Hearmeman24
49e5be766b
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.
22 lines
943 B
Python
22 lines
943 B
Python
"""Load block_surgeon.py by explicit file path.
|
|
|
|
ComfyUI requires an __init__.py at the pack root, and that __init__ imports
|
|
comfy.sd -- which does not exist outside a ComfyUI tree. Any import mechanism
|
|
that infers a parent package from the directory layout ends up importing that
|
|
__init__ and dying. Loading the one module under test by path sidesteps the whole
|
|
question, and is honest about the boundary: these tests cover the pure-math core
|
|
and never touch ComfyUI.
|
|
"""
|
|
|
|
import importlib.util
|
|
import pathlib
|
|
import sys
|
|
|
|
_root = pathlib.Path(__file__).resolve().parent.parent
|
|
_spec = importlib.util.spec_from_file_location("block_surgeon", _root / "block_surgeon.py")
|
|
_mod = importlib.util.module_from_spec(_spec)
|
|
# Registered BEFORE exec_module: @dataclass resolves sys.modules[cls.__module__]
|
|
# while the class body executes, and gets None if the module is not there yet.
|
|
sys.modules["block_surgeon"] = _mod
|
|
_spec.loader.exec_module(_mod)
|