tahamajs/RR / Docs /visualizer /components /armor_pruner.json
tahamajs's picture
download
raw
2.65 kB
{
"id": "armor_pruner",
"name": "ARMOR: Adaptive Resilient Matrix Orthogonal Rank Pruning (2:4 Sparse Core +",
"file": "Codes/framework/compression_engine/armor_pruner.py",
"type": "Compression Engine",
"category": "Low-Rank SVD & Factorization",
"paradigm": "svd",
"icon": "\u26a1",
"description": "ARMOR: Adaptive Resilient Matrix Orthogonal Rank Pruning (2:4 Sparse Core + Low-Rank Correction). Registered in COMPRESSOR_REGISTRY as 'armor_pruner', 'armor', and 'armor_compressor'. Mathematical Formulation: 1. Resilient First-Order Taylor Saliency (handling zero-grad, NaN/Inf, and absent activations): $$S_{ij} = \\begin{cases} |W_{ij} \\odot g_{ij}| & \\text{if } g \\text{ is valid and } \\|g\\|",
"classes": [
"ARMORLinear",
"ARMORPruner"
],
"methods_count": 14,
"math_latex": "W = W_{\\text{sparse}} + R",
"components": [
{
"id": "v_factor",
"name": "Factor V (Low-Rank Subspace Projection)",
"role": "Projection to Rank r",
"shape_transform": "[r \u00d7 d_in]",
"color": "#3acbb8",
"math_latex": "z_t = x_t V_r^\\top, \\quad V_r \\in \\mathbb{R}^{r \\times d_{in}}"
},
{
"id": "u_factor",
"name": "Factor U (Full-Rank Expansion Projection)",
"role": "Expansion to Output Space",
"shape_transform": "[d_out \u00d7 r]",
"color": "#2dd4bf",
"math_latex": "y_t = z_t (\\Sigma_r U_r^\\top) + b, \\quad U_r \\in \\mathbb{R}^{d_{out} \\times r}"
},
{
"id": "adapter",
"name": "Residual LoRA / Norm Calibration",
"role": "Fidelity Compensation",
"shape_transform": "[d_out \u00d7 d_in] Residual",
"color": "#5eead4",
"math_latex": "\\Delta y_t = x_t (B A)^\\top \\cdot \\frac{\\alpha}{r}, \\quad A \\in \\mathbb{R}^{r \\times d_{in}}, \\; B \\in \\mathbb{R}^{d_{out} \\times r}"
}
],
"model_projections": {
"distilbert": {
"hidden_size": 768,
"intermediate_size": 3072,
"baseline_layer_params": 6488064,
"compressed_layer_params": 786432,
"savings_ratio": 8.25
},
"roberta": {
"hidden_size": 768,
"intermediate_size": 3072,
"baseline_layer_params": 6488064,
"compressed_layer_params": 786432,
"savings_ratio": 8.25
},
"gpt2": {
"hidden_size": 768,
"intermediate_size": 3072,
"baseline_layer_params": 6488064,
"compressed_layer_params": 786432,
"savings_ratio": 8.25
},
"qwen": {
"hidden_size": 896,
"intermediate_size": 4864,
"baseline_layer_params": 11124736,
"compressed_layer_params": 1081344,
"savings_ratio": 10.29
}
}
}

Xet Storage Details

Size:
2.65 kB
·
Xet hash:
3639bdcdc90610c174c1c73ca5e4409501a77d18aad117da442ef433369f9aa8

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.