| """Depth recurrence without new parameters: rebuild a GGUF whose layer sequence revisits a block of layers. | |
| e.g. --order 0-13,10-13,14-23 runs layers 10..13 twice (the second pass reuses the same weights; it gets its own | |
| KV/conv state because llama.cpp sees it as a separate layer). Per-layer metadata arrays are remapped. | |
| usage: loop_layers.py in.gguf out.gguf --order 0-13,10-13,14-23""" | |
| import argparse, re, numpy as np | |
| import gguf | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("inp"); ap.add_argument("out"); ap.add_argument("--order", default="0-23") | |
| ap.add_argument("--window", type=int, default=None, help="override the sliding-attention window") | |
| ap.add_argument("--global-swa", action="store_true", help="turn every sliding-attention layer into full attention") | |
| a = ap.parse_args() | |
| order = [] | |
| for part in a.order.split(","): | |
| lo, hi = map(int, part.split("-")) if "-" in part else (int(part), int(part)) | |
| order += list(range(lo, hi + 1)) | |
| r = gguf.GGUFReader(a.inp) | |
| arch = r.fields["general.architecture"].contents() | |
| w = gguf.GGUFWriter(a.out, arch) | |
| per_layer = {f"{arch}.attention.head_count_kv", f"{arch}.attention.sliding_window_pattern"} | |
| for f in r.fields.values(): | |
| if f.name == "general.architecture" or f.name.startswith("GGUF."): | |
| continue | |
| vt = f.types[0] | |
| st = f.types[-1] if vt == gguf.GGUFValueType.ARRAY else None | |
| v = f.contents() | |
| if f.name == f"{arch}.block_count": | |
| v = len(order) | |
| elif f.name in per_layer: | |
| v = [v[i] for i in order] | |
| if a.global_swa and f.name.endswith("sliding_window_pattern"): | |
| v = [False] * len(v) | |
| elif f.name == f"{arch}.attention.sliding_window" and a.window: | |
| v = a.window | |
| w.add_key_value(f.name, v, vt, sub_type=st) | |
| w.add_string("general.layer_order", a.order) | |
| src = {t.name: t for t in r.tensors} | |
| out = [] | |
| for t in r.tensors: | |
| if not t.name.startswith("blk."): | |
| out.append((t.name, t)) | |
| for new_i, old_i in enumerate(order): | |
| for name, t in src.items(): | |
| m = re.match(rf"blk\.{old_i}\.(.+)", name) | |
| if m: | |
| out.append((f"blk.{new_i}.{m.group(1)}", t)) | |
| for name, t in out: | |
| w.add_tensor_info(name, t.data.shape, t.data.dtype, t.data.nbytes, t.tensor_type) | |
| w.write_header_to_file(); w.write_kv_data_to_file(); w.write_ti_data_to_file() | |
| for name, t in out: | |
| w.write_tensor_data(t.data, tensor_endianess=r.endianess) | |
| w.close() | |
| print("wrote", a.out, len(order), "layers:", order) | |