| { | |
| "id": "come", | |
| "name": "CoMe: Compress LLMs through Layer Concatenation and Shedding with Learnable", | |
| "file": "Codes/framework/compression_engine/come.py", | |
| "type": "Compression Engine", | |
| "category": "Low-Rank SVD & Factorization", | |
| "paradigm": "svd", | |
| "icon": "\u26a1", | |
| "description": "CoMe: Compress LLMs through Layer Concatenation and Shedding with Learnable Inter-Layer Interpolation Adapters. Registered in COMPRESSOR_REGISTRY as 'come' and 'come_compressor'. Mathematical Formulation: 1. Layer Shedding & Uniform Strided Selection: $$\\mathcal{I}_{\\text{keep}} = \\{i_0, i_1, \\dots, i_{K-1}\\}, \\quad i_k = \\left\\lfloor k \\cdot \\frac{L-1}{K-1} \\right\\rceil$$ 2. Learnable Linear", | |
| "classes": [ | |
| "CoMeBridge", | |
| "CoMeBridgedLayer", | |
| "CoMeLoRALinear", | |
| "CoMeCompressor" | |
| ], | |
| "methods_count": 19, | |
| "math_latex": "M_{\\text{compressed}} = X W_{\\text{bridge}}, \\quad W_{\\text{bridge}} = U_r \\Sigma_r V_r^\\top", | |
| "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:
- 2.76 kB
- Xet hash:
- 579cfd41c2bd25eebe723ba7b7aeeec2bda17c896d242e51e9d31ef943f95b31
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.