| { | |
| "id": "factorised_matrix_state", | |
| "name": "Factorised Matrix State Attention Module.", | |
| "file": "Codes/framework/attentions/factorised_matrix_state.py", | |
| "type": "Attention Module", | |
| "category": "Low-Rank SVD & Factorization", | |
| "paradigm": "svd", | |
| "icon": "\u26a1", | |
| "description": "Factorised Matrix State Attention Module. Implements multi-head bidirectional/causal associative matrix memory updates combined with learnable decay gating, denominator normalization, low-rank factorized output projection, and pretrained weight transfer. Hardened for production across DistilBERT, RoBERTa/BERT, GPT-2, and Qwen/LLaMA architectures: - Robust pretrained weight extraction (GPT-2 fuse", | |
| "classes": [ | |
| "LoRALinear", | |
| "FactorisedMatrixState" | |
| ], | |
| "methods_count": 28, | |
| "math_latex": "Q = X W_q, \\quad K = X W_k, \\quad V = X W_v \\quad (W_q, W_k, W_v \\in \\mathbb{R}^{D \\times D})", | |
| "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:
- 2.71 kB
- Xet hash:
- eaac2cda09bd090d758ce629f8e76c0f9e33215cecb3adcd88a8fefaa999c53a
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.