Two ComfyUI nodes. Measure the Frobenius norm of the effective weight delta per transformer block, then apply only the blocks you choose. The .safetensors is opened read-only and never rewritten. Measures up@down (or kron(w1,w2) for LoKr), not the stored factors: LoRA has a free scale, so factor norms alone carry no information. Both fast paths are exact and avoid materialising the delta, and both are checked against explicitly materialised products in the tests. Unsupported formats are reported as NOT MEASURED, never folded into a zero -- a silent zero would make a block look prunable when it was merely not understood. 39 tests, no network, no GPU, no ComfyUI import.
3 lines
27 B
INI
3 lines
27 B
INI
[pytest]
|
|
testpaths = tests
|