| { | |
| "module_name": "kronecker_linear", | |
| "class_name": "KroneckerLinear", | |
| "category": "compression", | |
| "family": "compression", | |
| "params": [ | |
| "in_features", | |
| "out_features", | |
| "bias", | |
| "rank", | |
| "m1", | |
| "n1", | |
| "m2", | |
| "n2", | |
| "dtype", | |
| "device" | |
| ], | |
| "ops": [ | |
| "reshape", | |
| "matmul", | |
| "linear", | |
| "transpose" | |
| ], | |
| "docstring": "Linear layer whose weight matrix is parameterised as the Kronecker product\nof two smaller factor matrices: W = A \u2297 B.\n\nFor W in R^{out_features x in_features}, with out_features = m1 * m2 and\nin_features = n1 * n2:\n- A in R^{m1 x n1}\n- B in R^{m2 x n2}\n\nForward contraction uses the vectorization identity:\n(A \u2297 B) x = vec(B X A^T) without materializing the full (out x in) matrix,\nreducing FLOPs from O(B * in * out) to O(B * (in * m2 + out * n1)).", | |
| "rank": 32, | |
| "d_state": 1, | |
| "num_experts": 1, | |
| "top_k": 1, | |
| "sparsity_pct": 0.0, | |
| "bits": 4, | |
| "hash_val": 215010877849182823680577184959469542096 | |
| } |
Xet Storage Details
- Size:
- 996 Bytes
- Xet hash:
- ad07b2444893ac3e7e045c1c6af18aa97a72a5465dc6e8e30245c89922dac8b6
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.