| { | |
| "module_name": "adasvd_compressor", | |
| "class_name": "AdaSVDCompressor", | |
| "category": "compression", | |
| "family": "compression", | |
| "params": [ | |
| "model", | |
| "rank", | |
| "target_param_ratio", | |
| "num_iters", | |
| "target_modules" | |
| ], | |
| "ops": [ | |
| "linear", | |
| "matmul", | |
| "norm", | |
| "softmax" | |
| ], | |
| "docstring": "Adaptive SVD Compressor (AdaSVD).\n\nAssigns SVD ranks dynamically across layers proportional to spectral importance\n(nuclear norm ||W||_* = sum(sigma_i)), initializes factorized matrices with\nsymmetric square-root singular value distribution (A = U * sqrt(Sigma), B = sqrt(Sigma) * V^T),\nrefines factors via Alternating Least Squares (ALS), and replaces dense nn.Linear layers\nwith factorized SVDLinear modules.\n\nArgs:\n model: PyTorch neural network to compress.\n rank: Optional baseline rank for all layers or global target rank.\n target_param_ratio: Target parameter retention ratio (e.g. 0.7 for 30% reduction).\n num_iters: Number of alternating least squares refinement iterations.\n target_modules: Optional list of module name substrings to target.", | |
| "rank": 32, | |
| "d_state": 1, | |
| "num_experts": 1, | |
| "top_k": 1, | |
| "sparsity_pct": 0.0, | |
| "bits": 4, | |
| "hash_val": 240951343454300061049532962383599806077 | |
| } |
Xet Storage Details
- Size:
- 1.26 kB
- Xet hash:
- b5f6fb47497dd36b6ffc33aa6be61343920456097d727d1e2bf64071d0973f78
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.