Spaces:
Running on Zero
Running on Zero
| """Apply a sparse weight patch (edited down_proj columns) to an f16 GGUF. | |
| The patch is a .npz with, for each edited layer L, `L{L}_cols` (column indices of the MLP down projection) and | |
| `L{L}_vals` (the new column values, shape [hidden, n_cols]). The patched file is a copy; the input is never modified. | |
| """ | |
| import re | |
| import shutil | |
| from pathlib import Path | |
| import numpy as np | |
| def apply_patch(gguf_in: Path, patch_npz: Path, gguf_out: Path) -> list[int]: | |
| """Copy gguf_in to gguf_out, write the patched columns into it, and return the patched layer numbers.""" | |
| from gguf import GGUFReader # lazy: only the patch step needs it | |
| shutil.copyfile(gguf_in, gguf_out) | |
| patch = np.load(patch_npz) | |
| layers = sorted({int(m.group(1)) for k in patch.files if (m := re.fullmatch(r"L(\d+)_cols", k))}) | |
| reader = GGUFReader(str(gguf_out), "r+") | |
| by_name = {t.name: t for t in reader.tensors} | |
| for layer in layers: | |
| tensor = by_name[f"blk.{layer}.ffn_down.weight"] | |
| cols, vals = patch[f"L{layer}_cols"], patch[f"L{layer}_vals"] | |
| data = tensor.data # row-major [hidden, intermediate], f16 | |
| if data.dtype != np.float16 or data.shape[0] != vals.shape[0] or cols.max() >= data.shape[1]: | |
| raise ValueError(f"layer {layer}: tensor {data.dtype}{data.shape} does not fit the patch {vals.shape}") | |
| data[:, cols] = vals.astype(np.float16) | |
| if hasattr(reader.data, "flush"): | |
| reader.data.flush() | |
| return layers | |