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,964B, BPFP=1.0744 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,496B, BPFP=2.2895 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,808B, BPFP=1.2399 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,720B, BPFP=2.3096 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,672B, BPFP=1.5869 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,024B, BPFP=2.2471 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,228B, BPFP=0.9185 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,860B, BPFP=2.3222 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,228B, BPFP=0.8582 -⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.423s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03212153 10.19901951 - layer.0.v_cache 0.00000027 0.00026529 - layer.1.k_cache 0.00343529 1.51863800 - layer.1.v_cache 0.00000080 0.00090218 - layer.2.k_cache 0.00113462 0.65618265 - layer.2.v_cache 0.00000107 0.00135623 - layer.3.k_cache 0.00129755 0.74790043 - layer.3.v_cache 0.00000202 0.00221947 - layer.4.k_cache 0.00322320 1.52021842 - layer.4.v_cache 0.00000301 0.00371462 - layer.4.output 0.00017088 0.11762284 - ------------------------------------------------------------------------------------- - TOTAL 0.00299306 1.08006487 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 219060 -BPFP 1.4051 bits/point -EBPFP 2.8102 equivalent bits/point -MSE 1.080065 ----------------------- -------------------------------------------------------- -Time: 3.159s Load: 0.006s, Pack+Encode: 1.730s, Decode+Unpack: 1.423s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0801 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample60-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample60-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-layer4-item1.zst (72/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,892B, BPFP=0.1642 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,868B, BPFP=1.9851 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,740B, BPFP=1.1059 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,336B, BPFP=2.1993 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,892B, BPFP=1.1191 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,512B, BPFP=2.2146 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,724B, BPFP=1.6253 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,028B, BPFP=2.1726 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,968B, BPFP=0.8653 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,620B, BPFP=2.2240 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 38,968B, BPFP=0.8457 -⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.243s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03068415 9.68293457 - layer.0.v_cache 0.00000027 0.00026961 - layer.1.k_cache 0.00335417 1.55900913 - layer.1.v_cache 0.00000084 0.00091436 - layer.2.k_cache 0.00111816 0.66869028 - layer.2.v_cache 0.00000105 0.00135109 - layer.3.k_cache 0.00134433 0.77950490 - layer.3.v_cache 0.00000196 0.00220178 - layer.4.k_cache 0.00329790 1.47540453 - layer.4.v_cache 0.00000303 0.00368907 - layer.4.output 0.00016086 0.10947108 - ------------------------------------------------------------------------------------- - TOTAL 0.00288924 1.04370383 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 219548 -BPFP 1.3613 bits/point -EBPFP 2.7226 equivalent bits/point -MSE 1.043704 ----------------------- -------------------------------------------------------- -Time: 2.984s Load: 0.006s, Pack+Encode: 1.735s, Decode+Unpack: 1.243s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0437 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample62-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample62-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-layer4-item1.zst (73/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,896B, BPFP=0.1646 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,112B, BPFP=2.0063 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,496B, BPFP=1.0847 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,224B, BPFP=2.1896 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,204B, BPFP=1.1462 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,432B, BPFP=2.2076 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,924B, BPFP=1.6427 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,808B, BPFP=2.1535 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,368B, BPFP=0.8132 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,524B, BPFP=2.2156 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,540B, BPFP=0.8581 -⌛️ [2/4] FRONTEND: Frontend time: 1.813s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.344s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03145831 10.07800361 - layer.0.v_cache 0.00000027 0.00027218 - layer.1.k_cache 0.00344612 1.46011353 - layer.1.v_cache 0.00000080 0.00093608 - layer.2.k_cache 0.00114283 0.66085773 - layer.2.v_cache 0.00000108 0.00136201 - layer.3.k_cache 0.00133751 0.78934462 - layer.3.v_cache 0.00000207 0.00226131 - layer.4.k_cache 0.00330099 1.46531474 - layer.4.v_cache 0.00000295 0.00372658 - layer.4.output 0.00016962 0.11197233 - ------------------------------------------------------------------------------------- - TOTAL 0.00295510 1.06500584 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 219528 -BPFP 1.3612 bits/point -EBPFP 2.7223 equivalent bits/point -MSE 1.065006 ----------------------- -------------------------------------------------------- -Time: 3.163s Load: 0.006s, Pack+Encode: 1.813s, Decode+Unpack: 1.344s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0650 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample63-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample63-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst (74/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,924B, BPFP=0.1652 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,916B, BPFP=1.9674 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,252B, BPFP=1.1377 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,020B, BPFP=2.1480 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,900B, BPFP=1.1075 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,372B, BPFP=2.1782 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,812B, BPFP=1.6150 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,644B, BPFP=2.1157 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,660B, BPFP=0.8293 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,524B, BPFP=2.1913 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,228B, BPFP=0.9278 -⌛️ [2/4] FRONTEND: Frontend time: 1.705s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.357s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03132481 9.85455926 - layer.0.v_cache 0.00000027 0.00027228 - layer.1.k_cache 0.00345143 1.55269849 - layer.1.v_cache 0.00000080 0.00092066 - layer.2.k_cache 0.00115294 0.66805091 - layer.2.v_cache 0.00000107 0.00137038 - layer.3.k_cache 0.00135241 0.78378413 - layer.3.v_cache 0.00000203 0.00224704 - layer.4.k_cache 0.00326810 1.51418380 - layer.4.v_cache 0.00000305 0.00375153 - layer.4.output 0.00025628 0.12225046 - ------------------------------------------------------------------------------------- - TOTAL 0.00297015 1.06220288 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 223252 -BPFP 1.3690 bits/point -EBPFP 2.7381 equivalent bits/point -MSE 1.062203 ----------------------- -------------------------------------------------------- -Time: 3.069s Load: 0.007s, Pack+Encode: 1.705s, Decode+Unpack: 1.357s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0622 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample65-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample65-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-layer4-item1.zst (75/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1650 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,936B, BPFP=1.9267 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,816B, BPFP=1.1606 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,896B, BPFP=2.0914 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,936B, BPFP=1.4227 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,520B, BPFP=2.1438 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,472B, BPFP=1.5517 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,736B, BPFP=2.0780 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,792B, BPFP=1.0746 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,472B, BPFP=2.1398 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,628B, BPFP=0.9372 -⌛️ [2/4] FRONTEND: Frontend time: 1.823s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.231s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03759774 10.05340839 - layer.0.v_cache 0.00000028 0.00026926 - layer.1.k_cache 0.00332062 1.50304667 - layer.1.v_cache 0.00000078 0.00093759 - layer.2.k_cache 0.00112828 0.69289193 - layer.2.v_cache 0.00000108 0.00144218 - layer.3.k_cache 0.00133686 0.77060068 - layer.3.v_cache 0.00000204 0.00232621 - layer.4.k_cache 0.00329063 1.58214315 - layer.4.v_cache 0.00000307 0.00384074 - layer.4.output 0.00016515 0.10609500 - ------------------------------------------------------------------------------------- - TOTAL 0.00338157 1.07394906 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 232168 -BPFP 1.3931 bits/point -EBPFP 2.7862 equivalent bits/point -MSE 1.073949 ----------------------- -------------------------------------------------------- -Time: 3.061s Load: 0.007s, Pack+Encode: 1.823s, Decode+Unpack: 1.231s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0739 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample66-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample66-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-layer4-item1.zst (76/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,964B, BPFP=0.1632 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,604B, BPFP=1.8787 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,532B, BPFP=1.1247 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,784B, BPFP=2.0598 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,584B, BPFP=1.1290 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,076B, BPFP=2.0841 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,680B, BPFP=1.5525 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,680B, BPFP=2.0512 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,200B, BPFP=0.9309 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,412B, BPFP=2.1120 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,724B, BPFP=0.8462 -⌛️ [2/4] FRONTEND: Frontend time: 1.770s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.459s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03150173 9.50624117 - layer.0.v_cache 0.00000027 0.00027224 - layer.1.k_cache 0.00348105 1.66264181 - layer.1.v_cache 0.00000083 0.00092261 - layer.2.k_cache 0.00114206 0.66964523 - layer.2.v_cache 0.00000103 0.00134114 - layer.3.k_cache 0.00133122 0.79695284 - layer.3.v_cache 0.00000211 0.00230928 - layer.4.k_cache 0.00332532 1.53191100 - layer.4.v_cache 0.00000306 0.00372274 - layer.4.output 0.00018383 0.12879816 - ------------------------------------------------------------------------------------- - TOTAL 0.00296600 1.04936805 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 222240 -BPFP 1.3193 bits/point -EBPFP 2.6387 equivalent bits/point -MSE 1.049368 ----------------------- -------------------------------------------------------- -Time: 3.236s Load: 0.007s, Pack+Encode: 1.770s, Decode+Unpack: 1.459s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0494 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample67-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample67-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst (77/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.008s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,916B, BPFP=0.1645 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,640B, BPFP=1.9437 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,428B, BPFP=0.9811 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,656B, BPFP=2.1168 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,744B, BPFP=1.1799 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,100B, BPFP=2.1549 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,532B, BPFP=1.5910 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,488B, BPFP=2.1023 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,592B, BPFP=0.8235 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,268B, BPFP=2.1693 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,952B, BPFP=0.8789 -⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.272s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03172423 9.59399615 - layer.0.v_cache 0.00000027 0.00027346 - layer.1.k_cache 0.00338285 1.50758496 - layer.1.v_cache 0.00000085 0.00092824 - layer.2.k_cache 0.00114373 0.65577689 - layer.2.v_cache 0.00000106 0.00135353 - layer.3.k_cache 0.00134976 0.78627266 - layer.3.v_cache 0.00000205 0.00229024 - layer.4.k_cache 0.00321980 1.46728734 - layer.4.v_cache 0.00000306 0.00380494 - layer.4.output 0.00016531 0.11191616 - ------------------------------------------------------------------------------------- - TOTAL 0.00296349 1.03337379 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 218316 -BPFP 1.3388 bits/point -EBPFP 2.6775 equivalent bits/point -MSE 1.033374 ----------------------- -------------------------------------------------------- -Time: 2.987s Load: 0.008s, Pack+Encode: 1.707s, Decode+Unpack: 1.272s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0334 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample68-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample68-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-layer4-item1.zst (78/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1633 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,672B, BPFP=1.9046 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,316B, BPFP=1.0346 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,992B, BPFP=2.0995 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,828B, BPFP=1.1616 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,288B, BPFP=2.1243 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,548B, BPFP=1.5581 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,500B, BPFP=2.0581 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,084B, BPFP=0.9311 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,380B, BPFP=2.1321 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,120B, BPFP=0.8426 -⌛️ [2/4] FRONTEND: Frontend time: 1.814s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03162601 10.15271521 - layer.0.v_cache 0.00000026 0.00026554 - layer.1.k_cache 0.00335281 1.58993251 - layer.1.v_cache 0.00000078 0.00091617 - layer.2.k_cache 0.00111993 0.70883876 - layer.2.v_cache 0.00000106 0.00138068 - layer.3.k_cache 0.00131014 0.77975685 - layer.3.v_cache 0.00000197 0.00226930 - layer.4.k_cache 0.00327871 1.64035526 - layer.4.v_cache 0.00000299 0.00382384 - layer.4.output 0.00016230 0.10305278 - ------------------------------------------------------------------------------------- - TOTAL 0.00295313 1.09231895 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 220672 -BPFP 1.3241 bits/point -EBPFP 2.6482 equivalent bits/point -MSE 1.092319 ----------------------- -------------------------------------------------------- -Time: 3.110s Load: 0.007s, Pack+Encode: 1.814s, Decode+Unpack: 1.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0923 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample7-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample7-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-layer4-item1.zst (79/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 94, 128) -Output shape: (1, 94, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) -> torch.Size([1, 1, 94, 1024]) - layer.4.output: torch.Size([1, 94, 4096]) -> torch.Size([1, 1, 94, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,972B, BPFP=0.1639 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,144B, BPFP=1.9235 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,880B, BPFP=1.0705 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,348B, BPFP=2.1067 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 16,560B, BPFP=1.3763 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,736B, BPFP=2.1390 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,756B, BPFP=1.5588 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,944B, BPFP=2.0731 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,348B, BPFP=1.0263 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,844B, BPFP=2.1479 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 50,044B, BPFP=1.0398 -⌛️ [2/4] FRONTEND: Frontend time: 1.689s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.403s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 94, 128]) - layer.0.v_cache: torch.Size([1, 8, 94, 128]) - layer.1.k_cache: torch.Size([1, 8, 94, 128]) - layer.1.v_cache: torch.Size([1, 8, 94, 128]) - layer.2.k_cache: torch.Size([1, 8, 94, 128]) - layer.2.v_cache: torch.Size([1, 8, 94, 128]) - layer.3.k_cache: torch.Size([1, 8, 94, 128]) - layer.3.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.k_cache: torch.Size([1, 8, 94, 128]) - layer.4.v_cache: torch.Size([1, 8, 94, 128]) - layer.4.output: torch.Size([1, 94, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03152096 9.35270918 - layer.0.v_cache 0.00000028 0.00027979 - layer.1.k_cache 0.00374224 1.51141163 - layer.1.v_cache 0.00000083 0.00096269 - layer.2.k_cache 0.00112812 0.68724174 - layer.2.v_cache 0.00000115 0.00146872 - layer.3.k_cache 0.00131826 0.78787524 - layer.3.v_cache 0.00000212 0.00233512 - layer.4.k_cache 0.00331939 1.48727920 - layer.4.v_cache 0.00000319 0.00393683 - layer.4.output 0.00017462 0.10669443 - ------------------------------------------------------------------------------------- - TOTAL 0.00298107 1.01873413 - (elements=1,347,584) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1347584 -Total Bytes 237576 -BPFP 1.4104 bits/point -EBPFP 2.8208 equivalent bits/point -MSE 1.018734 ----------------------- -------------------------------------------------------- -Time: 3.098s Load: 0.006s, Pack+Encode: 1.689s, Decode+Unpack: 1.403s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 94, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 94, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0187 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample70-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample70-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-layer4-item1.zst (80/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1644 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,852B, BPFP=1.9406 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,724B, BPFP=1.0805 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,252B, BPFP=2.1444 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,536B, BPFP=1.2344 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,660B, BPFP=2.1790 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,144B, BPFP=1.6257 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,856B, BPFP=2.1107 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,732B, BPFP=0.9113 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,464B, BPFP=2.1624 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,604B, BPFP=0.9257 -⌛️ [2/4] FRONTEND: Frontend time: 1.764s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.254s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03010495 9.92499177 - layer.0.v_cache 0.00000027 0.00027361 - layer.1.k_cache 0.00345006 1.40225452 - layer.1.v_cache 0.00000091 0.00096048 - layer.2.k_cache 0.00113922 0.68353172 - layer.2.v_cache 0.00000110 0.00142072 - layer.3.k_cache 0.00129336 0.78114369 - layer.3.v_cache 0.00000204 0.00234670 - layer.4.k_cache 0.00332675 1.49128458 - layer.4.v_cache 0.00000304 0.00386672 - layer.4.output 0.00015992 0.12312533 - ------------------------------------------------------------------------------------- - TOTAL 0.00285438 1.05604113 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 226760 -BPFP 1.3754 bits/point -EBPFP 2.7509 equivalent bits/point -MSE 1.056041 ----------------------- -------------------------------------------------------- -Time: 3.024s Load: 0.006s, Pack+Encode: 1.764s, Decode+Unpack: 1.254s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0560 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample71-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample71-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-layer4-item1.zst (81/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,948B, BPFP=0.1636 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,700B, BPFP=1.9069 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,088B, BPFP=1.0995 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,156B, BPFP=2.1132 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,492B, BPFP=1.3014 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,332B, BPFP=2.1280 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,012B, BPFP=1.5971 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,688B, BPFP=2.0739 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,520B, BPFP=0.9677 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,412B, BPFP=2.1347 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,720B, BPFP=0.8762 -⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.419s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03144996 9.82060258 - layer.0.v_cache 0.00000027 0.00026614 - layer.1.k_cache 0.00343081 1.61440105 - layer.1.v_cache 0.00000081 0.00095344 - layer.2.k_cache 0.00115410 0.65957322 - layer.2.v_cache 0.00000107 0.00139777 - layer.3.k_cache 0.00131217 0.77982060 - layer.3.v_cache 0.00000205 0.00227654 - layer.4.k_cache 0.00325876 1.63426619 - layer.4.v_cache 0.00000310 0.00380444 - layer.4.output 0.00018378 0.11648877 - ------------------------------------------------------------------------------------- - TOTAL 0.00295345 1.07023693 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 226068 -BPFP 1.3565 bits/point -EBPFP 2.7130 equivalent bits/point -MSE 1.070237 ----------------------- -------------------------------------------------------- -Time: 3.188s Load: 0.006s, Pack+Encode: 1.763s, Decode+Unpack: 1.419s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0702 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample72-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample72-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-layer4-item1.zst (82/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1724 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,308B, BPFP=2.1423 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,444B, BPFP=1.2357 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,524B, BPFP=2.3460 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,180B, BPFP=1.1195 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,940B, BPFP=2.3842 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,848B, BPFP=1.6404 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,016B, BPFP=2.2993 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,364B, BPFP=0.8607 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,772B, BPFP=2.3687 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,304B, BPFP=0.9491 -⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.310s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03191245 10.81795726 - layer.0.v_cache 0.00000028 0.00026880 - layer.1.k_cache 0.00355479 1.50679806 - layer.1.v_cache 0.00000082 0.00094874 - layer.2.k_cache 0.00115713 0.65309650 - layer.2.v_cache 0.00000109 0.00138954 - layer.3.k_cache 0.00128442 0.77152333 - layer.3.v_cache 0.00000206 0.00233601 - layer.4.k_cache 0.00320041 1.46850694 - layer.4.v_cache 0.00000307 0.00387339 - layer.4.output 0.00016954 0.10573408 - ------------------------------------------------------------------------------------- - TOTAL 0.00298533 1.11783107 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 221576 -BPFP 1.4547 bits/point -EBPFP 2.9093 equivalent bits/point -MSE 1.117831 ----------------------- -------------------------------------------------------- -Time: 3.030s Load: 0.006s, Pack+Encode: 1.714s, Decode+Unpack: 1.310s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1178 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample73-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample73-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-layer4-item1.zst (83/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,928B, BPFP=0.1674 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,468B, BPFP=2.0372 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,640B, BPFP=1.1840 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,420B, BPFP=2.2066 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,428B, BPFP=1.2524 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,644B, BPFP=2.2260 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,808B, BPFP=1.6326 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,280B, BPFP=2.1944 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,144B, BPFP=0.7937 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,908B, BPFP=2.2490 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 52,136B, BPFP=1.1314 -⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.270s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03149115 9.32766859 - layer.0.v_cache 0.00000027 0.00027508 - layer.1.k_cache 0.00351934 1.56425900 - layer.1.v_cache 0.00000089 0.00093789 - layer.2.k_cache 0.00114079 0.67827496 - layer.2.v_cache 0.00000110 0.00139171 - layer.3.k_cache 0.00129789 0.79273504 - layer.3.v_cache 0.00000226 0.00239631 - layer.4.k_cache 0.00328524 1.42091658 - layer.4.v_cache 0.00000318 0.00386335 - layer.4.output 0.00020520 0.11897575 - ------------------------------------------------------------------------------------- - TOTAL 0.00296878 1.01918725 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 235804 -BPFP 1.4621 bits/point -EBPFP 2.9242 equivalent bits/point -MSE 1.019187 ----------------------- -------------------------------------------------------- -Time: 3.112s Load: 0.007s, Pack+Encode: 1.835s, Decode+Unpack: 1.270s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0192 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample74-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample74-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-layer4-item1.zst (84/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,952B, BPFP=0.1658 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,028B, BPFP=1.9555 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,804B, BPFP=1.1722 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,080B, BPFP=2.1298 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,660B, BPFP=1.3298 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,356B, BPFP=2.1532 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,336B, BPFP=1.6420 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,728B, BPFP=2.0999 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,524B, BPFP=0.9786 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,316B, BPFP=2.1498 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 43,180B, BPFP=0.9167 -⌛️ [2/4] FRONTEND: Frontend time: 1.689s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.434s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03262251 9.77823738 - layer.0.v_cache 0.00000028 0.00026970 - layer.1.k_cache 0.00349940 1.62464590 - layer.1.v_cache 0.00000082 0.00095938 - layer.2.k_cache 0.00113634 0.66725909 - layer.2.v_cache 0.00000107 0.00139889 - layer.3.k_cache 0.00132950 0.78137904 - layer.3.v_cache 0.00000207 0.00235703 - layer.4.k_cache 0.00323306 1.52150544 - layer.4.v_cache 0.00000312 0.00392129 - layer.4.output 0.00016774 0.10855995 - ------------------------------------------------------------------------------------- - TOTAL 0.00303565 1.05829807 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 228964 -BPFP 1.3888 bits/point -EBPFP 2.7776 equivalent bits/point -MSE 1.058298 ----------------------- -------------------------------------------------------- -Time: 3.128s Load: 0.005s, Pack+Encode: 1.689s, Decode+Unpack: 1.434s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0583 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample75-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample75-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst (85/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1662 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,096B, BPFP=1.9828 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,692B, BPFP=1.0896 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,048B, BPFP=2.1504 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,908B, BPFP=1.1082 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,284B, BPFP=2.1707 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,684B, BPFP=1.6041 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,596B, BPFP=2.1116 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,588B, BPFP=0.9090 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,692B, BPFP=2.2057 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,360B, BPFP=0.8877 -⌛️ [2/4] FRONTEND: Frontend time: 1.769s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.258s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03087558 9.68603180 - layer.0.v_cache 0.00000028 0.00027372 - layer.1.k_cache 0.00335485 1.40183074 - layer.1.v_cache 0.00000080 0.00092690 - layer.2.k_cache 0.00115595 0.68358377 - layer.2.v_cache 0.00000105 0.00134104 - layer.3.k_cache 0.00133087 0.77191011 - layer.3.v_cache 0.00000204 0.00224131 - layer.4.k_cache 0.00324117 1.62279619 - layer.4.v_cache 0.00000319 0.00383261 - layer.4.output 0.00016760 0.11497396 - ------------------------------------------------------------------------------------- - TOTAL 0.00290258 1.04533315 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 221884 -BPFP 1.3607 bits/point -EBPFP 2.7213 equivalent bits/point -MSE 1.045333 ----------------------- -------------------------------------------------------- -Time: 3.034s Load: 0.007s, Pack+Encode: 1.769s, Decode+Unpack: 1.258s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0453 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample78-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample78-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-layer4-item1.zst (86/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 97, 128) -Output shape: (1, 97, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) -> torch.Size([1, 1, 97, 1024]) - layer.4.output: torch.Size([1, 97, 4096]) -> torch.Size([1, 1, 97, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,044B, BPFP=0.1646 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,980B, BPFP=1.8508 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,152B, BPFP=1.1398 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,368B, BPFP=2.0432 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,772B, BPFP=1.1898 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,408B, BPFP=2.0464 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,652B, BPFP=1.5023 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,896B, BPFP=2.0052 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,396B, BPFP=0.9984 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,716B, BPFP=2.0712 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,208B, BPFP=0.9304 -⌛️ [2/4] FRONTEND: Frontend time: 1.821s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.380s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 97, 128]) - layer.0.v_cache: torch.Size([1, 8, 97, 128]) - layer.1.k_cache: torch.Size([1, 8, 97, 128]) - layer.1.v_cache: torch.Size([1, 8, 97, 128]) - layer.2.k_cache: torch.Size([1, 8, 97, 128]) - layer.2.v_cache: torch.Size([1, 8, 97, 128]) - layer.3.k_cache: torch.Size([1, 8, 97, 128]) - layer.3.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.k_cache: torch.Size([1, 8, 97, 128]) - layer.4.v_cache: torch.Size([1, 8, 97, 128]) - layer.4.output: torch.Size([1, 97, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03000178 9.78758759 - layer.0.v_cache 0.00000027 0.00026921 - layer.1.k_cache 0.00327340 1.76574235 - layer.1.v_cache 0.00000092 0.00094809 - layer.2.k_cache 0.00113100 0.65238898 - layer.2.v_cache 0.00000104 0.00137323 - layer.3.k_cache 0.00131735 0.78640818 - layer.3.v_cache 0.00000202 0.00226472 - layer.4.k_cache 0.00333574 1.60288183 - layer.4.v_cache 0.00000307 0.00379652 - layer.4.output 0.00017322 0.10692506 - ------------------------------------------------------------------------------------- - TOTAL 0.00283996 1.07366864 - (elements=1,390,592) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1390592 -Total Bytes 232592 -BPFP 1.3381 bits/point -EBPFP 2.6762 equivalent bits/point -MSE 1.073669 ----------------------- -------------------------------------------------------- -Time: 3.207s Load: 0.007s, Pack+Encode: 1.821s, Decode+Unpack: 1.380s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 97, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 97, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0737 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample8-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample8-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst (87/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 91, 128) -Output shape: (1, 91, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) -> torch.Size([1, 1, 91, 1024]) - layer.4.output: torch.Size([1, 91, 4096]) -> torch.Size([1, 1, 91, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,920B, BPFP=0.1648 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,880B, BPFP=1.9643 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,752B, BPFP=1.0948 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,960B, BPFP=2.1429 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,816B, BPFP=1.1861 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,344B, BPFP=2.1758 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,492B, BPFP=1.5876 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,552B, BPFP=2.1078 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,476B, BPFP=0.8135 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,448B, BPFP=2.1848 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,916B, BPFP=0.8567 -⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.330s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 91, 128]) - layer.0.v_cache: torch.Size([1, 8, 91, 128]) - layer.1.k_cache: torch.Size([1, 8, 91, 128]) - layer.1.v_cache: torch.Size([1, 8, 91, 128]) - layer.2.k_cache: torch.Size([1, 8, 91, 128]) - layer.2.v_cache: torch.Size([1, 8, 91, 128]) - layer.3.k_cache: torch.Size([1, 8, 91, 128]) - layer.3.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.k_cache: torch.Size([1, 8, 91, 128]) - layer.4.v_cache: torch.Size([1, 8, 91, 128]) - layer.4.output: torch.Size([1, 91, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03041337 9.42033571 - layer.0.v_cache 0.00000027 0.00027263 - layer.1.k_cache 0.00341539 1.49805995 - layer.1.v_cache 0.00000079 0.00092376 - layer.2.k_cache 0.00115498 0.67935445 - layer.2.v_cache 0.00000106 0.00138155 - layer.3.k_cache 0.00132454 0.78628087 - layer.3.v_cache 0.00000206 0.00225468 - layer.4.k_cache 0.00329878 1.53859652 - layer.4.v_cache 0.00000306 0.00370134 - layer.4.output 0.00018052 0.11403713 - ------------------------------------------------------------------------------------- - TOTAL 0.00288117 1.02766500 - (elements=1,304,576) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1304576 -Total Bytes 219556 -BPFP 1.3464 bits/point -EBPFP 2.6927 equivalent bits/point -MSE 1.027665 ----------------------- -------------------------------------------------------- -Time: 3.038s Load: 0.006s, Pack+Encode: 1.702s, Decode+Unpack: 1.330s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 91, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 91, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0277 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample80-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample80-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst (88/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 88, 128) -Output shape: (1, 88, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) -> torch.Size([1, 1, 88, 1024]) - layer.4.output: torch.Size([1, 88, 4096]) -> torch.Size([1, 1, 88, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,836B, BPFP=0.1630 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,220B, BPFP=2.0614 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,444B, BPFP=1.0160 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,368B, BPFP=2.2521 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,184B, BPFP=1.2592 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,720B, BPFP=2.2834 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,308B, BPFP=1.6254 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,984B, BPFP=2.2180 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,928B, BPFP=0.7926 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,748B, BPFP=2.2859 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,828B, BPFP=0.9949 -⌛️ [2/4] FRONTEND: Frontend time: 1.798s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.227s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 88, 128]) - layer.0.v_cache: torch.Size([1, 8, 88, 128]) - layer.1.k_cache: torch.Size([1, 8, 88, 128]) - layer.1.v_cache: torch.Size([1, 8, 88, 128]) - layer.2.k_cache: torch.Size([1, 8, 88, 128]) - layer.2.v_cache: torch.Size([1, 8, 88, 128]) - layer.3.k_cache: torch.Size([1, 8, 88, 128]) - layer.3.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.k_cache: torch.Size([1, 8, 88, 128]) - layer.4.v_cache: torch.Size([1, 8, 88, 128]) - layer.4.output: torch.Size([1, 88, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03108732 10.23322643 - layer.0.v_cache 0.00000028 0.00027439 - layer.1.k_cache 0.00347999 1.35238856 - layer.1.v_cache 0.00000081 0.00095415 - layer.2.k_cache 0.00113734 0.67862372 - layer.2.v_cache 0.00000108 0.00139502 - layer.3.k_cache 0.00129639 0.76987102 - layer.3.v_cache 0.00000204 0.00230065 - layer.4.k_cache 0.00326982 1.53661485 - layer.4.v_cache 0.00000302 0.00388092 - layer.4.output 0.00017394 0.11081351 - ------------------------------------------------------------------------------------- - TOTAL 0.00292670 1.07305598 - (elements=1,261,568) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1261568 -Total Bytes 224568 -BPFP 1.4241 bits/point -EBPFP 2.8481 equivalent bits/point -MSE 1.073056 ----------------------- -------------------------------------------------------- -Time: 3.031s Load: 0.005s, Pack+Encode: 1.798s, Decode+Unpack: 1.227s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 88, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 88, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0731 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample81-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample81-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-layer4-item1.zst (89/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,936B, BPFP=0.1644 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,900B, BPFP=1.9446 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,724B, BPFP=1.0805 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,140B, BPFP=2.1349 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,964B, BPFP=1.1858 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,572B, BPFP=2.1715 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 16,824B, BPFP=1.4287 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,916B, BPFP=2.1158 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,736B, BPFP=0.8268 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,696B, BPFP=2.1821 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,344B, BPFP=0.9839 -⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.455s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03045331 9.90543001 - layer.0.v_cache 0.00000027 0.00026690 - layer.1.k_cache 0.00344094 1.47241410 - layer.1.v_cache 0.00000081 0.00094495 - layer.2.k_cache 0.00113827 0.69087410 - layer.2.v_cache 0.00000106 0.00138172 - layer.3.k_cache 0.00131047 0.77261518 - layer.3.v_cache 0.00000202 0.00231413 - layer.4.k_cache 0.00338451 1.44695680 - layer.4.v_cache 0.00000307 0.00383765 - layer.4.output 0.00016907 0.11888818 - ------------------------------------------------------------------------------------- - TOTAL 0.00288650 1.05518488 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 225752 -BPFP 1.3693 bits/point -EBPFP 2.7386 equivalent bits/point -MSE 1.055185 ----------------------- -------------------------------------------------------- -Time: 3.186s Load: 0.006s, Pack+Encode: 1.725s, Decode+Unpack: 1.455s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0552 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample82-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample82-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-layer4-item1.zst (90/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,928B, BPFP=0.1637 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,872B, BPFP=1.9423 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,668B, BPFP=1.0757 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,156B, BPFP=2.1362 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,704B, BPFP=1.0788 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,636B, BPFP=2.1770 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,576B, BPFP=1.4925 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,872B, BPFP=2.1121 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,656B, BPFP=0.8200 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,620B, BPFP=2.1756 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 45,344B, BPFP=0.9626 -⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.233s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03019064 9.71980750 - layer.0.v_cache 0.00000027 0.00026810 - layer.1.k_cache 0.00347154 1.46663599 - layer.1.v_cache 0.00000080 0.00094068 - layer.2.k_cache 0.00112864 0.67611338 - layer.2.v_cache 0.00000106 0.00138810 - layer.3.k_cache 0.00131097 0.77646836 - layer.3.v_cache 0.00000203 0.00231406 - layer.4.k_cache 0.00338659 1.46506268 - layer.4.v_cache 0.00000308 0.00382568 - layer.4.output 0.00016864 0.12200378 - ------------------------------------------------------------------------------------- - TOTAL 0.00286930 1.04291712 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 224032 -BPFP 1.3589 bits/point -EBPFP 2.7178 equivalent bits/point -MSE 1.042917 ----------------------- -------------------------------------------------------- -Time: 2.977s Load: 0.007s, Pack+Encode: 1.737s, Decode+Unpack: 1.233s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0429 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample83-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample83-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-layer4-item1.zst (91/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 96, 128) -Output shape: (1, 96, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) -> torch.Size([1, 1, 96, 1024]) - layer.4.output: torch.Size([1, 96, 4096]) -> torch.Size([1, 1, 96, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 2,012B, BPFP=0.1637 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,028B, BPFP=1.8740 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 13,004B, BPFP=1.0583 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,892B, BPFP=2.0257 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,580B, BPFP=1.1051 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,288B, BPFP=2.0579 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,304B, BPFP=1.4896 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,664B, BPFP=2.0072 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 14,060B, BPFP=1.1442 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,376B, BPFP=2.0651 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 42,380B, BPFP=0.8622 -⌛️ [2/4] FRONTEND: Frontend time: 1.802s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.355s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 96, 128]) - layer.0.v_cache: torch.Size([1, 8, 96, 128]) - layer.1.k_cache: torch.Size([1, 8, 96, 128]) - layer.1.v_cache: torch.Size([1, 8, 96, 128]) - layer.2.k_cache: torch.Size([1, 8, 96, 128]) - layer.2.v_cache: torch.Size([1, 8, 96, 128]) - layer.3.k_cache: torch.Size([1, 8, 96, 128]) - layer.3.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.k_cache: torch.Size([1, 8, 96, 128]) - layer.4.v_cache: torch.Size([1, 8, 96, 128]) - layer.4.output: torch.Size([1, 96, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03109797 9.44980558 - layer.0.v_cache 0.00000027 0.00026689 - layer.1.k_cache 0.00337694 1.65569544 - layer.1.v_cache 0.00000077 0.00093192 - layer.2.k_cache 0.00112303 0.68440278 - layer.2.v_cache 0.00000110 0.00139291 - layer.3.k_cache 0.00132768 0.79840159 - layer.3.v_cache 0.00000201 0.00226964 - layer.4.k_cache 0.00329028 1.54762125 - layer.4.v_cache 0.00000306 0.00381974 - layer.4.output 0.00016342 0.11281665 - ------------------------------------------------------------------------------------- - TOTAL 0.00291977 1.04256245 - (elements=1,376,256) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1376256 -Total Bytes 226588 -BPFP 1.3171 bits/point -EBPFP 2.6343 equivalent bits/point -MSE 1.042562 ----------------------- -------------------------------------------------------- -Time: 3.163s Load: 0.006s, Pack+Encode: 1.802s, Decode+Unpack: 1.355s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 96, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 96, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0426 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample85-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample85-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-layer4-item1.zst (92/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,960B, BPFP=0.1647 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,860B, BPFP=1.9204 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,196B, BPFP=1.0245 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,032B, BPFP=2.1028 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,828B, BPFP=1.3296 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,328B, BPFP=2.1277 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,880B, BPFP=1.5860 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,552B, BPFP=2.0625 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 12,324B, BPFP=1.0353 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,208B, BPFP=2.1176 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 41,312B, BPFP=0.8676 -⌛️ [2/4] FRONTEND: Frontend time: 1.732s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.368s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03084236 9.82214749 - layer.0.v_cache 0.00000028 0.00026769 - layer.1.k_cache 0.00343721 1.51431373 - layer.1.v_cache 0.00000078 0.00094436 - layer.2.k_cache 0.00115435 0.67207542 - layer.2.v_cache 0.00000107 0.00141022 - layer.3.k_cache 0.00133139 0.77246159 - layer.3.v_cache 0.00000204 0.00234701 - layer.4.k_cache 0.00330969 1.66983853 - layer.4.v_cache 0.00000309 0.00385588 - layer.4.output 0.00020550 0.10511532 - ------------------------------------------------------------------------------------- - TOTAL 0.00292173 1.06286594 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 225480 -BPFP 1.3530 bits/point -EBPFP 2.7059 equivalent bits/point -MSE 1.062866 ----------------------- -------------------------------------------------------- -Time: 3.105s Load: 0.006s, Pack+Encode: 1.732s, Decode+Unpack: 1.368s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0629 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample88-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample88-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-layer4-item1.zst (93/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1647 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,480B, BPFP=2.0611 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,492B, BPFP=1.0088 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,444B, BPFP=2.2335 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,296B, BPFP=1.1671 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,792B, BPFP=2.2640 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,660B, BPFP=1.7258 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,144B, BPFP=2.2072 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,740B, BPFP=1.0305 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,864B, BPFP=2.2704 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 45,904B, BPFP=1.0074 -⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.242s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03125147 10.03081795 - layer.0.v_cache 0.00000028 0.00027163 - layer.1.k_cache 0.00348025 1.58848246 - layer.1.v_cache 0.00000083 0.00095012 - layer.2.k_cache 0.00111536 0.66481074 - layer.2.v_cache 0.00000108 0.00140311 - layer.3.k_cache 0.00131582 0.78305782 - layer.3.v_cache 0.00000206 0.00229767 - layer.4.k_cache 0.00329313 1.48464623 - layer.4.v_cache 0.00000300 0.00387198 - layer.4.output 0.00017131 0.11168400 - ------------------------------------------------------------------------------------- - TOTAL 0.00293918 1.07195326 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 229692 -BPFP 1.4402 bits/point -EBPFP 2.8804 equivalent bits/point -MSE 1.071953 ----------------------- -------------------------------------------------------- -Time: 3.073s Load: 0.005s, Pack+Encode: 1.825s, Decode+Unpack: 1.242s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0720 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample90-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample90-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-layer4-item1.zst (94/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 93, 128) -Output shape: (1, 93, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) -> torch.Size([1, 1, 93, 1024]) - layer.4.output: torch.Size([1, 93, 4096]) -> torch.Size([1, 1, 93, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,944B, BPFP=0.1633 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,784B, BPFP=1.9140 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,492B, BPFP=1.0494 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 24,960B, BPFP=2.0968 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,152B, BPFP=1.2728 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,264B, BPFP=2.1223 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,216B, BPFP=1.5302 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,476B, BPFP=2.0561 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 13,088B, BPFP=1.0995 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,340B, BPFP=2.1287 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 46,600B, BPFP=0.9787 -⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.400s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 93, 128]) - layer.0.v_cache: torch.Size([1, 8, 93, 128]) - layer.1.k_cache: torch.Size([1, 8, 93, 128]) - layer.1.v_cache: torch.Size([1, 8, 93, 128]) - layer.2.k_cache: torch.Size([1, 8, 93, 128]) - layer.2.v_cache: torch.Size([1, 8, 93, 128]) - layer.3.k_cache: torch.Size([1, 8, 93, 128]) - layer.3.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.k_cache: torch.Size([1, 8, 93, 128]) - layer.4.v_cache: torch.Size([1, 8, 93, 128]) - layer.4.output: torch.Size([1, 93, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03101371 9.83344490 - layer.0.v_cache 0.00000028 0.00026710 - layer.1.k_cache 0.00364020 1.51376129 - layer.1.v_cache 0.00000086 0.00092847 - layer.2.k_cache 0.00114853 0.67499198 - layer.2.v_cache 0.00000108 0.00139615 - layer.3.k_cache 0.00131058 0.79658164 - layer.3.v_cache 0.00000204 0.00232650 - layer.4.k_cache 0.00330942 1.52141481 - layer.4.v_cache 0.00000298 0.00386138 - layer.4.output 0.00017313 0.12030846 - ------------------------------------------------------------------------------------- - TOTAL 0.00293730 1.05930057 - (elements=1,333,248) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1333248 -Total Bytes 230316 -BPFP 1.3820 bits/point -EBPFP 2.7640 equivalent bits/point -MSE 1.059301 ----------------------- -------------------------------------------------------- -Time: 3.151s Load: 0.007s, Pack+Encode: 1.744s, Decode+Unpack: 1.400s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 93, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 93, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0593 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample92-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample92-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-layer4-item1.zst (95/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 90, 128) -Output shape: (1, 90, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) -> torch.Size([1, 1, 90, 1024]) - layer.4.output: torch.Size([1, 90, 4096]) -> torch.Size([1, 1, 90, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,912B, BPFP=0.1660 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 22,972B, BPFP=1.9941 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,600B, BPFP=1.0938 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,108B, BPFP=2.1795 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,924B, BPFP=1.2955 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,340B, BPFP=2.1997 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,160B, BPFP=1.6632 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,728B, BPFP=2.1465 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,652B, BPFP=0.8378 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,584B, BPFP=2.2208 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,424B, BPFP=0.8556 -⌛️ [2/4] FRONTEND: Frontend time: 1.762s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.245s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 90, 128]) - layer.0.v_cache: torch.Size([1, 8, 90, 128]) - layer.1.k_cache: torch.Size([1, 8, 90, 128]) - layer.1.v_cache: torch.Size([1, 8, 90, 128]) - layer.2.k_cache: torch.Size([1, 8, 90, 128]) - layer.2.v_cache: torch.Size([1, 8, 90, 128]) - layer.3.k_cache: torch.Size([1, 8, 90, 128]) - layer.3.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.k_cache: torch.Size([1, 8, 90, 128]) - layer.4.v_cache: torch.Size([1, 8, 90, 128]) - layer.4.output: torch.Size([1, 90, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03189978 9.85012682 - layer.0.v_cache 0.00000027 0.00027202 - layer.1.k_cache 0.00346617 1.40216404 - layer.1.v_cache 0.00000078 0.00090257 - layer.2.k_cache 0.00115535 0.67166778 - layer.2.v_cache 0.00000103 0.00131419 - layer.3.k_cache 0.00134436 0.78478995 - layer.3.v_cache 0.00000198 0.00218022 - layer.4.k_cache 0.00332606 1.51614787 - layer.4.v_cache 0.00000301 0.00371366 - layer.4.output 0.00018488 0.11322877 - ------------------------------------------------------------------------------------- - TOTAL 0.00299559 1.04901387 - (elements=1,290,240) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1290240 -Total Bytes 221404 -BPFP 1.3728 bits/point -EBPFP 2.7456 equivalent bits/point -MSE 1.049014 ----------------------- -------------------------------------------------------- -Time: 3.012s Load: 0.005s, Pack+Encode: 1.762s, Decode+Unpack: 1.245s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 90, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 90, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0490 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample93-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample93-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-layer4-item1.zst (96/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.007s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,876B, BPFP=0.1647 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,096B, BPFP=2.0274 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,612B, BPFP=1.1071 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,308B, BPFP=2.2216 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 15,028B, BPFP=1.3192 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,600B, BPFP=2.2472 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,952B, BPFP=1.5758 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,880B, BPFP=2.1840 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,936B, BPFP=0.7844 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,476B, BPFP=2.2363 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,428B, BPFP=0.8872 -⌛️ [2/4] FRONTEND: Frontend time: 1.866s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.290s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03210033 9.78827110 - layer.0.v_cache 0.00000027 0.00026629 - layer.1.k_cache 0.00338447 1.51529531 - layer.1.v_cache 0.00000081 0.00096166 - layer.2.k_cache 0.00116617 0.65686374 - layer.2.v_cache 0.00000108 0.00140525 - layer.3.k_cache 0.00134809 0.79427020 - layer.3.v_cache 0.00000208 0.00237134 - layer.4.k_cache 0.00324151 1.43693440 - layer.4.v_cache 0.00000296 0.00378473 - layer.4.output 0.00019744 0.11641431 - ------------------------------------------------------------------------------------- - TOTAL 0.00300268 1.04757723 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 221192 -BPFP 1.3869 bits/point -EBPFP 2.7738 equivalent bits/point -MSE 1.047577 ----------------------- -------------------------------------------------------- -Time: 3.162s Load: 0.007s, Pack+Encode: 1.866s, Decode+Unpack: 1.290s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0476 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample94-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample94-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-layer4-item1.zst (97/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 89, 128) -Output shape: (1, 89, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) -> torch.Size([1, 1, 89, 1024]) - layer.4.output: torch.Size([1, 89, 4096]) -> torch.Size([1, 1, 89, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,888B, BPFP=0.1657 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,192B, BPFP=2.0358 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,596B, BPFP=1.1057 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,328B, BPFP=2.2233 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,124B, BPFP=1.2398 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,616B, BPFP=2.2486 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,864B, BPFP=1.5681 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,988B, BPFP=2.1935 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 8,980B, BPFP=0.7883 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,448B, BPFP=2.2338 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 40,584B, BPFP=0.8906 -⌛️ [2/4] FRONTEND: Frontend time: 1.686s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.414s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 89, 128]) - layer.0.v_cache: torch.Size([1, 8, 89, 128]) - layer.1.k_cache: torch.Size([1, 8, 89, 128]) - layer.1.v_cache: torch.Size([1, 8, 89, 128]) - layer.2.k_cache: torch.Size([1, 8, 89, 128]) - layer.2.v_cache: torch.Size([1, 8, 89, 128]) - layer.3.k_cache: torch.Size([1, 8, 89, 128]) - layer.3.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.k_cache: torch.Size([1, 8, 89, 128]) - layer.4.v_cache: torch.Size([1, 8, 89, 128]) - layer.4.output: torch.Size([1, 89, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03204785 9.77776208 - layer.0.v_cache 0.00000027 0.00026525 - layer.1.k_cache 0.00337303 1.49138856 - layer.1.v_cache 0.00000080 0.00095161 - layer.2.k_cache 0.00115601 0.66753833 - layer.2.v_cache 0.00000108 0.00140644 - layer.3.k_cache 0.00135783 0.79173433 - layer.3.v_cache 0.00000209 0.00238887 - layer.4.k_cache 0.00324176 1.46158231 - layer.4.v_cache 0.00000298 0.00378954 - layer.4.output 0.00019437 0.12026417 - ------------------------------------------------------------------------------------- - TOTAL 0.00299723 1.04856172 - (elements=1,275,904) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1275904 -Total Bytes 220608 -BPFP 1.3832 bits/point -EBPFP 2.7665 equivalent bits/point -MSE 1.048562 ----------------------- -------------------------------------------------------- -Time: 3.106s Load: 0.006s, Pack+Encode: 1.686s, Decode+Unpack: 1.414s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 89, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 89, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0486 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample95-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample95-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-layer4-item1.zst (98/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.005s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 85, 128) -Output shape: (1, 85, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) -> torch.Size([1, 1, 85, 1024]) - layer.4.output: torch.Size([1, 85, 4096]) -> torch.Size([1, 1, 85, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,836B, BPFP=0.1688 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,532B, BPFP=2.1629 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 14,012B, BPFP=1.2879 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,384B, BPFP=2.3331 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 12,204B, BPFP=1.1217 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,888B, BPFP=2.3794 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 19,388B, BPFP=1.7820 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,128B, BPFP=2.3096 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 9,428B, BPFP=0.8665 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,792B, BPFP=2.3706 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 39,320B, BPFP=0.9035 -⌛️ [2/4] FRONTEND: Frontend time: 1.773s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.248s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 85, 128]) - layer.0.v_cache: torch.Size([1, 8, 85, 128]) - layer.1.k_cache: torch.Size([1, 8, 85, 128]) - layer.1.v_cache: torch.Size([1, 8, 85, 128]) - layer.2.k_cache: torch.Size([1, 8, 85, 128]) - layer.2.v_cache: torch.Size([1, 8, 85, 128]) - layer.3.k_cache: torch.Size([1, 8, 85, 128]) - layer.3.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.k_cache: torch.Size([1, 8, 85, 128]) - layer.4.v_cache: torch.Size([1, 8, 85, 128]) - layer.4.output: torch.Size([1, 85, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03332241 11.08607250 - layer.0.v_cache 0.00000028 0.00027102 - layer.1.k_cache 0.00345674 1.55864725 - layer.1.v_cache 0.00000080 0.00094159 - layer.2.k_cache 0.00115406 0.65427031 - layer.2.v_cache 0.00000108 0.00141739 - layer.3.k_cache 0.00129640 0.77367895 - layer.3.v_cache 0.00000208 0.00229262 - layer.4.k_cache 0.00325461 1.44675706 - layer.4.v_cache 0.00000299 0.00381175 - layer.4.output 0.00016656 0.11400572 - ------------------------------------------------------------------------------------- - TOTAL 0.00308269 1.14172738 - (elements=1,218,560) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1218560 -Total Bytes 221912 -BPFP 1.4569 bits/point -EBPFP 2.9138 equivalent bits/point -MSE 1.141727 ----------------------- -------------------------------------------------------- -Time: 3.026s Load: 0.005s, Pack+Encode: 1.773s, Decode+Unpack: 1.248s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 85, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 85, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.1417 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample96-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample96-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-layer4-item1.zst (99/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 92, 128) -Output shape: (1, 92, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) -> torch.Size([1, 1, 92, 1024]) - layer.4.output: torch.Size([1, 92, 4096]) -> torch.Size([1, 1, 92, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,952B, BPFP=0.1658 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,232B, BPFP=1.9728 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 12,424B, BPFP=1.0550 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,196B, BPFP=2.1396 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 14,416B, BPFP=1.2242 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,580B, BPFP=2.1722 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 18,276B, BPFP=1.5520 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 24,780B, BPFP=2.1043 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 11,228B, BPFP=0.9535 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,696B, BPFP=2.1821 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,512B, BPFP=0.9450 -⌛️ [2/4] FRONTEND: Frontend time: 1.773s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.422s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 92, 128]) - layer.0.v_cache: torch.Size([1, 8, 92, 128]) - layer.1.k_cache: torch.Size([1, 8, 92, 128]) - layer.1.v_cache: torch.Size([1, 8, 92, 128]) - layer.2.k_cache: torch.Size([1, 8, 92, 128]) - layer.2.v_cache: torch.Size([1, 8, 92, 128]) - layer.3.k_cache: torch.Size([1, 8, 92, 128]) - layer.3.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.k_cache: torch.Size([1, 8, 92, 128]) - layer.4.v_cache: torch.Size([1, 8, 92, 128]) - layer.4.output: torch.Size([1, 92, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03086128 9.66599970 - layer.0.v_cache 0.00000027 0.00026939 - layer.1.k_cache 0.00344194 1.59714475 - layer.1.v_cache 0.00000079 0.00093061 - layer.2.k_cache 0.00114743 0.64958361 - layer.2.v_cache 0.00000107 0.00137131 - layer.3.k_cache 0.00130642 0.76171112 - layer.3.v_cache 0.00000205 0.00226330 - layer.4.k_cache 0.00331827 1.45350017 - layer.4.v_cache 0.00000300 0.00375339 - layer.4.output 0.00017802 0.11928613 - ------------------------------------------------------------------------------------- - TOTAL 0.00291390 1.04383371 - (elements=1,318,912) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1318912 -Total Bytes 227292 -BPFP 1.3787 bits/point -EBPFP 2.7573 equivalent bits/point -MSE 1.043834 ----------------------- -------------------------------------------------------- -Time: 3.201s Load: 0.006s, Pack+Encode: 1.773s, Decode+Unpack: 1.422s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 92, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 92, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0438 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample98-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample98-layer4-item1.zst - - 💪 Processing: ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-layer4-item1.zst (100/100) - -[1/4] FRONTEND: Loading features from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-layer4-item1.zst... - -Original data structure: -root: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -⌛️ [1/4] FRONTEND: Load time: 0.006s - ------------------------------------------------------------- -Qwen Features Summary ------------------------------------------------------------- -Number of layers: 5 -Layer indices: [0, 1, 2, 3, 4] -Last layer index: 4 -Cache shape: (1, 8, 87, 128) -Output shape: (1, 87, 4096) -Data type: torch.bfloat16 ------------------------------------------------------------- - -[2/4] FRONTEND: Pack + Encode (strategy: individual)... - IndividualPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) -> torch.Size([1, 1, 87, 1024]) - layer.4.output: torch.Size([1, 87, 4096]) -> torch.Size([1, 1, 87, 4096]) - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - layer.0.k_cache: 1,860B, BPFP=0.1670 - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - layer.0.v_cache: 23,576B, BPFP=2.1171 - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - layer.1.k_cache: 11,552B, BPFP=1.0374 - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - layer.1.v_cache: 25,608B, BPFP=2.2996 - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - layer.2.k_cache: 13,800B, BPFP=1.2392 - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - layer.2.v_cache: 25,824B, BPFP=2.3190 - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - layer.3.k_cache: 17,044B, BPFP=1.5305 - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - layer.3.v_cache: 25,428B, BPFP=2.2834 - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - layer.4.k_cache: 10,752B, BPFP=0.9655 - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - layer.4.v_cache: 25,996B, BPFP=2.3344 - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - layer.4.output: 44,812B, BPFP=1.0060 -⌛️ [2/4] FRONTEND: Frontend time: 1.690s (Pack+Encode) - -[3/4] BACKEND: Decode + Unpack... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) -⌛️ [3/4] BACKEND: Backend time: 1.259s - -[4/4] METRICS: Computing MSE Breakdown... - Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache - Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache - Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache - Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache - Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache - Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache - Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache - Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache - Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache - Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache - Using per-key quantization points (layer.4.output: torch.Size([256])) for layer.4.output - IndividualUnPacker: - layer.0.k_cache: torch.Size([1, 8, 87, 128]) - layer.0.v_cache: torch.Size([1, 8, 87, 128]) - layer.1.k_cache: torch.Size([1, 8, 87, 128]) - layer.1.v_cache: torch.Size([1, 8, 87, 128]) - layer.2.k_cache: torch.Size([1, 8, 87, 128]) - layer.2.v_cache: torch.Size([1, 8, 87, 128]) - layer.3.k_cache: torch.Size([1, 8, 87, 128]) - layer.3.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.k_cache: torch.Size([1, 8, 87, 128]) - layer.4.v_cache: torch.Size([1, 8, 87, 128]) - layer.4.output: torch.Size([1, 87, 4096]) - Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): - Key Quant-MSE Total-MSE - ------------------------------------------------------------------------------------- - layer.0.k_cache 0.03104359 9.55133828 - layer.0.v_cache 0.00000028 0.00026971 - layer.1.k_cache 0.00347075 1.45753005 - layer.1.v_cache 0.00000085 0.00094546 - layer.2.k_cache 0.00119839 0.65761259 - layer.2.v_cache 0.00000112 0.00140224 - layer.3.k_cache 0.00133526 0.75898953 - layer.3.v_cache 0.00000211 0.00229530 - layer.4.k_cache 0.00323497 1.45235320 - layer.4.v_cache 0.00000315 0.00386204 - layer.4.output 0.00017081 0.11251525 - ------------------------------------------------------------------------------------- - TOTAL 0.00292669 1.02404710 - (elements=1,247,232) ----------------------- -------------------------------------------------------- -SAMPLE-WISE STATISTICS ----------------------- -------------------------------------------------------- -Handler qwen -Strategy individual -Architecture elic-featurecoding ----------------------- -------------------------------------------------------- -Total Elements 1247232 -Total Bytes 226252 -BPFP 1.4512 bits/point -EBPFP 2.9025 equivalent bits/point -MSE 1.024047 ----------------------- -------------------------------------------------------- -Time: 2.955s Load: 0.006s, Pack+Encode: 1.690s, Decode+Unpack: 1.259s ----------------------- -------------------------------------------------------- -Restored Feature Format: [dict] with 3 keys - key['output']: [Tensor] shape=torch.Size([1, 87, 4096]), dtype=torch.bfloat16, device=cpu - key['key']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - key['value']: [list] with 5 items - item[0]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[1]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[2]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[3]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu - item[4]: [Tensor] shape=torch.Size([1, 8, 87, 128]), dtype=torch.bfloat16, device=cpu -💾 Converting with 1.0240 MSE: - from ../datasets/Qwen3-500features-L5wCache/Qwen3-500features-L5wCache/qwen/qwen3-8b/fc_winogrande/sample99-layer4-item1.zst - to output-fixed/qwen/lambda0.004/elic-featurecoding-8bit-individual/fc_winogrande/sample99-layer4-item1.zst ------------------------- ---------------------------- -TOTAL PROCESSING SUMMARY ------------------------- ---------------------------- -Total files 100 -Avg BPFP 1.3747 bits/point -Avg EBPFP 2.7493 equivalent bits/point -Avg MSE 1.055189 -Avg Time 3.093s ------------------------- ---------------------------- +version https://git-lfs.github.com/spec/v1 +oid sha256:16fafc95d3c7bb0dba784ea2a058a0c640b85f0d70b4fa941f88242ca1b272e8 +size 1117356