tahamajs/RR / Docs /module_metadata /cur_linear.json
tahamajs's picture
download
raw
817 Bytes
{
"module_name": "cur_linear",
"class_name": "CURLinear",
"category": "compression",
"family": "compression",
"params": [
"in_features",
"out_features",
"rank",
"bias",
"method",
"dtype",
"device"
],
"ops": [
"reshape",
"linear",
"view"
],
"docstring": "Linear layer whose weight matrix is factorized via CUR cross-approximation:\nW \u2248 C @ U @ R\n- C in R^{out_features x rank}: selected columns of W\n- U in R^{rank x rank}: Moore-Penrose pseudo-inverse core matrix\n- R in R^{rank x in_features}: selected rows of W\n\nExecutes forward pass in factorized form:\ny = ((x @ R^T) @ U^T) @ C^T + bias",
"rank": 32,
"d_state": 1,
"num_experts": 1,
"top_k": 1,
"sparsity_pct": 0.0,
"bits": 4,
"hash_val": 99168362477302272877579185487811989308
}

Xet Storage Details

Size:
817 Bytes
·
Xet hash:
d836d3ebdc66ba6b5ffe4d261870cf7faec4caabbf6544db195decc5e893ad33

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